keras-io/video-vision-transformer

Keras Implementation of Video Vision Transformer on medmnist

This repo contains the model to this Keras example on Video Vision Transformer.

Background Information

This example implements ViViT: A Video Vision Transformer by Arnab et al., a pure Transformer-based model for video classification. The authors propose a novel embedding scheme and a number of Transformer variants to model video clips.

Datasets

We use the MedMNIST v2: A Large-Scale Lightweight Benchmark for 2D and 3D Biomedical Image Classification dataset.

Training Parameters

# DATA
DATASET_NAME = "organmnist3d"
BATCH_SIZE = 32
AUTO = tf.data.AUTOTUNE
INPUT_SHAPE = (28, 28, 28, 1)
NUM_CLASSES = 11

# OPTIMIZER
LEARNING_RATE = 1e-4
WEIGHT_DECAY = 1e-5

# TRAINING
EPOCHS = 80

# TUBELET EMBEDDING
PATCH_SIZE = (8, 8, 8)
NUM_PATCHES = (INPUT_SHAPE[0] // PATCH_SIZE[0]) ** 2

# ViViT ARCHITECTURE
LAYER_NORM_EPS = 1e-6
PROJECTION_DIM = 128
NUM_HEADS = 8
NUM_LAYERS = 8