FashionBert - Generate image features

[StefanoSalvatori]

Why does the image_feature_extract.py code for patch feature generation uses num_patches = (4, 4) at line 475? shouldn't it be (8,8)?

Also the paper states that patch features are generated using resnext101 while the script you give uses resnet 50, what is the correct one?

I'm asking because I did some experiments and i tried to generate features from the fashiongen validation dataset using your image_feature_extract.py code and i get strange results: given a certain product, the features i get for its image does not match with the features that you have in the csv evalution files. Maybe there is some preprocessing to do to images? The fashion gen validation dataset has 256x256 RGB images, are these the correct sizes?

I'm using this this code to extract features:

feature_extractor = PatchFeatureExtractor("/content/resnet_v1_50")
results_batch = feature_extractor.predict([img], batch_size=1)
feature = results_batch[0]['feature']

Thank you for our help

[HowieMa]

Same question. May I ask have you fixed it yet? Thanks!