jimmy-ren / vlstm Goto Github PK

I have a question that is the accuracy I trained the face_only condition is higher than that in face-audio condition. What's the problem?
Isn't face-audio high be 97%? but I achieved only 88%.

best wishes!

Hi,
I am interested in your vLSTM code?
As you stated that, you use the single-model LSTM with one hidden memory cell layer to train a sequence classifier. The node of this hidden layer is 512.
I want to use such LSTM with one hidden layer with 64 nodes, one LSTM layer with 64 memory cells.
How to revise your code to realize it?
Sorry for this trouble.

Hi,
when I run vLSTM/applications/speaker-naming/audio_only/sn_audio_train.m, or vLSTM/applications/speaker-naming/face_only/sn_face_train.m, there exists an error as described below.

没有为类 'gpuArray' 找到具有匹配签名的方法 'cumprod'。

出错 lstm_core_v52 (line 117)
cum_prod = cumprod(mem.forget_gate{L}{s}(:,:,2:t), 3, 'reverse');

出错 sn_audio_train (line 59)
lstm_core_v52(in, label);

My platform is Windows 7 and Matlab 2014a. When i implement "help cumprod", the cumprod can be found.

help cumprod
cumprod - Cumulative product

This MATLAB function returns an array the same size as the array A containing
the cumulative product.If A is a vector, then cumprod(A) returns a vector
containing the cumulative product of the elements of A.If A is a matrix, then
cumprod(A) returns a matrix containing the cumulative products for each column
of A.If A is a multidimensional array, then cumprod(A) acts along the first
nonsingleton dimension.

B = cumprod(A)
B = cumprod(A,dim)

I do not know why.
Could you tell me how to fix this problem.