Shouldn't the kernel size be equal to kernel_size. It seems to be explicitly assigned to 3. Just commenting because I am using part of your code for a different project, not replicating the paper.

for e in xrange(max_epoch): for time in xrange(num_seqs): h_next = model(input_gru[time], None) err += MSE_criterion(h_next [0], target_gru[time]) print(err.data[0])

It seems to me that you assume that previous states are all None. But after time step 1, the previous state is now h_t-1 so I think the code should be like below:

prev_state = None for e in xrange(max_epoch): for time in xrange(num_seqs): h_next = model(input_gru[time], prev_state) err += MSE_criterion(h_next [0], target_gru[time]) prev_state = h_next print(err.data[0])

In your implementation part.
In your implementation process, Equation (8) uses the following code:
next_h = torch.mul(update_gate,hidden) + (1-update_gate)*ct

According to the original calculation process, it should be the following code organization:
next_h = torch.mul(update_gate,ct) + (1-update_gate)*hidden