4 Commits

Author SHA1 Message Date
b1e4211e6a more work on the rust runner 2024-05-06 12:48:02 +01:00
e22df8adc9 started working on runner 2024-05-06 01:10:58 +01:00
7d346ba2ce closes #101 2024-05-02 16:58:01 +01:00
29b69deaf6 Fixed the model not training and running forever 2024-04-19 22:03:14 +01:00
61 changed files with 4333 additions and 3293 deletions

View File

@@ -1,5 +0,0 @@
tmp/
testData/
savedData/
!savedData/.keep
fyp

View File

@@ -1,19 +1,30 @@
FROM docker.io/nvidia/cuda:11.8.0-devel-ubuntu22.04
FROM docker.io/nvidia/cuda:12.3.2-devel-ubuntu22.04
ENV DEBIAN_FRONTEND=noninteractive
# Sometimes you have to get update twice because ?
RUN apt-get update
RUN apt-get update
RUN apt-get install -y wget sudo pkg-config libopencv-dev unzip python3-pip vim
RUN pip install torch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0
RUN mkdir /go
ENV GOPATH=/go
RUN apt-get install -y wget unzip python3-pip vim python3 python3-pip curl
RUN wget https://go.dev/dl/go1.22.2.linux-amd64.tar.gz
RUN tar -xvf go1.22.2.linux-amd64.tar.gz -C /usr/local
ENV PATH=$PATH:/usr/local/go/bin
ENV GOPATH=/go
RUN bash -c 'curl -L "https://storage.googleapis.com/tensorflow/libtensorflow/libtensorflow-cpu-linux-x86_64-2.9.1.tar.gz" | tar -C /usr/local -xz'
# RUN bash -c 'curl -L "https://storage.googleapis.com/tensorflow/libtensorflow/libtensorflow-cpu-linux-x86_64-2.13.1.tar.gz" | tar -C /usr/local -xz'
RUN ldconfig
RUN ln -s /usr/bin/python3 /usr/bin/python
RUN python -m pip install nvidia-pyindex
ADD requirements.txt .
RUN python -m pip install -r requirements.txt
ENV CUDNN_PATH=/usr/local/lib/python3.10/dist-packages/nvidia/cudnn
ENV LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/lib/python3.10/dist-packages/nvidia/cudnn/lib
RUN mkdir /app
WORKDIR /app
ADD go.mod .
@@ -23,38 +34,4 @@ ADD logic logic
RUN go install || true
WORKDIR /root
RUN wget https://git.andr3h3nriqu3s.com/andr3/gotch/raw/commit/22e75becf0432cda41a7c055a4d60ea435f76599/setup-libtorch.sh
RUN chmod +x setup-libtorch.sh
ENV CUDA_VER=11.8
ENV GOTCH_VER=v0.9.2
RUN bash setup-libtorch.sh
ENV GOTCH_LIBTORCH="/usr/local/lib/libtorch"
ENV REFRESH_SETUP=0
ENV LIBRARY_PATH="$LIBRARY_PATH:$GOTCH_LIBTORCH/lib"
ENV export CPATH="$CPATH:$GOTCH_LIBTORCH/lib:$GOTCH_LIBTORCH/include:$GOTCH_LIBTORCH/include/torch/csrc/api/include"
ENV LD_LIBRARY_PATH="$LD_LIBRARY_PATH:$GOTCH_LIBTORCH/lib:/usr/lib64-nvidia:/usr/local/cuda-${CUDA_VERSION}/lib64"
RUN wget https://git.andr3h3nriqu3s.com/andr3/gotch/raw/branch/master/setup-gotch.sh
RUN chmod +x setup-gotch.sh
RUN echo 'root ALL=(ALL) NOPASSWD:ALL' >> /etc/sudoers
RUN bash setup-gotch.sh
RUN ln -s /usr/local/lib/libtorch/include/torch/csrc /usr/local/lib/libtorch/include/torch/csrc/api/include/torch
RUN mkdir -p /go/pkg/mod/git.andr3h3nriqu3s.com/andr3/gotch@v0.9.2/libtch/libtorch/include/torch/csrc/api
RUN find /usr/local/lib/libtorch/include -maxdepth 1 -type d | tail -n +2 | grep -ve 'torch$' | xargs -I{} ln -s {} /go/pkg/mod/git.andr3h3nriqu3s.com/andr3/gotch@v0.9.2/libtch/libtorch/include
RUN ln -s /usr/local/lib/libtorch/include/torch/csrc/api/include /go/pkg/mod/git.andr3h3nriqu3s.com/andr3/gotch@v0.9.2/libtch/libtorch/include/torch/csrc/api/include
RUN find /usr/local/lib/libtorch/include/torch -maxdepth 1 -type f | xargs -I{} ln -s {} /go/pkg/mod/git.andr3h3nriqu3s.com/andr3/gotch@v0.9.2/libtch/libtorch/include/torch
RUN ln -s /usr/local/lib/libtorch/lib/libcudnn.so.8 /usr/local/lib/libcudnn.so
WORKDIR /app
ENV CGO_CXXFLAGS="-I/usr/local/lib/libtorch/include/torch/csrc/api/include/ -I/usr/local/lib/libtorch/include"
ENV CGO_CFLAGS="-I/usr/local/lib/libtorch/include/torch/csrc/api/include/ -I/usr/local/lib/libtorch/include"
ADD . .
RUN go build -x || true
CMD ["bash", "-c", "go run ."]
CMD ["go", "run", "."]

3
go.mod
View File

@@ -4,10 +4,11 @@ go 1.21
require (
github.com/charmbracelet/log v0.3.1
github.com/galeone/tensorflow/tensorflow/go v0.0.0-20240119075110-6ad3cf65adfe
github.com/galeone/tfgo v0.0.0-20230715013254-16113111dc99
github.com/google/uuid v1.6.0
github.com/lib/pq v1.10.9
golang.org/x/crypto v0.19.0
git.andr3h3nriqu3s.com/andr3/gotch v0.9.2
)
require (

12
go.sum
View File

@@ -1,7 +1,3 @@
git.andr3h3nriqu3s.com/andr3/gotch v0.9.1 h1:1q34JKV8cX80n7LXbJswlXCiRtNbzcvJ/vbgb6an1tA=
git.andr3h3nriqu3s.com/andr3/gotch v0.9.1/go.mod h1:FXusE3CHt8NLf5wynUGaHtIbToRuYifsZaC5EZH0pJY=
git.andr3h3nriqu3s.com/andr3/gotch v0.9.2 h1:aZcsPgDVGVhrEFoer0upSkzPqJWNMxdUHRktP4s6MSc=
git.andr3h3nriqu3s.com/andr3/gotch v0.9.2/go.mod h1:FXusE3CHt8NLf5wynUGaHtIbToRuYifsZaC5EZH0pJY=
github.com/BurntSushi/toml v1.3.2 h1:o7IhLm0Msx3BaB+n3Ag7L8EVlByGnpq14C4YWiu/gL8=
github.com/BurntSushi/toml v1.3.2/go.mod h1:CxXYINrC8qIiEnFrOxCa7Jy5BFHlXnUU2pbicEuybxQ=
github.com/aymanbagabas/go-osc52/v2 v2.0.1 h1:HwpRHbFMcZLEVr42D4p7XBqjyuxQH5SMiErDT4WkJ2k=
@@ -17,6 +13,12 @@ github.com/charmbracelet/log v0.3.1/go.mod h1:OR4E1hutLsax3ZKpXbgUqPtTjQfrh1pG3z
github.com/davecgh/go-spew v1.1.0/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
github.com/gabriel-vasile/mimetype v1.4.3 h1:in2uUcidCuFcDKtdcBxlR0rJ1+fsokWf+uqxgUFjbI0=
github.com/gabriel-vasile/mimetype v1.4.3/go.mod h1:d8uq/6HKRL6CGdk+aubisF/M5GcPfT7nKyLpA0lbSSk=
github.com/galeone/tensorflow/tensorflow/go v0.0.0-20221023090153-6b7fa0680c3e h1:9+2AEFZymTi25FIIcDwuzcOPH04z9+fV6XeLiGORPDI=
github.com/galeone/tensorflow/tensorflow/go v0.0.0-20221023090153-6b7fa0680c3e/go.mod h1:TelZuq26kz2jysARBwOrTv16629hyUsHmIoj54QqyFo=
github.com/galeone/tensorflow/tensorflow/go v0.0.0-20240119075110-6ad3cf65adfe h1:7yELf1NFEwECpXMGowkoftcInMlVtLTCdwWLmxKgzNM=
github.com/galeone/tensorflow/tensorflow/go v0.0.0-20240119075110-6ad3cf65adfe/go.mod h1:TelZuq26kz2jysARBwOrTv16629hyUsHmIoj54QqyFo=
github.com/galeone/tfgo v0.0.0-20230715013254-16113111dc99 h1:8Bt1P/zy1gb37L4n8CGgp1qmFwBV5729kxVfj0sqhJk=
github.com/galeone/tfgo v0.0.0-20230715013254-16113111dc99/go.mod h1:3YgYBeIX42t83uP27Bd4bSMxTnQhSbxl0pYSkCDB1tc=
github.com/go-logfmt/logfmt v0.6.0 h1:wGYYu3uicYdqXVgoYbvnkrPVXkuLM1p1ifugDMEdRi4=
github.com/go-logfmt/logfmt v0.6.0/go.mod h1:WYhtIu8zTZfxdn5+rREduYbwxfcBr/Vr6KEVveWlfTs=
github.com/go-playground/locales v0.14.1 h1:EWaQ/wswjilfKLTECiXz7Rh+3BjFhfDFKv/oXslEjJA=
@@ -72,9 +74,7 @@ github.com/rivo/uniseg v0.4.6 h1:Sovz9sDSwbOz9tgUy8JpT+KgCkPYJEN/oYzlJiYTNLg=
github.com/rivo/uniseg v0.4.6/go.mod h1:FN3SvrM+Zdj16jyLfmOkMNblXMcoc8DfTHruCPUcx88=
github.com/stretchr/objx v0.1.0/go.mod h1:HFkY916IF+rwdDfMAkV7OtwuqBVzrE8GR6GFx+wExME=
github.com/stretchr/testify v1.3.0/go.mod h1:M5WIy9Dh21IEIfnGCwXGc5bZfKNJtfHm1UVUgZn+9EI=
github.com/stretchr/testify v1.6.1/go.mod h1:6Fq8oRcR53rry900zMqJjRRixrwX3KX962/h/Wwjteg=
github.com/stretchr/testify v1.7.0/go.mod h1:6Fq8oRcR53rry900zMqJjRRixrwX3KX962/h/Wwjteg=
github.com/x448/float16 v0.8.4/go.mod h1:14CWIYCyZA/cWjXOioeEpHeN/83MdbZDRQHoFcYsOfg=
golang.org/x/crypto v0.13.0 h1:mvySKfSWJ+UKUii46M40LOvyWfN0s2U+46/jDd0e6Ck=
golang.org/x/crypto v0.13.0/go.mod h1:y6Z2r+Rw4iayiXXAIxJIDAJ1zMW4yaTpebo8fPOliYc=
golang.org/x/crypto v0.18.0 h1:PGVlW0xEltQnzFZ55hkuX5+KLyrMYhHld1YHO4AKcdc=

1
lib
View File

@@ -1 +0,0 @@
/usr/local/lib

View File

@@ -1,10 +0,0 @@
package libtch
// #cgo LDFLAGS: -lstdc++ -ltorch -lc10 -ltorch_cpu -L${SRCDIR}/libtorch/lib
// #cgo LDFLAGS: -L/usr/local/cuda/lib64 -lcuda -lcudart -lcublas -lcudnn -lcaffe2_nvrtc -lnvrtc-builtins -lnvrtc -lnvToolsExt -lc10_cuda -ltorch_cuda
// #cgo CFLAGS: -I${SRCDIR} -O3 -Wall -Wno-unused-variable -Wno-deprecated-declarations -Wno-c++11-narrowing -g -Wno-sign-compare -Wno-unused-function
// #cgo CFLAGS: -D_GLIBCXX_USE_CXX11_ABI=0
// #cgo CFLAGS: -I/usr/local/cuda/include
// #cgo CXXFLAGS: -std=c++17 -I${SRCDIR} -g -O3
// #cgo CXXFLAGS: -I${SRCDIR}/libtorch/lib -I${SRCDIR}/libtorch/include -I${SRCDIR}/libtorch/include/torch/csrc/api/include -I/opt/libtorch/include/torch/csrc/api/include
import "C"

View File

@@ -16,9 +16,9 @@ const (
)
type ModelClass struct {
Id string `db:"mc.id"`
ModelId string `db:"mc.model_id"`
Name string `db:"mc.name"`
ClassOrder int `db:"mc.class_order"`
Status int `db:"mc.status"`
Id string `db:"mc.id" json:"id"`
ModelId string `db:"mc.model_id" json:"model_id"`
Name string `db:"mc.name" json:"name"`
ClassOrder int `db:"mc.class_order" json:"class_order"`
Status int `db:"mc.status" json:"status"`
}

View File

@@ -20,14 +20,14 @@ const (
)
type Definition struct {
Id string `db:"md.id"`
ModelId string `db:"md.model_id"`
Accuracy float64 `db:"md.accuracy"`
TargetAccuracy int `db:"md.target_accuracy"`
Epoch int `db:"md.epoch"`
Status int `db:"md.status"`
CreatedOn time.Time `db:"md.created_on"`
EpochProgress int `db:"md.epoch_progress"`
Id string `db:"md.id" json:"id"`
ModelId string `db:"md.model_id" json:"model_id"`
Accuracy float64 `db:"md.accuracy" json:"accuracy"`
TargetAccuracy int `db:"md.target_accuracy" json:"target_accuracy"`
Epoch int `db:"md.epoch" json:"epoch"`
Status int `db:"md.status" json:"status"`
CreatedOn time.Time `db:"md.created_on" json:"created"`
EpochProgress int `db:"md.epoch_progress" json:"epoch_progress"`
}
type SortByAccuracyDefinitions []*Definition

View File

@@ -16,12 +16,12 @@ const (
)
type Layer struct {
Id string `db:"mdl.id"`
DefinitionId string `db:"mdl.def_id"`
LayerOrder string `db:"mdl.layer_order"`
LayerType LayerType `db:"mdl.layer_type"`
Shape string `db:"mdl.shape"`
ExpType string `db:"mdl.exp_type"`
Id string `db:"mdl.id" json:"id"`
DefinitionId string `db:"mdl.def_id" json:"definition_id"`
LayerOrder string `db:"mdl.layer_order" json:"layer_order"`
LayerType LayerType `db:"mdl.layer_type" json:"layer_type"`
Shape string `db:"mdl.shape" json:"shape"`
ExpType string `db:"mdl.exp_type" json:"exp_type"`
}
func ShapeToString(args ...int) string {

View File

@@ -1,32 +1,45 @@
package dbtypes
import (
"errors"
"fmt"
"os"
"path"
"git.andr3h3nriqu3s.com/andr3/fyp/logic/db"
"github.com/jackc/pgx/v5"
)
const (
FAILED_TRAINING = -4
FAILED_PREPARING_TRAINING = -3
FAILED_PREPARING_ZIP_FILE = -2
FAILED_PREPARING = -1
type ModelStatus int
PREPARING = 1
CONFIRM_PRE_TRAINING = 2
PREPARING_ZIP_FILE = 3
TRAINING = 4
READY = 5
READY_ALTERATION = 6
READY_ALTERATION_FAILED = -6
const (
FAILED_TRAINING ModelStatus = -4
FAILED_PREPARING_TRAINING = -3
FAILED_PREPARING_ZIP_FILE = -2
FAILED_PREPARING = -1
PREPARING = 1
CONFIRM_PRE_TRAINING = 2
PREPARING_ZIP_FILE = 3
TRAINING = 4
READY = 5
READY_ALTERATION = 6
READY_ALTERATION_FAILED = -6
READY_RETRAIN = 7
READY_RETRAIN_FAILED = -7
)
type ModelDefinitionStatus int
const (
MODEL_DEFINITION_STATUS_CANCELD_TRAINING ModelDefinitionStatus = -4
MODEL_DEFINITION_STATUS_FAILED_TRAINING = -3
MODEL_DEFINITION_STATUS_PRE_INIT = 1
MODEL_DEFINITION_STATUS_INIT = 2
MODEL_DEFINITION_STATUS_TRAINING = 3
MODEL_DEFINITION_STATUS_PAUSED_TRAINING = 6
MODEL_DEFINITION_STATUS_TRANIED = 4
MODEL_DEFINITION_STATUS_READY = 5
)
type ModelHeadStatus int
const (
@@ -51,6 +64,8 @@ type BaseModel struct {
CanTrain int `db:"can_train"`
}
var ModelNotFoundError = errors.New("Model not found error")
func GetBaseModel(db db.Db, id string) (base *BaseModel, err error) {
var model BaseModel
err = GetDBOnce(db, &model, "models where id=$1", id)
@@ -68,36 +83,6 @@ func (m BaseModel) CanEval() bool {
return true
}
func (m BaseModel) removeFailedDataPoints(c BasePack) (err error) {
rows, err := c.GetDb().Query("select mdp.id from model_data_point as mdp join model_classes as mc on mc.id=mdp.class_id where mc.model_id=$1 and mdp.status=-1;", m.Id)
if err != nil {
return
}
defer rows.Close()
base_path := path.Join("savedData", m.Id, "data")
for rows.Next() {
var dataPointId string
err = rows.Scan(&dataPointId)
if err != nil {
return
}
p := path.Join(base_path, dataPointId+"."+m.Format)
c.GetLogger().Warn("Removing image", "path", p)
err = os.RemoveAll(p)
if err != nil {
return
}
}
_, err = c.GetDb().Exec("delete from model_data_point as mdp using model_classes as mc where mdp.class_id = mc.id and mc.model_id=$1 and mdp.status=-1;", m.Id)
return
}
// DO NOT Pass un filtered data on filters
func (m BaseModel) GetDefinitions(db db.Db, filters string, args ...any) ([]*Definition, error) {
n_args := []any{m.Id}
@@ -105,25 +90,20 @@ func (m BaseModel) GetDefinitions(db db.Db, filters string, args ...any) ([]*Def
return GetDbMultitple[Definition](db, fmt.Sprintf("model_definition as md where md.model_id=$1 %s", filters), n_args...)
}
// DO NOT Pass un filtered data on filters
func (m BaseModel) GetClasses(db db.Db, filters string, args ...any) ([]*ModelClass, error) {
n_args := []any{m.Id}
n_args = append(n_args, args...)
return GetDbMultitple[ModelClass](db, fmt.Sprintf("model_classes as mc where mc.model_id=$1 %s", filters), n_args...)
}
type DataPointIterator struct {
rows pgx.Rows
Model BaseModel
func (m *BaseModel) UpdateStatus(db db.Db, status ModelStatus) (err error) {
_, err = db.Exec("update models set status=$1 where id=$2", status, m.Id)
return
}
type DataPoint struct {
Class int
Path string
}
func (iter DataPointIterator) Close() {
iter.rows.Close()
Class int `json:"class"`
Path string `json:"path"`
}
func (m BaseModel) DataPoints(db db.Db, mode DATA_POINT_MODE) (data []DataPoint, err error) {
@@ -158,14 +138,11 @@ func (m BaseModel) DataPoints(db db.Db, mode DATA_POINT_MODE) (data []DataPoint,
return
}
const RGB string = "rgb"
const GRAY string = "greyscale"
func StringToImageMode(colorMode string) int {
switch colorMode {
case GRAY:
case "greyscale":
return 1
case RGB:
case "rgb":
return 3
default:
panic("unkown color mode")

View File

@@ -14,13 +14,11 @@ import (
"github.com/charmbracelet/log"
"github.com/google/uuid"
"github.com/jackc/pgx/v5"
"github.com/jackc/pgx/v5/pgconn"
db "git.andr3h3nriqu3s.com/andr3/fyp/logic/db"
)
type BasePack interface {
db.Db
GetDb() db.Db
GetLogger() *log.Logger
GetHost() string
@@ -44,18 +42,6 @@ func (b BasePackStruct) GetLogger() *log.Logger {
return b.Logger
}
func (c BasePackStruct) Query(query string, args ...any) (pgx.Rows, error) {
return c.Db.Query(query, args...)
}
func (c BasePackStruct) Exec(query string, args ...any) (pgconn.CommandTag, error) {
return c.Db.Exec(query, args...)
}
func (c BasePackStruct) Begin() (pgx.Tx, error) {
return c.Db.Begin()
}
func CheckEmpty(f url.Values, path string) bool {
return !f.Has(path) || f.Get(path) == ""
}

View File

@@ -14,7 +14,7 @@ type ModelClassJSON struct {
Status int `json:"status"`
}
func ListClassesJSON(c BasePack, model_id string) (cls []*ModelClassJSON, err error) {
func ListClasses(c BasePack, model_id string) (cls []*ModelClassJSON, err error) {
return GetDbMultitple[ModelClassJSON](c.GetDb(), "model_classes where model_id=$1", model_id)
}

View File

@@ -435,7 +435,7 @@ func handleDataUpload(handle *Handle) {
}
model, err := GetBaseModel(handle.Db, id)
if err == NotFoundError {
if err == ModelNotFoundError {
return c.SendJSONStatus(http.StatusNotFound, "Model not found")
} else if err != nil {
return c.Error500(err)
@@ -468,7 +468,7 @@ func handleDataUpload(handle *Handle) {
}
PostAuthJson(handle, "/models/data/class/new", User_Normal, func(c *Context, obj *CreateNewEmptyClass) *Error {
model, err := GetBaseModel(c.Db, obj.Id)
if err == NotFoundError {
if err == ModelNotFoundError {
return c.JsonBadRequest("Model not found")
} else if err != nil {
return c.E500M("Failed to get model information", err)
@@ -518,7 +518,7 @@ func handleDataUpload(handle *Handle) {
c.Logger.Info("model", "model", *model_id)
model, err := GetBaseModel(c.Db, *model_id)
if err == NotFoundError {
if err == ModelNotFoundError {
return c.JsonBadRequest("Could not find the model")
} else if err != nil {
return c.E500M("Error getting model information", err)
@@ -626,7 +626,7 @@ func handleDataUpload(handle *Handle) {
c.Logger.Info("Trying to expand model", "id", id)
model, err := GetBaseModel(handle.Db, id)
if err == NotFoundError {
if err == ModelNotFoundError {
return c.SendJSONStatus(http.StatusNotFound, "Model not found")
} else if err != nil {
return c.Error500(err)
@@ -670,7 +670,7 @@ func handleDataUpload(handle *Handle) {
}
model, err := GetBaseModel(handle.Db, dat.Id)
if err == NotFoundError {
if err == ModelNotFoundError {
return c.SendJSONStatus(http.StatusNotFound, "Model not found")
} else if err != nil {
return c.Error500(err)

View File

@@ -51,7 +51,7 @@ func handleDelete(handle *Handle) {
return c.E500M("Faield to get model", err)
}
switch model.Status {
switch ModelStatus(model.Status) {
case FAILED_TRAINING:
fallthrough
case FAILED_PREPARING_ZIP_FILE:

View File

@@ -24,7 +24,7 @@ func handleEdit(handle *Handle) {
return c.Error500(err)
}
cls, err := model_classes.ListClassesJSON(c, model.Id)
cls, err := model_classes.ListClasses(c, model.Id)
if err != nil {
return c.Error500(err)
}
@@ -109,7 +109,7 @@ func handleEdit(handle *Handle) {
layers := []layerdef{}
for _, def := range defs {
if def.Status == DEFINITION_STATUS_TRAINING {
if def.Status == MODEL_DEFINITION_STATUS_TRAINING {
rows, err := c.Db.Query("select id, layer_type, shape from model_definition_layer where def_id=$1 order by layer_order asc;", def.Id)
if err != nil {
return c.Error500(err)
@@ -166,7 +166,7 @@ func handleEdit(handle *Handle) {
for i, def := range defs {
var lay *[]layerdef = nil
if def.Status == DEFINITION_STATUS_TRAINING && !setLayers {
if def.Status == MODEL_DEFINITION_STATUS_TRAINING && !setLayers {
lay = &layers
setLayers = true
}

View File

@@ -8,6 +8,7 @@ import (
. "git.andr3h3nriqu3s.com/andr3/fyp/logic/db_types"
. "git.andr3h3nriqu3s.com/andr3/fyp/logic/tasks/utils"
"github.com/charmbracelet/log"
tf "github.com/galeone/tensorflow/tensorflow/go"
"github.com/galeone/tensorflow/tensorflow/go/op"
tg "github.com/galeone/tfgo"
@@ -19,6 +20,7 @@ func ReadPNG(scope *op.Scope, imagePath string, channels int64) *image.Image {
contents := op.ReadFile(scope.SubScope("ReadFile"), op.Const(scope.SubScope("filename"), imagePath))
output := op.DecodePng(scope.SubScope("DecodePng"), contents, op.DecodePngChannels(channels))
output = op.ExpandDims(scope.SubScope("ExpandDims"), output, op.Const(scope.SubScope("axis"), []int32{0}))
output = op.ExpandDims(scope.SubScope("Stack"), output, op.Const(scope.SubScope("axis"), []int32{1}))
image := &image.Image{
Tensor: tg.NewTensor(scope, output)}
return image.Scale(0, 255)
@@ -29,6 +31,7 @@ func ReadJPG(scope *op.Scope, imagePath string, channels int64) *image.Image {
contents := op.ReadFile(scope.SubScope("ReadFile"), op.Const(scope.SubScope("filename"), imagePath))
output := op.DecodePng(scope.SubScope("DecodeJpeg"), contents, op.DecodePngChannels(channels))
output = op.ExpandDims(scope.SubScope("ExpandDims"), output, op.Const(scope.SubScope("axis"), []int32{0}))
output = op.ExpandDims(scope.SubScope("Stack"), output, op.Const(scope.SubScope("axis"), []int32{1}))
image := &image.Image{
Tensor: tg.NewTensor(scope, output)}
return image.Scale(0, 255)
@@ -49,6 +52,8 @@ func runModelNormal(base BasePack, model *BaseModel, def_id string, inputImage *
var vmax float32 = 0.0
var predictions = results[0].Value().([][]float32)[0]
log.Info("preds", "preds", predictions)
for i, v := range predictions {
if v > vmax {
order = i
@@ -62,10 +67,13 @@ func runModelNormal(base BasePack, model *BaseModel, def_id string, inputImage *
}
func runModelExp(base BasePack, model *BaseModel, def_id string, inputImage *tf.Tensor) (order int, confidence float32, err error) {
log := base.GetLogger()
err = nil
order = 0
log.Info("Running base")
base_model := tg.LoadModel(path.Join("savedData", model.Id, "defs", def_id, "base", "model"), []string{"serve"}, nil)
//results := base_model.Exec([]tf.Output{
@@ -86,7 +94,7 @@ func runModelExp(base BasePack, model *BaseModel, def_id string, inputImage *tf.
return
}
base.GetLogger().Info("test", "count", len(heads))
log.Info("Running heads", "heads", heads)
var vmax float32 = 0.0

View File

@@ -0,0 +1,79 @@
package models_train
import (
"errors"
. "git.andr3h3nriqu3s.com/andr3/fyp/logic/db_types"
. "git.andr3h3nriqu3s.com/andr3/fyp/logic/tasks/utils"
. "git.andr3h3nriqu3s.com/andr3/fyp/logic/utils"
"github.com/goccy/go-json"
)
func PrepareTraining(handler *Handle, b BasePack, task Task, runner_id string) (err error) {
l := b.GetLogger()
model, err := GetBaseModel(b.GetDb(), *task.ModelId)
if err != nil {
task.UpdateStatusLog(b, TASK_FAILED_RUNNING, "Failed to get model information")
l.Error("Failed to get model information", "err", err)
return err
}
if model.Status != TRAINING {
task.UpdateStatusLog(b, TASK_FAILED_RUNNING, "Model not in the correct status for training")
return errors.New("Model not in the right status")
}
// TODO do this when the runner says it's OK
//task.UpdateStatusLog(b, TASK_RUNNING, "Training model")
// TODO move this to the runner part as well
var dat struct {
NumberOfModels int
Accuracy int
}
err = json.Unmarshal([]byte(task.ExtraTaskInfo), &dat)
if err != nil {
task.UpdateStatusLog(b, TASK_FAILED_RUNNING, "Failed to get model extra information")
}
if model.ModelType == 2 {
panic("TODO")
full_error := generateExpandableDefinitions(b, model, dat.Accuracy, dat.NumberOfModels)
if full_error != nil {
l.Error("Failed to generate defintions", "err", full_error)
task.UpdateStatusLog(b, TASK_FAILED_RUNNING, "Failed generate model")
return errors.New("Failed to generate definitions")
}
} else {
error := generateDefinitions(b, model, dat.Accuracy, dat.NumberOfModels)
if error != nil {
task.UpdateStatusLog(b, TASK_FAILED_RUNNING, "Failed generate model")
return errors.New("Failed to generate definitions")
}
}
runners := handler.DataMap["runners"].(map[string]interface{})
runner := runners[runner_id].(map[string]interface{})
runner["task"] = &task
runners[runner_id] = runner
handler.DataMap["runners"] = runners
return
}
func CleanUpFailed(b BasePack, task *Task) {
db := b.GetDb()
l := b.GetLogger()
model, err := GetBaseModel(db, *task.ModelId)
if err != nil {
l.Error("Failed to get model", "err", err)
} else {
err = model.UpdateStatus(db, FAILED_TRAINING)
if err != nil {
l.Error("Failed to get status", err)
}
}
}

View File

@@ -11,13 +11,13 @@ import (
func handleRest(handle *Handle) {
DeleteAuthJson(handle, "/models/train/reset", User_Normal, func(c *Context, dat *JustId) *Error {
model, err := GetBaseModel(c.Db, dat.Id)
if err == NotFoundError {
if err == ModelNotFoundError {
return c.JsonBadRequest("Model not found")
} else if err != nil {
return c.E500M("Failed to get model", err)
}
if model.Status != FAILED_PREPARING_TRAINING && model.Status != FAILED_TRAINING {
if model.Status != FAILED_PREPARING_TRAINING && model.Status != int(FAILED_TRAINING) {
return c.JsonBadRequest("Model is not in status that be reset")
}

View File

@@ -1,179 +0,0 @@
package imageloader
import (
"git.andr3h3nriqu3s.com/andr3/fyp/logic/db"
types "git.andr3h3nriqu3s.com/andr3/fyp/logic/db_types"
"git.andr3h3nriqu3s.com/andr3/gotch"
torch "git.andr3h3nriqu3s.com/andr3/gotch/ts"
"git.andr3h3nriqu3s.com/andr3/gotch/vision"
)
type Dataset struct {
TrainImages *torch.Tensor
TrainLabels *torch.Tensor
TestImages *torch.Tensor
TestLabels *torch.Tensor
TrainImagesSize int
TestImagesSize int
Device gotch.Device
}
func LoadImagesAndLables(db db.Db, m *types.BaseModel, mode types.DATA_POINT_MODE, classStart int, classEnd int) (imgs, labels *torch.Tensor, count int, err error) {
train_points, err := m.DataPoints(db, types.DATA_POINT_MODE_TRAINING)
if err != nil {
return
}
size := int64(classEnd - classStart + 1)
pimgs := []*torch.Tensor{}
plabels := []*torch.Tensor{}
for _, point := range train_points {
var img, label *torch.Tensor
img, err = vision.Load(point.Path)
if err != nil {
return
}
pimgs = append(pimgs, img)
t_label := make([]int, size)
if point.Class <= classEnd && point.Class >= classStart {
t_label[point.Class-classStart] = 1
}
label, err = torch.OfSlice(t_label)
if err != nil {
return
}
plabels = append(plabels, label)
}
imgs, err = torch.Concat(pimgs, 0)
if err != nil {
return
}
labels, err = torch.Stack(plabels, 0)
if err != nil {
return
}
count = len(pimgs)
imgs, err = torch.Stack(pimgs, 0)
if err != nil {
return
}
imgs, err = imgs.ToDtype(gotch.Float, false, false, true)
if err != nil {
return
}
labels, err = labels.ToDtype(gotch.Float, false, false, true)
if err != nil {
return
}
return
}
func NewDataset(db db.Db, m *types.BaseModel, classStart int, classEnd int) (ds *Dataset, err error) {
trainImages, trainLabels, train_count, err := LoadImagesAndLables(db, m, types.DATA_POINT_MODE_TRAINING, classStart, classEnd)
if err != nil {
return
}
testImages, testLabels, test_count, err := LoadImagesAndLables(db, m, types.DATA_POINT_MODE_TESTING, classStart, classEnd)
if err != nil {
return
}
ds = &Dataset{
TrainImages: trainImages,
TrainLabels: trainLabels,
TestImages: testImages,
TestLabels: testLabels,
TrainImagesSize: train_count,
TestImagesSize: test_count,
Device: gotch.CPU,
}
return
}
func (ds *Dataset) To(device gotch.Device) (err error) {
ds.TrainImages, err = ds.TrainImages.ToDevice(device, ds.TrainImages.DType(), device.IsCuda(), true, true)
if err != nil {
return
}
ds.TrainLabels, err = ds.TrainLabels.ToDevice(device, ds.TrainLabels.DType(), device.IsCuda(), true, true)
if err != nil {
return
}
ds.TestImages, err = ds.TestImages.ToDevice(device, ds.TestImages.DType(), device.IsCuda(), true, true)
if err != nil {
return
}
ds.TestLabels, err = ds.TestLabels.ToDevice(device, ds.TestLabels.DType(), device.IsCuda(), true, true)
if err != nil {
return
}
ds.Device = device
return
}
func (ds *Dataset) TestIter(batchSize int64) *torch.Iter2 {
return torch.MustNewIter2(ds.TestImages, ds.TestLabels, batchSize)
}
func (ds *Dataset) TrainIter(batchSize int64) (iter *torch.Iter2, err error) {
// Create a clone of the trainimages
size, err := ds.TrainImages.Size()
if err != nil {
return
}
train_images, err := torch.Zeros(size, gotch.Float, ds.Device)
if err != nil {
return
}
ds.TrainImages, err = ds.TrainImages.Clone(train_images, false)
if err != nil {
return
}
// Create a clone of the labels
size, err = ds.TrainLabels.Size()
if err != nil {
return
}
train_labels, err := torch.Zeros(size, gotch.Float, ds.Device)
if err != nil {
return
}
ds.TrainLabels, err = ds.TrainLabels.Clone(train_labels, false)
if err != nil {
return
}
iter, err = torch.NewIter2(train_images, train_labels, batchSize)
if err != nil {
return
}
return
}

View File

@@ -1,174 +0,0 @@
package my_nn
// linear is a fully-connected layer
import (
"math"
"git.andr3h3nriqu3s.com/andr3/gotch/nn"
"git.andr3h3nriqu3s.com/andr3/gotch/ts"
"github.com/charmbracelet/log"
)
// LinearConfig is a configuration for a linear layer
type LinearConfig struct {
WsInit nn.Init // iniital weights
BsInit nn.Init // optional initial bias
Bias bool
}
// DefaultLinearConfig creates default LinearConfig with
// weights initiated using KaimingUniform and Bias is set to true
func DefaultLinearConfig() *LinearConfig {
negSlope := math.Sqrt(5)
return &LinearConfig{
// NOTE. KaimingUniform cause mem leak due to ts.Uniform()!!!
// Avoid using it now.
WsInit: nn.NewKaimingUniformInit(nn.WithKaimingNegativeSlope(negSlope)),
BsInit: nil,
Bias: true,
}
}
// Linear is a linear fully-connected layer
type Linear struct {
Ws *ts.Tensor
weight_name string
Bs *ts.Tensor
bias_name string
}
// NewLinear creates a new linear layer
// y = x*wT + b
// inDim - input dimension (x) [input features - columns]
// outDim - output dimension (y) [output features - columns]
// NOTE: w will have shape{outDim, inDim}; b will have shape{outDim}
func NewLinear(vs *Path, inDim, outDim int64, c *LinearConfig) *Linear {
var bias_name string
var bs *ts.Tensor
var err error
if c.Bias {
switch {
case c.BsInit == nil:
shape := []int64{inDim, outDim}
fanIn, _, err := nn.CalculateFans(shape)
or_panic(err)
bound := 0.0
if fanIn > 0 {
bound = 1 / math.Sqrt(float64(fanIn))
}
bsInit := nn.NewUniformInit(-bound, bound)
bs, bias_name, err = vs.NewVarNamed("bias", []int64{outDim}, bsInit)
or_panic(err)
// Find better way to do this
bs, err = bs.T(true)
or_panic(err)
bs, err = bs.T(true)
or_panic(err)
bs, err = bs.SetRequiresGrad(true, true)
or_panic(err)
err = bs.RetainGrad(false)
or_panic(err)
vs.varstore.UpdateVarTensor(bias_name, bs, true)
case c.BsInit != nil:
bs, bias_name, err = vs.NewVarNamed("bias", []int64{outDim}, c.BsInit)
or_panic(err)
}
}
ws, weight_name, err := vs.NewVarNamed("weight", []int64{outDim, inDim}, c.WsInit)
or_panic(err)
ws, err = ws.T(true)
or_panic(err)
ws, err = ws.SetRequiresGrad(true, true)
or_panic(err)
err = ws.RetainGrad(false)
or_panic(err)
vs.varstore.UpdateVarTensor(weight_name, ws, true)
return &Linear{
Ws: ws,
weight_name: weight_name,
Bs: bs,
bias_name: bias_name,
}
}
func (l *Linear) Debug() {
log.Info("Ws", "ws", l.Ws.MustGrad(false).MustMax(false).Float64Values())
log.Info("Bs", "bs", l.Bs.MustGrad(false).MustMax(false).Float64Values())
}
func (l *Linear) ExtractFromVarstore(vs *VarStore) {
l.Ws = vs.GetTensorOfVar(l.weight_name)
l.Bs = vs.GetTensorOfVar(l.bias_name)
}
// Implement `Module` for `Linear` struct:
// =======================================
// Forward proceeds input node through linear layer.
// NOTE:
// - It assumes that node has dimensions of 2 (matrix).
// To make it work for matrix multiplication, input node should
// has same number of **column** as number of **column** in
// `LinearLayer` `Ws` property as weights matrix will be
// transposed before multiplied to input node. (They are all used `inDim`)
// - Input node should have shape of `shape{batch size, input features}`.
// (shape{batchSize, inDim}). The input features is `inDim` while the
// output feature is `outDim` in `LinearConfig` struct.
//
// Example:
//
// inDim := 3
// outDim := 2
// batchSize := 4
// weights: 2x3
// [ 1 1 1
// 1 1 1 ]
//
// input node: 3x4
// [ 1 1 1
// 1 1 1
// 1 1 1
// 1 1 1 ]
func (l *Linear) Forward(xs *ts.Tensor) (retVal *ts.Tensor) {
mul, err := xs.Matmul(l.Ws, false)
or_panic(err)
if l.Bs != nil {
mul, err = mul.Add(l.Bs, false)
or_panic(err)
}
out, err := mul.Relu(false)
or_panic(err)
return out
}
// ForwardT implements ModuleT interface for Linear layer.
//
// NOTE: train param will not be used.
func (l *Linear) ForwardT(xs *ts.Tensor, train bool) (retVal *ts.Tensor) {
mul, err := xs.Matmul(l.Ws, true)
or_panic(err)
mul, err = mul.Add(l.Bs, true)
or_panic(err)
out, err := mul.Relu(true)
or_panic(err)
return out
}

View File

@@ -1,603 +0,0 @@
package my_nn
// Optimizers to be used for gradient-descent based training.
import (
"fmt"
"math"
"github.com/charmbracelet/log"
"git.andr3h3nriqu3s.com/andr3/gotch/ts"
)
// Optimizer is a struct object to run gradient descent.
type Optimizer struct {
varstore *VarStore
opt *ts.COptimizer
// variablesInOptimizer uint8
variablesInOptimizer map[string]struct{}
config OptimizerConfig //interface{}
stepCount int
lr float64
}
func (o *Optimizer) Debug() {
for n, _ := range o.variablesInOptimizer {
v := o.varstore.GetVarOfName(n)
leaf, err := v.Tensor.IsLeaf(false)
or_panic(err)
retains, err := v.Tensor.RetainsGrad(false)
or_panic(err)
log.Info("[opt] var test", "n", n, "leaf", leaf, "retains", retains)
}
}
func (o *Optimizer) RefreshValues() (err error) {
opt, err := o.config.buildCOpt(o.lr)
if err != nil {
return
}
for name := range o.variablesInOptimizer {
v := o.varstore.GetVarOfName(name)
if v.Trainable {
if err = opt.AddParameter(v.Tensor, v.Group); err != nil {
err = fmt.Errorf("Optimizer defaultBuild - AddParameter failed: %w\n", err)
return
}
}
}
o.opt = opt
return
}
// OptimizerConfig defines Optimizer configurations. These configs can be used to build optimizer.
type OptimizerConfig interface {
buildCOpt(lr float64) (*ts.COptimizer, error)
// Build builds an optimizer with the specified learning rate handling variables stored in `vs`.
//
// NOTE: Build is a 'default' method. It can be called by wrapping
// 'DefaultBuild' function
// E.g. AdamOptimizerConfig struct have a method to fullfil `Build` method of
// OptimizerConfig by wrapping `DefaultBuild` like
// (config AdamOptimizerConfig) Build(vs VarStore, lr float64) (retVal Optimizer, err error){
// return defaultBuild(config, vs, lr)
// }
Build(vs *VarStore, lr float64) (*Optimizer, error)
}
// defaultBuild is `default` Build method for OptimizerConfig interface
func defaultBuild(config OptimizerConfig, vs *VarStore, lr float64) (*Optimizer, error) {
opt, err := config.buildCOpt(lr)
if err != nil {
return nil, err
}
names := make(map[string]struct{})
for name, v := range vs.vars {
if v.Trainable {
log.Info("Adding parameter", "name", name, "g", v.Group)
if err = opt.AddParameter(v.Tensor, v.Group); err != nil {
err = fmt.Errorf("Optimizer defaultBuild - AddParameter failed: %w\n", err)
return nil, err
}
}
names[name] = struct{}{}
}
return &Optimizer{
varstore: vs,
opt: opt,
variablesInOptimizer: names,
config: config,
stepCount: 0,
lr: 0,
}, nil
}
// SGD Optimizer:
//===============
// SGDConfig holds parameters for building the SGD (Stochastic Gradient Descent) optimizer.
type SGDConfig struct {
Momentum float64
Dampening float64
Wd float64
Nesterov bool
}
// DefaultSGDConfig creates SGDConfig with default values.
func DefaultSGDConfig() *SGDConfig {
return &SGDConfig{
Momentum: 0.0,
Dampening: 0.0,
Wd: 0.0,
Nesterov: false,
}
}
// NewSGD creates the configuration for a SGD optimizer with specified values
func NewSGDConfig(momentum, dampening, wd float64, nesterov bool) *SGDConfig {
return &SGDConfig{
Momentum: momentum,
Dampening: dampening,
Wd: wd,
Nesterov: nesterov,
}
}
// Implement OptimizerConfig interface for SGDConfig
func (c *SGDConfig) buildCOpt(lr float64) (*ts.COptimizer, error) {
return ts.Sgd(lr, c.Momentum, c.Dampening, c.Wd, c.Nesterov)
}
func (c *SGDConfig) Build(vs *VarStore, lr float64) (*Optimizer, error) {
return defaultBuild(c, vs, lr)
}
// Adam optimizer:
// ===============
type AdamConfig struct {
Beta1 float64
Beta2 float64
Wd float64
}
// DefaultAdamConfig creates AdamConfig with default values
func DefaultAdamConfig() *AdamConfig {
return &AdamConfig{
Beta1: 0.9,
Beta2: 0.999,
Wd: 0.0,
}
}
// NewAdamConfig creates AdamConfig with specified values
func NewAdamConfig(beta1, beta2, wd float64) *AdamConfig {
return &AdamConfig{
Beta1: beta1,
Beta2: beta2,
Wd: wd,
}
}
// Implement OptimizerConfig interface for AdamConfig
func (c *AdamConfig) buildCOpt(lr float64) (*ts.COptimizer, error) {
return ts.Adam(lr, c.Beta1, c.Beta2, c.Wd)
}
func (c *AdamConfig) Build(vs *VarStore, lr float64) (*Optimizer, error) {
return defaultBuild(c, vs, lr)
}
// AdamW optimizer:
// ===============
type AdamWConfig struct {
Beta1 float64
Beta2 float64
Wd float64
}
// DefaultAdamWConfig creates AdamWConfig with default values
func DefaultAdamWConfig() *AdamWConfig {
return &AdamWConfig{
Beta1: 0.9,
Beta2: 0.999,
Wd: 0.01,
}
}
// NewAdamWConfig creates AdamWConfig with specified values
func NewAdamWConfig(beta1, beta2, wd float64) *AdamWConfig {
return &AdamWConfig{
Beta1: beta1,
Beta2: beta2,
Wd: wd,
}
}
// Implement OptimizerConfig interface for AdamWConfig
func (c *AdamWConfig) buildCOpt(lr float64) (*ts.COptimizer, error) {
return ts.AdamW(lr, c.Beta1, c.Beta2, c.Wd)
}
// Build builds AdamW optimizer
func (c *AdamWConfig) Build(vs *VarStore, lr float64) (*Optimizer, error) {
return defaultBuild(c, vs, lr)
}
// RMSProp optimizer:
// ===============
type RMSPropConfig struct {
Alpha float64
Eps float64
Wd float64
Momentum float64
Centered bool
}
// DefaultAdamConfig creates AdamConfig with default values
func DefaultRMSPropConfig() *RMSPropConfig {
return &RMSPropConfig{
Alpha: 0.99,
Eps: 1e-8,
Wd: 0.0,
Momentum: 0.0,
Centered: false,
}
}
// NewRMSPropConfig creates RMSPropConfig with specified values
func NewRMSPropConfig(alpha, eps, wd, momentum float64, centered bool) *RMSPropConfig {
return &RMSPropConfig{
Alpha: alpha,
Eps: eps,
Wd: wd,
Momentum: momentum,
Centered: centered,
}
}
// Implement OptimizerConfig interface for RMSPropConfig
func (c *RMSPropConfig) buildCOpt(lr float64) (*ts.COptimizer, error) {
return ts.RmsProp(lr, c.Alpha, c.Eps, c.Wd, c.Momentum, c.Centered)
}
func (c *RMSPropConfig) Build(vs *VarStore, lr float64) (*Optimizer, error) {
return defaultBuild(c, vs, lr)
}
// Optimizer methods:
// ==================
func (opt *Optimizer) addMissingVariables() {
type param struct {
tensor *ts.Tensor
group uint
}
trainables := make(map[string]param)
for name, v := range opt.varstore.vars {
if v.Trainable {
trainables[name] = param{tensor: v.Tensor, group: v.Group}
}
}
missingVariables := len(trainables) - len(opt.variablesInOptimizer)
if missingVariables > 0 {
log.Info("INFO: Optimizer.addMissingVariables()...")
for name, x := range trainables {
if _, ok := opt.variablesInOptimizer[name]; !ok {
opt.opt.AddParameter(x.tensor, x.group)
opt.variablesInOptimizer[name] = struct{}{}
}
}
}
}
// ZeroGrad zeroes the gradient for the tensors tracked by this optimizer.
func (opt *Optimizer) ZeroGrad() error {
if err := opt.opt.ZeroGrad(); err != nil {
err = fmt.Errorf("Optimizer.ZeroGrad() failed: %w\n", err)
return err
}
return nil
}
// MustZeroGrad zeroes the gradient for the tensors tracked by this optimizer.
func (opt *Optimizer) MustZeroGrad() {
err := opt.ZeroGrad()
if err != nil {
log.Fatal(err)
}
}
// Clips gradient value at some specified maximum value.
func (opt *Optimizer) ClipGradValue(max float64) {
opt.varstore.Lock()
defer opt.varstore.Unlock()
for _, v := range opt.varstore.vars {
if v.Trainable {
// v.Tensor.MustGrad().Clamp_(ts.FloatScalar(-max), ts.FloatScalar(max))
gradTs := v.Tensor.MustGrad(false)
gradTs.Clamp_(ts.FloatScalar(-max), ts.FloatScalar(max))
}
}
}
// Step performs an optimization step, updating the tracked tensors based on their gradients.
func (opt *Optimizer) Step() error {
err := opt.opt.Step()
if err != nil {
err = fmt.Errorf("Optimizer.Step() failed: %w\n", err)
return err
}
opt.stepCount += 1
return nil
}
// MustStep performs an optimization step, updating the tracked tensors based on their gradients.
func (opt *Optimizer) MustStep() {
err := opt.Step()
if err != nil {
log.Fatal(err)
}
}
// ResetStepCount set step count to zero.
func (opt *Optimizer) ResetStepCount() {
opt.stepCount = 0
}
// StepCount get current step count.
func (opt *Optimizer) StepCount() int {
return opt.stepCount
}
// BackwardStep applies a backward step pass, update the gradients, and performs an optimization step.
func (opt *Optimizer) BackwardStep(loss *ts.Tensor) error {
err := opt.opt.ZeroGrad()
if err != nil {
err = fmt.Errorf("Optimizer.BackwardStep() failed: %w\n", err)
return err
}
loss.MustBackward()
err = opt.opt.Step()
if err != nil {
err = fmt.Errorf("Optimizer.BackwardStep() failed: %w\n", err)
return err
}
return nil
}
// MustBackwardStep applies a backward step pass, update the gradients, and performs an optimization step.
func (opt *Optimizer) MustBackwardStep(loss *ts.Tensor) {
err := opt.BackwardStep(loss)
if err != nil {
log.Fatal(err)
}
}
// BackwardStepClip applies a backward step pass, update the gradients, and performs an optimization step.
//
// The gradients are clipped based on `max` before being applied.
func (opt *Optimizer) BackwardStepClip(loss *ts.Tensor, max float64) error {
err := opt.opt.ZeroGrad()
if err != nil {
err = fmt.Errorf("Optimizer.BackwardStepClip() failed: %w\n", err)
return err
}
loss.MustBackward()
opt.ClipGradValue(max)
err = opt.opt.Step()
if err != nil {
err = fmt.Errorf("Optimizer.BackwardStepClip() failed: %w\n", err)
return err
}
return nil
}
// MustBackwardStepClip applies a backward step pass, update the gradients, and performs an optimization step.
//
// The gradients are clipped based on `max` before being applied.
func (opt *Optimizer) MustBackwardStepClip(loss *ts.Tensor, max float64) {
err := opt.BackwardStepClip(loss, max)
if err != nil {
log.Fatal(err)
}
}
type ClipOpts struct {
NormType float64
ErrorIfNonFinite bool
}
type ClipOpt func(*ClipOpts)
func defaultClipOpts() *ClipOpts {
return &ClipOpts{
NormType: 2.0,
ErrorIfNonFinite: false, // will switch to "true" in the future.
}
}
func WithNormType(v float64) ClipOpt {
return func(o *ClipOpts) {
o.NormType = v
}
}
func WithErrorIfNonFinite(v bool) ClipOpt {
return func(o *ClipOpts) {
o.ErrorIfNonFinite = v
}
}
// / Clips gradient L2 norm over all trainable parameters.
//
// The norm is computed over all gradients together, as if they were
// concatenated into a single vector.
//
// / Args:
// - max: max norm of the gradient
// - o.NormType. Type of the used p-norm, can be "inf" for infinity norm. Default= 2.0
// - o.ErrorIfNonFinite bool. If true, throw error if total norm of the gradients from paramters is "nan", "inf" or "-inf". Default=false
// Returns: total norm of the parameters (viewed as a single vector)
// ref. https://github.com/pytorch/pytorch/blob/cb4aeff7d8e4c70bb638cf159878c5204d0cc2da/torch/nn/utils/clip_grad.py#L59
func (opt *Optimizer) ClipGradNorm(max float64, opts ...ClipOpt) error {
o := defaultClipOpts()
for _, option := range opts {
option(o)
}
opt.varstore.Lock()
defer opt.varstore.Unlock()
parameters := opt.varstore.TrainableVariables()
if len(parameters) == 0 {
// return ts.MustOfSlice([]float64{0.0}), nil
return nil
}
var (
norms []*ts.Tensor
totalNorm *ts.Tensor
)
device := opt.varstore.device
// FIXME. What about mixed-precision?
dtype := parameters[0].DType()
if o.NormType == math.Inf(1) {
for _, v := range opt.varstore.vars {
n := v.Tensor.MustGrad(false).MustDetach(true).MustAbs(true).MustMax(true).MustTo(device, true)
norms = append(norms, n)
}
// total_norm = norms[0] if len(norms) == 1 else torch.max(torch.stack(norms))
totalNorm = ts.MustStack(norms, 0).MustMax(true)
} else {
for _, v := range opt.varstore.vars {
// x := v.Tensor.MustGrad(false).MustNorm(true)
// NOTE. tensor.Norm() is going to be deprecated. So use linalg_norm
// Ref. https://pytorch.org/docs/stable/generated/torch.linalg.norm.html#torch.linalg.norm
x := v.Tensor.MustGrad(false).MustDetach(true).MustLinalgNorm(ts.FloatScalar(o.NormType), nil, false, dtype, true)
norms = append(norms, x)
}
}
// totalNorm = ts.MustStack(norms, 0).MustNorm(true).MustAddScalar(ts.FloatScalar(1e-6), true)
// total_norm = torch.norm(torch.stack([torch.norm(p.grad.detach(), norm_type).to(device) for p in parameters]), norm_type)
totalNorm = ts.MustStack(norms, 0).MustLinalgNorm(ts.FloatScalar(o.NormType), nil, false, dtype, true)
for _, x := range norms {
x.MustDrop()
}
totalNormVal := totalNorm.Float64Values(true)[0]
// if error_if_nonfinite and torch.logical_or(total_norm.isnan(), total_norm.isinf()):
if o.ErrorIfNonFinite && (math.IsNaN(totalNormVal) || math.IsInf(totalNormVal, 1)) {
err := fmt.Errorf("The total norm of order (%v) for gradients from 'parameters' is non-finite, so it cannot be clipped. To disable this error and scale the gradients by the non-finite norm anyway, set option.ErrorIfNonFinite= false", o.NormType)
return err
}
// clip_coef = max_norm / (total_norm + 1e-6)
// clipCoefTs := ts.TensorFrom([]float64{max}).MustDiv(totalNorm, true)
clipCoef := max / (totalNormVal + 1e-6)
// NOTE: multiplying by the clamped coef is redundant when the coef is clamped to 1, but doing so
// avoids a `if clip_coef < 1:` conditional which can require a CPU <=> device synchronization
// when the gradients do not reside in CPU memory.
// clip_coef_clamped = torch.clamp(clip_coef, max=1.0)
if clipCoef > 1.0 {
clipCoef = 1.0
}
for _, v := range opt.varstore.vars {
if v.Trainable {
// p.grad.detach().mul_(clip_coef_clamped.to(p.grad.device))
// v.Tensor.MustGrad(false).MustDetach(true).MustMulScalar_(ts.FloatScalar(clipCoef))
v.Tensor.MustGrad(false).MustMulScalar_(ts.FloatScalar(clipCoef))
}
}
return nil
}
// BackwardStepClipNorm applies a backward step pass, update the gradients, and performs an optimization step.
//
// The gradients L2 norm is clipped based on `max`.
func (opt *Optimizer) BackwardStepClipNorm(loss *ts.Tensor, max float64, opts ...ClipOpt) error {
err := opt.opt.ZeroGrad()
if err != nil {
err := fmt.Errorf("Optimizer.BackwardStepClipNorm() failed: %w\n", err)
return err
}
err = loss.Backward()
if err != nil {
err := fmt.Errorf("Optimizer.BackwardStepClipNorm() failed: %w\n", err)
return err
}
err = opt.ClipGradNorm(max, opts...)
if err != nil {
err := fmt.Errorf("Optimizer.BackwardStepClipNorm() failed: %w\n", err)
return err
}
err = opt.Step()
if err != nil {
err := fmt.Errorf("Optimizer.BackwardStepClipNorm() failed: %w\n", err)
return err
}
return nil
}
// MustBackwardStepClipNorm applies a backward step pass, update the gradients, and performs an optimization step.
//
// The gradients L2 norm is clipped based on `max`.
func (opt *Optimizer) MustBackwardStepClipNorm(loss *ts.Tensor, max float64, opts ...ClipOpt) {
err := opt.BackwardStepClipNorm(loss, max, opts...)
if err != nil {
log.Fatal(err)
}
}
// SetLR sets the optimizer learning rate.
//
// NOTE. it sets a SINGLE value of learning rate for all parameter groups.
// Most of the time, there's one parameter group.
func (opt *Optimizer) SetLR(lr float64) {
err := opt.opt.SetLearningRate(lr)
if err != nil {
log.Fatalf("Optimizer - SetLR method call error: %v\n", err)
}
}
func (opt *Optimizer) GetLRs() []float64 {
lrs, err := opt.opt.GetLearningRates()
if err != nil {
log.Fatalf("Optimizer - GetLRs method call error: %v\n", err)
}
return lrs
}
// SetLRs sets learning rates for ALL parameter groups respectively.
func (opt *Optimizer) SetLRs(lrs []float64) {
err := opt.opt.SetLearningRates(lrs)
if err != nil {
log.Fatalf("Optimizer - SetLRs method call error: %v\n", err)
}
}
// SetMomentum sets the optimizer momentum.
func (opt *Optimizer) SetMomentum(m float64) {
err := opt.opt.SetMomentum(m)
if err != nil {
log.Fatalf("Optimizer - SetMomentum method call error: %v\n", err)
}
}
func (opt *Optimizer) ParamGroupNum() int {
ngroup, err := opt.opt.ParamGroupNum()
if err != nil {
log.Fatalf("Optimizer - ParamGroupNum method call error: %v\n", err)
}
return int(ngroup)
}
func (opt *Optimizer) AddParamGroup(tensors []*ts.Tensor) {
err := opt.opt.AddParamGroup(tensors)
if err != nil {
log.Fatalf("Optimizer - ParamGroupNum method call error: %v\n", err)
}
}

View File

@@ -1,18 +0,0 @@
package my_nn
import (
torch "git.andr3h3nriqu3s.com/andr3/gotch/ts"
)
func or_panic(err error) {
if err != nil {
panic(err)
}
}
type MyLayer interface {
torch.ModuleT
ExtractFromVarstore(vs *VarStore)
Debug()
}

File diff suppressed because it is too large Load Diff

View File

@@ -1,120 +0,0 @@
package train
import (
types "git.andr3h3nriqu3s.com/andr3/fyp/logic/db_types"
my_nn "git.andr3h3nriqu3s.com/andr3/fyp/logic/models/train/torch/nn"
"git.andr3h3nriqu3s.com/andr3/gotch"
"github.com/charmbracelet/log"
torch "git.andr3h3nriqu3s.com/andr3/gotch/ts"
)
type IForwardable interface {
Forward(xs *torch.Tensor) *torch.Tensor
}
// Container for a model
type ContainerModel struct {
Layers []my_nn.MyLayer
Vs *my_nn.VarStore
path *my_nn.Path
}
func (n *ContainerModel) ForwardT(x *torch.Tensor, train bool) *torch.Tensor {
if len(n.Layers) == 0 {
return x.MustShallowClone()
}
if len(n.Layers) == 1 {
log.Info("here")
return n.Layers[0].ForwardT(x, train)
}
// forward sequentially
outs := make([]*torch.Tensor, len(n.Layers))
for i := 0; i < len(n.Layers); i++ {
if i == 0 {
outs[0] = n.Layers[i].ForwardT(x, train)
//defer outs[0].MustDrop()
} else if i == len(n.Layers)-1 {
return n.Layers[i].ForwardT(outs[i-1], train)
} else {
outs[i] = n.Layers[i].ForwardT(outs[i-1], train)
//defer outs[i].MustDrop()
}
}
panic("Do not reach here")
}
func (n *ContainerModel) To(device gotch.Device) {
n.Vs.ToDevice(device)
for _, layer := range n.Layers {
layer.ExtractFromVarstore(n.Vs)
}
}
func (n *ContainerModel) Refresh() {
for _, layer := range n.Layers {
layer.ExtractFromVarstore(n.Vs)
}
}
func BuildModel(layers []*types.Layer, _lastLinearSize int64, addSigmoid bool) *ContainerModel {
base_vs := my_nn.NewVarStore(gotch.CPU)
vs := base_vs.Root()
m_layers := []my_nn.MyLayer{}
var lastLinearSize int64 = _lastLinearSize
lastLinearConv := []int64{}
for _, layer := range layers {
if layer.LayerType == types.LAYER_INPUT {
lastLinearConv = layer.GetShape()
log.Info("Input: ", "In:", lastLinearConv)
} else if layer.LayerType == types.LAYER_DENSE {
shape := layer.GetShape()
log.Info("New Dense: ", "In:", lastLinearSize, "out:", shape[0])
m_layers = append(m_layers, NewLinear(vs, lastLinearSize, shape[0]))
lastLinearSize = shape[0]
} else if layer.LayerType == types.LAYER_FLATTEN {
m_layers = append(m_layers, NewFlatten())
lastLinearSize = 1
for _, i := range lastLinearConv {
lastLinearSize *= i
}
log.Info("Flatten: ", "In:", lastLinearConv, "out:", lastLinearSize)
} else if layer.LayerType == types.LAYER_SIMPLE_BLOCK {
panic("TODO")
log.Info("New Block: ", "In:", lastLinearConv, "out:", []int64{lastLinearConv[1] / 2, lastLinearConv[2] / 2, 128})
//m_layers = append(m_layers, NewSimpleBlock(vs, lastLinearConv[0]))
lastLinearConv[0] = 128
lastLinearConv[1] /= 2
lastLinearConv[2] /= 2
}
}
if addSigmoid {
m_layers = append(m_layers, NewSigmoid())
}
b := &ContainerModel{
Layers: m_layers,
Vs: base_vs,
path: vs,
}
return b
}
func (model *ContainerModel) Debug() {
for _, v := range model.Layers {
v.Debug()
}
}
func SaveModel(model *ContainerModel, modelFn string) (err error) {
model.Vs.ToDevice(gotch.CPU)
return model.Vs.Save(modelFn)
}

View File

@@ -1,152 +0,0 @@
package train
import (
"unsafe"
my_nn "git.andr3h3nriqu3s.com/andr3/fyp/logic/models/train/torch/nn"
"github.com/charmbracelet/log"
"git.andr3h3nriqu3s.com/andr3/gotch/nn"
torch "git.andr3h3nriqu3s.com/andr3/gotch/ts"
)
func or_panic(err error) {
if err != nil {
log.Fatal(err)
}
}
type SimpleBlock struct {
C1, C2 *nn.Conv2D
BN1 *nn.BatchNorm
}
// BasicBlock returns a BasicBlockModule instance
func NewSimpleBlock(_vs *my_nn.Path, inplanes int64) *SimpleBlock {
vs := (*nn.Path)(unsafe.Pointer(_vs))
conf1 := nn.DefaultConv2DConfig()
conf1.Stride = []int64{2, 2}
conf2 := nn.DefaultConv2DConfig()
conf2.Padding = []int64{2, 2}
b := &SimpleBlock{
C1: nn.NewConv2D(vs, inplanes, 128, 3, conf1),
C2: nn.NewConv2D(vs, 128, 128, 3, conf2),
BN1: nn.NewBatchNorm(vs, 2, 128, nn.DefaultBatchNormConfig()),
}
return b
}
// Forward method
func (b *SimpleBlock) Forward(x *torch.Tensor) *torch.Tensor {
identity := x
out := b.C1.Forward(x)
out = out.MustRelu(false)
out = b.C2.Forward(out)
out = out.MustRelu(false)
shape, err := out.Size()
or_panic(err)
out, err = out.AdaptiveAvgPool2d(shape, false)
or_panic(err)
out = b.BN1.Forward(out)
out, err = out.LeakyRelu(false)
or_panic(err)
out = out.MustAdd(identity, false)
out = out.MustRelu(false)
return out
}
func (b *SimpleBlock) ForwardT(x *torch.Tensor, train bool) *torch.Tensor {
identity := x
out := b.C1.ForwardT(x, train)
out = out.MustRelu(false)
out = b.C2.ForwardT(out, train)
out = out.MustRelu(false)
shape, err := out.Size()
or_panic(err)
out, err = out.AdaptiveAvgPool2d(shape, false)
or_panic(err)
out = b.BN1.ForwardT(out, train)
out, err = out.LeakyRelu(false)
or_panic(err)
out = out.MustAdd(identity, false)
out = out.MustRelu(false)
return out
}
// BasicBlock returns a BasicBlockModule instance
func NewLinear(vs *my_nn.Path, in, out int64) *my_nn.Linear {
config := my_nn.DefaultLinearConfig()
return my_nn.NewLinear(vs, in, out, config)
}
type Flatten struct{}
// BasicBlock returns a BasicBlockModule instance
func NewFlatten() *Flatten {
return &Flatten{}
}
// The flatten layer does not to move anything to the device
func (b *Flatten) ExtractFromVarstore(vs *my_nn.VarStore) {}
func (b *Flatten) Debug() {}
// Forward method
func (b *Flatten) Forward(x *torch.Tensor) *torch.Tensor {
out, err := x.Flatten(1, -1, false)
or_panic(err)
return out
}
func (b *Flatten) ForwardT(x *torch.Tensor, train bool) *torch.Tensor {
out, err := x.Flatten(1, -1, false)
or_panic(err)
return out
}
type Sigmoid struct{}
func NewSigmoid() *Sigmoid {
return &Sigmoid{}
}
// The sigmoid layer does not need to move anything to another device
func (b *Sigmoid) ExtractFromVarstore(vs *my_nn.VarStore) {}
func (b *Sigmoid) Debug() {}
func (b *Sigmoid) Forward(x *torch.Tensor) *torch.Tensor {
out, err := x.Sigmoid(false)
or_panic(err)
return out
}
func (b *Sigmoid) ForwardT(x *torch.Tensor, train bool) *torch.Tensor {
out, err := x.Sigmoid(false)
or_panic(err)
return out
}

View File

@@ -14,7 +14,7 @@ func handleTasksStats(handle *Handle) {
}
PostAuthJson(handle, "/stats/task/model/day", User_Normal, func(c *Context, dat *ModelTasksStatsRequest) *Error {
model, err := GetBaseModel(c, dat.ModelId)
if err == NotFoundError {
if err == ModelNotFoundError {
return c.JsonBadRequest("Model not found!")
} else if err != nil {
return c.E500M("Failed to get model", err)

7
logic/tasks/README.md Normal file
View File

@@ -0,0 +1,7 @@
# Runner Protocol
```
/----\
\/ |
Register -> Init -> Active ---> Ready -> Info
```

View File

@@ -14,7 +14,7 @@ func handleRequests(x *Handle) {
PostAuthJson(x, "/task/agreement", User_Normal, func(c *Context, dat *AgreementRequest) *Error {
var task Task
err := GetDBOnce(c, &task, "tasks where id=$1", dat.Id)
if err == NotFoundError {
if err == ModelNotFoundError {
return c.JsonBadRequest("Model not found")
} else if err != nil {
return c.E500M("Failed to get task data", err)

View File

@@ -8,4 +8,5 @@ func HandleTasks(handle *Handle) {
handleUpload(handle)
handleList(handle)
handleRequests(handle)
handleRemoteRunner(handle)
}

View File

@@ -46,7 +46,7 @@ func handleList(handler *Handle) {
if requestData.ModelId != "" {
_, err := GetBaseModel(c.Db, requestData.ModelId)
if err == NotFoundError {
if err == ModelNotFoundError {
return c.SendJSONStatus(404, "Model not found!")
} else if err != nil {
return c.Error500(err)

386
logic/tasks/runner.go Normal file
View File

@@ -0,0 +1,386 @@
package tasks
import (
"sync"
"time"
. "git.andr3h3nriqu3s.com/andr3/fyp/logic/db_types"
. "git.andr3h3nriqu3s.com/andr3/fyp/logic/models/train"
. "git.andr3h3nriqu3s.com/andr3/fyp/logic/tasks/utils"
. "git.andr3h3nriqu3s.com/andr3/fyp/logic/utils"
)
func verifyRunner(c *Context, dat *JustId) (runner *Runner, e *Error) {
runner, err := GetRunner(c, dat.Id)
if err == NotFoundError {
e = c.JsonBadRequest("Could not find runner, please register runner first")
return
} else if err != nil {
e = c.E500M("Failed to get information about the runner", err)
return
}
if runner.Token != *c.Token {
return nil, c.SendJSONStatus(401, "Only runners can use this funcion")
}
return
}
type VerifyTask struct {
Id string `json:"id" validate:"required"`
TaskId string `json:"taskId" validate:"required"`
}
func verifyTask(x *Handle, c *Context, dat *VerifyTask) (task *Task, error *Error) {
mutex := x.DataMap["runners_mutex"].(*sync.Mutex)
mutex.Lock()
defer mutex.Unlock()
var runners map[string]interface{} = x.DataMap["runners"].(map[string]interface{})
if runners[dat.Id] == nil {
return nil, c.JsonBadRequest("Runner not active")
}
var runner_data map[string]interface{} = runners[dat.Id].(map[string]interface{})
if runner_data["task"] == nil {
return nil, c.SendJSONStatus(404, "No active task")
}
return runner_data["task"].(*Task), nil
}
func handleRemoteRunner(x *Handle) {
type RegisterRunner struct {
Token string `json:"token" validate:"required"`
Type RunnerType `json:"type" validate:"required"`
}
PostAuthJson(x, "/tasks/runner/register", User_Normal, func(c *Context, dat *RegisterRunner) *Error {
if *c.Token != dat.Token {
// TODO do admin
return c.E500M("Please make sure that the token is the same that is being registered", nil)
}
c.Logger.Info("test", "dat", dat)
var runner Runner
err := GetDBOnce(c, &runner, "remote_runner as ru where token=$1", dat.Token)
if err != NotFoundError && err != nil {
return c.E500M("Failed to get information remote runners", err)
}
if err != NotFoundError {
return c.JsonBadRequest("Token is already registered by a runner")
}
// TODO get id from token passed by when doing admin
var userId = c.User.Id
var new_runner = struct {
Type RunnerType
UserId string `db:"user_id"`
Token string
}{
Type: dat.Type,
Token: dat.Token,
UserId: userId,
}
id, err := InsertReturnId(c, &new_runner, "remote_runner", "id")
if err != nil {
return c.E500M("Failed to create remote runner", err)
}
return c.SendJSON(struct {
Id string `json:"id"`
}{
Id: id,
})
})
// TODO remove runner
PostAuthJson(x, "/tasks/runner/init", User_Normal, func(c *Context, dat *JustId) *Error {
runner, error := verifyRunner(c, dat)
if error != nil {
return error
}
mutex := x.DataMap["runners_mutex"].(*sync.Mutex)
mutex.Lock()
defer mutex.Unlock()
var runners map[string]interface{} = x.DataMap["runners"].(map[string]interface{})
if runners[dat.Id] != nil {
c.Logger.Info("Logger trying to register but already registerd")
c.ShowMessage = false
return c.SendJSON("Ok")
}
var new_runner = map[string]interface{}{}
new_runner["last_time_check"] = time.Now()
new_runner["runner_info"] = runner
runners[dat.Id] = new_runner
x.DataMap["runners"] = runners
return c.SendJSON("Ok")
})
PostAuthJson(x, "/tasks/runner/active", User_Normal, func(c *Context, dat *JustId) *Error {
_, error := verifyRunner(c, dat)
if error != nil {
return error
}
mutex := x.DataMap["runners_mutex"].(*sync.Mutex)
mutex.Lock()
defer mutex.Unlock()
var runners map[string]interface{} = x.DataMap["runners"].(map[string]interface{})
if runners[dat.Id] == nil {
return c.JsonBadRequest("Runner not active")
}
var runner_data map[string]interface{} = runners[dat.Id].(map[string]interface{})
if runner_data["task"] == nil {
c.ShowMessage = false
return c.SendJSONStatus(404, "No active task")
}
c.ShowMessage = false
// This should be a task obj
return c.SendJSON(runner_data["task"])
})
PostAuthJson(x, "/tasks/runner/ready", User_Normal, func(c *Context, dat *VerifyTask) *Error {
_, error := verifyRunner(c, &JustId{Id: dat.Id})
if error != nil {
return error
}
task, error := verifyTask(x, c, dat)
if error != nil {
return error
}
err := task.UpdateStatus(c, TASK_RUNNING, "Task Running on Runner")
if err != nil {
return c.E500M("Failed to set task status", err)
}
return c.SendJSON("Ok")
})
type TaskFail struct {
Id string `json:"id" validate:"required"`
TaskId string `json:"taskId" validate:"required"`
Reason string `json:"reason" validate:"required"`
}
PostAuthJson(x, "/tasks/runner/fail", User_Normal, func(c *Context, dat *TaskFail) *Error {
_, error := verifyRunner(c, &JustId{Id: dat.Id})
if error != nil {
return error
}
task, error := verifyTask(x, c, &VerifyTask{Id: dat.Id, TaskId: dat.TaskId})
if error != nil {
return error
}
err := task.UpdateStatus(c, TASK_FAILED_RUNNING, dat.Reason)
if err != nil {
return c.E500M("Failed to set task status", err)
}
// Do extra clean up on tasks
switch task.TaskType {
case int(TASK_TYPE_TRAINING):
CleanUpFailed(c, task)
default:
panic("Do not know how to handle this")
}
mutex := x.DataMap["runners_mutex"].(*sync.Mutex)
mutex.Lock()
defer mutex.Unlock()
var runners map[string]interface{} = x.DataMap["runners"].(map[string]interface{})
var runner_data map[string]interface{} = runners[dat.Id].(map[string]interface{})
runner_data["task"] = nil
runners[dat.Id] = runner_data
x.DataMap["runners"] = runners
return c.SendJSON("Ok")
})
PostAuthJson(x, "/tasks/runner/train/defs", User_Normal, func(c *Context, dat *VerifyTask) *Error {
_, error := verifyRunner(c, &JustId{Id: dat.Id})
if error != nil {
return error
}
task, error := verifyTask(x, c, dat)
if error != nil {
return error
}
if task.TaskType != int(TASK_TYPE_TRAINING) {
c.Logger.Error("Task not is not the right type to get the definitions", "task type", task.TaskType)
return c.JsonBadRequest("Task is not the right type go get the definitions")
}
model, err := GetBaseModel(c, *task.ModelId)
if err != nil {
return c.E500M("Failed to get model information", err)
}
defs, err := model.GetDefinitions(c, "and md.status=$2", DEFINITION_STATUS_INIT)
if err != nil {
return c.E500M("Failed to get the model definitions", err)
}
return c.SendJSON(defs)
})
PostAuthJson(x, "/tasks/runner/train/classes", User_Normal, func(c *Context, dat *VerifyTask) *Error {
_, error := verifyRunner(c, &JustId{Id: dat.Id})
if error != nil {
return error
}
task, error := verifyTask(x, c, dat)
if error != nil {
return error
}
if task.TaskType != int(TASK_TYPE_TRAINING) {
c.Logger.Error("Task not is not the right type to get the definitions", "task type", task.TaskType)
return c.JsonBadRequest("Task is not the right type go get the definitions")
}
model, err := GetBaseModel(c, *task.ModelId)
if err != nil {
return c.E500M("Failed to get model information", err)
}
classes, err := model.GetClasses(c, "and status=$2 order by mc.class_order asc", CLASS_STATUS_TO_TRAIN)
if err != nil {
return c.E500M("Failed to get the model classes", err)
}
return c.SendJSON(classes)
})
type RunnerTrainDefStatus struct {
Id string `json:"id" validate:"required"`
TaskId string `json:"taskId" validate:"required"`
DefId string `json:"defId" validate:"required"`
Status DefinitionStatus `json:"status" validate:"required"`
}
PostAuthJson(x, "/tasks/runner/train/def/status", User_Normal, func(c *Context, dat *RunnerTrainDefStatus) *Error {
_, error := verifyRunner(c, &JustId{Id: dat.Id})
if error != nil {
return error
}
task, error := verifyTask(x, c, &VerifyTask{Id: dat.Id, TaskId: dat.TaskId})
if error != nil {
return error
}
if task.TaskType != int(TASK_TYPE_TRAINING) {
c.Logger.Error("Task not is not the right type to get the definitions", "task type", task.TaskType)
return c.JsonBadRequest("Task is not the right type go get the definitions")
}
def, err := GetDefinition(c, dat.DefId)
if err != nil {
return c.E500M("Failed to get definition information", err)
}
err = def.UpdateStatus(c, dat.Status)
if err != nil {
return c.E500M("Failed to update model status", err)
}
return c.SendJSON("Ok")
})
type RunnerTrainDefLayers struct {
Id string `json:"id" validate:"required"`
TaskId string `json:"taskId" validate:"required"`
DefId string `json:"defId" validate:"required"`
}
PostAuthJson(x, "/tasks/runner/train/def/layers", User_Normal, func(c *Context, dat *RunnerTrainDefLayers) *Error {
_, error := verifyRunner(c, &JustId{Id: dat.Id})
if error != nil {
return error
}
task, error := verifyTask(x, c, &VerifyTask{Id: dat.Id, TaskId: dat.TaskId})
if error != nil {
return error
}
if task.TaskType != int(TASK_TYPE_TRAINING) {
c.Logger.Error("Task not is not the right type to get the definitions", "task type", task.TaskType)
return c.JsonBadRequest("Task is not the right type go get the definitions")
}
def, err := GetDefinition(c, dat.DefId)
if err != nil {
return c.E500M("Failed to get definition information", err)
}
layers, err := def.GetLayers(c, " order by layer_order asc")
if err != nil {
return c.E500M("Failed to get layers", err)
}
return c.SendJSON(layers)
})
PostAuthJson(x, "/tasks/runner/train/datapoints", User_Normal, func(c *Context, dat *VerifyTask) *Error {
_, error := verifyRunner(c, &JustId{Id: dat.Id})
if error != nil {
return error
}
task, error := verifyTask(x, c, dat)
if error != nil {
return error
}
if task.TaskType != int(TASK_TYPE_TRAINING) {
c.Logger.Error("Task not is not the right type to get the definitions", "task type", task.TaskType)
return c.JsonBadRequest("Task is not the right type go get the definitions")
}
model, err := GetBaseModel(c, *task.ModelId)
if err != nil {
return c.E500M("Failed to get model information", err)
}
training_points, err := model.DataPoints(c, DATA_POINT_MODE_TRAINING)
if err != nil {
return c.E500M("Failed to get the model classes", err)
}
testing_points, err := model.DataPoints(c, DATA_POINT_MODE_TRAINING)
if err != nil {
return c.E500M("Failed to get the model classes", err)
}
return c.SendJSON(struct {
Testing []DataPoint `json:"testing"`
Training []DataPoint `json:"training"`
}{
Testing: testing_points,
Training: training_points,
})
})
}

View File

@@ -5,14 +5,14 @@ import (
"math"
"os"
"runtime/debug"
"sync"
"time"
"github.com/charmbracelet/log"
"git.andr3h3nriqu3s.com/andr3/fyp/logic/db"
. "git.andr3h3nriqu3s.com/andr3/fyp/logic/db_types"
// . "git.andr3h3nriqu3s.com/andr3/fyp/logic/models"
. "git.andr3h3nriqu3s.com/andr3/fyp/logic/models"
. "git.andr3h3nriqu3s.com/andr3/fyp/logic/models/train"
. "git.andr3h3nriqu3s.com/andr3/fyp/logic/tasks/utils"
. "git.andr3h3nriqu3s.com/andr3/fyp/logic/users"
@@ -53,10 +53,9 @@ func runner(config Config, db db.Db, task_channel chan Task, index int, back_cha
if task.TaskType == int(TASK_TYPE_CLASSIFICATION) {
logger.Info("Classification Task")
/*if err = ClassifyTask(base, task); err != nil {
if err = ClassifyTask(base, task); err != nil {
logger.Error("Classification task failed", "error", err)
}*/
task.UpdateStatusLog(base, TASK_FAILED_RUNNING, "TODO move tasks to pytorch")
}
back_channel <- index
continue
@@ -92,6 +91,45 @@ func runner(config Config, db db.Db, task_channel chan Task, index int, back_cha
}
}
/**
* Handle remote runner
*/
func handleRemoteTask(handler *Handle, base BasePack, runner_id string, task Task) {
logger := log.NewWithOptions(os.Stdout, log.Options{
ReportCaller: true,
ReportTimestamp: true,
TimeFormat: time.Kitchen,
Prefix: fmt.Sprintf("Runner pre %s", runner_id),
})
defer func() {
if r := recover(); r != nil {
logger.Error("Runner failed to setup for runner", "due to", r, "stack", string(debug.Stack()))
// TODO maybe create better failed task
task.UpdateStatusLog(base, TASK_FAILED_RUNNING, "Failed to setup task for runner")
}
}()
err := task.UpdateStatus(base, TASK_PICKED_UP, "Failed to setup task for runner")
if err != nil {
logger.Error("Failed to mark task as PICK UP")
return
}
mutex := handler.DataMap["runners_mutex"].(*sync.Mutex)
mutex.Lock()
defer mutex.Unlock()
switch task.TaskType {
case int(TASK_TYPE_TRAINING):
if err := PrepareTraining(handler, base, task, runner_id); err != nil {
logger.Error("Failed to prepare for training", "err", err)
}
default:
logger.Error("Not sure what to do panicing", "taskType", task.TaskType)
panic("not sure what to do")
}
}
/**
* Tells the orcchestator to look at the task list from time to time
*/
@@ -127,7 +165,7 @@ func attentionSeeker(config Config, back_channel chan int) {
/**
* Manages what worker should to Work
*/
func RunnerOrchestrator(db db.Db, config Config) {
func RunnerOrchestrator(db db.Db, config Config, handler *Handle) {
logger := log.NewWithOptions(os.Stdout, log.Options{
ReportCaller: true,
ReportTimestamp: true,
@@ -135,6 +173,16 @@ func RunnerOrchestrator(db db.Db, config Config) {
Prefix: "Runner Orchestrator Logger",
})
// Setup vars
handler.DataMap["runners"] = map[string]interface{}{}
handler.DataMap["runners_mutex"] = &sync.Mutex{}
base := BasePackStruct{
Db: db,
Logger: logger,
Host: config.Hostname,
}
gpu_workers := config.GpuWorker.NumberOfWorkers
logger.Info("Starting runners")
@@ -151,7 +199,7 @@ func RunnerOrchestrator(db db.Db, config Config) {
close(task_runners[x])
}
close(back_channel)
go RunnerOrchestrator(db, config)
go RunnerOrchestrator(db, config, handler)
}
}()
@@ -200,19 +248,45 @@ func RunnerOrchestrator(db db.Db, config Config) {
}
if task_to_dispatch != nil {
for i := 0; i < len(task_runners_used); i += 1 {
if !task_runners_used[i] {
task_runners[i] <- *task_to_dispatch
task_runners_used[i] = true
// Only let CPU tasks be done by the local users
if task_to_dispatch.TaskType == int(TASK_TYPE_DELETE_USER) {
for i := 0; i < len(task_runners_used); i += 1 {
if !task_runners_used[i] {
task_runners[i] <- *task_to_dispatch
task_runners_used[i] = true
task_to_dispatch = nil
break
}
}
continue
}
mutex := handler.DataMap["runners_mutex"].(*sync.Mutex)
mutex.Lock()
remote_runners := handler.DataMap["runners"].(map[string]interface{})
for k, v := range remote_runners {
runner_data := v.(map[string]interface{})
runner_info := runner_data["runner_info"].(*Runner)
if runner_data["task"] != nil {
continue
}
if runner_info.UserId == task_to_dispatch.UserId {
go handleRemoteTask(handler, base, k, *task_to_dispatch)
task_to_dispatch = nil
break
}
}
mutex.Unlock()
}
}
}
func StartRunners(db db.Db, config Config) {
go RunnerOrchestrator(db, config)
func StartRunners(db db.Db, config Config, handler *Handle) {
go RunnerOrchestrator(db, config, handler)
}

View File

@@ -0,0 +1,29 @@
package tasks_utils
import (
"time"
"git.andr3h3nriqu3s.com/andr3/fyp/logic/db"
dbtypes "git.andr3h3nriqu3s.com/andr3/fyp/logic/db_types"
)
type RunnerType int64
const (
RUNNER_TYPE_GPU RunnerType = iota + 1
)
type Runner struct {
Id string `json:"id" db:"ru.id"`
UserId string `json:"user_id" db:"ru.user_id"`
Token string `json:"token" db:"ru.token"`
Type RunnerType `json:"type" db:"ru.type"`
CreateOn time.Time `json:"createOn" db:"ru.created_on"`
}
func GetRunner(db db.Db, id string) (ru *Runner, err error) {
var runner Runner
err = dbtypes.GetDBOnce(db, &runner, "remote_runner as ru where ru.id=$1", id)
ru = &runner
return
}

View File

@@ -412,6 +412,18 @@ func UsersEndpints(db db.Db, handle *Handle) {
return c.SendJSON("Ok")
})
handle.DeleteAuth("/user/token/logoff", User_Normal, func(c *Context) *Error {
if c.Token == nil {
return c.JsonBadRequest("Failed to get token")
}
_, err := c.Db.Exec("delete from tokens where token=$1;", c.Token)
if err != nil {
return c.E500M("Failed to delete token", err)
}
return c.SendJSON("OK")
})
type DeleteUser struct {
Id string `json:"id" validate:"required"`
Password string `json:"password" validate:"required"`

View File

@@ -175,7 +175,7 @@ func (x *Handle) DeleteAuth(path string, authLevel dbtypes.UserType, fn func(c *
}
return fn(c)
}
x.posts = append(x.posts, HandleFunc{path, inner_fn})
x.deletes = append(x.deletes, HandleFunc{path, inner_fn})
}
func DeleteAuthJson[T interface{}](x *Handle, path string, authLevel dbtypes.UserType, fn func(c *Context, obj *T) *Error) {
@@ -374,7 +374,7 @@ func (c Context) JsonBadRequest(dat any) *Error {
c.SetReportCaller(true)
c.Logger.Warn("Request failed with a bad request", "dat", dat)
c.SetReportCaller(false)
return c.ErrorCode(nil, 404, dat)
return c.SendJSONStatus(http.StatusBadRequest, dat)
}
func (c Context) JsonErrorBadRequest(err error, dat any) *Error {
@@ -392,7 +392,7 @@ func (c *Context) GetModelFromId(id_path string) (*dbtypes.BaseModel, *Error) {
}
model, err := dbtypes.GetBaseModel(c.Db, id)
if err == dbtypes.NotFoundError {
if err == dbtypes.ModelNotFoundError {
return nil, c.SendJSONStatus(http.StatusNotFound, "Model not found")
} else if err != nil {
return nil, c.Error500(err)

View File

@@ -23,7 +23,7 @@ const (
dbname = "aistuff"
)
func main_() {
func main() {
psqlInfo := fmt.Sprintf("host=%s port=%d user=%s "+
"password=%s dbname=%s sslmode=disable",
@@ -36,11 +36,11 @@ func main_() {
log.Info("Config loaded!", "config", config)
config.GenerateToken(db)
StartRunners(db, config)
//TODO check if file structure exists to save data
handle := NewHandler(db, config)
StartRunners(db, config, handle)
config.Cleanup(db)
// TODO Handle this in other way

View File

@@ -13,7 +13,7 @@ http {
server {
listen 8000;
client_max_body_size 1G;
client_max_body_size 5G;
location / {
proxy_http_version 1.1;

6
requirements.txt Normal file
View File

@@ -0,0 +1,6 @@
# tensorflow[and-cuda] == 2.15.1
tensorflow[and-cuda] == 2.9.1
pandas
# Make sure to install the nvidia pyindex first
# nvidia-pyindex
nvidia-cudnn

5
run.sh
View File

@@ -1,3 +1,2 @@
#!/bin/fish
podman run --rm --network host --gpus all -ti -v (pwd):/app -e "TERM=xterm-256color" fyp-server bash
#!/bin/bash
podman run --rm --network host --gpus all -ti -v $(pwd):/app -e "TERM=xterm-256color" fyp-server bash

1
runner/.gitignore vendored Normal file
View File

@@ -0,0 +1 @@
target/

1936
runner/Cargo.lock generated Normal file

File diff suppressed because it is too large Load Diff

17
runner/Cargo.toml Normal file
View File

@@ -0,0 +1,17 @@
[package]
name = "runner"
version = "0.1.0"
edition = "2021"
# See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html
[dependencies]
anyhow = "1.0.82"
serde = { version = "1.0.200", features = ["derive"] }
toml = "0.8.12"
reqwest = { version = "0.12", features = ["json"] }
tokio = { version = "1", features = ["full"] }
serde_json = "1.0.116"
serde_repr = "0.1"
tch = { version = "0.16.0", features = ["download-libtorch"] }
rand = "0.8.5"

12
runner/Dockerfile Normal file
View File

@@ -0,0 +1,12 @@
FROM docker.io/nvidia/cuda:11.7.1-devel-ubuntu22.04
RUN apt-get update
RUN apt-get install -y curl
RUN curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y
ENV PATH="$PATH:/root/.cargo/bin"
RUN rustup toolchain install stable
RUN apt-get install -y pkg-config libssl-dev
WORKDIR /app

3
runner/config.toml Normal file
View File

@@ -0,0 +1,3 @@
hostname = "https://testing.andr3h3nriqu3s.com/api"
token = "d2bc41e8293937bcd9397870c98f97acc9603f742924b518e193cd1013e45d57897aa302b364001c72b458afcfb34239dfaf38a66b318e5cbc973eea"
data_path = "/home/andr3/Documents/my-repos/fyp"

1
runner/data.toml Normal file
View File

@@ -0,0 +1 @@
id = "a7cec9e9-1d05-4633-8bc5-6faabe4fd5a3"

2
runner/run.sh Executable file
View File

@@ -0,0 +1,2 @@
#!/bin/bash
podman run --rm --network host --gpus all -ti -v $(pwd):/app -e "TERM=xterm-256color" fyp-runner bash

115
runner/src/dataloader.rs Normal file
View File

@@ -0,0 +1,115 @@
use crate::{model::DataPoint, settings::ConfigFile};
use std::{path::Path, sync::Arc};
use tch::Tensor;
pub struct DataLoader {
pub batch_size: i64,
pub len: usize,
pub inputs: Vec<Tensor>,
pub labels: Vec<Tensor>,
pub pos: usize,
}
fn import_image(
item: &DataPoint,
base_path: &Path,
classes_len: i64,
inputs: &mut Vec<Tensor>,
labels: &mut Vec<Tensor>,
) {
inputs.push(
tch::vision::image::load(base_path.join(&item.path))
.ok()
.unwrap()
.unsqueeze(0),
);
if item.class >= 0 {
let t = tch::Tensor::from_slice(&[item.class]).onehot(classes_len as i64);
labels.push(t);
} else {
labels.push(tch::Tensor::zeros(
[1, classes_len as i64],
(tch::Kind::Float, tch::Device::Cpu),
))
}
}
impl DataLoader {
pub fn new(
config: Arc<ConfigFile>,
data: Vec<DataPoint>,
classes_len: i64,
batch_size: i64,
) -> DataLoader {
let len: f64 = (data.len() as f64) / (batch_size as f64);
let min_len: i64 = len.floor() as i64;
let max_len: i64 = len.ceil() as i64;
println!(
"Creating dataloader data len: {} len: {} min_len: {} max_len:{}",
data.len(),
len,
min_len,
max_len
);
let base_path = Path::new(&config.data_path);
let mut inputs: Vec<Tensor> = Vec::new();
let mut all_labels: Vec<Tensor> = Vec::new();
for batch in 0..min_len {
let mut batch_acc: Vec<Tensor> = Vec::new();
let mut labels: Vec<Tensor> = Vec::new();
for image in 0..batch_size {
let i: usize = (batch * batch_size + image).try_into().unwrap();
let item = &data[i];
import_image(item, base_path, classes_len, &mut batch_acc, &mut labels)
}
inputs.push(tch::Tensor::cat(&batch_acc[0..], 0));
all_labels.push(tch::Tensor::cat(&labels[0..], 0));
}
// Import the last batch that has irregular sizing
if min_len != max_len {
let mut batch_acc: Vec<Tensor> = Vec::new();
let mut labels: Vec<Tensor> = Vec::new();
for image in 0..(data.len() - (batch_size * min_len) as usize) {
let i: usize = (min_len * batch_size + (image as i64)) as usize;
let item = &data[i];
import_image(item, base_path, classes_len, &mut batch_acc, &mut labels);
}
inputs.push(tch::Tensor::cat(&batch_acc[0..], 0));
all_labels.push(tch::Tensor::cat(&labels[0..], 0));
}
println!("ins shape: {:?}", inputs[0].size());
return DataLoader {
batch_size,
inputs,
labels: all_labels,
len: max_len as usize,
pos: 0,
};
}
pub fn restart(self: &mut DataLoader) {
self.pos = 0;
}
pub fn next(self: &mut DataLoader) -> Option<(Tensor, Tensor)> {
if self.pos >= self.len {
return None;
}
let input = self.inputs[self.pos].empty_like();
self.inputs[self.pos] = self.inputs[self.pos].clone(&input);
let label = self.labels[self.pos].empty_like();
self.labels[self.pos] = self.labels[self.pos].clone(&label);
self.pos += 1;
return Some((input, label));
}
}

206
runner/src/main.rs Normal file
View File

@@ -0,0 +1,206 @@
mod dataloader;
mod model;
mod settings;
mod tasks;
mod training;
mod types;
use crate::settings::*;
use crate::tasks::{fail_task, Task, TaskType};
use crate::training::handle_train;
use anyhow::{bail, Result};
use reqwest::StatusCode;
use serde_json::json;
use std::{fs, process::exit, sync::Arc, time::Duration};
enum ResultAlive {
Ok,
Error,
NotInit,
}
async fn send_keep_alive_message(
config: Arc<ConfigFile>,
runner_data: Arc<RunnerData>,
) -> ResultAlive {
let client = reqwest::Client::new();
let to_send = json!({
"id": runner_data.id,
});
let resp = client
.post(format!("{}/tasks/runner/beat", config.hostname))
.header("token", &config.token)
.body(to_send.to_string())
.send()
.await;
if resp.is_err() {
return ResultAlive::Error;
}
let resp = resp.ok();
if resp.is_none() {
return ResultAlive::Error;
}
let resp = resp.unwrap();
// TODO see if the message is related to not being inited
if resp.status() != 200 {
println!("Could not connect with the status");
return ResultAlive::Error;
}
ResultAlive::Ok
}
async fn keep_alive(config: Arc<ConfigFile>, runner_data: Arc<RunnerData>) -> Result<()> {
let mut failed = 0;
loop {
match send_keep_alive_message(config.clone(), runner_data.clone()).await {
ResultAlive::Error => failed += 1,
ResultAlive::NotInit => {
println!("Runner not inited! Restarting!");
exit(1)
}
ResultAlive::Ok => failed = 0,
}
// TODO move to config
if failed > 20 {
println!("Failed to connect to API! More than 20 times in a row stoping");
exit(1)
}
tokio::time::sleep(Duration::from_secs(1)).await;
}
}
async fn handle_task(
task: Task,
config: Arc<ConfigFile>,
runner_data: Arc<RunnerData>,
) -> Result<()> {
let res = match task.task_type {
TaskType::Training => handle_train(&task, config.clone(), runner_data.clone()).await,
_ => {
println!("Do not know how to handle this task #{:?}", task);
bail!("Failed")
}
};
if res.is_err() {
println!("task failed #{:?}", res);
fail_task(
&task,
config,
runner_data,
"Do not know how to handle this kind of task",
)
.await?
}
Ok(())
}
#[tokio::main]
async fn main() -> Result<()> {
// Load config file
let config_data = fs::read_to_string("./config.toml")?;
let mut config: ConfigFile = toml::from_str(&config_data)?;
let client = reqwest::Client::new();
if config.config_path == None {
config.config_path = Some(String::from("./data.toml"))
}
let runner_data: RunnerData = load_runner_data(&config).await?;
let to_send = json!({
"id": runner_data.id,
});
// Inform the server that the runner is available
let resp = client
.post(format!("{}/tasks/runner/init", config.hostname))
.header("token", &config.token)
.body(to_send.to_string())
.send()
.await?;
if resp.status() != 200 {
println!(
"Could not connect with the api: status {} body {}",
resp.status(),
resp.text().await?
);
return Ok(());
}
let res = resp.json::<String>().await?;
if res != "Ok" {
print!("Unexpected problem: {}", res);
return Ok(());
}
let config = Arc::new(config);
let runner_data = Arc::new(runner_data);
let config_alive = config.clone();
let runner_data_alive = runner_data.clone();
std::thread::spawn(move || keep_alive(config_alive.clone(), runner_data_alive.clone()));
println!("Started main loop");
loop {
//TODO move time to config
tokio::time::sleep(Duration::from_secs(1)).await;
let to_send = json!({ "id": runner_data.id });
let resp = client
.post(format!("{}/tasks/runner/active", config.hostname))
.header("token", &config.token)
.body(to_send.to_string())
.send()
.await;
if resp.is_err() || resp.as_ref().ok().is_none() {
println!("Failed to get info from server {:?}", resp);
continue;
}
let resp = resp?;
match resp.status() {
// No active task
StatusCode::NOT_FOUND => (),
StatusCode::OK => {
println!("Found task!");
let task: Result<Task, reqwest::Error> = resp.json().await;
if task.is_err() || task.as_ref().ok().is_none() {
println!("Failed to resolve the json {:?}", task);
continue;
}
let task = task?;
let res = handle_task(task, config.clone(), runner_data.clone()).await;
if res.is_err() || res.as_ref().ok().is_none() {
println!("Failed to run the task");
}
_ = res;
()
}
_ => {
println!("Unexpected error #{:?}", resp);
exit(1)
}
}
}
}

117
runner/src/model/mod.rs Normal file
View File

@@ -0,0 +1,117 @@
use anyhow::bail;
use serde::{Deserialize, Serialize};
use serde_repr::{Deserialize_repr, Serialize_repr};
use tch::{
nn::{self, Module},
Device,
};
#[derive(Debug)]
pub struct Model {
pub vs: nn::VarStore,
pub seq: nn::Sequential,
pub layers: Vec<Layer>,
}
#[derive(Debug, Clone, Copy, Serialize_repr, Deserialize_repr)]
#[repr(i8)]
pub enum LayerType {
Input = 1,
Dense = 2,
Flatten = 3,
SimpleBlock = 4,
}
#[derive(Debug, Serialize, Deserialize)]
pub struct Layer {
pub id: String,
pub definition_id: String,
pub layer_order: String,
pub layer_type: LayerType,
pub shape: String,
pub exp_type: String,
}
#[derive(Debug, Serialize, Deserialize)]
pub struct DataPoint {
pub class: i64,
pub path: String,
}
pub fn build_model(layers: Vec<Layer>, last_linear_size: i64, add_sigmoid: bool) -> Model {
let vs = nn::VarStore::new(Device::Cuda(0));
let mut seq = nn::seq();
let mut last_linear_size = last_linear_size;
let mut last_linear_conv: Vec<i64> = Vec::new();
for layer in layers.iter() {
match layer.layer_type {
LayerType::Input => {
last_linear_conv = serde_json::from_str(&layer.shape).ok().unwrap();
println!("Layer: Input, In: {:?}", last_linear_conv);
}
LayerType::Dense => {
let shape: Vec<i64> = serde_json::from_str(&layer.shape).ok().unwrap();
println!("Layer: Dense, In: {}, Out: {}", last_linear_size, shape[0]);
seq = seq
.add(nn::linear(
&vs.root(),
last_linear_size,
shape[0],
Default::default(),
))
.add_fn(|xs| xs.relu());
last_linear_size = shape[0];
}
LayerType::Flatten => {
seq = seq.add_fn(|xs| xs.flatten(1, -1));
last_linear_size = 1;
for i in &last_linear_conv {
last_linear_size *= i;
}
println!(
"Layer: flatten, In: {:?}, Out: {}",
last_linear_conv, last_linear_size
)
}
LayerType::SimpleBlock => {
let new_last_linear_conv =
vec![128, last_linear_conv[1] / 2, last_linear_conv[2] / 2];
println!(
"Layer: block, In: {:?}, Put: {:?}",
last_linear_conv, new_last_linear_conv,
);
let out_size = vec![new_last_linear_conv[1], new_last_linear_conv[2]];
seq = seq
.add(nn::conv2d(
&vs.root(),
last_linear_conv[0],
128,
3,
nn::ConvConfig::default(),
))
.add_fn(|xs| xs.relu())
.add(nn::conv2d(
&vs.root(),
128,
128,
3,
nn::ConvConfig::default(),
))
.add_fn(|xs| xs.relu())
.add_fn(move |xs| xs.adaptive_avg_pool2d([out_size[1], out_size[1]]))
.add_fn(|xs| xs.leaky_relu());
//m_layers = append(m_layers, NewSimpleBlock(vs, lastLinearConv[0]))
last_linear_conv = new_last_linear_conv;
}
}
}
if add_sigmoid {
seq = seq.add_fn(|xs| xs.sigmoid());
}
return Model { vs, layers, seq };
}

57
runner/src/settings.rs Normal file
View File

@@ -0,0 +1,57 @@
use anyhow::{bail, Result};
use serde::{Deserialize, Serialize};
use serde_json::json;
use std::{fs, path};
#[derive(Deserialize)]
pub struct ConfigFile {
// Hostname to connect with the api
pub hostname: String,
// Token used in the api to authenticate
pub token: String,
// Path to where to store some generated configuration values
// defaults to ./data.toml
pub config_path: Option<String>,
// Data Path
// Path to where the data is mounted
pub data_path: String,
}
#[derive(Deserialize, Serialize)]
pub struct RunnerData {
pub id: String,
}
pub async fn load_runner_data(config: &ConfigFile) -> Result<RunnerData> {
let data_path = config.config_path.as_ref().unwrap();
let data_path = path::Path::new(&*data_path);
if data_path.exists() {
let runner_data = fs::read_to_string(data_path)?;
Ok(toml::from_str(&runner_data)?)
} else {
let client = reqwest::Client::new();
let to_send = json!({
"token": config.token,
"type": 1,
});
let register_resp = client
.post(format!("{}/tasks/runner/register", config.hostname))
.header("token", &config.token)
.body(to_send.to_string())
.send()
.await?;
if register_resp.status() != 200 {
bail!(format!("Could not create runner {:#?}", register_resp));
}
let runner_data: RunnerData = register_resp.json().await?;
fs::write(data_path, toml::to_string(&runner_data)?)
.expect("Faield to save data for runner");
Ok(runner_data)
}
}

90
runner/src/tasks.rs Normal file
View File

@@ -0,0 +1,90 @@
use std::sync::Arc;
use anyhow::{bail, Result};
use serde::Deserialize;
use serde_json::json;
use serde_repr::Deserialize_repr;
use crate::{ConfigFile, RunnerData};
#[derive(Clone, Copy, Deserialize_repr, Debug)]
#[repr(i8)]
pub enum TaskStatus {
FailedRunning = -2,
FailedCreation = -1,
Preparing = 0,
Todo = 1,
PickedUp = 2,
Running = 3,
Done = 4,
}
#[derive(Clone, Copy, Deserialize_repr, Debug)]
#[repr(i8)]
pub enum TaskType {
Classification = 1,
Training = 2,
Retraining = 3,
DeleteUser = 4,
}
#[derive(Deserialize, Debug)]
pub struct Task {
pub id: String,
pub user_id: String,
pub model_id: String,
pub status: TaskStatus,
pub status_message: String,
pub user_confirmed: i8,
pub compacted: i8,
#[serde(alias = "type")]
pub task_type: TaskType,
pub extra_task_info: String,
pub result: String,
pub created: String,
}
pub async fn fail_task(
task: &Task,
config: Arc<ConfigFile>,
runner_data: Arc<RunnerData>,
reason: &str,
) -> Result<()> {
println!("Marking Task as failed");
let client = reqwest::Client::new();
let to_send = json!({
"id": runner_data.id,
"taskId": task.id,
"reason": reason
});
let resp = client
.post(format!("{}/tasks/runner/fail", config.hostname))
.header("token", &config.token)
.body(to_send.to_string())
.send()
.await?;
if resp.status() != 200 {
println!("Failed to update status of task");
bail!("Failed to update status of task");
}
Ok(())
}
impl Task {
pub async fn fail(
self: &mut Task,
config: Arc<ConfigFile>,
runner_data: Arc<RunnerData>,
reason: &str,
) -> Result<()> {
fail_task(self, config, runner_data, reason).await?;
self.status = TaskStatus::FailedRunning;
self.status_message = reason.to_string();
Ok(())
}
}

599
runner/src/training.rs Normal file
View File

@@ -0,0 +1,599 @@
use crate::{
dataloader::DataLoader,
model::{self, build_model},
settings::{ConfigFile, RunnerData},
tasks::{fail_task, Task},
types::{DataPointRequest, Definition, ModelClass},
};
use std::{
io::{self, Write},
sync::Arc,
};
use anyhow::Result;
use rand::{seq::SliceRandom, thread_rng};
use serde_json::json;
use tch::{
nn::{self, Module, OptimizerConfig},
Cuda, Tensor,
};
pub async fn handle_train(
task: &Task,
config: Arc<ConfigFile>,
runner_data: Arc<RunnerData>,
) -> Result<()> {
let client = reqwest::Client::new();
println!("Preparing to train a model");
let to_send = json!({
"id": runner_data.id,
"taskId": task.id,
});
let mut defs: Vec<Definition> = client
.post(format!("{}/tasks/runner/train/defs", config.hostname))
.header("token", &config.token)
.body(to_send.to_string())
.send()
.await?
.json()
.await?;
if defs.len() == 0 {
println!("No defs found");
fail_task(task, config, runner_data, "No definitions found").await?;
return Ok(());
}
let classes: Vec<ModelClass> = client
.post(format!("{}/tasks/runner/train/classes", config.hostname))
.header("token", &config.token)
.body(to_send.to_string())
.send()
.await?
.json()
.await?;
let data: DataPointRequest = client
.post(format!("{}/tasks/runner/train/datapoints", config.hostname))
.header("token", &config.token)
.body(to_send.to_string())
.send()
.await?
.json()
.await?;
let mut testing = data.testing;
testing.shuffle(&mut thread_rng());
let mut data_loader = DataLoader::new(config.clone(), testing, classes.len() as i64, 64);
// TODO make this a vec
let mut model: Option<model::Model> = None;
loop {
let config = config.clone();
let runner_data = runner_data.clone();
let mut to_remove: Vec<usize> = Vec::new();
let mut def_iter = defs.iter_mut();
let mut i: usize = 0;
while let Some(def) = def_iter.next() {
def.updateStatus(
task,
config.clone(),
runner_data.clone(),
crate::types::DefinitionStatus::Training,
)
.await?;
let model_err = train_definition(
def,
&mut data_loader,
model,
config.clone(),
runner_data.clone(),
&task,
)
.await;
if model_err.is_err() {
println!("Failed to create model {:?}", model_err);
model = None;
to_remove.push(i);
continue;
}
model = model_err?;
i += 1;
}
defs = defs
.into_iter()
.enumerate()
.filter(|&(i, _)| to_remove.iter().any(|b| *b == i))
.map(|(_, e)| e)
.collect();
break;
}
fail_task(task, config, runner_data, "TODO").await?;
Ok(())
/*
for {
// Keep track of definitions that did not train fast enough
var toRemove ToRemoveList = []int{}
for i, def := range definitions {
accuracy, ml_model, err := trainDefinition(c, model, def, models[def.Id], classes)
if err != nil {
log.Error("Failed to train definition!Err:", "err", err)
def.UpdateStatus(c, DEFINITION_STATUS_FAILED_TRAINING)
toRemove = append(toRemove, i)
continue
}
models[def.Id] = ml_model
if accuracy >= float64(def.TargetAccuracy) {
log.Info("Found a definition that reaches target_accuracy!")
_, err = db.Exec("update model_definition set accuracy=$1, status=$2, epoch=$3 where id=$4", accuracy, DEFINITION_STATUS_TRANIED, def.Epoch, def.Id)
if err != nil {
log.Error("Failed to train definition!Err:\n", "err", err)
ModelUpdateStatus(c, model.Id, FAILED_TRAINING)
return err
}
_, err = db.Exec("update model_definition set status=$1 where id!=$2 and model_id=$3 and status!=$4", DEFINITION_STATUS_CANCELD_TRAINING, def.Id, model.Id, DEFINITION_STATUS_FAILED_TRAINING)
if err != nil {
log.Error("Failed to train definition!Err:\n", "err", err)
ModelUpdateStatus(c, model.Id, FAILED_TRAINING)
return err
}
finished = true
break
}
if def.Epoch > MAX_EPOCH {
fmt.Printf("Failed to train definition! Accuracy less %f < %d\n", accuracy, def.TargetAccuracy)
def.UpdateStatus(c, DEFINITION_STATUS_FAILED_TRAINING)
toRemove = append(toRemove, i)
continue
}
_, err = db.Exec("update model_definition set accuracy=$1, epoch=$2, status=$3 where id=$4", accuracy, def.Epoch, DEFINITION_STATUS_PAUSED_TRAINING, def.Id)
if err != nil {
log.Error("Failed to train definition!Err:\n", "err", err)
ModelUpdateStatus(c, model.Id, FAILED_TRAINING)
return err
}
}
if finished {
break
}
sort.Sort(sort.Reverse(toRemove))
log.Info("Round done", "toRemove", toRemove)
for _, n := range toRemove {
// Clean up unsed models
models[definitions[n].Id] = nil
definitions = remove(definitions, n)
}
len_def := len(definitions)
if len_def == 0 {
break
}
if len_def == 1 {
continue
}
sort.Sort(sort.Reverse(definitions))
acc := definitions[0].Accuracy - 20.0
log.Info("Training models, Highest acc", "acc", definitions[0].Accuracy, "mod_acc", acc)
toRemove = []int{}
for i, def := range definitions {
if def.Accuracy < acc {
toRemove = append(toRemove, i)
}
}
log.Info("Removing due to accuracy", "toRemove", toRemove)
sort.Sort(sort.Reverse(toRemove))
for _, n := range toRemove {
log.Warn("Removing definition not fast enough learning", "n", n)
definitions[n].UpdateStatus(c, DEFINITION_STATUS_FAILED_TRAINING)
models[definitions[n].Id] = nil
definitions = remove(definitions, n)
}
}
var def Definition
err = GetDBOnce(c, &def, "model_definition as md where md.model_id=$1 and md.status=$2 order by md.accuracy desc limit 1;", model.Id, DEFINITION_STATUS_TRANIED)
if err != nil {
if err == NotFoundError {
log.Error("All definitions failed to train!")
} else {
log.Error("DB: failed to read definition", "err", err)
}
ModelUpdateStatus(c, model.Id, FAILED_TRAINING)
return
}
if err = def.UpdateStatus(c, DEFINITION_STATUS_READY); err != nil {
log.Error("Failed to update model definition", "err", err)
ModelUpdateStatus(c, model.Id, FAILED_TRAINING)
return
}
to_delete, err := db.Query("select id from model_definition where status != $1 and model_id=$2", DEFINITION_STATUS_READY, model.Id)
if err != nil {
log.Error("Failed to select model_definition to delete")
log.Error(err)
ModelUpdateStatus(c, model.Id, FAILED_TRAINING)
return
}
defer to_delete.Close()
for to_delete.Next() {
var id string
if err = to_delete.Scan(&id); err != nil {
log.Error("Failed to scan the id of a model_definition to delete", "err", err)
ModelUpdateStatus(c, model.Id, FAILED_TRAINING)
return
}
os.RemoveAll(path.Join("savedData", model.Id, "defs", id))
}
// TODO Check if returning also works here
if _, err = db.Exec("delete from model_definition where status!=$1 and model_id=$2;", DEFINITION_STATUS_READY, model.Id); err != nil {
log.Error("Failed to delete model_definition")
log.Error(err)
ModelUpdateStatus(c, model.Id, FAILED_TRAINING)
return
}
ModelUpdateStatus(c, model.Id, READY)
return
*/
}
async fn train_definition(
def: &Definition,
data_loader: &mut DataLoader,
model: Option<model::Model>,
config: Arc<ConfigFile>,
runner_data: Arc<RunnerData>,
task: &Task,
) -> Result<Option<model::Model>> {
let client = reqwest::Client::new();
println!("About to start training definition");
let mut accuracy = 0;
let model = model.unwrap_or({
let layers: Vec<model::Layer> = client
.post(format!("{}/tasks/runner/train/def/layers", config.hostname))
.header("token", &config.token)
.body(
json!({
"id": runner_data.id,
"taskId": task.id,
"defId": def.id,
})
.to_string(),
)
.send()
.await?
.json()
.await?;
build_model(layers, 0, true)
});
// TODO CUDA
// get device
// Move model to cuda
let mut opt = nn::Adam::default().build(&model.vs, 1e-3)?;
let mut last_acc = 0.0;
for epoch in 1..40 {
data_loader.restart();
let mut mean_loss: f64 = 0.0;
let mut mean_acc: f64 = 0.0;
while let Some((inputs, labels)) = data_loader.next() {
let inputs = inputs
.to_kind(tch::Kind::Float)
.to_device(tch::Device::Cuda(0));
let labels = labels
.to_kind(tch::Kind::Float)
.to_device(tch::Device::Cuda(0));
let out = model.seq.forward(&inputs);
let weight: Option<Tensor> = None;
let loss = out.binary_cross_entropy(&labels, weight, tch::Reduction::Mean);
opt.backward_step(&loss);
mean_loss += loss
.to_device(tch::Device::Cpu)
.unsqueeze(0)
.double_value(&[0]);
let out = out.to_device(tch::Device::Cpu);
let test = out.empty_like();
_ = out.clone(&test);
let out = test.argmax(1, true);
let mut labels = labels.to_device(tch::Device::Cpu);
labels = labels.unsqueeze(-1);
let size = out.size()[0];
let mut acc = 0;
for i in 0..size {
let res = out.double_value(&[i]);
let exp = labels.double_value(&[i, res as i64]);
if exp == 1.0 {
acc += 1;
}
}
mean_acc += acc as f64 / size as f64;
last_acc = acc as f64 / size as f64;
}
print!(
"\repoch: {} loss: {} acc: {} l acc: {} ",
epoch,
mean_loss / data_loader.len as f64,
mean_acc / data_loader.len as f64,
last_acc
);
io::stdout().flush().expect("Unable to flush stdout");
}
println!("\nlast acc: {}", last_acc);
return Ok(Some(model));
/*
opt, err := my_nn.DefaultAdamConfig().Build(model.Vs, 0.001)
if err != nil {
return
}
for epoch := 0; epoch < EPOCH_PER_RUN; epoch++ {
var trainIter *torch.Iter2
trainIter, err = ds.TrainIter(32)
if err != nil {
return
}
// trainIter.ToDevice(device)
log.Info("epoch", "epoch", epoch)
var trainLoss float64 = 0
var trainCorrect float64 = 0
ok := true
for ok {
var item torch.Iter2Item
var loss *torch.Tensor
item, ok = trainIter.Next()
if !ok {
continue
}
data := item.Data
data, err = data.ToDevice(device, gotch.Float, false, true, false)
if err != nil {
return
}
var size []int64
size, err = data.Size()
if err != nil {
return
}
var zeros *torch.Tensor
zeros, err = torch.Zeros(size, gotch.Float, device)
if err != nil {
return
}
data, err = zeros.Add(data, true)
if err != nil {
return
}
log.Info("\n\nhere 1, data\n\n", "retains", data.MustRetainsGrad(false), "requires", data.MustRequiresGrad())
data, err = data.SetRequiresGrad(true, false)
if err != nil {
return
}
log.Info("\n\nhere 2, data\n\n", "retains", data.MustRetainsGrad(false), "requires", data.MustRequiresGrad())
err = data.RetainGrad(false)
if err != nil {
return
}
log.Info("\n\nhere 2, data\n\n", "retains", data.MustRetainsGrad(false), "requires", data.MustRequiresGrad())
pred := model.ForwardT(data, true)
pred, err = pred.SetRequiresGrad(true, true)
if err != nil {
return
}
err = pred.RetainGrad(false)
if err != nil {
return
}
label := item.Label
label, err = label.ToDevice(device, gotch.Float, false, true, false)
if err != nil {
return
}
label, err = label.SetRequiresGrad(true, true)
if err != nil {
return
}
err = label.RetainGrad(false)
if err != nil {
return
}
// Calculate loss
loss, err = pred.BinaryCrossEntropyWithLogits(label, &torch.Tensor{}, &torch.Tensor{}, 2, false)
if err != nil {
return
}
loss, err = loss.SetRequiresGrad(true, false)
if err != nil {
return
}
err = loss.RetainGrad(false)
if err != nil {
return
}
err = opt.ZeroGrad()
if err != nil {
return
}
err = loss.Backward()
if err != nil {
return
}
log.Info("pred grad", "pred", pred.MustGrad(false).MustMax(false).Float64Values())
log.Info("pred grad", "outs", label.MustGrad(false).MustMax(false).Float64Values())
log.Info("pred grad", "data", data.MustGrad(false).MustMax(false).Float64Values(), "lol", data.MustRetainsGrad(false))
vars := model.Vs.Variables()
for k, v := range vars {
log.Info("[grad check]", "k", k, "grad", v.MustGrad(false).MustMax(false).Float64Values(), "lol", v.MustRetainsGrad(false))
}
model.Debug()
err = opt.Step()
if err != nil {
return
}
trainLoss = loss.Float64Values()[0]
// Calculate accuracy
/ *var p_pred, p_labels *torch.Tensor
p_pred, err = pred.Argmax([]int64{1}, true, false)
if err != nil {
return
}
p_labels, err = item.Label.Argmax([]int64{1}, true, false)
if err != nil {
return
}
floats := p_pred.Float64Values()
floats_labels := p_labels.Float64Values()
for i := range floats {
if floats[i] == floats_labels[i] {
trainCorrect += 1
}
} * /
panic("fornow")
}
//v := []float64{}
log.Info("model training epoch done loss", "loss", trainLoss, "correct", trainCorrect, "out", ds.TrainImagesSize, "accuracy", trainCorrect/float64(ds.TrainImagesSize))
/ *correct := int64(0)
//torch.NoGrad(func() {
ok = true
testIter := ds.TestIter(64)
for ok {
var item torch.Iter2Item
item, ok = testIter.Next()
if !ok {
continue
}
output := model.Forward(item.Data)
var pred, labels *torch.Tensor
pred, err = output.Argmax([]int64{1}, true, false)
if err != nil {
return
}
labels, err = item.Label.Argmax([]int64{1}, true, false)
if err != nil {
return
}
floats := pred.Float64Values()
floats_labels := labels.Float64Values()
for i := range floats {
if floats[i] == floats_labels[i] {
correct += 1
}
}
}
accuracy = float64(correct) / float64(ds.TestImagesSize)
log.Info("Eval accuracy", "accuracy", accuracy)
err = def.UpdateAfterEpoch(db, accuracy*100)
if err != nil {
return
}* /
//})
}
result_path := path.Join(getDir(), "savedData", m.Id, "defs", def.Id)
err = os.MkdirAll(result_path, os.ModePerm)
if err != nil {
return
}
err = my_torch.SaveModel(model, path.Join(result_path, "model.dat"))
if err != nil {
return
}
log.Info("Model finished training!", "accuracy", accuracy)
return
*/
}

89
runner/src/types.rs Normal file
View File

@@ -0,0 +1,89 @@
use crate::{model, tasks::Task, ConfigFile, RunnerData};
use anyhow::{bail, Result};
use serde::Deserialize;
use serde_json::json;
use serde_repr::{Deserialize_repr, Serialize_repr};
use std::sync::Arc;
#[derive(Clone, Copy, Deserialize_repr, Serialize_repr, Debug)]
#[repr(i8)]
pub enum DefinitionStatus {
CanceldTraining = -4,
FailedTraining = -3,
PreInit = 1,
Init = 2,
Training = 3,
PausedTraining = 6,
Tranied = 4,
Ready = 5,
}
#[derive(Deserialize, Debug)]
pub struct Definition {
pub id: String,
pub model_id: String,
pub accuracy: f64,
pub target_accuracy: i64,
pub epoch: i64,
pub status: i64,
pub created: String,
pub epoch_progress: i64,
}
impl Definition {
pub async fn updateStatus(
self: &mut Definition,
task: &Task,
config: Arc<ConfigFile>,
runner_data: Arc<RunnerData>,
status: DefinitionStatus,
) -> Result<()> {
println!("Marking Task as faield");
let client = reqwest::Client::new();
let to_send = json!({
"id": runner_data.id,
"taskId": task.id,
"defId": self.id,
"status": status,
});
let resp = client
.post(format!("{}/tasks/runner/train/def/status", config.hostname))
.header("token", &config.token)
.body(to_send.to_string())
.send()
.await?;
if resp.status() != 200 {
println!("Failed to update status of task");
bail!("Failed to update status of task");
}
Ok(())
}
}
#[derive(Clone, Copy, Deserialize_repr, Debug)]
#[repr(i8)]
pub enum ModelClassStatus {
ToTrain = 1,
Training = 2,
Trained = 3,
}
#[derive(Deserialize, Debug)]
pub struct ModelClass {
pub id: String,
pub model_id: String,
pub name: String,
pub class_order: i64,
pub status: ModelClassStatus,
}
#[derive(Deserialize, Debug)]
pub struct DataPointRequest {
pub testing: Vec<model::DataPoint>,
pub training: Vec<model::DataPoint>,
}

View File

@@ -38,3 +38,14 @@ create table if not exists tasks_dependencies (
main_id uuid references tasks (id) on delete cascade not null,
dependent_id uuid references tasks (id) on delete cascade not null
);
create table if not exists remote_runner (
id uuid primary key default gen_random_uuid(),
user_id uuid references users (id) on delete cascade not null,
token text not null,
-- 1: GPU
type integer,
created_on timestamp default current_timestamp
);

120
test.go
View File

@@ -1,120 +0,0 @@
package main
import (
"git.andr3h3nriqu3s.com/andr3/gotch"
dbtypes "git.andr3h3nriqu3s.com/andr3/fyp/logic/db_types"
"git.andr3h3nriqu3s.com/andr3/fyp/logic/models/train/torch"
//my_nn "git.andr3h3nriqu3s.com/andr3/fyp/logic/models/train/torch/nn"
torch "git.andr3h3nriqu3s.com/andr3/gotch/ts"
"github.com/charmbracelet/log"
)
func main() {
log.Info("Hello world")
m := train.BuildModel([]*dbtypes.Layer{
&dbtypes.Layer{
LayerType: dbtypes.LAYER_INPUT,
Shape: "[ 3, 28, 28 ]",
},
&dbtypes.Layer{
LayerType: dbtypes.LAYER_FLATTEN,
},
&dbtypes.Layer{
LayerType: dbtypes.LAYER_DENSE,
Shape: "[ 27 ]",
},
&dbtypes.Layer{
LayerType: dbtypes.LAYER_DENSE,
Shape: "[ 18 ]",
},
// &dbtypes.Layer{
// LayerType: dbtypes.LAYER_DENSE,
// Shape: "[ 9 ]",
// },
}, 0, true)
//var err error
d := gotch.CudaIfAvailable()
log.Info("device", "d", d)
m.To(d)
var count = 0
// vars1 := m.Vs.Variables()
//
// for k, v := range vars1 {
// ones := torch.MustOnes(v.MustSize(), gotch.Float, d)
// v := ones.MustSetRequiresGrad(true, false)
// v.MustDrop()
// ones.RetainGrad(false)
//
// m.Vs.UpdateVarTensor(k, ones, true)
// m.Refresh()
// }
//
// opt, err := my_nn.DefaultAdamConfig().Build(m.Vs, 0.001)
// if err != nil {
// return
// }
log.Info("start")
for count < 100 {
ones := torch.MustOnes([]int64{1, 3, 28, 28}, gotch.Float, d)
// ones = ones.MustSetRequiresGrad(true, true)
// ones.RetainGrad(false)
res := m.ForwardT(ones, true)
//res = res.MustSetRequiresGrad(true, true)
//res.RetainGrad(false)
outs := torch.MustZeros([]int64{1, 18}, gotch.Float, d)
loss, err := res.BinaryCrossEntropyWithLogits(outs, &torch.Tensor{}, &torch.Tensor{}, 2, false)
if err != nil {
log.Fatal(err)
}
// loss = loss.MustSetRequiresGrad(true, true)
//opt.ZeroGrad()
log.Info("loss", "loss", loss.Float64Values())
loss.MustBackward()
//opt.Step()
// log.Info(mean.MustGrad(false).Float64Values())
//ones_grad = ones.MustGrad(true).MustMax(true).Float64Values()[0]
// log.Info(res.MustGrad(true).MustMax(true).Float64Values())
// log.Info(ones_grad)
vars := m.Vs.Variables()
for k, v := range vars {
log.Info("[grad check]", "k", k, "grad", v.MustGrad(false).MustMax(true).Float64Values())
}
m.Debug()
outs.MustDrop()
count += 1
log.Fatal("grad zero")
}
log.Warn("out")
}

View File

@@ -9,9 +9,9 @@ import requests
class NotifyServerCallback(tf.keras.callbacks.Callback):
def on_epoch_end(self, epoch, log, *args, **kwargs):
{{ if .HeadId }}
requests.get(f'{{ .Host }}/api/model/head/epoch/update?epoch={epoch + 1}&accuracy={log["accuracy"]}&head_id={{.HeadId}}')
requests.get(f'{{ .Host }}/api/model/head/epoch/update?epoch={epoch + 1}&accuracy={log["val_accuracy"]}&head_id={{.HeadId}}')
{{ else }}
requests.get(f'{{ .Host }}/api/model/epoch/update?model_id={{.Model.Id}}&epoch={epoch + 1}&accuracy={log["accuracy"]}&definition={{.DefId}}')
requests.get(f'{{ .Host }}/api/model/epoch/update?model_id={{.Model.Id}}&epoch={epoch + 1}&accuracy={log["val_accuracy"]}&definition={{.DefId}}')
{{end}}
@@ -82,7 +82,7 @@ def prepare_dataset(ds: tf.data.Dataset, size: int) -> tf.data.Dataset:
def filterDataset(path):
path = tf.strings.regex_replace(path, DATA_DIR_PREPARE, "")
{{ if eq .Model.Format "png" }}
path = tf.strings.regex_replace(path, ".png", "")
{{ else if eq .Model.Format "jpeg" }}
@@ -90,7 +90,7 @@ def filterDataset(path):
{{ else }}
ERROR
{{ end }}
return tf.reshape(table.lookup(tf.strings.as_string([path])), []) != -1
seed = random.randint(0, 100000000)
@@ -135,9 +135,9 @@ def addBlock(
model.add(layers.ReLU())
if top:
if pooling_same:
model.add(pool_func(padding="same", strides=(1, 1)))
model.add(pool_func(pool_size=(2,2), padding="same", strides=(1, 1)))
else:
model.add(pool_func())
model.add(pool_func(pool_size=(2,2)))
model.add(layers.BatchNormalization())
model.add(layers.LeakyReLU())
model.add(layers.Dropout(0.4))
@@ -172,7 +172,7 @@ model.compile(
his = model.fit(dataset, validation_data= dataset_validation, epochs={{.EPOCH_PER_RUN}}, callbacks=[
NotifyServerCallback(),
tf.keras.callbacks.EarlyStopping("loss", mode="min", patience=5)], use_multiprocessing = True)
tf.keras.callbacks.EarlyStopping("loss", mode="min", patience=5)])
acc = his.history["accuracy"]

View File

@@ -10,7 +10,7 @@ import numpy as np
class NotifyServerCallback(tf.keras.callbacks.Callback):
def on_epoch_end(self, epoch, log, *args, **kwargs):
requests.get(f'{{ .Host }}/api/model/head/epoch/update?epoch={epoch + 1}&accuracy={log["accuracy"]}&head_id={{.HeadId}}')
requests.get(f'{{ .Host }}/api/model/head/epoch/update?epoch={epoch + 1}&accuracy={log["val_accuracy"]}&head_id={{.HeadId}}')
DATA_DIR = "{{ .DataDir }}"

View File

@@ -1,4 +1,5 @@
import { goto } from '$app/navigation';
import { rdelete } from '$lib/requests.svelte';
type User = {
token: string;
@@ -33,6 +34,10 @@ export function createUserStore() {
if (value) {
localStorage.setItem('user', JSON.stringify(value));
} else {
if (user) {
// Request the deletion of the token
rdelete('/user/token/logoff', {});
}
localStorage.removeItem('user');
}
user = value;

View File

@@ -215,7 +215,7 @@
</div>
{:else if m.status == -3 || m.status == -4}
<BaseModelInfo model={m} />
<form on:submit|preventDefault={resetModel}>
<form on:submit={resetModel}>
Failed Prepare for training.<br />
<div class="spacer"></div>
<MessageSimple bind:this={resetMessages} />