ratikumari227 / Convolution_Variants

Reimplementing SOTA convolution variants with Tensorflow 2.0.

Convolution Variants

This repository replicates various convolution layers from SOTA papers.

This repository currently includes:

• Attention Augmented (AA) Convolution Layer
• Mixed Depthwise Convolution Layer
• Drop Block
• Efficient Channel Attention (ECA) Layer
• Convolutional Block Attention Module (CBAM) Layer

Attention Augmented Convolution Layer

For other implementations in:

• Pytorch: leaderj1001
• Keras: titu1994
• TensorFlow 1.0: gan3sh500

Notes

• This implementation does not yet include relative positional encodings.

Mixed Depthwise Convolution Layer

For other implementations in:

• TensorFlow 1.0: Tensorflow

Notes

• This implementation combines depthwise convolution with pointwise convolution. The original implementation only used depthwise convolutions.

Drop Block

For other implementations in:

• TensorFlow 1.0: Tensorflow
• Pytorch: miguelvr

Efficient Channel Attention Layer

For other implementations in:

• Pytorch: BangguWu

Convolutional Block Attention Module Layer

For other implementations in:

• Pytorch: Jongchan

Usage

Here is an example of how to use one of the layers:

import tensorflow as tf
from convVariants import AAConv

aaConv = AAConv(
    channels_out=32,
    kernel_size=3,
    depth_k=8, 
    depth_v=8, 
    num_heads=4)

The layer can be treated like any other tf.keras.layers class.

model = tf.keras.models.Sequential([
    aaConv,
    tf.keras.layers.Flatten(),
    tf.keras.layers.Dense(128, activation='relu'),
    tf.keras.layers.Dropout(0.2),
    tf.keras.layers.Dense(10, activation='softmax')
    ])

model.compile(
    optimizer='adam',
    loss='sparse_categorical_crossentropy',
    metrics=['accuracy'])

model.fit(x_train, y_train, epochs=5)

Tests

Test cases are located here.

To run tests:

cd Convolution_Variants
python tests.py

Requirements

• tensorflow 2.0.0 with GPU

Caveats

• These layers are only tested to work for input format: NCHW.

Acknowledgements

Links to the original papers:

• Attention Augmented Convolutional Networks
• Mixed Depthwise Convolution Layer
• Drop Block Layer
• Efficient Channel Attention Layer
• Convolutional Block Attention Module Layer