PCK mismatch for ResNet

[annabalabaeva]

Hi!
Thanks for the source code!

I'm trying to train the model with ResNet101 feature extractor and affine transformations. I'd like to get the same PCK as the model mentioned in commit f24a83a has.

I use the following command:
python3 train.py --num-epochs=20 --feature-extraction-cnn=resnet101 --random-sample=True --trained-models-dir=trained_models/

But the max PCK that I managed to achieve is 0.5388 (vs 0.559 that your model gives).

What am I doing wrong and how can I obtain the same results as your model gives?