fyp/logic/models/train/train.go
2023-10-25 14:50:58 +01:00

762 lines
20 KiB
Go

package models_train
import (
"database/sql"
"errors"
"fmt"
"io"
"math"
"net/http"
"os"
"os/exec"
"path"
"sort"
"strconv"
"text/template"
model_classes "git.andr3h3nriqu3s.com/andr3/fyp/logic/models/classes"
. "git.andr3h3nriqu3s.com/andr3/fyp/logic/models/utils"
. "git.andr3h3nriqu3s.com/andr3/fyp/logic/utils"
)
const EPOCH_PER_RUN = 20
const MAX_EPOCH = 100
func MakeDefenition(db *sql.DB, model_id string, target_accuracy int) (id string, err error) {
id = ""
rows, err := db.Query("insert into model_definition (model_id, target_accuracy) values ($1, $2) returning id;", model_id, target_accuracy)
if err != nil {
return
}
defer rows.Close()
if !rows.Next() {
return id, errors.New("Something wrong!")
}
err = rows.Scan(&id)
return
}
func ModelDefinitionUpdateStatus(c *Context, id string, status ModelDefinitionStatus) (err error) {
_, err = c.Db.Exec("update model_definition set status = $1 where id = $2", status, id)
return
}
func MakeLayer(db *sql.DB, def_id string, layer_order int, layer_type LayerType, shape string) (err error) {
_, err = db.Exec("insert into model_definition_layer (def_id, layer_order, layer_type, shape) values ($1, $2, $3, $4)", def_id, layer_order, layer_type, shape)
return
}
func generateCvs(c *Context, run_path string, model_id string) (count int, err error) {
classes, err := c.Db.Query("select count(*) from model_classes where model_id=$1;", model_id)
if err != nil {
return
}
defer classes.Close()
if !classes.Next() {
return
}
if err = classes.Scan(&count); err != nil {
return
}
data, err := c.Db.Query("select mdp.id, mc.class_order, mdp.file_path from model_data_point as mdp inner join model_classes as mc on mc.id = mdp.class_id where mc.model_id = $1 and mdp.model_mode=$2;", model_id, model_classes.DATA_POINT_MODE_TRAINING)
if err != nil {
return
}
defer data.Close()
f, err := os.Create(path.Join(run_path, "train.csv"))
if err != nil {
return
}
defer f.Close()
f.Write([]byte("Id,Index\n"))
for data.Next() {
var id string
var class_order int
var file_path string
if err = data.Scan(&id, &class_order, &file_path); err != nil {
return
}
if file_path == "id://" {
f.Write([]byte(id + "," + strconv.Itoa(class_order) + "\n"))
} else {
return count, errors.New("TODO generateCvs to file_path " + file_path)
}
}
return
}
func trainDefinition(c *Context, model *BaseModel, definition_id string, load_prev bool) (accuracy float64, err error) {
c.Logger.Warn("About to start training definition")
accuracy = 0
layers, err := c.Db.Query("select layer_type, shape from model_definition_layer where def_id=$1 order by layer_order asc;", definition_id)
if err != nil {
return
}
defer layers.Close()
type layerrow struct {
LayerType int
Shape string
}
got := []layerrow{}
for layers.Next() {
var row = layerrow{}
if err = layers.Scan(&row.LayerType, &row.Shape); err != nil {
return
}
row.Shape = shapeToSize(row.Shape)
got = append(got, row)
}
// Generate run folder
run_path := path.Join("/tmp", model.Id, "defs", definition_id)
err = os.MkdirAll(run_path, os.ModePerm)
if err != nil {
return
}
defer os.RemoveAll(run_path)
_, err = generateCvs(c, run_path, model.Id)
if err != nil {
return
}
// Create python script
f, err := os.Create(path.Join(run_path, "run.py"))
if err != nil {
return
}
defer f.Close()
tmpl, err := template.New("python_model_template.py").ParseFiles("views/py/python_model_template.py")
if err != nil {
return
}
// Copy result around
result_path := path.Join("savedData", model.Id, "defs", definition_id)
if err = tmpl.Execute(f, AnyMap{
"Layers": got,
"Size": got[0].Shape,
"DataDir": path.Join(getDir(), "savedData", model.Id, "data"),
"RunPath": run_path,
"ColorMode": model.ImageMode,
"Model": model,
"EPOCH_PER_RUN": EPOCH_PER_RUN,
"DefId": definition_id,
"LoadPrev": load_prev,
"LastModelRunPath": path.Join(getDir(), result_path, "model.keras"),
"SaveModelPath": path.Join(getDir(), result_path),
}); err != nil {
return
}
// Run the command
out, err := exec.Command("bash", "-c", fmt.Sprintf("cd %s && python run.py", run_path)).CombinedOutput()
if err != nil {
c.Logger.Debug(string(out))
return
}
c.Logger.Info("Python finished running")
if err = os.MkdirAll(result_path, os.ModePerm); err != nil {
return
}
accuracy_file, err := os.Open(path.Join(run_path, "accuracy.val"))
if err != nil {
return
}
defer accuracy_file.Close()
accuracy_file_bytes, err := io.ReadAll(accuracy_file)
if err != nil {
return
}
accuracy, err = strconv.ParseFloat(string(accuracy_file_bytes), 64)
if err != nil {
return
}
c.Logger.Info("Model finished training!", "accuracy", accuracy)
return
}
func remove[T interface{}](lst []T, i int) []T {
lng := len(lst)
if i >= lng {
return []T{}
}
if i+1 >= lng {
return lst[:lng-1]
}
if i == 0 {
return lst[1:]
}
return append(lst[:i], lst[i+1:]...)
}
type TrainModelRow struct {
id string
target_accuracy int
epoch int
acuracy float64
}
type TraingModelRowDefinitions []TrainModelRow
func (nf TraingModelRowDefinitions) Len() int { return len(nf) }
func (nf TraingModelRowDefinitions) Swap(i, j int) { nf[i], nf[j] = nf[j], nf[i] }
func (nf TraingModelRowDefinitions) Less(i, j int) bool {
return nf[i].acuracy < nf[j].acuracy
}
type ToRemoveList []int
func (nf ToRemoveList) Len() int { return len(nf) }
func (nf ToRemoveList) Swap(i, j int) { nf[i], nf[j] = nf[j], nf[i] }
func (nf ToRemoveList) Less(i, j int) bool {
return nf[i] < nf[j]
}
func trainModel(c *Context, model *BaseModel) {
definitionsRows, err := c.Db.Query("select id, target_accuracy, epoch from model_definition where status=$1 and model_id=$2", MODEL_DEFINITION_STATUS_INIT, model.Id)
if err != nil {
c.Logger.Error("Failed to trainModel!Err:")
c.Logger.Error(err)
ModelUpdateStatus(c, model.Id, FAILED_TRAINING)
return
}
defer definitionsRows.Close()
var definitions TraingModelRowDefinitions = []TrainModelRow{}
for definitionsRows.Next() {
var rowv TrainModelRow
rowv.acuracy = 0
if err = definitionsRows.Scan(&rowv.id, &rowv.target_accuracy, &rowv.epoch); err != nil {
c.Logger.Error("Failed to train Model Could not read definition from db!Err:")
c.Logger.Error(err)
ModelUpdateStatus(c, model.Id, FAILED_TRAINING)
return
}
definitions = append(definitions, rowv)
}
if len(definitions) == 0 {
c.Logger.Error("No Definitions defined!")
ModelUpdateStatus(c, model.Id, FAILED_TRAINING)
return
}
firstRound := true
finished := false
for {
var toRemove ToRemoveList = []int{}
for i, def := range definitions {
ModelDefinitionUpdateStatus(c, def.id, MODEL_DEFINITION_STATUS_TRAINING)
accuracy, err := trainDefinition(c, model, def.id, !firstRound)
if err != nil {
c.Logger.Error("Failed to train definition!Err:", "err", err)
ModelDefinitionUpdateStatus(c, def.id, MODEL_DEFINITION_STATUS_FAILED_TRAINING)
toRemove = append(toRemove, i)
continue
}
def.epoch += EPOCH_PER_RUN
accuracy = accuracy * 100
def.acuracy = float64(accuracy)
definitions[i].epoch += EPOCH_PER_RUN
definitions[i].acuracy = accuracy
if accuracy >= float64(def.target_accuracy) {
c.Logger.Info("Found a definition that reaches target_accuracy!")
_, err = c.Db.Exec("update model_definition set accuracy=$1, status=$2, epoch=$3 where id=$4", accuracy, MODEL_DEFINITION_STATUS_TRANIED, def.epoch, def.id)
if err != nil {
c.Logger.Error("Failed to train definition!Err:\n", "err", err)
ModelUpdateStatus(c, model.Id, FAILED_TRAINING)
return
}
_, err = c.Db.Exec("update model_definition set status=$1 where id!=$2 and model_id=$3 and status!=$4", MODEL_DEFINITION_STATUS_CANCELD_TRAINING, def.id, model.Id, MODEL_DEFINITION_STATUS_FAILED_TRAINING)
if err != nil {
c.Logger.Error("Failed to train definition!Err:\n", "err", err)
ModelUpdateStatus(c, model.Id, FAILED_TRAINING)
return
}
finished = true
break
}
if def.epoch > MAX_EPOCH {
fmt.Printf("Failed to train definition! Accuracy less %f < %d\n", accuracy, def.target_accuracy)
ModelDefinitionUpdateStatus(c, def.id, MODEL_DEFINITION_STATUS_FAILED_TRAINING)
toRemove = append(toRemove, i)
continue
}
_, err = c.Db.Exec("update model_definition set accuracy=$1, epoch=$2, status=$3 where id=$4", accuracy, def.epoch, MODEL_DEFINITION_STATUS_PAUSED_TRAINING, def.id)
if err != nil {
c.Logger.Error("Failed to train definition!Err:\n", "err", err)
ModelUpdateStatus(c, model.Id, FAILED_TRAINING)
return
}
}
firstRound = false
if finished {
break
}
sort.Sort(sort.Reverse(toRemove))
c.Logger.Info("Round done", "toRemove", toRemove)
for _, n := range toRemove {
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].acuracy - 20.0
c.Logger.Info("Training models, Highest acc", "acc", definitions[0].acuracy, "mod_acc", acc)
toRemove = []int{}
for i, def := range definitions {
if def.acuracy < acc {
toRemove = append(toRemove, i)
}
}
c.Logger.Info("Removing due to accuracy", "toRemove", toRemove)
sort.Sort(sort.Reverse(toRemove))
for _, n := range toRemove {
c.Logger.Warn("Removing definition not fast enough learning", "n", n)
ModelDefinitionUpdateStatus(c, definitions[n].id, MODEL_DEFINITION_STATUS_FAILED_TRAINING)
definitions = remove(definitions, n)
}
}
rows, err := c.Db.Query("select id from model_definition where model_id=$1 and status=$2 order by accuracy desc limit 1;", model.Id, MODEL_DEFINITION_STATUS_TRANIED)
if err != nil {
c.Logger.Error("DB: failed to read definition")
c.Logger.Error(err)
ModelUpdateStatus(c, model.Id, FAILED_TRAINING)
return
}
defer rows.Close()
if !rows.Next() {
// TODO Make the Model status have a message
c.Logger.Error("All definitions failed to train!")
ModelUpdateStatus(c, model.Id, FAILED_TRAINING)
return
}
var id string
if err = rows.Scan(&id); err != nil {
c.Logger.Error("Failed to read id:")
c.Logger.Error(err)
ModelUpdateStatus(c, model.Id, FAILED_TRAINING)
return
}
if _, err = c.Db.Exec("update model_definition set status=$1 where id=$2;", MODEL_DEFINITION_STATUS_READY, id); err != nil {
c.Logger.Error("Failed to update model definition")
c.Logger.Error(err)
ModelUpdateStatus(c, model.Id, FAILED_TRAINING)
return
}
to_delete, err := c.Db.Query("select id from model_definition where status != $1 and model_id=$2", MODEL_DEFINITION_STATUS_READY, model.Id)
if err != nil {
c.Logger.Error("Failed to select model_definition to delete")
c.Logger.Error(err)
ModelUpdateStatus(c, model.Id, FAILED_TRAINING)
return
}
defer to_delete.Close()
for to_delete.Next() {
var id string
if to_delete.Scan(&id); err != nil {
c.Logger.Error("Failed to scan the id of a model_definition to delete")
c.Logger.Error(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 = c.Db.Exec("delete from model_definition where status!=$1 and model_id=$2;", MODEL_DEFINITION_STATUS_READY, model.Id); err != nil {
c.Logger.Error("Failed to delete model_definition")
c.Logger.Error(err)
ModelUpdateStatus(c, model.Id, FAILED_TRAINING)
return
}
ModelUpdateStatus(c, model.Id, READY)
}
func removeFailedDataPoints(c *Context, model *BaseModel) (err error) {
rows, err := c.Db.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;", model.Id)
if err != nil {
return
}
defer rows.Close()
base_path := path.Join("savedData", model.Id, "data")
for rows.Next() {
var dataPointId string
err = rows.Scan(&dataPointId)
if err != nil {
return
}
p := path.Join(base_path, dataPointId+"."+model.Format)
c.Logger.Warn("Removing image", "path", p)
err = os.RemoveAll(p)
if err != nil {
return
}
}
_, err = c.Db.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;", model.Id)
return
}
// This generates a definition
func generateDefinition(c *Context, model *BaseModel, target_accuracy int, number_of_classes int, complexity int) *Error {
var err error = nil
failed := func() *Error {
ModelUpdateStatus(c, model.Id, FAILED_PREPARING_TRAINING)
// TODO improve this response
return c.Error500(err)
}
def_id, err := MakeDefenition(c.Db, model.Id, target_accuracy)
if err != nil {
return failed()
}
order := 1
// Note the shape of the first layer defines the import size
if complexity == 2 {
// Note the shape for now is no used
width := int(math.Pow(2, math.Floor(math.Log(float64(model.Width))/math.Log(2.0))))
height := int(math.Pow(2, math.Floor(math.Log(float64(model.Height))/math.Log(2.0))))
c.Logger.Warn("Complexity 2 creating model with smaller size", "width", width, "height", height)
err = MakeLayer(c.Db, def_id, order, LAYER_INPUT, fmt.Sprintf("%d,%d,1", width, height))
if err != nil {
return failed()
}
order++
} else {
err = MakeLayer(c.Db, def_id, order, LAYER_INPUT, fmt.Sprintf("%d,%d,1", model.Width, model.Height))
if err != nil {
return failed()
}
order++
}
if complexity == 0 {
err = MakeLayer(c.Db, def_id, order, LAYER_FLATTEN, "")
if err != nil {
return failed()
}
order++
loop := int(math.Log2(float64(number_of_classes)))
for i := 0; i < loop; i++ {
err = MakeLayer(c.Db, def_id, order, LAYER_DENSE, fmt.Sprintf("%d,1", number_of_classes*(loop-i)))
order++
if err != nil {
ModelUpdateStatus(c, model.Id, FAILED_PREPARING_TRAINING)
// TODO improve this response
return c.Error500(err)
}
}
} else if complexity == 1 {
loop := int((math.Log(float64(model.Width)) / math.Log(float64(10))))
if loop == 0 {
loop = 1
}
for i := 0; i < loop; i++ {
err = MakeLayer(c.Db, def_id, order, LAYER_SIMPLE_BLOCK, "")
order++
if err != nil {
return failed()
}
}
err = MakeLayer(c.Db, def_id, order, LAYER_FLATTEN, "")
if err != nil {
return failed()
}
order++
loop = int((math.Log(float64(number_of_classes)) / math.Log(float64(10))) / 2)
if loop == 0 {
loop = 1
}
for i := 0; i < loop; i++ {
err = MakeLayer(c.Db, def_id, order, LAYER_DENSE, fmt.Sprintf("%d,1", number_of_classes*(loop-i)))
order++
if err != nil {
return failed()
}
}
} else if complexity == 2 {
loop := int((math.Log(float64(model.Width)) / math.Log(float64(10))))
if loop == 0 {
loop = 1
}
for i := 0; i < loop; i++ {
err = MakeLayer(c.Db, def_id, order, LAYER_SIMPLE_BLOCK, "")
order++
if err != nil {
return failed()
}
}
err = MakeLayer(c.Db, def_id, order, LAYER_FLATTEN, "")
if err != nil {
return failed()
}
order++
loop = int((math.Log(float64(number_of_classes)) / math.Log(float64(10))) / 2)
if loop == 0 {
loop = 1
}
for i := 0; i < loop; i++ {
err = MakeLayer(c.Db, def_id, order, LAYER_DENSE, fmt.Sprintf("%d,1", number_of_classes*(loop-i)))
order++
if err != nil {
return failed()
}
}
} else {
c.Logger.Error("Unkown complexity", "complexity", complexity)
return failed()
}
err = ModelDefinitionUpdateStatus(c, def_id, MODEL_DEFINITION_STATUS_INIT)
if err != nil {
return failed()
}
return nil
}
func generateDefinitions(c *Context, model *BaseModel, target_accuracy int, number_of_models int) *Error {
cls, err := model_classes.ListClasses(c.Db, model.Id)
if err != nil {
ModelUpdateStatus(c, model.Id, FAILED_PREPARING_TRAINING)
// TODO improve this response
return c.Error500(err)
}
err = removeFailedDataPoints(c, model)
if err != nil {
return c.Error500(err)
}
cls_len := len(cls)
if number_of_models == 1 {
if model.Width < 100 && model.Height < 100 && cls_len < 30 {
generateDefinition(c, model, target_accuracy, cls_len, 0)
} else if model.Width > 100 && model.Height > 100 {
generateDefinition(c, model, target_accuracy, cls_len, 2)
} else {
generateDefinition(c, model, target_accuracy, cls_len, 1)
}
} else if number_of_models == 3 {
for i := 0; i < number_of_models; i++ {
generateDefinition(c, model, target_accuracy, cls_len, i)
}
} else {
// TODO handle incrisea the complexity
for i := 0; i < number_of_models; i++ {
generateDefinition(c, model, target_accuracy, cls_len, 0)
}
}
return nil
}
func handleTrain(handle *Handle) {
handle.Post("/models/train", func(w http.ResponseWriter, r *http.Request, c *Context) *Error {
if !CheckAuthLevel(1, w, r, c) {
return nil
}
if c.Mode == JSON {
panic("TODO /models/train JSON")
}
r.ParseForm()
f := r.Form
number_of_models := 0
accuracy := 0
if !CheckId(f, "id") || CheckEmpty(f, "model_type") || !CheckNumber(f, "number_of_models", &number_of_models) || !CheckNumber(f, "accuracy", &accuracy) {
fmt.Println(
!CheckId(f, "id"), CheckEmpty(f, "model_type"), !CheckNumber(f, "number_of_models", &number_of_models), !CheckNumber(f, "accuracy", &accuracy),
)
// TODO improve this response
return ErrorCode(nil, 400, c.AddMap(nil))
}
id := f.Get("id")
model_type := f.Get("model_type")
// Its not used rn
_ = model_type
// TODO check if the model has data
/*rows, err := handle.Db.Query("select mc.name, mdp.file_path from model_classes as mc join model_data_point as mdp on mdp.class_id = mc.id where mdp.model_mode = 1 and mc.model_id = $1 limit 1;", id)
if err != nil {
return Error500(err)
}
defer rows.Close()
if !rows.Next() {
return Error500(err)
}
var name string
var file_path string
err = rows.Scan(&name, &file_path)
if err != nil {
return Error500(err)
}*/
model, err := GetBaseModel(handle.Db, id)
if err == ModelNotFoundError {
return ErrorCode(nil, http.StatusNotFound, c.AddMap(AnyMap{
"NotFoundMessage": "Model not found",
"GoBackLink": "/models",
}))
} else if err != nil {
// TODO improve this response
return Error500(err)
}
if model.Status != CONFIRM_PRE_TRAINING {
// TODO improve this response
return ErrorCode(nil, 400, c.AddMap(nil))
}
full_error := generateDefinitions(c, model, accuracy, number_of_models)
if full_error != nil {
return full_error
}
go trainModel(c, model)
ModelUpdateStatus(c, model.Id, TRAINING)
Redirect("/models/edit?id="+model.Id, c.Mode, w, r)
return nil
})
handle.Get("/model/epoch/update", func(w http.ResponseWriter, r *http.Request, c *Context) *Error {
// TODO check auth level
if c.Mode != NORMAL {
// This should only handle normal requests
c.Logger.Warn("This function only works with normal")
return c.UnsafeErrorCode(nil, 400, nil)
}
f := r.URL.Query()
accuracy := 0.0
if !CheckId(f, "model_id") || !CheckId(f, "definition") || CheckEmpty(f, "epoch") || !CheckFloat64(f, "accuracy", &accuracy) {
c.Logger.Warn("Invalid: model_id or definition or epoch or accuracy")
return c.UnsafeErrorCode(nil, 400, nil)
}
accuracy = accuracy * 100
model_id := f.Get("model_id")
def_id := f.Get("definition")
epoch, err := strconv.Atoi(f.Get("epoch"))
if err != nil {
c.Logger.Warn("Epoch is not a number")
// No need to improve message because this function is only called internaly
return c.UnsafeErrorCode(nil, 400, nil)
}
rows, err := c.Db.Query("select md.status from model_definition as md where md.model_id=$1 and md.id=$2", model_id, def_id)
if err != nil {
return c.Error500(err)
}
defer rows.Close()
if !rows.Next() {
c.Logger.Error("Could not get status of model definition")
return c.Error500(nil)
}
var status int
err = rows.Scan(&status)
if err != nil {
return c.Error500(err)
}
if status != 3 {
c.Logger.Warn("Definition not on status 3(training)", "status", status)
// No need to improve message because this function is only called internaly
return c.UnsafeErrorCode(nil, 400, nil)
}
c.Logger.Info("Updated model_definition!", "model", model_id, "progress", epoch, "accuracy", accuracy)
_, err = c.Db.Exec("update model_definition set epoch_progress=$1, accuracy=$2 where id=$3", epoch, accuracy, def_id)
if err != nil {
return c.Error500(err)
}
return nil
})
}