vseledkin / gortex Goto Github PK

Hi,

lovely library ! I need to make a LSTM or GRU with a time series...I don't know how to start.
I have tryed gorgonia, lake of documentation too. We found only exemple for words, not numbers.

Thanks for your help.

i am sorry for contacting you here about your another project,but i send an email to your alibaba address but no response ,i am wondering whether you still use that email so i finad another way contact you .
about your spn project ,i am running it ,its likelihood is about 42 in 11th iteration but return to 439 in 12th iteration ,i don't know whether it is right .and i found maybe there is a bug in your SPN.initUnitRegion(), where you records r.counts[],i think it should be r.counts[step] = len(values[lowerIndex : upperIndex]),if i am worng ,please let me know .thank you .

Hi there,
First of all thank you for building this project for Go.
Im trying to use it as base of my project for RNN/LSTM networks and struggle to get things going.

Here is a toy structure that I think should work but I have few questions:

1. Can encoder step perform all training samples at once?
2. Can MSE calculate all costs at once ?
3. It seems that running backprop doesnt update DW as they are initially set to 0 and matrix functions such as Tanh/Add/Mul will user CopyAs to create new matrix and return it with 0 DW.
   From the example below the cost doenst change between epochs.

import (
	"fmt"
	g "github.com/vseledkin/gortex"
)

func main() {
	optimizer := g.NewOptimizer(g.OpOp{Method: g.SGD, LearningRate: 0.01, Momentum: 0.0, Clip: 4, Debug: false})

	var encoder *g.RNN

	encoder = g.MakeRNN(10, 10, 10)

	Who := g.RandMat(10, 1)

	// define model parameters
	encoderModel := encoder.GetParameters("Encoder")

	training := g.Mat(10, 1)
	training.W = []float32{1, 2, 3, 4, 5, 6, 7, 8, 9, 10}

	yLabels := g.Mat(10, 1)
	yLabels.W = []float32{2, 3, 4, 5, 6, 7, 8, 9, 10, 11}

	var epoch int
	grapth := g.Graph{NeedsBackprop: true}
	var output *g.Matrix

	for epoch < 100 {
		//can encoder step perform all training samples at once?
		Who, output = encoder.Step(&grapth, training, Who)
		//can MSE calculate all costs at once ?
		cost := grapth.MSE(yLabels, output)
		
		grapth.Backward()

		optimizer.Step(encoderModel)

		fmt.Printf("Epoch %v cost : %v\n", epoch, cost)
		
		epoch++
	}
}