README.md

Models

ML model weights and trainable codes for Basars

Prepare Dependencies

pip install tensorflow requests pandas opencv-python
pip install git+https://github.com/Basars/trans-unet.git
pip install git+https://github.com/Basars/basars-addons.git

Training Your Own Basars

Make sure you've installed Python >= 3.8.

usage: python -m basars.train 
                [-h] --type {stairs,proj}
                [--num-classes NUM_CLASSES] [--epochs EPOCHS] 
                [--batch_size BATCH_SIZE] [--buffer_size BUFFER_SIZE]
                [--multiprocessing-workers MULTIPROCESSING_WORKERS] [--cache-dataset CACHE_DATASET]

Polyp Segmentation and Phase Classification from Endoscopic Images

optional arguments:
  -h, --help            show this help message and exit
  --type {stairs,proj}  The type of transformer model. Default value is 'stairs'
  --num-classes NUM_CLASSES
                        Number of classes to be classified. Default value is 5
  --epochs EPOCHS       Epochs that how many times the model would be trained. Default value is 1290
  --batch_size BATCH_SIZE
                        The batch size. Default value is 64
  --buffer_size BUFFER_SIZE
                        The buffer size for shuffling datasets. Default value is 1024
  --multiprocessing-workers MULTIPROCESSING_WORKERS
                        Number of workers for prefetching datasets. Default value is 64
  --cache-dataset CACHE_DATASET
                        True to cache datasets on memory otherwise don't. Default value is True

Training Sample

python -m basars.train --type proj --epochs 1290

Configuration Guide

Refer the repository: final-experiments

Weights

• basars-cls5-stairs.h5
• basars-cls5-proj.h5

You can find out the weights in Releases.

Naming Convention and Meaning

stairs model have conv3x3 (256, 128, 64, 32, 16) → conv1x1 (5) upsamples

proj model have conv3x3 (256, 128, 64, 16) → conv1x1 (5, 5) upsamples

Model Definition

stairs model:

model = Sequential(name='ViT-stairs', layers=[
    VisionTransformer(input_shape=(224, 224, 3),
                      upsample_channels=(256, 128, 64, 32),
                      output_kernel_size=3, num_classes=16),
    Conv2D(5, kernel_size=(1, 1), padding='same', activation='softmax', use_bias=False)
])
model.load_weights('basars-cls5-stairs.h5')

proj model:

Sequential(name='ViT-proj', layers=[
    VisionTransformer(input_shape=(224, 224, 3), num_classes=5),
    Conv2D(5, kernel_size=(1, 1), padding='same', activation='softmax', use_bias=False)
])
model.load_weights('basars-cls5-proj.h5')

Model Architecture

Samples