GNR 638 Mini Project 1

Fine-grained classification:

1. Train a CNN model with an upper limit of 10M parameters.

2. Please download CUB dataset from here: https://data.caltech.edu/records/65de6-vp158/files/CUB_200_2011.tgz?download=1

3. Use default train-test split for this task.

4. Submissions will be evaluated based on parameter efficiency, training time efficiency (no. of iterations) and accuracy.

5. Submissions should include report, code and final model checkpoint. Drive link is fine for checkpoint.

6. Report must include architecture and traiining details. Training loss and accuracy curves should also be incorporated along with final results.

7. External models: Only ImageNet pretrained models are allowed.

8. Please use moodle forum or mail to gnr638@googlegroups.com to post queries.

9. Deadline: March 5 2024, 11:59 PM

Link to saved checkpoint and report

Report and saved checkpoint for all commits are uploaded here

How to use?

• Download the dataset from here
• Clone/Download the repo
• Extract the .tgz file in the repo folder
• cnn_classification.ipynb have all the util functions and hyperparams
• By default we use default number of train/test dataset
• For data augmentation, call data_augmentation() function before calling train_mode function
• To call the train function use this

train_model(model,
            criterion = nn.CrossEntropyLoss(),
            learning_rate=learning_rate,
            optimizer = None,
            schedular=None,
            num_epoch=num_epoch,
            save_checkpoint=False,
            time_start_from=0):

How to use pre-trained model like Efficient-Net_v2?

• Keep above points in mind
• Open efficientnet_no_data_augmentation.ipynb
• Run all cells of the .ipynb file

Architecture Used: EfficientNet_B1_Weights.IMAGENET1K_V2

The exact architecture can be found in efficientnet_no_data_augmentation.ipynb

Model Performance on IMAGENET dataset

Weight	Acc@1	Acc@5	Params	GFLOPS
EfficientNet_B1_Weights.IMAGENET1K_V2	79.838	94.934	7.8M	0.69

Changes in classifier layer of EfficientNet

Sequential(
  (0): Dropout(p=0.4, inplace=True)
  (1): Linear(in_features=1280, out_features=200, bias=True)
)

Training our model

1. Fine Tune the classifier (for other parameters requires_grad=False)

Number of training parameters = 256200

2. Transfer Learning (all parameters requires_grad=True)

Number of training parameters = 6769384

3. Fine Tune the classifier (for other paramters requires_grad=False)

Number of training parameters = 256200

Final Accuracy and Loss graph

Results

• Top 1 Accuracy: 71.48%
• Time required for training: 26 min 54 sec
• Total Number of Parameters: 6,769,384

Fine Tuning:

• Total Number of Parameters: 256,200

Transfer Learning:

• Total Number of Parameters: 6,769,384