yechengxi / lightnet Goto Github PK

Hi,

I run into trouble of running Main_Char_LSTM.m.
I've fixed the issue with these changes:
train_lstm.m line 94-97

[ net{1},res.Gate,opts ] = opts.parameters.learning_method( net{1},res.Gate,opts); [ net{2},res.Input,opts ] = opts.parameters.learning_method (net{2},res.Input,opts ); [ net{3},res.Cell,opts ] = opts.parameters.learning_method( net{3},res.Cell,opts ); [ net{4},res.Fit,opts ] = opts.parameters.learning_method( net{4},res.Fit,opts );
changed to:
[ net{1},res.Gate,opts ] = feval(opts.parameters.learning_method,net{1},res.Gate,opts); [ net{2},res.Input,opts ] = feval(opts.parameters.learning_method,net{2},res.Input,opts ); [ net{3},res.Cell,opts ] = feval(opts.parameters.learning_method, net{3},res.Cell,opts ); [ net{4},res.Fit,opts ] = feval(opts.parameters.learning_method,net{4},res.Fit,opts );

I want to use the QRNN based on your code to solve the squence to squence problem. how can i write the code based on lightnet?
Thankyou!

Hello,
I tried running your examples (unaltered) within MATLAB R2015b x64. Unfortunately, I get errors like the following for all types of the example networks:

clear all;
cd('./RNN');
disp('Testing training an LSTM.')
Main_Char_RNN();
cd ..
Testing training an LSTM.
Error using  + 
Matrix dimensions must agree.

Error in linear_layer (line 56)
        y=y+bias;

Error in net_ff (line 56)
                [res(layer+1).x,~,~,opts] = linear_layer(
                res(layer).x,net.layers{layer}.weights{1},net.layers{layer}.weights{2},[], opts );

Error in lstm_ff (line 45)
        [ net{1},res.Gates{f},opts ] = net_ff( net{1},res.Gates{f},opts );

Error in train_rnn (line 29)
            [ net,res,opts ] = lstm_ff( net,opts );

Error in Main_Char_RNN (line 146)
    [net,opts]=train_rnn(net,opts);

I know that you specified the MATLAB-Version to be R2016b or higher. Unfortunately, I do not have that version available yet.
So I tried R2015b. As far as I was able to trace back the code, I could not see that the dimensions would be different in MATLAB R2016b. Can the newer MATLAB version simply handle the mismatching dimensions? I would have guessed no. Are the Example data maybe faulty?

Every help is appreciated!

Is it possible to create split custom layer-wise layers? If so, would you provide an example?
Here's my intended implementation:

PrepareData_MNIST_CNN and net_init_cnn_mnist are missing and there maybe more. Please check the commented parts in the Main_Template file. %% Also can you please update your comments (especially for the RL example.)

Great job tho!

When I run Main_char_RNN.m, some error happend.
错误使用 +
矩阵维度必须一致。

出错 linear_layer (line 56)
y=y+bias;

出错 net_ff (line 56)
[res(layer+1).x,,,opts] = linear_layer(
res(layer).x,net.layers{layer}.weights{1},net.layers{layer}.weights{2},[],
opts );

出错 rnn_ff (line 46)
[ net{1},res.Input{f},opts ] = net_ff( net{1},res.Input{f},opts
);

出错 train_rnn (line 36)
[ net,res,opts ] = rnn_ff( net,opts );

出错 Main_Char_RNN (line 146)
[net,opts]=train_rnn(net,opts);

the shape of y is [90,97],the bias's shape is [90,1]

Hi guys,
great work! thanks for sharing.

Just one question: what's the different of lightnet and matconvnet? better wrapper function? more convenient choices?

Thanks for your time,
peng

Hello,
I have been trying to see if I can get more use out of LightNet than I have so far gotten out of the NNToolbox. I am using MatLab R2016b, with Neural Network toolbox installed. My PC is a 16-core Xeon system, w/ GTX 1070 GPU.

One issue I have run into is that when I try to run the provided example script "Main_Char_RNN" with opts.use_nntoolbox=1 and use_gpu=0, an error occurs. I have traced the source of the error to the function "linear_layer.m", line 26:

y=fc_nntb.forward(I);

What appears to be happening is that the 'forward' method of the 2dconvolutional layer class returns a 4-D gpuArray, and the remainder of the function is not expecting a gpuArray because use_gpu is set to 0.

I have found that changing this to y=gather(fc_nntb.forward(I)); fixes the problem, but it does not solve the problem of what to do if you don't want to use GPU at all. I have found that it is substantially faster simply to set 'use_nntoolbox' to 0.

Also, is there any more documentation available for the LSTM / GRU modules? I want to apply them to neural spike train data (as binary time series). I am having some difficulty understanding the data prep and network initialization functions.

Thanks for any help you can provide!

Hi,
I am trying use your LightNet to implement RNN/CNN in Matlab2015a, but failed. It seems that there is no successful demo.
Could you please tell me why?
The error information of performing Main_Char_RNN:

%%%%%%%%%%%%%%%%%%%%%%%%
Error using +
Matrix dimensions must agree.
Error in linear_layer (line 56)
y=y+bias;
Error in net_ff (line 40)
[res(layer+1).x,,,opts] = linear_layer(
res(layer).x,net.layers{1,layer}.weights{1},net.layers{1,layer}.weights{2},[], opts);

Error in lstm_ff (line 45)
[ net{1},res.Gates{f},opts ] = net_ff( net{1},res.Gates{f},opts );

Error in train_rnn (line 29)
[ net,res,opts ] = lstm_ff( net,opts );

Error in Main_Char_RNN (line 146)
[net,opts]=train_rnn(net,opts);

%%%%%%%%%%%%%%%%%%%%%%%
The error information of performing Main_CNN_ImageNet_minimal
Error using +
Matrix dimensions must agree.
Error in conv_layer_2d (line 103)
y=y+bias_p;

Error in net_ff (line 37)
[res(layer+1).x,,,opts] = conv_layer_2d(
res(layer).x,net.layers{1,layer}.weights{1},net.layers{1,layer}.weights{2},net.layers{1,layer}.stride,net.layers{1,layer}.pad,[],opts
);

Error in Main_CNN_ImageNet_minimal (line 33)
[ net,res,opts ] = net_ff( net,res,opts );
%%%%%%%%%%%%%%%

Main_Cart_Pole_Q_Network
Error using zeros
Leading inputs must be numeric.

Error in Main_Cart_Pole_Q_Network (line 63)
InputBatch=zeros(4,MaxUpdateDelay,'like',state);

what is problem?

Matlab has error
zeros(4,MaxUpdateDelay,'like',state);
Error using zeros
Leading inputs must be numeric.

or
I run runall

RunAll
Testing training a multilayer perceptron.

use_gpu =

 0

opts =

use_nntoolbox: 1

Undefined variable "nnet" or class "nnet.internal.cnn.layer.Convolution2D".

Error in linear_layer (line 14)
opts.layer{opts.current_layer}.fc_nntb=nnet.internal.cnn.layer.Convolution2D('conv2d_nntb',
[1,1], in ,out, [1,1], [0,0]);

Error in net_ff (line 54)
[res(layer+1).x,,,opts] = linear_layer(
res(layer).x,net.layers{1,layer}.weights{1},net.layers{1,layer}.weights{2},[], opts);

Error in select_learning_rate (line 38)
[ net,res,opts ] = net_ff( net,res,opts );

Error in selective_sgd (line 13)
[lr_best]=select_learning_rate(net,opts);

Error in train_net (line 20)
[ net,opts ] = selective_sgd( net,opts );

Error in TrainingScript (line 51)
[net,opts]=train_net(net,opts);