Spaces:
Paused
Paused
| import torch | |
| from torch import nn, einsum | |
| import torch.nn.functional as F | |
| from einops import rearrange, repeat | |
| from einops.layers.torch import Rearrange | |
| from module import Attention, PreNorm, FeedForward, CrossAttention, SSTransformer | |
| import numpy as np | |
| class SSTTransformerEncoder(nn.Module): | |
| def __init__(self, dim, depth, heads, dim_head, mlp_dim, b_dim, b_depth, b_heads, b_dim_head, b_mlp_head, num_patches, cross_attn_depth=3, cross_attn_heads=8, dropout = 0): | |
| super().__init__() | |
| self.transformer = SSTransformer(dim, depth, heads, dim_head, mlp_dim, b_dim, b_depth, b_heads, b_dim_head, b_mlp_head, num_patches, dropout) | |
| self.cross_attn_layers = nn.ModuleList([]) | |
| for _ in range(cross_attn_depth): | |
| self.cross_attn_layers.append(PreNorm(b_dim, CrossAttention(b_dim, heads = cross_attn_heads, dim_head=dim_head, dropout=0))) | |
| def forward(self, x1, x2): | |
| x1 = self.transformer(x1) | |
| x2 = self.transformer(x2) | |
| for cross_attn in self.cross_attn_layers: | |
| x1_class = x1[:, 0] | |
| x1 = x1[:, 1:] | |
| x2_class = x2[:, 0] | |
| x2 = x2[:, 1:] | |
| # Cross Attn | |
| cat1_q = x1_class.unsqueeze(1) | |
| cat1_qkv = torch.cat((cat1_q, x2), dim=1) | |
| cat1_out = cat1_q+cross_attn(cat1_qkv) | |
| x1 = torch.cat((cat1_out, x1), dim=1) | |
| cat2_q = x2_class.unsqueeze(1) | |
| cat2_qkv = torch.cat((cat2_q, x1), dim=1) | |
| cat2_out = cat2_q+cross_attn(cat2_qkv) | |
| x2 = torch.cat((cat2_out, x2), dim=1) | |
| return cat1_out, cat2_out | |
| class SSTViT(nn.Module): | |
| def __init__(self, image_size, near_band, num_patches, num_classes, dim, depth, heads, mlp_dim, b_dim, b_depth, b_heads, b_dim_head, b_mlp_head, pool='cls', channels=1, dim_head = 16, dropout=0., emb_dropout=0., multi_scale_enc_depth=1): | |
| super().__init__() | |
| patch_dim = image_size ** 2 * near_band | |
| self.num_patches = num_patches+1 | |
| self.pos_embedding = nn.Parameter(torch.randn(1, self.num_patches, dim)) | |
| self.patch_to_embedding = nn.Linear(patch_dim, dim) | |
| self.cls_token_t1 = nn.Parameter(torch.randn(1, 1, dim)) | |
| self.cls_token_t2 = nn.Parameter(torch.randn(1, 1, dim)) | |
| self.dropout = nn.Dropout(emb_dropout) | |
| self.multi_scale_transformers = nn.ModuleList([]) | |
| for _ in range(multi_scale_enc_depth): | |
| self.multi_scale_transformers.append(SSTTransformerEncoder(dim, depth, heads, dim_head, mlp_dim,b_dim, b_depth, b_heads, b_dim_head, b_mlp_head, self.num_patches, | |
| dropout = 0.)) | |
| self.pool = pool | |
| self.to_latent = nn.Identity() | |
| self.mlp_head = nn.Sequential( | |
| nn.LayerNorm(b_dim), | |
| nn.Linear(b_dim, num_classes) | |
| ) | |
| def forward(self, x1, x2): | |
| # patchs[batch, patch_num, patch_size*patch_size*c] [batch,200,145*145] | |
| # x = rearrange(x, 'b c h w -> b c (h w)') | |
| ## embedding every patch vector to embedding size: [batch, patch_num, embedding_size] | |
| x1 = self.patch_to_embedding(x1) #[b,n,dim] | |
| x2 = self.patch_to_embedding(x2) | |
| b, n, _ = x1.shape | |
| # add position embedding | |
| cls_tokens_t1 = repeat(self.cls_token_t1, '() n d -> b n d', b = b) #[b,1,dim] | |
| cls_tokens_t2 = repeat(self.cls_token_t2, '() n d -> b n d', b = b) | |
| x1 = torch.cat((cls_tokens_t1, x1), dim = 1) #[b,n+1,dim] | |
| x1 += self.pos_embedding[:, :(n + 1)] | |
| x1 = self.dropout(x1) | |
| x2 = torch.cat((cls_tokens_t2, x2), dim = 1) #[b,n+1,dim] | |
| x2 += self.pos_embedding[:, :(n + 1)] | |
| x2 = self.dropout(x2) | |
| # transformer: x[b,n + 1,dim] -> x[b,n + 1,dim] | |
| for multi_scale_transformer in self.multi_scale_transformers: | |
| out1, out2 = multi_scale_transformer(x1, x2) | |
| # classification: using cls_token output | |
| out1 = self.to_latent(out1[:,0]) | |
| out2 = self.to_latent(out2[:,0]) | |
| out = out1+out2 | |
| # MLP classification layer | |
| return self.mlp_head(out) | |