ERA V1 Assignment 7

Model 1:

Target:

- Getting a working modular setup. We will use assignment 5 code.
- Model.py has models and utils.py has train, test and plotting functions.

Results:

- Model has 5.9M parameters.
- Train Accuracy = 99.95
- Test Accuracy = 99.35

Analysis:

- Our setup is running correctly.
- Model is very large and overfitting
- Closer look at images tell us that a receptive field of 5 pixels is enough to catch edges.

Model 2:

Target:

- Getting the model skeleton right.

Results:

- Model has 174k parameters.
- Train Accuracy = 99.7
- Test Accuracy = 99.34

Analysis:

- Simple model gradually increasing channels to 256.
- Model is still large and overfitting.

Model 3:

Target:

- Making the model as light as possible.

Results:

- Model has 7781 parameters.
- Train Accuracy = 99.11
- Test Accuracy = 98.96

Analysis:

- Simple model gradually increasing channels to 32.
- Overfitting reduced significantly. Even though test accuracy is less than train accuracy they are not diverging.

Model 4:

Target:

- Add Batch Normalisation to increase efficiency.

Results:

- Model has 7979 parameters.
- Train Accuracy = 99.34
- Test Accuracy = 99.24

Analysis:

- Model has started overfitting again slightly as train and test accuracies are diverging.

Model 5:

Target:

- Add Dropout of 1%.

Results:

- Model has 7979 parameters.
- Train Accuracy = 99.23
- Test Accuracy = 99.26

Analysis:

- No longer overfitting. Infact it's the opposite.

• Here Model changes are done. In next two files we will use Model5 itself for augmentation and LR changes.

Model 6:

Target:

- Add image augmentations random rotation and random perspective.

Results:

- Model has 7979 parameters.
- Train Accuracy = 98.69
- Test Accuracy = 99.53

Analysis:

- Crossed 99.4 for first time and test loss is fluctuating.
- Time to add an LR scheduler to stabilise test loss.

Model 7:

Target:

- Add LR Scheduler ReduceLROnPlateau.

Results:

- Model has 7979 parameters.
- Train Accuracy = 98.85
- Test Accuracy = 99.44

Analysis:

- Target achieved: crossed 99.4% validation accuracy 3 times (epochs 11, 14 and 15)

Receptive Field Calculations

	kernel	padding	stride	Jout	Nout	RF
input				1	28	1
conv	3	1	1	1	28	3
conv	3	1	1	1	28	5
maxpool	2	0	2	2	14	6
conv	3	1	1	2	14	10
conv	3	1	1	2	14	14
maxpool	2	0	2	4	7	16
conv	3	1	1	4	7	24
avgpool	7	0	7	28	1	48
conv	1	0	1	28	1	48
conv	1	0	1	28	1	48