Coding a vision transformer from scratch using python's pytorch framework. The Vision Transformer Architecture:
Let's break down the subparts of the transformer...
-
Initial patch embeddings: The input image is broken down into distinguishable patches of size nxn (n=16) and these patches are embedded using convolutional layers.
-
Conversion to projected vectors: These patches are then converted to key, query and value projection vectors using linear layers
-
Attention blocks: Consists of the self-attention layers to assign weights and context to patches
-
Multi-layer perceptron: A simple network with linear layers for image classification. Uses Gaussian Error Linear Rectified Unit activation.
-
Final classification layer: Linear layer which maps the output to respective class probabilities.
The vision transformer can be used in inference mode as from the file inference.py with necessary arguments. The result will be of the shape: (batch_size, nclasses) and can be changed to suit the application involved. The model and inputs can be sent to a pre-existing GPU using PyTorch's cuda capabilities.
Run
python inference.py --depth --proj_dropout --attn_dropout --gpu --weights- PyTorch (preferable with CUDA integration)
- numpy