| |
| |
| |
| |
| |
| |
| |
| from __future__ import annotations |
|
|
| import math |
|
|
| import torch |
| import torch.nn as nn |
| import torch.nn.functional as F |
|
|
| try: |
| from .jit_model_util import RMSNorm, VisionRotaryEmbeddingFast, get_2d_sincos_pos_embed |
| except ImportError: |
| from jit_model_util import RMSNorm, VisionRotaryEmbeddingFast, get_2d_sincos_pos_embed |
|
|
|
|
| def modulate(x, shift, scale): |
| return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1) |
|
|
|
|
| class BottleneckPatchEmbed(nn.Module): |
| """Image to patch embedding.""" |
|
|
| def __init__( |
| self, |
| img_size=224, |
| patch_size=16, |
| in_chans=3, |
| pca_dim=768, |
| embed_dim=768, |
| bias=True, |
| ): |
| super().__init__() |
| img_size = (img_size, img_size) |
| patch_size = (patch_size, patch_size) |
| num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0]) |
| self.img_size = img_size |
| self.patch_size = patch_size |
| self.num_patches = num_patches |
|
|
| self.proj1 = nn.Conv2d(in_chans, pca_dim, kernel_size=patch_size, stride=patch_size, bias=False) |
| self.proj2 = nn.Conv2d(pca_dim, embed_dim, kernel_size=1, stride=1, bias=bias) |
|
|
| def forward(self, x): |
| b, c, h, w = x.shape |
| assert h == self.img_size[0] and w == self.img_size[1], ( |
| f"Input image size ({h}*{w}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})." |
| ) |
|
|
| x = self.proj2(self.proj1(x)).flatten(2).transpose(1, 2) |
| return x |
|
|
|
|
| class TimestepEmbedder(nn.Module): |
| """Embeds scalar timesteps into vector representations.""" |
|
|
| def __init__(self, hidden_size, frequency_embedding_size=256): |
| super().__init__() |
| self.mlp = nn.Sequential( |
| nn.Linear(frequency_embedding_size, hidden_size, bias=True), |
| nn.SiLU(), |
| nn.Linear(hidden_size, hidden_size, bias=True), |
| ) |
| self.frequency_embedding_size = frequency_embedding_size |
|
|
| @staticmethod |
| def timestep_embedding(t, dim, max_period=10000): |
| half = dim // 2 |
| freqs = torch.exp( |
| -math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32, device=t.device) / half |
| ) |
| args = t[:, None].float() * freqs[None] |
| embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1) |
| if dim % 2: |
| embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1) |
| return embedding |
|
|
| def forward(self, t): |
| t_freq = self.timestep_embedding(t, self.frequency_embedding_size) |
| return self.mlp(t_freq) |
|
|
|
|
| def scaled_dot_product_attention( |
| query: torch.Tensor, |
| key: torch.Tensor, |
| value: torch.Tensor, |
| dropout_p: float = 0.0, |
| training: bool = True, |
| ) -> torch.Tensor: |
| scale_factor = 1 / math.sqrt(query.size(-1)) |
| with torch.cuda.amp.autocast(enabled=False): |
| attn_weight = query.float() @ key.float().transpose(-2, -1) * scale_factor |
| attn_weight = torch.softmax(attn_weight, dim=-1) |
| attn_weight = torch.dropout(attn_weight, dropout_p, train=training) |
| return attn_weight @ value |
|
|
|
|
| class Attention(nn.Module): |
| def __init__(self, dim, num_heads=8, qkv_bias=True, qk_norm=True, attn_drop=0.0, proj_drop=0.0): |
| super().__init__() |
| self.num_heads = num_heads |
| head_dim = dim // num_heads |
|
|
| self.q_norm = RMSNorm(head_dim) if qk_norm else nn.Identity() |
| self.k_norm = RMSNorm(head_dim) if qk_norm else nn.Identity() |
|
|
| self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias) |
| self.attn_drop = nn.Dropout(attn_drop) |
| self.proj = nn.Linear(dim, dim) |
| self.proj_drop = nn.Dropout(proj_drop) |
|
|
| def forward(self, x, rope): |
| b, n, c = x.shape |
| qkv = self.qkv(x).reshape(b, n, 3, self.num_heads, c // self.num_heads).permute(2, 0, 3, 1, 4) |
| q, k, v = qkv[0], qkv[1], qkv[2] |
|
|
| q = self.q_norm(q) |
| k = self.k_norm(k) |
|
|
| q = rope(q) |
| k = rope(k) |
|
|
| x = scaled_dot_product_attention( |
| q, |
| k, |
| v, |
| dropout_p=self.attn_drop.p if self.training else 0.0, |
| training=self.training, |
| ) |
|
|
| x = x.transpose(1, 2).reshape(b, n, c) |
|
|
| x = self.proj(x) |
| x = self.proj_drop(x) |
| return x |
|
|
|
|
| class SwiGLUFFN(nn.Module): |
| def __init__(self, dim: int, hidden_dim: int, drop=0.0, bias=True) -> None: |
| super().__init__() |
| hidden_dim = int(hidden_dim * 2 / 3) |
| self.w12 = nn.Linear(dim, 2 * hidden_dim, bias=bias) |
| self.w3 = nn.Linear(hidden_dim, dim, bias=bias) |
| self.ffn_dropout = nn.Dropout(drop) |
|
|
| def forward(self, x): |
| x12 = self.w12(x) |
| x1, x2 = x12.chunk(2, dim=-1) |
| hidden = F.silu(x1) * x2 |
| return self.w3(self.ffn_dropout(hidden)) |
|
|
|
|
| class FinalLayer(nn.Module): |
| """The final layer of JiT.""" |
|
|
| def __init__(self, hidden_size, patch_size, out_channels): |
| super().__init__() |
| self.norm_final = RMSNorm(hidden_size) |
| self.linear = nn.Linear(hidden_size, patch_size * patch_size * out_channels, bias=True) |
| self.adaLN_modulation = nn.Sequential(nn.SiLU(), nn.Linear(hidden_size, 2 * hidden_size, bias=True)) |
|
|
| def forward(self, x, c): |
| shift, scale = self.adaLN_modulation(c).chunk(2, dim=1) |
| x = modulate(self.norm_final(x), shift, scale) |
| return self.linear(x) |
|
|
|
|
| class JiTBlock(nn.Module): |
| def __init__(self, hidden_size, num_heads, mlp_ratio=4.0, attn_drop=0.0, proj_drop=0.0): |
| super().__init__() |
| self.norm1 = RMSNorm(hidden_size, eps=1e-6) |
| self.attn = Attention( |
| hidden_size, |
| num_heads=num_heads, |
| qkv_bias=True, |
| qk_norm=True, |
| attn_drop=attn_drop, |
| proj_drop=proj_drop, |
| ) |
| self.norm2 = RMSNorm(hidden_size, eps=1e-6) |
| mlp_hidden_dim = int(hidden_size * mlp_ratio) |
| self.mlp = SwiGLUFFN(hidden_size, mlp_hidden_dim, drop=proj_drop) |
| self.adaLN_modulation = nn.Sequential(nn.SiLU(), nn.Linear(hidden_size, 6 * hidden_size, bias=True)) |
|
|
| def forward(self, x, c, feat_rope=None): |
| shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(c).chunk(6, dim=-1) |
| x = x + gate_msa.unsqueeze(1) * self.attn( |
| modulate(self.norm1(x), shift_msa, scale_msa), rope=feat_rope |
| ) |
| x = x + gate_mlp.unsqueeze(1) * self.mlp(modulate(self.norm2(x), shift_mlp, scale_mlp)) |
| return x |
|
|
|
|
| class JiT(nn.Module): |
| """ |
| Just image Transformer — unconditional (time embedding only). |
| """ |
|
|
| def __init__( |
| self, |
| input_size=256, |
| patch_size=16, |
| in_channels=3, |
| hidden_size=1024, |
| depth=24, |
| num_heads=16, |
| mlp_ratio=4.0, |
| attn_drop=0.0, |
| proj_drop=0.0, |
| bottleneck_dim=128, |
| in_context_len=32, |
| in_context_start=8, |
| ): |
| super().__init__() |
| self.in_channels = in_channels |
| self.out_channels = in_channels |
| self.patch_size = patch_size |
| self.num_heads = num_heads |
| self.hidden_size = hidden_size |
| self.input_size = input_size |
| self.in_context_len = in_context_len |
| self.in_context_start = in_context_start |
|
|
| self.t_embedder = TimestepEmbedder(hidden_size) |
|
|
| self.x_embedder = BottleneckPatchEmbed( |
| input_size, patch_size, in_channels, bottleneck_dim, hidden_size, bias=True |
| ) |
|
|
| num_patches = self.x_embedder.num_patches |
| self.pos_embed = nn.Parameter(torch.zeros(1, num_patches, hidden_size), requires_grad=False) |
|
|
| if self.in_context_len > 0: |
| self.in_context_posemb = nn.Parameter(torch.zeros(1, self.in_context_len, hidden_size)) |
| torch.nn.init.normal_(self.in_context_posemb, std=0.02) |
|
|
| half_head_dim = hidden_size // num_heads // 2 |
| hw_seq_len = input_size // patch_size |
| self.feat_rope = VisionRotaryEmbeddingFast( |
| dim=half_head_dim, pt_seq_len=hw_seq_len, num_cls_token=0 |
| ) |
| self.feat_rope_incontext = VisionRotaryEmbeddingFast( |
| dim=half_head_dim, pt_seq_len=hw_seq_len, num_cls_token=self.in_context_len |
| ) |
|
|
| self.blocks = nn.ModuleList( |
| [ |
| JiTBlock( |
| hidden_size, |
| num_heads, |
| mlp_ratio=mlp_ratio, |
| attn_drop=attn_drop if (depth // 4 * 3 > i >= depth // 4) else 0.0, |
| proj_drop=proj_drop if (depth // 4 * 3 > i >= depth // 4) else 0.0, |
| ) |
| for i in range(depth) |
| ] |
| ) |
|
|
| self.final_layer = FinalLayer(hidden_size, patch_size, self.out_channels) |
|
|
| self.initialize_weights() |
|
|
| def initialize_weights(self): |
| def _basic_init(module): |
| if isinstance(module, nn.Linear): |
| torch.nn.init.xavier_uniform_(module.weight) |
| if module.bias is not None: |
| nn.init.constant_(module.bias, 0) |
|
|
| self.apply(_basic_init) |
|
|
| pos_embed = get_2d_sincos_pos_embed(self.pos_embed.shape[-1], int(self.x_embedder.num_patches**0.5)) |
| self.pos_embed.data.copy_(torch.from_numpy(pos_embed).float().unsqueeze(0)) |
|
|
| w1 = self.x_embedder.proj1.weight.data |
| nn.init.xavier_uniform_(w1.view([w1.shape[0], -1])) |
| w2 = self.x_embedder.proj2.weight.data |
| nn.init.xavier_uniform_(w2.view([w2.shape[0], -1])) |
| nn.init.constant_(self.x_embedder.proj2.bias, 0) |
|
|
| nn.init.normal_(self.t_embedder.mlp[0].weight, std=0.02) |
| nn.init.normal_(self.t_embedder.mlp[2].weight, std=0.02) |
|
|
| for block in self.blocks: |
| nn.init.constant_(block.adaLN_modulation[-1].weight, 0) |
| nn.init.constant_(block.adaLN_modulation[-1].bias, 0) |
|
|
| nn.init.constant_(self.final_layer.adaLN_modulation[-1].weight, 0) |
| nn.init.constant_(self.final_layer.adaLN_modulation[-1].bias, 0) |
|
|
| nn.init.constant_(self.final_layer.linear.weight, 0) |
| nn.init.constant_(self.final_layer.linear.bias, 0) |
|
|
| def unpatchify(self, x, p): |
| c = self.out_channels |
| h = w = int(x.shape[1] ** 0.5) |
| assert h * w == x.shape[1] |
|
|
| x = x.reshape(shape=(x.shape[0], h, w, p, p, c)) |
| x = torch.einsum("nhwpqc->nchpwq", x) |
| imgs = x.reshape(shape=(x.shape[0], c, h * p, h * p)) |
| return imgs |
|
|
| def forward(self, x: torch.Tensor, t: torch.Tensor) -> torch.Tensor: |
| """ |
| Args: |
| x: (N, C, H, W) |
| t: (N,) timesteps in [0, 1] (or arbitrary floats, as in upstream) |
| Returns: |
| (N, C, H, W) predicted velocity / noise depending on training objective |
| """ |
| c_emb = self.t_embedder(t) |
|
|
| x = self.x_embedder(x) |
| x = x + self.pos_embed |
|
|
| for i, block in enumerate(self.blocks): |
| if self.in_context_len > 0 and i == self.in_context_start: |
| b = x.shape[0] |
| in_context_tokens = self.in_context_posemb.expand(b, self.in_context_len, -1) |
| x = torch.cat([in_context_tokens, x], dim=1) |
| x = block(x, c_emb, self.feat_rope if i < self.in_context_start else self.feat_rope_incontext) |
|
|
| x = x[:, self.in_context_len :] |
|
|
| x = self.final_layer(x, c_emb) |
| return self.unpatchify(x, self.patch_size) |
|
|
|
|
| def JiT_B_16(**kwargs): |
| return JiT( |
| depth=12, |
| hidden_size=768, |
| num_heads=12, |
| bottleneck_dim=128, |
| in_context_len=32, |
| in_context_start=4, |
| patch_size=16, |
| **kwargs, |
| ) |
|
|
|
|
| def JiT_B_32(**kwargs): |
| return JiT( |
| depth=12, |
| hidden_size=768, |
| num_heads=12, |
| bottleneck_dim=128, |
| in_context_len=32, |
| in_context_start=4, |
| patch_size=32, |
| **kwargs, |
| ) |
|
|
|
|
| def JiT_L_16(**kwargs): |
| return JiT( |
| depth=24, |
| hidden_size=1024, |
| num_heads=16, |
| bottleneck_dim=128, |
| in_context_len=32, |
| in_context_start=8, |
| patch_size=16, |
| **kwargs, |
| ) |
|
|
|
|
| def JiT_L_32(**kwargs): |
| return JiT( |
| depth=24, |
| hidden_size=1024, |
| num_heads=16, |
| bottleneck_dim=128, |
| in_context_len=32, |
| in_context_start=8, |
| patch_size=32, |
| **kwargs, |
| ) |
|
|
|
|
| def JiT_H_16(**kwargs): |
| return JiT( |
| depth=32, |
| hidden_size=1280, |
| num_heads=16, |
| bottleneck_dim=256, |
| in_context_len=32, |
| in_context_start=10, |
| patch_size=16, |
| **kwargs, |
| ) |
|
|
|
|
| def JiT_H_32(**kwargs): |
| return JiT( |
| depth=32, |
| hidden_size=1280, |
| num_heads=16, |
| bottleneck_dim=256, |
| in_context_len=32, |
| in_context_start=10, |
| patch_size=32, |
| **kwargs, |
| ) |
|
|
|
|
| JiT_models = { |
| "JiT-B/16": JiT_B_16, |
| "JiT-B/32": JiT_B_32, |
| "JiT-L/16": JiT_L_16, |
| "JiT-L/32": JiT_L_32, |
| "JiT-H/16": JiT_H_16, |
| "JiT-H/32": JiT_H_32, |
| } |
|
|
