Description

Weights of the VGG16 convolutional layers (include_top=False) ported from PyTorch to Keras.

The VGG16 model is described in the following paper (configuration D in Table 1):

Simonyan K. & Zisserman A., Very Deep Convolutional Networks for Large-Scale Image Recognition, ICLR 2015.

Why do you need this?

The VGG16 model in Keras comes with weights ported from the original Caffe implementation. It expects the following image pre-processing:

1. convert the images from RGB to BGR,
2. subtract [103.939, 116.779, 123.68] from the BGR channels, respectively.

The VGG16 model in PyTorch comes with a different set of weights and expects another pre-processing:

1. divide the image by 255,
2. subtract [0.485, 0.456, 0.406] from the RGB channels, respectively,
3. divide the RGB channels by [0.229, 0.224, 0.225], respectively.

The scale of features extracted from the VGG16 model using the two sets of weights is different. This can be important in some applications.

Script used

import numpy as np
import torchvision.models as models
from keras.applications.vgg16 import VGG16

pytorch_model = models.vgg16(pretrained=True)

# select weights in the conv2d layers and transpose them to keras dim ordering:
wblist_torch = list(pytorch_model.parameters())[:26]
wblist_keras = []
for i in range(len(wblist_torch)):
    if wblist_torch[i].dim() == 4:
        w = np.transpose(wblist_torch[i].detach().numpy(), axes=[2, 3, 1, 0])
        wblist_keras.append(w)
    elif wblist_torch[i].dim() == 1:
        b = wblist_torch[i].detach().numpy()
        wblist_keras.append(b)
    else:
        raise Exception('Fully connected layers are not implemented.')

keras_model = VGG16(include_top=False, weights=None)
keras_model.set_weights(wblist_keras)
keras_model.save_weights('output_path/vgg16_pytorch2keras.h5')