fyp/logic/models/train/torch/nn/linear.go

175 lines
4.2 KiB
Go

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
}