thainamhoang/resnet50_fatima_challenge

Fatima Challenge

• Dataset: I'm using the cats_vs_dogs dataset from kaggle instead of huggingface datasets since cats_vs_dogs in datasets have checksum error. The original dataset is from Microsoft. However, some files are corrupted and it hasn't been divided into train and test folders.
• Model: I'm using ResNet50 with pretrained=True from pytorch. The choice of optimizer is Adam with lr=1e-4. I had tried AdamW before coming back to Adam. I also added the lr_scheduler to reconfigure the learning rate.
• Number of epochs: I'm using 10 epochs in total. I had tried with 20, however, only the train_acc increases by ~1%, so it does not worth trying for an increasing amount of training time.

Model: