`KeyError` with nested transposed convolution in `nn.Sequential`

Describe the bug When converting my model, I get a KeyError for {layer_name}.weight when it gets to a ConvTranspose2d. I’ve managed to dumb the model down to a simple reproduction case, see code below.

To Reproduce This fails:

class TestNet(nn.Module):
    def __init__(self):
        super().__init__()
        c1x1 = nn.Sequential(
            nn.Conv2d(in_channels=6, out_channels=1, kernel_size=1, padding=0, bias=False),
            nn.BatchNorm2d(num_features=1)
        )
        self.my_op = nn.Sequential(
            c1x1,
            nn.ConvTranspose2d(1, 1, kernel_size=4, stride=2, padding=1)
        )
    
    def forward(self, x):
        return self.my_op(x)

my_x = torch.ones(1, 6, 128, 128)
my_test_net = TestNet()
print(my_test_net(my_x).shape) # torch.Size([1, 1, 256, 256])
km = pytorch_to_keras(my_test_net, my_x, [(6, 128, 128)], verbose=True)

with:

[....]
Converting transposed convolution ...
---------------------------------------------------------------------------
KeyError                                  Traceback (most recent call last)
<ipython-input-39-10087659f3d4> in <module>
----> 1 km = pytorch_to_keras(my_test_net, my_x, [(6, 128, 128)], verbose=True)

~/anaconda3/envs/general/lib/python3.6/site-packages/pytorch2keras/converter.py in pytorch_to_keras(model, args, input_shapes, change_ordering, training, verbose, names)
    313             node_input_names,
    314             layers, state_dict,
--> 315             names
    316         )
    317         if node_id in graph_outputs:

~/anaconda3/envs/general/lib/python3.6/site-packages/pytorch2keras/convolution_layers.py in convert_convtranspose(params, w_name, scope_name, inputs, layers, weights, names)
    244     weights_name = '{0}.weight'.format(w_name)
    245 
--> 246     if len(weights[weights_name].numpy().shape) == 4:
    247         W = weights[weights_name].numpy().transpose(2, 3, 1, 0)
    248         height, width, n_filters, channels = W.shape

KeyError: 'my_op.0.weight'

Rewriting to reduce the nn.Sequential nesting however converts successfully:

class TestNet(nn.Module):
    def __init__(self):
        super().__init__()
        self.my_op = nn.Sequential(
            nn.Conv2d(in_channels=6, out_channels=1, kernel_size=1, padding=0, bias=False),
            nn.BatchNorm2d(num_features=1),
            nn.ConvTranspose2d(1, 1, kernel_size=4, stride=2, padding=1)
        )
    
    def forward(self, x):
        return self.my_op(x)

Of course the real model is quite a bit more complex than this, so un-nesting isn’t so trivial.

Expected behavior I would expect both models to convert just fine, but maybe I am missing something, in which case please let me know 😃.

Logs I hope the provided error suffices.

Environment (please complete the following information):

• OS: MacOS
• Python 3.7
• Version v0.1.15

Additional context

[gmalivenko]

Hello @ElteHupkes. Yes, it’s known issue related to nn.Sequential / nn.ModuleList. You can try to unroll your submodules into the model class and convert it with the converter. Actually, I plan to rewrite graph-traverse functions and I hope the converter will be able to convert nn.Sequential / nn.ModuleList as well.

[ElteHupkes]

Thanks, good to know! I don’t have time to test this right now, but when I do I’ll close this.