Use the following command to train a model
python train.py -a resnet18 -d dataset_folder_path --epochs num_epochs -b batch_size --finetune --seed 2022 --exp exp_name --entity wandb_username --sample_interval sample_interval --cuda --mv- Replace dataset_folder_path with path to dataset folder
- Replace num_epochs (integer)
- Replace batch_size with number of examples to load for each gradient descent step (integer)
- Replace exp_name with experiment name. This option is used to track the experiment in Weights and Biases https://wandb.ai
- Replace wandb_username with your Weights & Biases username
- Replace sample_interval with number of intervals to sample predictions on validation set. This will be uploaded to Weights and Biases.
--finetune is used to load pretrained model. Its certain layers are frozen, and rest of the network is trained on the shoe dataset. --mv is used to train in the multi-view mode. --cuda option is used to train on gpu. If you want to train on different gpu, change the gpu id in 'cuda:0' of line 69 to the required gpu id as 'cuda:gpu_id'. Explore the code and other parameters as well.
After training, the predictions on test set are stored in test_{exp_name}.json file in the format of 'idx':[predictions, ground_truth]. 'idx' refers to an index of the dataset. To view predictions, follow the code in test.ipynb Try generating plots like the right hand side plots in slide number 16.
Dataset can be found in: https://drive.google.com/drive/folders/1VQZc9kWp-_MYTv8p62Db4JjXElV4v3xh?usp=sharing