Passing model from keras to Pytorch

1

Hello, I am new to Pytorch (third day using it) and I am finding some trouble converting a model I had in keras. The model is the following:
X_input = Input(shape = (resized_image_size, resized_image_size, 3), name = ‘X_input’)

X = Conv2D(8, (3, 3), strides = (1,1) , padding = 'valid')(X_input)
X = MaxPooling2D(pool_size = (2, 2))(X)

X = Conv2D(16, (3, 3), strides = (1,1) , padding = 'valid')(X)
X = Conv2D(32, (3, 3), strides = (1,1) , padding = 'valid')(X)
X = MaxPooling2D(pool_size = (2, 2))(X)

X = Conv2D(64, (3, 3), strides = (1,1) , padding = 'valid')(X)
X = Conv2D(128, (3, 3), strides = (1,1) , padding = 'valid')(X)
X = MaxPooling2D(pool_size = (2, 2))(X)

X = Flatten()(X)

X = Dense(num_classes * 8 + 20, input_shape = (29 * 29 * 16, 1))(X)
X = Activation('relu')(X)
X = Dropout(0.4)(X)

X = Dense(num_classes * 4 + 10)(X)
X = Activation('relu')(X)

X = Dense(num_classes)(X)
X = Activation('softmax')(X)

keras model
keras model680×745 14.8 KB

What I have so far in Pytorch is the following:
class NeuralNetwork(nn.Module):
def init(self):
super(NeuralNetwork, self).init()
self.flatten = nn.Flatten()
self.l1 = nn.Sequential(
nn.Conv2d(1,8,(3,3),1,1),
nn.MaxPool2d(2,2))
self.l2 = nn.Sequential(
nn.Conv2d(8,16,(3,3),1,1),
nn.Conv2d(16,32,(3,3),1,1),
nn.MaxPool2d(2,2))
self.l3 = nn.Sequential(
nn.Conv2d(32,64,(3,3),1,1),
nn.Conv2d(64,128,(3,3),1,1),
nn.MaxPool2d(2,2))
self.l4 = nn.Sequential(
nn.Linear(,),
nn.ReLU(),
nn.Dropout(0.4))
self.l5 = nn.Sequential(
nn.Linear(,),
nn.ReLU())
self.l6 = nn.Sequential(
nn.Linear(,),
nn.Softmax())

def forward(self, x):
    x = self.l1(x)
    x = self.l2(x)
    x = self.l3(x)
    x = self.flatten(x)
    x = self.l4(x)
    x = self.l5(x)
    x = self.l6(x)
    return x

model = NeuralNetwork().to(device)

I would like to know if the first layers (with the convolutions) are ok and how to calculate the features for Linear since I read that it is the substitute to keras’ Dense.

2

The first conv layers look alright (your code is a bit hard to read and you can post code snippets by wrapping them into three backticks ``` :wink: ).

You could either calculate them manually, print the activation shape in forward before passing it to the first linear layer and adapt the in_features based on this shape, or use the nn.Lazy... modules.

1 Like
3

Forgot to mention one more potential issue: I don’t know which criterion you are using, but in case you are working on a multi-class classification use case with nn.CrossEntropyLoss, remove the last nn.Softmax layer in your PyTorch model as this criterion expects raw logits.

1 Like