Re-implementation of Vision Transformer

• Image Preprocessing
• Patching
• Flattening
• Linear Projection
• Position Embedding
• CLS Token
• Transformer
  • Attention
  • Feedforward
• MLP Head Classifier

• MNIST (98%)
• CIFAR-10 (75.25%)
• Tiny ImageNet (44%) (Same param count as ResNet18)

• Patch + Position Embedding (Extra learnable class embedding)
• Linear Projection of Flattened Patches
• Transformer Encoder
• MLP Head (Contains GeLU)
• Class (Bird, Ball, Car)

• Embedded Patches
• Norm (Layer Norm?)
• MHA
• Add (Identity + Activation)
• Norm (Layer Norm?)
• MLP
• Add (Identity + Activation)

• Reshape image x { R(HxWxC) into eq of flattened 2d patches xp { R(N*(P2 * C)) where (H, W) is res, C is channel count, (P, P) is res of each image patch, N = HW / P2 is resulting number of patches, also servces as effective input seq len for Transformer. Transformer uses constant latent vector size D through all layers, patches flattened to map to D dimensions with trainable linear projection, output of this projection are patch embeddings.