A bunch of codes and experiments
Project 1 : Kaggle cassava leaf diseases classification
This is a week map for all the experiments I do.
-
Pytorch-lightning setup
-
Configure data module & Model
-
Hyperparams tried:
-
Training on : Nvidia GTX 1060 (6.1k MB VRAM)
- num_epochs = 10
- lr = 3e-4
- resize = 400
- img_h = 350
- img_w = 350
- weight_decay = .01
- eps = 1e-8
- train_batch_size = 32
- test_batch_size = 32
- base_model = 'resnet34'
- seed_val = 2021
- Adam Optimizer with weight decay
- ReduceLrOnPlateau (monitoring val_accuracy)
- Using auto mixed precision (fp16)
-
Results (Local):
- Validation loss : 0.65
- Validation accuracy : 0.79
-
Results (Kaggle):
- Test accuracy : 0.838
-
Pytorch-lightning setup
-
Configure data module & Model
-
Hyperparams tried:
-
Training on : Nvidia Tesla P100 from kaggle kernels (16k MB VRAM)
- num_epochs = 15
- lr = 2e-2
- resize = 600
- img_h = 512
- img_w = 512
- weight_decay = .01
- eps = 1e-8
- train_batch_size = 32
- test_batch_size = 32
- base_model = 'resnet34'
- seed_val = 2021
- Adam Optimizer with weight decay
- ReduceLrOnPlateau (monitoring val_accuracy)
- Using auto mixed precision (fp16)
-
Results (Local):
- Validation loss : NaN
- Validation accuracy : NaN
-
Results (Kaggle):
- Test accuracy : NaN
Also tried
-
Pytorch-lightning setup
-
Configure data module & Model
-
Hyperparams tried:
-
Training on : Nvidia Tesla T4 (16k MB VRAM)
- num_epochs = 25
- lr = 2e-2
- resize = 600
- img_h = 512
- img_w = 512
- weight_decay = .01
- eps = 1e-8
- train_batch_size = 128
- test_batch_size = 64
- base_model = 'resnet34'
- seed_val = 2021
- Adam Optimizer with weight decay
- ReduceLrOnPlateau (monitoring val_accuracy)
- Using auto mixed precision (fp16)
-
Results (Local):
- Validation loss : NaN
- Validation accuracy : NaN
-
Results (Kaggle):
- Test accuracy : NaN