I have extracted features using the ResNet50 network pre-trained on ImageNet (using Keras framework), the features were extracted from the average pooling layer (2048 dimens