@automan000 Thanks a lot for you work! I am a little confused that in code:

41 def init_hidden(self, batch_size, hidden, shape): 42 ¦ self.Wci = Variable(torch.zeros(1, hidden, shape[0], shape[1])).cuda() 43 ¦ self.Wcf = Variable(torch.zeros(1, hidden, shape[0], shape[1])).cuda() 44 ¦ self.Wco = Variable(torch.zeros(1, hidden, shape[0], shape[1])).cuda() 45 ¦ return (Variable(torch.zeros(batch_size, hidden, shape[0], shape[1])).cuda(), 46 ¦ ¦ ¦ Variable(torch.zeros(batch_size, hidden, shape[0], shape[1])).cuda())

Why Wci, Wcf, Wco should be initialized?

https://arxiv.org/pdf/1506.04214.pdf in this parper the LSTM compute i(t),f(t),o(t) is use x(t),h(t-1) and c(t-1). but in your code convolution_lstm.py line 24: combined = torch.cat((input, h), dim=1)

you just use x and h, why?

Default code snippet ran in pytorch 0.40, it hangs a bit then this error shows up.

Complete Error:

`RuntimeError Traceback (most recent call last) in () 101 output = convlstm(input) 102 output = output[0][0] --> 103 res = torch.autograd.gradcheck(loss_fn, (output, target), raise_exception=True) 104 print(res)

C:\Anaconda3\lib\site-packages\torch\autograd\gradcheck.py in gradcheck(func, inputs, eps, atol, rtol, raise_exception) 190 if not ((a - n).abs() <= (atol + rtol * n.abs())).all(): 191 return fail_test('Jacobian mismatch for output %d with respect to input %d,\n' --> 192 'numerical:%s\nanalytical:%s\n' % (i, j, n, a)) 193 194 if not reentrant:

C:\Anaconda3\lib\site-packages\torch\autograd\gradcheck.py in fail_test(msg) 170 def fail_test(msg): 171 if raise_exception: --> 172 raise RuntimeError(msg) 173 return False 174

RuntimeError: Jacobian mismatch for output 0 with respect to input 0, numerical:tensor(1.00000e-02 * [[ 0.0000], [ 0.0000], [ 0.0000], ..., [ 0.0000], [ 0.0000], [ 0.0000]]) analytical:tensor([[-3.6467e-05], [ 2.2621e-06], [-3.9878e-05], ..., [-4.3014e-05], [ 1.2278e-05], [ 3.6285e-06]])`

代码的输入数据定义为: input = Variable(torch.randn(1, 512, 64, 32)).cuda() 参数分别为(batch_size, channels, height, weight) 在输入convlstm网络进行计算时有下面这个for循环： for step in range(self.step): x= input for i in range(self.num_layers): ...

请问这个循环每次的输入x都是相同的吗？序列体现在哪里呢？有理解不对的地方希望能得到您的指正，谢谢！

This is a question rather than an issue, sorry for bothering. > + I have questions about input/output shape, as well as the meaning of "input_channel". According to #gradient check part of the code, the input is in the shape of (1, 512, 64, 32), while the output shrinks to (1, 32, 64, 32). I assume for the input, 1 is batch size, 512 is input_channel, and the image is of size 64*32.

The questions are: What does these channels mean? Are they filters as in Keras convLSTM library(https://keras.io/layers/recurrent/#convlstm2dcell)? How do we input a sequence of 5 images? And why is output channel smaller than the input?

I'm very happy to see that you have modified the previous question, but I suggest that your ConvLSTMCell forward process take the following calculations: ci = torch.sigmoid(self.Wxi(input) + self.Whi(hidden_state) + c * self.Wci) cf = torch.sigmoid(self.Wxf(input) + self.Whf(hidden_state) + c * self.Wcf) new_c = cf * c + ci * torch.tanh(self.Wxc(input) + self.Whc(hidden_state)) co = torch.sigmoid(self.Wxo(input) + self.Who(hidden_state) + new_c * self.Wco) new_h = co * torch.tanh(new_c)

Thanks for your implementation of conv-lstm. However, there may exist some bugs in the code. I use the code as a part of my project. The convolutional features are extracted from images and are passed to the conv-lstm. Following the conv-lstm are a fully-connected layer and the loss. But the loss.backward() will report an error, which tells that the 'retain_graph' parameter should be true. However, setting retain_graph=True will consume more and more memory and slow down the program.

Hi, I'm trying to apply your code to a sequential image dataset.

However, whenever I tried to input concatenated images (Batch x Timeseries x Channels x Width x Height), it gives following error.

<ipython-input-40-c130ba752793> in forward(self, input, h, c)
     29         input = input.cuda()
     30 
---> 31         combined = torch.cat((input, h), dim=1)
     32 
     33         A = self.conv(combined)

TypeError: cat received an invalid combination of arguments - got (tuple, dim=int), but expected one of:
 * (sequence[torch.cuda.FloatTensor] seq)
 * (sequence[torch.cuda.FloatTensor] seq, int dim)
      didn't match because some of the arguments have invalid types: (tuple, dim=int)

It seems like something is wrong with .cuda() declaration.

So I looked into those two Variables inputand h

input has torch.FloatTensor type,

h has torch.autograd.variable.Variable type.

From this article, those two Variables have to be changed to same data type.

My question is,

i) Have you undergone the same issue like this?

ii) I've tried to change the data type of h by h.cuda() didn't work. I'm not used to Pytorch. So is there any advice I can get?

BTW thank you for your update. Gave me a lot of help

The LSTM paper defines a specific rule for gradient updates of the 'peephole' connections. Specifically:

[...] during learning no error signals are propagated back from gates via peephole connections to CEC

Based on my understanding of the code the way these 3 variables are initialized (as asked in Issue 17) is an attempt at implementing this update rule, but I don't see how does initializing them as Variables helps. From my understanding of the quoted part of the LSTM paper, the peephole connections should be updated but the gradient that updates them should stop there and not flow any further. If that is the case then this implementation is incorrect, although it might be that Pytorch does not support such an operation as .detach() is not suitable for the job.

The first problem is that in ConvLSTM.forward, the code is using the same x = input in multiple timesteps. I guess the input shape of forward func. shall be changed to

[sequence, bsize, channel, x, y] 

instead of the original

[bsize, channel, x, y]

And, x=input line shall be changed to

x=input[step]

for different steps. I am still studying if it's appropriate to loop layers within loops of timesteps, but after training your current code(with the change I mentioned above), I can get decent outcomes.

The second problem is that in ConvLSTMCell, there're no biases. For example in

ci = torch.sigmoid(self.Wxi(x) + self.Whi(h) + c * self.Wci)

While it should be something like

ci = torch.sigmoid(self.Wxi(x) + self.Whi(h) + c * self.Wci + self.Bci)

But I don't know if such constants would affect the backward phase.

P.S. I'm myself a beginner so maybe I'm wrong. Please reply :)