https://github.com/dudeperf3ct/diamond-clarity
https://github.com/dudeperf3ct/diamond-clarity
Last synced: 3 months ago
JSON representation
- Host: GitHub
- URL: https://github.com/dudeperf3ct/diamond-clarity
- Owner: dudeperf3ct
- Created: 2021-11-10T16:35:09.000Z (over 3 years ago)
- Default Branch: master
- Last Pushed: 2021-11-20T07:59:45.000Z (over 3 years ago)
- Last Synced: 2024-12-31T10:17:36.383Z (5 months ago)
- Language: Jupyter Notebook
- Size: 312 KB
- Stars: 0
- Watchers: 2
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: Readme.md
Awesome Lists containing this project
README
## Diamond Clarity Classification
There are 2 classes present in the dataset.
classes : [0, 1] => [275, 225]. The baseline accuracy to beat will be 55%. All models should perform better than this baseline.
## Installation
```bash
docker build -t dc .
./run_container.sh
python main.py
```
Inside `main.py` change the path of model_dir as requiired to either `cnn3d`, `cnnlstm` or `cnn3dlstm` and any of the submodels.
## Notebooks
Run the notebook on colab present in `notebooks` folder.
### Training
We use a lot of SOTA approaches like OneCycleLR, AdamW, FP16 training, Stratified splits for 90%-10% train val dataset, logging to wandb, augmentations using albumentation library.
### Experiments
**CNN 3d**
- Resnet-18
- Resnet-10
- 6-layer simple cnn 3d architecture
Experiments can be tracked here : https://wandb.ai/dudeperf3ct/cnn3d
Analysis :
- We overfit the model with 3 samples, only `custom_simple_cnn3d` overfits the 3 samples of each class. Other 2 model fail to overfit.
-----
**CNN 3d + LSTM**
-----
**CNN 3d + GRU**