adithya-samavedhi / CSE_251B_PA3

CSE_251B_PA3

Code structure summary

1. basic_fcn.py has the code for the baseline model.

2. Dataset Download: run download.py to download the data. The dataset will be downloaded in 'data' directory

3. train.py: This file contains the training loop with early stopping and we define loss criterion, optimizer, cosine annealing learning rate scheduler.

4. voc.py : This file creates the datset using Pytorch's dataset class. Input transformations can be applied by providing the argument --transform true.

5. util.py: We calculate the iOU and the mean pixel accuracy. It also contains the code to generate plots.

6. custom.py: contains the custom architecture that builds upon our baseline model.

7. resnet34.py: contains the transfer learning architecture using resnet 34 as the base.

8. unet_architecture.py: contains the u-net architecture.

Instructions to run

python CSE_251B_PA3/PA3_starter/download.py # One-time dataset download

python CSE_251B_PA3/PA3_starter/train.py [args]

The arguments and their corresponding values are as follows:

1. scheduler: normal (default) , cosine
2. model: normal(default), transfer_learning, unet
3. filepath:
4. early-stop: True(default), False
5. early_stop-epoch: 3 (default) [This defines the patience]