Implementation of the U-Net architecture trained on the Carvana Image masking dataset. Trained on single nvidia A100 PCIe GPU. Bellow are the results with plain settings using train_val.py i.e. no weight decay, scheduler, data augmentation.
Task Summary 64x64 images |
|
|
Train Losses 
|
Validation Losses 
|
git clone git@github.com:Efesasa0/carvana-unet.git
cd carvana-unet
pip install -r requirements.txt
python train_val.pyDownloaded from kaggle page for the Carvana dataset. Organized as follows under ./data directory. ./manual_test and manual_test_masks consists of only few I specifically selected for fast sanity checks.
.
├── manual_test'
│ ├── 0cdf5b5d0ce1_01.jpg
│ ├── ...
│ └── 0cdf5b5d0ce1_05.jpg
├── manual_test_masks
│ ├── 0cdf5b5d0ce1_01_mask.gif
│ ├── ...
│ └── 0cdf5b5d0ce1_05_mask.gif
├── train
└── train_masks