How can I add new layers on pre-trained model with PyTorch? (Keras example given)

Question

I am working with Keras and trying to analyze the effects on accuracy that models which are built with some layers with meaningful weights, and some layers with random initializations.

Keras:

I load VGG19 pre-trained model with include_top = False parameter on load method.

model = keras.applications.VGG19(include_top=False, weights="imagenet", input_shape=(img_width, img_height, 3))

PyTorch:

I load VGG19 pre-trained model until the same layer with the previous model which loaded with Keras.

model = torch.hub.load('pytorch/vision:v0.6.0', 'vgg19', pretrained=True)
new_base =  (list(model.children())[:-2])[0]

After loaded models following images shows summary of them. (Pytorch, Keras)

So far there is no problem. After that, I want to add a Flatten layer and a Fully connected layer on these pre-trained models. I did it with Keras but I couldn't with PyTorch.

The output of new_model.summary() is that:

My question is, how can I add a new layer in PyTorch?

Fiducial

new_base.add_module ? – Lupe

Does this answer your question? How to remove the last FC layer from a ResNet model in PyTorch? – Mcdonald

Answer 1

If all you want to do is to replace the classifier section, you can simply do so. That is :

model = torch.hub.load('pytorch/vision:v0.6.0', 'vgg19', pretrained=True)
model.classifier = nn.Linear(model.classifier[0].in_features, 4096)
print(model)

will give you:

Before:

VGG(
  (features): Sequential(
    (0): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (1): ReLU(inplace=True)
    (2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (3): ReLU(inplace=True)
    (4): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
    (5): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (6): ReLU(inplace=True)
    (7): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (8): ReLU(inplace=True)
    (9): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
    (10): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (11): ReLU(inplace=True)
    (12): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (13): ReLU(inplace=True)
    (14): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (15): ReLU(inplace=True)
    (16): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (17): ReLU(inplace=True)
    (18): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
    (19): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (20): ReLU(inplace=True)
    (21): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (22): ReLU(inplace=True)
    (23): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (24): ReLU(inplace=True)
    (25): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (26): ReLU(inplace=True)
    (27): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
    (28): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (29): ReLU(inplace=True)
    (30): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (31): ReLU(inplace=True)
    (32): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (33): ReLU(inplace=True)
    (34): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (35): ReLU(inplace=True)
    (36): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
  )
  (avgpool): AdaptiveAvgPool2d(output_size=(7, 7))
  (classifier): Sequential(
    (0): Linear(in_features=25088, out_features=4096, bias=True)
    (1): ReLU(inplace=True)
    (2): Dropout(p=0.5, inplace=False)
    (3): Linear(in_features=4096, out_features=4096, bias=True)
    (4): ReLU(inplace=True)
    (5): Dropout(p=0.5, inplace=False)
    (6): Linear(in_features=4096, out_features=1000, bias=True)
  )
)

After:

VGG(
  (features): Sequential(
    (0): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (1): ReLU(inplace=True)
    (2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (3): ReLU(inplace=True)
    (4): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
    (5): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (6): ReLU(inplace=True)
    (7): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (8): ReLU(inplace=True)
    (9): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
    (10): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (11): ReLU(inplace=True)
    (12): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (13): ReLU(inplace=True)
    (14): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (15): ReLU(inplace=True)
    (16): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (17): ReLU(inplace=True)
    (18): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
    (19): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (20): ReLU(inplace=True)
    (21): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (22): ReLU(inplace=True)
    (23): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (24): ReLU(inplace=True)
    (25): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (26): ReLU(inplace=True)
    (27): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
    (28): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (29): ReLU(inplace=True)
    (30): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (31): ReLU(inplace=True)
    (32): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (33): ReLU(inplace=True)
    (34): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (35): ReLU(inplace=True)
    (36): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
  )
  (avgpool): AdaptiveAvgPool2d(output_size=(7, 7))
  (classifier): Linear(in_features=25088, out_features=4096, bias=True)
)

Also note that when you want to alter an existing architecture, you have two phases. You first get the modules you want (that's what you have done there) and then you must wrap that in a nn.Sequential because your list does not implement a forward() and thus you can't really feed it anything. It's just a collection of modules.

So you need to do something like this in general (as an example):

features = nn.ModuleList(your_model.children())[:-1]
model = nn.Sequential(*features) 
# carry on with what other changes you want to perform on your model

Note that if you want to create a new model and you intend on using it like:

output = model(imgs)

You need to wrap your features and new layers in a second sequential. That is, do something like this:

features = nn.ModuleList(your_model.children())[:-1]
model_features = nn.Sequential(*features) 
some_more_layers = nn.Sequential(Layer1,
                                 Layer2, 
                                 ... ) 
                                 
model = nn.Sequential(model_features, 
                      some_more_layers)
#
output = model(imgs)

otherwise you had to do something like :

features_output = model.features(imgs)
output = model.classifier(features_output)

Napalm

Actually I don't want to use the model as classifier, I will use the model as feature extractor and I need extract (1,4096) feature vectors for each image (from the first FC layer). On the other hand, while I do this, I want to add FC layers without meaningful weights ( not belongs to imagenet), FC layers should be has default weights which defined in PyTorch. For example, FC layer which had added on model in Keras has weights which are initialize with He_initialization not imagenet. – Sire

I didnt say you want to use it as a classifier, I said, if you want to replace the classifier its easy. if you need the features prior to the classifier, just use model.features. if you need to add a new layer, just do it the way I did. simply add a new layer. its weights are uninitialized. for layer initialization see this. see my last update as well. – Napalm

Answer 2

From the PyTorch tutorial "Finetuning TorchVision Models":

Torchvision offers eight versions of VGG with various lengths and some that have batch normalizations layers. Here we use VGG-11 with batch normalization. The output layer is similar to Alexnet, i.e.

(classifier): Sequential(
   ...
   (6): Linear(in_features=4096, out_features=1000, bias=True)
)

Therefore, we use the same technique to modify the output layer

model.classifier[6] = nn.Linear(4096,num_classes)

Mcdonald