use tch::{IndexOp, Tensor, nn, nn::ModuleT};
use crate::{
config::{Pool, ViTConfig},
layers::Transformer,
tensor::{num_patches, patch_dim, patchify_2d, repeat_token},
};
#[derive(Debug)]
pub struct PatchEmbedding {
norm1: nn::LayerNorm,
linear: nn::Linear,
norm2: nn::LayerNorm,
patch_height: i64,
patch_width: i64,
}
impl PatchEmbedding {
pub fn new(
vs: &nn::Path,
patch_dim: i64,
dim: i64,
patch_height: i64,
patch_width: i64,
) -> Self {
let norm1 = nn::layer_norm(vs / "norm1", vec![patch_dim], Default::default());
let linear = nn::linear(vs / "linear", patch_dim, dim, Default::default());
let norm2 = nn::layer_norm(vs / "norm2", vec![dim], Default::default());
Self {
norm1,
linear,
norm2,
patch_height,
patch_width,
}
}
pub fn forward_patches(&self, patches: &Tensor) -> Tensor {
patches
.apply(&self.norm1)
.apply(&self.linear)
.apply(&self.norm2)
}
}
impl nn::Module for PatchEmbedding {
fn forward(&self, xs: &Tensor) -> Tensor {
self.forward_patches(&patchify_2d(xs, self.patch_height, self.patch_width))
}
}
#[derive(Debug)]
pub struct ViT {
patch_embedding: PatchEmbedding,
cls_token: Tensor,
pos_embedding: Tensor,
transformer: Transformer,
mlp_head: Option<nn::Linear>,
pool: Pool,
emb_dropout: f64,
}
impl ViT {
pub fn new(vs: &nn::Path, config: ViTConfig) -> Self {
let image_height = config.image_size.height;
let image_width = config.image_size.width;
let patch_height = config.patch_size.height;
let patch_width = config.patch_size.width;
let num_patches = num_patches(image_height, image_width, patch_height, patch_width);
let patch_dim = patch_dim(config.channels, patch_height, patch_width);
let num_cls_tokens = match config.pool {
Pool::Cls => 1,
Pool::Mean => 0,
};
let patch_embedding = PatchEmbedding::new(
&(vs / "patch_embedding"),
patch_dim,
config.dim,
patch_height,
patch_width,
);
let cls_token = vs.var(
"cls_token",
&[num_cls_tokens, config.dim],
nn::Init::Randn {
mean: 0.0,
stdev: 1.0,
},
);
let pos_embedding = vs.var(
"pos_embedding",
&[num_patches + num_cls_tokens, config.dim],
nn::Init::Randn {
mean: 0.0,
stdev: 1.0,
},
);
let transformer = Transformer::new(
&(vs / "transformer"),
config.dim,
config.depth,
config.heads,
config.dim_head,
config.mlp_dim,
config.dropout,
);
let mlp_head = if config.num_classes > 0 {
Some(nn::linear(
vs / "mlp_head",
config.dim,
config.num_classes,
Default::default(),
))
} else {
None
};
Self {
patch_embedding,
cls_token,
pos_embedding,
transformer,
mlp_head,
pool: config.pool,
emb_dropout: config.emb_dropout,
}
}
pub fn embeddings_t(&self, xs: &Tensor, train: bool) -> Tensor {
let batch = xs.size()[0];
let mut xs = xs.apply(&self.patch_embedding);
if self.pool == Pool::Cls {
let cls_tokens = repeat_token(&self.cls_token, batch);
xs = Tensor::cat(&[cls_tokens, xs], 1);
}
let seq_len = xs.size()[1];
let pos_embedding = self.pos_embedding.i(0..seq_len).unsqueeze(0);
xs = (xs + pos_embedding).dropout(self.emb_dropout, train);
self.transformer.forward_t(&xs, train)
}
}
impl nn::ModuleT for ViT {
fn forward_t(&self, xs: &Tensor, train: bool) -> Tensor {
let xs = self.embeddings_t(xs, train);
let xs = match self.pool {
Pool::Mean => xs.mean_dim(1, false, xs.kind()),
Pool::Cls => xs.i((.., 0)),
};
match &self.mlp_head {
Some(mlp_head) => xs.apply(mlp_head),
None => xs,
}
}
}