The pytorch implemenation for convolutional rnn is alreaedy exisitng other than my module, for example.
However, there are no modules supporting neither variable length tensor nor bidirectional rnn.
I implemented AutogradConvRNN by referring to AutogradRNN at https://github.com/pytorch/pytorch/blob/master/torch/nn/_functions/rnn.py, so my convolutional RNN modules have similar structure to torch.nn.RNN and supports the above features as it has.
The benefit of using AutogradConvRNN is not only that it enables my modules to have the same interface as torch.nn.RNN, but makes it very easy to implement many kinds of CRNN, such as CLSTM, CGRU.
- python3 (Not supporting python2 because I prefer type annotation)
- pytorch>=0.4
- Implemented at python level, without any additional CUDA kernel, c++ codes.
- Convolutional RNN, Convolutional LSTM, Convolutional Peephole LSTM, Convolutional GRU
- Unidirectional, Bidirectional
- 1d, 2d, 3d
- Supporting PackedSequence (Supporting variable length tensor)
- Supporting nlayers RNN and RNN Cell, both.
- Not supporting different hidden sizes for each layers (But, it is very easy to implement it by stacking 1-layer-CRNNs)
import torch
import convolutional_rnn
from torch.nn.utils.rnn import pack_padded_sequence
net = convolutional_rnn.Conv3dGRU(in_channels=2, # Corresponds to input size
out_channels=5, # Corresponds to hidden size
kernel_size=(3, 4, 6), # Int or List[int]
num_layers=2,
bidirectional=True,
dilation=2, stride=2, dropout=0.5)
x = pack_padded_sequence(torch.randn(3, 2, 2, 10, 14, 18), [3, 1])
print(net)
y, h = net(x)
print(y.data.shape)
cell = convolutional_rnn.Conv2dLSTMCell(in_channels=3, out_channels=5, kernel_size=3).cuda()
time = 6
input = torch.randn(time, 16, 3, 10, 10).cuda()
output = []
for i in range(time):
if i == 0:
hx, cx = cell(input[i])
else:
hx, cx = cell(input[i], (hx, cx))
output.append(hx)