视觉变形金刚

A plain Vision Transformer starts by dividing an image into non-overlapping patches. Each patch is flattened and linearly projected into an embedding, producing a sequence for a transformer encoder. The original paper’s title refers to 16-by-16 patches, but the method and reported variants should not be treated as having one mandatory patch size. Dense self-attention allows each allowed patch token to interact with other tokens, supporting relationships beyond a local convolutional window. Position embeddings add information about sequence positions. In the original implementation, a learned classification token is prepended, and its final encoder representation supplies the classification head. That is a design choice rather than a requirement that all vision transformers use the same pooling. Training assumptions also matter. The original work compared plain ViT with convolutional networks under particular data and regularization setups. It identified fewer convolution-style locality and translation-equivariance biases and found improved results with large-scale pretraining and transfer. It did not establish that every small-data vision task requires enormous datasets or that a transformer always beats a CNN. Patch geometry affects sequence length: a 224-by-224 image partitioned into 16-by-16 non-overlapping patches has fourteen patches along each axis, giving 196 patch tokens. A separate class token is not included in that patch-only count. Increasing the number of tokens increases the pairwise work of ordinary dense attention, although the complete layer has other computational terms. The paper and repository provide an architecture and research results, not evidence about every current product that may use a transformer.