- This chapter should come before training!
- The order of profiling options should be:
- The Cell Painting CNN first
- ImageNet pre-trained second
- Your own trained model third
Also, is it true that we strongly recommend to prepare the dataset (compression) before profiling? I thought DeepProfiler was able to compute features for images without compression (which is desirable if the dataset is too large). Our model is robust to illumination changes, so this should not affect downstream performance. I recommend to remove this and instead clarify that DeepProfiler does not need compression to profile (I added a Note with this comment in Chapter 3).