How to transform conv1d from keras?

1

I’m trying to implement this network from keras to pytorch, not sure if it’s workable. The target model in keras is as following:

input_7 (InputLayer)             (None, 200, 76)       0                                            
____________________________________________________________________________________________________
conv_1 (Convolution1D)           (None, 192, 10)       6850        input_7[0][0]                    
____________________________________________________________________________________________________
conv_2 (Convolution1D)           (None, 184, 10)       910         conv_1[0][0]                     
____________________________________________________________________________________________________
conv_3 (Convolution1D)           (None, 174, 20)       2220        conv_2[0][0]                     
____________________________________________________________________________________________________

But the PyTorch seems to have a different API for conv1d. I tried this:

nn.Sequential(
            nn.Conv1d(in_channels=200, out_channels=192, kernel_size=10),
            nn.ReLU(),
            nn.Conv1d(in_channels=192, out_channels=184, kernel_size=10),
            nn.ReLU(),
            nn.Conv1d(in_channels=184, out_channels=174, kernel_size=20),
            nn.ReLU(),
        )

which seems not working properly.

The kernel_size in keras Convolution1D doesn’t seem very simple to transfer to conv1d in PyTorch.

Appreciate any help.

2

I can’t see any difference between kernel_size in Keras and kernel_size in PyTorch.
Nor can I see any errors in your tiny code snippet.

What errors are you getting? What isn’t working as expected?

3

Hi Thanks.

So here’s my output in pytorch:

(500L, 192L, 67L)
(500L, 192L, 67L)
(500L, 184L, 58L)
(500L, 184L, 58L)
(500L, 174L, 39L)
(500L, 174L, 39L)

The first dimension is batch-size, so you can simply ignore it. The duplicate one corresponds to relu.

4

What code produces those numbers?

5 
input_7 (InputLayer)             (None, 200, 76)       0                                            
____________________________________________________________________________________________________
conv_1 (Convolution1D)           (None, 192, 10)       6850        input_7[0][0]                    
____________________________________________________________________________________________________
conv_2 (Convolution1D)           (None, 184, 10)       910         conv_1[0][0]                     
____________________________________________________________________________________________________
conv_3 (Convolution1D)           (None, 174, 20)       2220        conv_2[0][0]                     
______________________________________________________________________________

Can you post the keras snippet code which produces this output shape?

6

@jpeg729 @ashwin.raju93
Sorry for late response. Here are the keras codes:

x = Input(shape=(200, 76))
h = Convolution1D(10, 9, activation = 'relu', name='conv_1')(x)
h = Convolution1D(10, 9, activation = 'relu', name='conv_2')(h)
h = Convolution1D(20, 11, activation = 'relu', name='conv_3')(h)

model = Model(x, h)
print model.summary()

And the output is

____________________________________________________________________________________________________
Layer (type)                     Output Shape          Param #     Connected to                     
====================================================================================================
input_6 (InputLayer)             (None, 200, 76)       0                                            
____________________________________________________________________________________________________
conv_1 (Convolution1D)           (None, 192, 10)       6850        input_6[0][0]                    
____________________________________________________________________________________________________
conv_2 (Convolution1D)           (None, 184, 10)       910         conv_1[0][0]                     
____________________________________________________________________________________________________
conv_3 (Convolution1D)           (None, 174, 20)       2220        conv_2[0][0]

And corresponding pytorch code is:

nn.Sequential(
            nn.Conv1d(in_channels=200, out_channels=192, kernel_size=10),
            nn.ReLU(),
            nn.Conv1d(in_channels=192, out_channels=184, kernel_size=10),
            nn.ReLU(),
            nn.Conv1d(in_channels=184, out_channels=174, kernel_size=20),
            nn.ReLU(),
        )

But the model structure is

(500L, 192L, 67L)
(500L, 192L, 67L)
(500L, 184L, 58L)
(500L, 184L, 58L)
(500L, 174L, 39L)
(500L, 174L, 39L)
1 Like
7

Your keras model defines 10 filters with kernel_size=9 in the first conv layer, while in your PyTorch model you define 192 filters with kernel_size=10. Keras’ filters is equal to out_channels.

EDIT: Also it seems to me that your Keras input hat 76 channels. You should transpose the input to get similar results to [batch, channels, length].

1 Like
8

Thanks for the explanation. Just to check it again.

So in keras, the input data is [batch, timestep/length, filter/channel], while in pytorch it becomes [batch, filter/channel, timestep/length] right?

If that’s the case, I guess I should do sth like transpose((0,2,1)).

9

Yes, if you are using the default settings with the tensorflow backend.
There is some option like channel_first to change this behavior in keras.

1 Like
10

Now the baseline model is using keras for VAE, and I want to reimplement them in pytorch. I can fix the input part by this transformation. But not sure how to do this for the decoder output.

To be more specific, in the pytorch implementation, the input is [batch, filter/channel, timestep/length]. And if I want to follow the structure of baseline models, which outputs [batch, timestep/length, filter/channel] (in keras), do you think it’s reasonable to use this transpose on the output for loss calculation?

11

Could you please post the complete Keras and PyTorch code, if that’s possible?
If you cannot post the complete code, a small snippet of the encoder and decoder part will be sufficient.

The decoder should keep the dimensions it gets from the encoder.
I don’t really understand the issue, but a look at the code will most likely change that. :wink: