v0

This version automobile_classifier_v0 will classify the given photo into one of this:

• autorikshaw
• bike
• car
• fighterjet
• tank
• truck

This was a testing version to test out ResNet50 on my own. The classes are very different from each other so training needed and data quality needed is low.

v1

The automobile_classifier_v1 will focus on KIA models seltos and carnival only

• The goal is to predict the KIA models with 90%+ accuracy.
• The main task?
  • The dataset has to be very clean, with no noise, and no faulty data.
  • The classification classes are going to be very close as KIA follows similar designs among cars.
• Unfroze the layer4 with the model.fc

First run:

• Carnival dataset had many U.S.A. variants of the carnival which looks very different from the Indian
• image size was set

Second run:

• Seltos dataset had noise -> CLEARED

Third run:

• Image size was changed to 384, and batch size to 32
• 20 epochs trained

Final Run:

• Increased the image size to 500, batch size to 64
• 20 epochs
• Dataset was highly cleaned - 400 to 500 images per class

v2

• Added Kia sonet
• Added grad cam

Models:

The models for each version are given in this drive link: Automobile Classifier Models

v3

• Added Kia EV6 data
• Increaed dataset from ~400 images to ~900 iamges for each class
• total ~3600 images combined
• Increased number of epochs from 25 to 40
• Decreased learning rate from 1e-5 to 4e-6

• Time taken to train: 73mins - but I forced stop at 39th epoch as it started mild overfitting

--------------------------------------------------
Epoch [38/40]
Train Loss: 0.0358 | Train Accuracy: 99.27%
Val Loss:   0.0997 | Val Accuracy:   96.94%
--------------------------------------------------

v4

No model was trained, this version uses the final_model_v3 from version 3

• Restructured repository and code
• Added checkpoints while training to notebook.py
• Made streamlit use full width instead of "centered"
• Added Inference Timing in streamlit page
• Added Prediction logs to track record each prediction made
• Added top-k prediction - Instead of prediction, it gives 1 top prediction, and 2nd and 3rd predictions according to confidence
• Model Statistics page added as a button to app.py

v5

This version is specifically designed to test out different models and compare them with metrics. A README file is given in the folder automobile_classifier_v5 explaining the structure of the code.

v6

The final version using resnet101, chosen from the results from v5. resnet101 showed the most accuracy, precision and confidence with various tests.