CVPR 2020 - FGVC ( Plant Pathology )

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)

License

This project has been released under the MIT license.

Citing the Pipeline

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} }

Performance

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.

Pipeline Structure

.
|-- auto_aug_utils
|-- config
|-- data
|   `-- fgvc7
|       `-- images
|-- dataset
|-- examples
|-- loss
|-- model
|-- optimiser
|-- results
|-- scheduler
|-- transformer
|-- utils

pipe strucutre

Plans

[x] Dataloaders
- [x] Stratified k fold
[x] Models
- [x] Efficientnet B7
- [x] Efficientnet B5
- [x] Efficientnet B4
- [x] Densenet 161
[x] Loss
- [x] Cross Entropy Loss
- [x] Focal loss
- [x] ArcFace loss
- [x] Mean ROC AUC loss ( Kaggle )
[x] Optimizer
- [x] RMSProp
- [X] Adam
- [x] AdamW
[X] Scheduler
- [x] [X] CosineAnelingLR with Warmup
- [X] CosineAnelingLR
- [X] StepLR

Config File structure

Training

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) )

Test

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 )

How to use the pipeline

1. Using it for train/test

simply modify the run.sh with the required commands:

python main.py "train_dummy_1.yml"
python main.py "test_dummy.yml"

2. Using it for augmentation search

Add -a flag with the training config that you are going to use for scoring:

python main.py -a 'train_dummy_1.yml'

3. Publish mode while training

Add -p flag with the training config to publish results to W&B and Telegram:

python main.py -p 'train_dummy_1.yml'

Lib dependency

pip install --user torch seaborn wandb torchvision catalyst albumentations pandas scikit-image tqdm scikit-learn pyyaml blessed pyzmq pretrainedmodels tensorboard transformers

Utility commands

This command will monitor GPU status and keep updating the terminal every 0.5 seconds

watch -n0.5 nvidia-smi

Pipe-Runner-Lab / CVPR2020-FGVC7

readme