Diabetic Retinopathy Detection

This Repo contains my scripts for the EEL 6825: Pattern Recogntion Spring 2021 Course.

Papers

• Skin lesion classification with ensemble of squeeze-and-excitation networks and semi-supervised learning
• Self-training with Noisy Student improves ImageNet classification
• Bottleneck Transformers for Visual Recognition
• Rotate to Attend: Convolutional Triplet Attention Module

Features

• ☑ Balanced Sampler

• ☑ Mixed Precision

• ☑ Gradient Accumulation

• ☑ Optimum Learning Rate Finder [LR Finder Suggestion is terrible. I just observed the learning rate at which loss starts to diverge and set learning_rate = learning rate at diverging loss/100. No particular intention behind it.]

• ☑ TTA

Resources

• Margin Focal Loss
• APTOS Gold Medal Solutions: Although data type is different but it might be helpful.
• Melanoma Recognition via Visual Attention

Can be useful

• Deep Metric Learning Solution For MVTec Anomaly Detection Dataset
• Ugly Duckling Concept
• Humpback Whale Classification 1st place solution
• Attention model for feature extraction

How to run

• Run git clone https://github.com/tahsin314/Diabetic_Retinopathy-Detection-Pattern-Recognition-Course-Project
• Run conda env create -f environment.yml
• Run train.py. Change parameters according to your preferences from the dr_config.ini file before training.
• dr_config parameters:

  n_fold = Total number of folds
  fold = fold that you want to keep as your validation set
  SEED = Seed value. This value will be use for random initialization
  batch_size 
  sz = Image Dimension
  learning_rate 
  patience = patience for Learning Rate Scheduler
  accum_step = Gradient Accumulation steps 
  num_class = If regression, it should be 1 otherwise 5
  gpu_ids = GPUs to use
  mixed_precision 
  target_type = Regression, Classification or Ordinal Regression
  pretrained_model = model name
  model_type = Normal, TripletAttention, CBAttention
  cam_layer_name = Class Activation Mapping Layer
  model_dir = Directory where models will be saved
  distributed_backend = None, dp or ddp. Currently ddp does not work
  mode = train or lr_finder
  load_model = 0 if False else 1
  imagenet_stats = mean and standard deviation of imagenet data
  crop = False if 0 i.e, circle crop won't be applied 
  ben_color = Ben Graham's preprocessing technique. True if set to 1
  n_epochs = number of epochs to train
  TTA = Test Time Augmentation
  balanced_sampler 
  data_dir 
  image_path 
  test_image_path 
  

One important thing about EfficientNet

EfficientNet's are designed to take in to account input image dimensions.

So if you want to squeeze every last droplet from your model make sure to use same image resolutions as described below:

Efficientnet-B0 : 224
Efficientnet-B1 : 240
Efficientnet-B2 : 260
Efficientnet-B3 : 300
Efficientnet-B4 : 380
Efficientnet-B5 : 456
Efficientnet-B6 : 528
Efficientnet-B7 : 600

Issues

• Currently muli gpu training only works with distributed_backend='dp'. However, Stochastic Weight Averaging breaks it.
• distributed_backend='dp' also fails to gather data from multiple GPUs and returns predictions and labels with shape (batch_num, ) where it is supposed to return data with shape (num_samples, ) (num_samples ≈ num_batches * batch_size). This is probably a pytorch-lightning bug. It is advised to comment out distributed_backend='dp' and training on single GPU at this moment.

Alternative: Try running inference on test data on single GPU or set batch_size=1 during test.

Update [20/04/2021]

Added an ugly hack for gathering predictions and labels with distributed_backend='dp'. Now data can be gathered from multiple GPUs. Hope that pytorch-Lightning will fix their gather issue.