The objective of this challenge is to train a model from images that can:
- Classify leaves into diseased / healthy class ( Mutually exclusive )
- A leaf may be either of the 4 classes
- healthy
- multiple_diseases
- rust
- scab
| Healthy (0) | Multiple Diseas (1) | Rust (2) | Scab (3) |
|---|---|---|---|
 |
 |
 |
 |
This project has been released under the MIT license.
If you use this project in your research or wish to refer to the baseline results published on W&B, please use the following BibTeX entry.
@misc{cvpr2020-fgvc7,
author = {Mallik et al},
title = {Pytorch Pipeline v 1.0},
howpublished =
{
\url{https://www.linkedin.com/in/rajat-modi-54377877/}
&
\url{https://www.linkedin.com/in/aakash-mallik-5b8198132/}
},
year = {2020}
}
The experiments conducted on this dataset have been well catalogued on W&B. Few reports that we prepared bases on the best results are also available.
The prominent results are as follows:
A few key points:
- The best model so far was Efficient net b5 (with image resolution of 475x475). This consumed the GPU memory (16GBs) completely. This was the tradeoff we went for.
- We opted for a fixed set of transformations rather than using auto augment.
- TTA improved our results from 0.971 to 0.976.
.
|-- auto_aug_utils
|-- config
|-- data
| `-- fgvc7
| `-- images
|-- dataset
|-- examples
|-- loss
|-- model
|-- optimiser
|-- results
|-- scheduler
|-- transformer
|-- utils
- Dataloaders
- Stratified k fold
- Models
- Efficientnet B7
- Efficientnet B5
- Efficientnet B4
- Densenet 161
- Loss
- Cross Entropy Loss
- Focal loss
- ArcFace loss
- Mean ROC AUC loss ( Kaggle )
- Optimizer
- RMSProp
- Adam
- AdamW
- Scheduler
- [X] CosineAnelingLR with Warmup
- CosineAnelingLR
- StepLR
Sample files ( train_dummy_1, train_dummy_2 ) are provided in the examples directory
mode: train
seed: (optional)
validation_frequency: ( optional )
epochs:
batch_size: ( optional )
num_classes:
train_dataset:
name:
fold: ( optional )
resize_dims:
transform: (optional, uses default if not specified)
val_dataset:
name:
fold: ( optional )
resize_dims:
transform: (optional, uses default if not specified)
model:
name:
pred_type: regression/classification/mixed
tuning_type: feature-extraction/fine-tuning
hyper_params: ( depends on the model )
key_1: ( depends on the model )
key_2: ( depends on the model )
pre_trained_path: ( optional key to resume training from another experiment )
weight_type: ( optional - type of weight - best_val_loss/best_val_kaggle_metric )
( will use weights from last epoch if nothing is mentioned )
optimiser:
name:
hyper_params:
key_1: ( depends on the optimiser )
key_2: ( depends on the optimiser )
scheduler:
name:
hyper_params:
key_1: ( depends on the scheduler )
key_2: ( depends on the scheduler )
loss_function:
name: ( in case of mixed, this will be the regression loss )
hyper_params:
classification_loss: ( used in case of mixed prediction )
classification_coefficient: ( in case of mixed - (0, 1) )
Sample files ( test_dummy ) are provided in the examples directory
mode: test
test_dataset:
name:
resize_dims:
tta: True/False
transform: (optional, uses default if not specified)
(A proper transformer is needed in case of TTA)
ensemble: True/False
num_classes:
experiment_list:
- experiment:
path: ( valid train config file name )
weight_type: ( optional - type of weight - best_val_kaggle_metric/best_val_roc )
resize_dims: ( optional override - original/integer)
tta: ( optional override )
- experiment:
path: ( valid train config file name )
weight_type: ( optional - type of weight - best_val_kaggle_metric/best_val_roc )
simply modify the run.sh with the required commands:
python main.py "train_dummy_1.yml"
python main.py "test_dummy.yml"
Add -a flag with the training config that you are going to use for scoring:
python main.py -a 'train_dummy_1.yml'
Add -p flag with the training config to publish results to W&B and Telegram:
python main.py -p 'train_dummy_1.yml'
pip install --user torch seaborn wandb torchvision catalyst albumentations pandas scikit-image tqdm scikit-learn pyyaml blessed pyzmq pretrainedmodels tensorboard transformersThis command will monitor GPU status and keep updating the terminal every 0.5 seconds
watch -n0.5 nvidia-smi