Open correiav opened 6 years ago
Yes, you are right img_norm
is defined here
https://github.com/iamaaditya/VQA_Keras/blob/d6999208fa52d6259f98518027d99d813e05bb12/utils/get_data.py#L53
Image normalization is required because the CNN (VGG16) was trained with normalized images. Thus, in order to get the maximum benefit from the classifier test images should also be normalized. Think about what the biases do in the neural network. If your bias is wrong even if the neurons are doing the correct computation the overall score will be bad.
Could you do the fix for the get_test_data
and send a pull request. I will be happy to accept it.
Hi @iamaaditya, I tried to run the
train.py
however I got the following error:NameError: name 'img_norm' is not defined
Which refers to the lines 53 and, possibly, 98 of theutils/get_data.py
script, where you state image normalisation must occur. Could you briefly explain why to normalise image features and how to fix this issue?Also, the
train.py
should importget_test_data
fromutils.get_data
however the function on utils is actually defined asget_data_test
, so I believe it's just a small typo to fix.Thank you so much!