hugging-apps/echo-memory
0
1import torch2from .attention import Attention3from .sd_unet import ResnetBlock, UpSampler4from .tiler import TileWorker5from einops import rearrange, repeat6 7 8class VAEAttentionBlock(torch.nn.Module):9 10 def __init__(self, num_attention_heads, attention_head_dim, in_channels, num_layers=1, norm_num_groups=32, eps=1e-5):11 super().__init__()12 inner_dim = num_attention_heads * attention_head_dim13 14 self.norm = torch.nn.GroupNorm(num_groups=norm_num_groups, num_channels=in_channels, eps=eps, affine=True)15 16 self.transformer_blocks = torch.nn.ModuleList([17 Attention(18 inner_dim,19 num_attention_heads,20 attention_head_dim,21 bias_q=True,22 bias_kv=True,23 bias_out=True24 )25 for d in range(num_layers)26 ])27 28 def forward(self, hidden_states, time_emb, text_emb, res_stack):29 batch, _, height, width = hidden_states.shape30 residual = hidden_states31 32 hidden_states = self.norm(hidden_states)33 inner_dim = hidden_states.shape[1]34 hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * width, inner_dim)35 36 for block in self.transformer_blocks:37 hidden_states = block(hidden_states)38 39 hidden_states = hidden_states.reshape(batch, height, width, inner_dim).permute(0, 3, 1, 2).contiguous()40 hidden_states = hidden_states + residual41 42 return hidden_states, time_emb, text_emb, res_stack43 44 45class TemporalResnetBlock(torch.nn.Module):46 47 def __init__(self, in_channels, out_channels, groups=32, eps=1e-5):48 super().__init__()49 self.norm1 = torch.nn.GroupNorm(num_groups=groups, num_channels=in_channels, eps=eps, affine=True)50 self.conv1 = torch.nn.Conv3d(in_channels, out_channels, kernel_size=(3, 1, 1), stride=1, padding=(1, 0, 0))51 self.norm2 = torch.nn.GroupNorm(num_groups=groups, num_channels=out_channels, eps=eps, affine=True)52 self.conv2 = torch.nn.Conv3d(out_channels, out_channels, kernel_size=(3, 1, 1), stride=1, padding=(1, 0, 0))53 self.nonlinearity = torch.nn.SiLU()54 self.mix_factor = torch.nn.Parameter(torch.Tensor([0.5]))55 56 def forward(self, hidden_states, time_emb, text_emb, res_stack, **kwargs):57 x_spatial = hidden_states58 x = rearrange(hidden_states, "T C H W -> 1 C T H W")59 x = self.norm1(x)60 x = self.nonlinearity(x)61 x = self.conv1(x)62 x = self.norm2(x)63 x = self.nonlinearity(x)64 x = self.conv2(x)65 x_temporal = hidden_states + x[0].permute(1, 0, 2, 3)66 alpha = torch.sigmoid(self.mix_factor)67 hidden_states = alpha * x_temporal + (1 - alpha) * x_spatial68 return hidden_states, time_emb, text_emb, res_stack69 70 71class SVDVAEDecoder(torch.nn.Module):72 def __init__(self):73 super().__init__()74 self.scaling_factor = 0.1821575 self.conv_in = torch.nn.Conv2d(4, 512, kernel_size=3, padding=1)76 77 self.blocks = torch.nn.ModuleList([78 # UNetMidBlock79 ResnetBlock(512, 512, eps=1e-6),80 TemporalResnetBlock(512, 512, eps=1e-6),81 VAEAttentionBlock(1, 512, 512, 1, eps=1e-6),82 ResnetBlock(512, 512, eps=1e-6),83 TemporalResnetBlock(512, 512, eps=1e-6),84 # UpDecoderBlock85 ResnetBlock(512, 512, eps=1e-6),86 TemporalResnetBlock(512, 512, eps=1e-6),87 ResnetBlock(512, 512, eps=1e-6),88 TemporalResnetBlock(512, 512, eps=1e-6),89 ResnetBlock(512, 512, eps=1e-6),90 TemporalResnetBlock(512, 512, eps=1e-6),91 UpSampler(512),92 # UpDecoderBlock93 ResnetBlock(512, 512, eps=1e-6),94 TemporalResnetBlock(512, 512, eps=1e-6),95 ResnetBlock(512, 512, eps=1e-6),96 TemporalResnetBlock(512, 512, eps=1e-6),97 ResnetBlock(512, 512, eps=1e-6),98 TemporalResnetBlock(512, 512, eps=1e-6),99 UpSampler(512),100 # UpDecoderBlock101 ResnetBlock(512, 256, eps=1e-6),102 TemporalResnetBlock(256, 256, eps=1e-6),103 ResnetBlock(256, 256, eps=1e-6),104 TemporalResnetBlock(256, 256, eps=1e-6),105 ResnetBlock(256, 256, eps=1e-6),106 TemporalResnetBlock(256, 256, eps=1e-6),107 UpSampler(256),108 # UpDecoderBlock109 ResnetBlock(256, 128, eps=1e-6),110 TemporalResnetBlock(128, 128, eps=1e-6),111 ResnetBlock(128, 128, eps=1e-6),112 TemporalResnetBlock(128, 128, eps=1e-6),113 ResnetBlock(128, 128, eps=1e-6),114 TemporalResnetBlock(128, 128, eps=1e-6),115 ])116 117 self.conv_norm_out = torch.nn.GroupNorm(num_channels=128, num_groups=32, eps=1e-5)118 self.conv_act = torch.nn.SiLU()119 self.conv_out = torch.nn.Conv2d(128, 3, kernel_size=3, padding=1)120 self.time_conv_out = torch.nn.Conv3d(3, 3, kernel_size=(3, 1, 1), padding=(1, 0, 0))121 122 123 def forward(self, sample):124 # 1. pre-process125 hidden_states = rearrange(sample, "C T H W -> T C H W")126 hidden_states = hidden_states / self.scaling_factor127 hidden_states = self.conv_in(hidden_states)128 time_emb, text_emb, res_stack = None, None, None129 130 # 2. blocks131 for i, block in enumerate(self.blocks):132 hidden_states, time_emb, text_emb, res_stack = block(hidden_states, time_emb, text_emb, res_stack)133 134 # 3. output135 hidden_states = self.conv_norm_out(hidden_states)136 hidden_states = self.conv_act(hidden_states)137 hidden_states = self.conv_out(hidden_states)138 hidden_states = rearrange(hidden_states, "T C H W -> C T H W")139 hidden_states = self.time_conv_out(hidden_states)140 141 return hidden_states142 143 144 def build_mask(self, data, is_bound):145 _, T, H, W = data.shape146 t = repeat(torch.arange(T), "T -> T H W", T=T, H=H, W=W)147 h = repeat(torch.arange(H), "H -> T H W", T=T, H=H, W=W)148 w = repeat(torch.arange(W), "W -> T H W", T=T, H=H, W=W)149 border_width = (T + H + W) // 6150 pad = torch.ones_like(t) * border_width151 mask = torch.stack([152 pad if is_bound[0] else t + 1,153 pad if is_bound[1] else T - t,154 pad if is_bound[2] else h + 1,155 pad if is_bound[3] else H - h,156 pad if is_bound[4] else w + 1,157 pad if is_bound[5] else W - w158 ]).min(dim=0).values159 mask = mask.clip(1, border_width)160 mask = (mask / border_width).to(dtype=data.dtype, device=data.device)161 mask = rearrange(mask, "T H W -> 1 T H W")162 return mask163 164 165 def decode_video(166 self, sample,167 batch_time=8, batch_height=128, batch_width=128,168 stride_time=4, stride_height=32, stride_width=32,169 progress_bar=lambda x:x170 ):171 sample = sample.permute(1, 0, 2, 3)172 data_device = sample.device173 computation_device = self.conv_in.weight.device174 torch_dtype = sample.dtype175 _, T, H, W = sample.shape176 177 weight = torch.zeros((1, T, H*8, W*8), dtype=torch_dtype, device=data_device)178 values = torch.zeros((3, T, H*8, W*8), dtype=torch_dtype, device=data_device)179 180 # Split tasks181 tasks = []182 for t in range(0, T, stride_time):183 for h in range(0, H, stride_height):184 for w in range(0, W, stride_width):185 if (t-stride_time >= 0 and t-stride_time+batch_time >= T)\186 or (h-stride_height >= 0 and h-stride_height+batch_height >= H)\187 or (w-stride_width >= 0 and w-stride_width+batch_width >= W):188 continue189 tasks.append((t, t+batch_time, h, h+batch_height, w, w+batch_width))190 191 # Run192 for tl, tr, hl, hr, wl, wr in progress_bar(tasks):193 sample_batch = sample[:, tl:tr, hl:hr, wl:wr].to(computation_device)194 sample_batch = self.forward(sample_batch).to(data_device)195 mask = self.build_mask(sample_batch, is_bound=(tl==0, tr>=T, hl==0, hr>=H, wl==0, wr>=W))196 values[:, tl:tr, hl*8:hr*8, wl*8:wr*8] += sample_batch * mask197 weight[:, tl:tr, hl*8:hr*8, wl*8:wr*8] += mask198 values /= weight199 return values200 201 202 @staticmethod203 def state_dict_converter():204 return SVDVAEDecoderStateDictConverter()205 206 207class SVDVAEDecoderStateDictConverter:208 def __init__(self):209 pass210 211 def from_diffusers(self, state_dict):212 static_rename_dict = {213 "decoder.conv_in": "conv_in",214 "decoder.mid_block.attentions.0.group_norm": "blocks.2.norm",215 "decoder.mid_block.attentions.0.to_q": "blocks.2.transformer_blocks.0.to_q",216 "decoder.mid_block.attentions.0.to_k": "blocks.2.transformer_blocks.0.to_k",217 "decoder.mid_block.attentions.0.to_v": "blocks.2.transformer_blocks.0.to_v",218 "decoder.mid_block.attentions.0.to_out.0": "blocks.2.transformer_blocks.0.to_out",219 "decoder.up_blocks.0.upsamplers.0.conv": "blocks.11.conv",220 "decoder.up_blocks.1.upsamplers.0.conv": "blocks.18.conv",221 "decoder.up_blocks.2.upsamplers.0.conv": "blocks.25.conv",222 "decoder.conv_norm_out": "conv_norm_out",223 "decoder.conv_out": "conv_out",224 "decoder.time_conv_out": "time_conv_out"225 }226 prefix_rename_dict = {227 "decoder.mid_block.resnets.0.spatial_res_block": "blocks.0",228 "decoder.mid_block.resnets.0.temporal_res_block": "blocks.1",229 "decoder.mid_block.resnets.0.time_mixer": "blocks.1",230 "decoder.mid_block.resnets.1.spatial_res_block": "blocks.3",231 "decoder.mid_block.resnets.1.temporal_res_block": "blocks.4",232 "decoder.mid_block.resnets.1.time_mixer": "blocks.4",233 234 "decoder.up_blocks.0.resnets.0.spatial_res_block": "blocks.5",235 "decoder.up_blocks.0.resnets.0.temporal_res_block": "blocks.6",236 "decoder.up_blocks.0.resnets.0.time_mixer": "blocks.6",237 "decoder.up_blocks.0.resnets.1.spatial_res_block": "blocks.7",238 "decoder.up_blocks.0.resnets.1.temporal_res_block": "blocks.8",239 "decoder.up_blocks.0.resnets.1.time_mixer": "blocks.8",240 "decoder.up_blocks.0.resnets.2.spatial_res_block": "blocks.9",241 "decoder.up_blocks.0.resnets.2.temporal_res_block": "blocks.10",242 "decoder.up_blocks.0.resnets.2.time_mixer": "blocks.10",243 244 "decoder.up_blocks.1.resnets.0.spatial_res_block": "blocks.12",245 "decoder.up_blocks.1.resnets.0.temporal_res_block": "blocks.13",246 "decoder.up_blocks.1.resnets.0.time_mixer": "blocks.13",247 "decoder.up_blocks.1.resnets.1.spatial_res_block": "blocks.14",248 "decoder.up_blocks.1.resnets.1.temporal_res_block": "blocks.15",249 "decoder.up_blocks.1.resnets.1.time_mixer": "blocks.15",250 "decoder.up_blocks.1.resnets.2.spatial_res_block": "blocks.16",251 "decoder.up_blocks.1.resnets.2.temporal_res_block": "blocks.17",252 "decoder.up_blocks.1.resnets.2.time_mixer": "blocks.17",253 254 "decoder.up_blocks.2.resnets.0.spatial_res_block": "blocks.19",255 "decoder.up_blocks.2.resnets.0.temporal_res_block": "blocks.20",256 "decoder.up_blocks.2.resnets.0.time_mixer": "blocks.20",257 "decoder.up_blocks.2.resnets.1.spatial_res_block": "blocks.21",258 "decoder.up_blocks.2.resnets.1.temporal_res_block": "blocks.22",259 "decoder.up_blocks.2.resnets.1.time_mixer": "blocks.22",260 "decoder.up_blocks.2.resnets.2.spatial_res_block": "blocks.23",261 "decoder.up_blocks.2.resnets.2.temporal_res_block": "blocks.24",262 "decoder.up_blocks.2.resnets.2.time_mixer": "blocks.24",263 264 "decoder.up_blocks.3.resnets.0.spatial_res_block": "blocks.26",265 "decoder.up_blocks.3.resnets.0.temporal_res_block": "blocks.27",266 "decoder.up_blocks.3.resnets.0.time_mixer": "blocks.27",267 "decoder.up_blocks.3.resnets.1.spatial_res_block": "blocks.28",268 "decoder.up_blocks.3.resnets.1.temporal_res_block": "blocks.29",269 "decoder.up_blocks.3.resnets.1.time_mixer": "blocks.29",270 "decoder.up_blocks.3.resnets.2.spatial_res_block": "blocks.30",271 "decoder.up_blocks.3.resnets.2.temporal_res_block": "blocks.31",272 "decoder.up_blocks.3.resnets.2.time_mixer": "blocks.31",273 }274 suffix_rename_dict = {275 "norm1.weight": "norm1.weight",276 "conv1.weight": "conv1.weight",277 "norm2.weight": "norm2.weight",278 "conv2.weight": "conv2.weight",279 "conv_shortcut.weight": "conv_shortcut.weight",280 "norm1.bias": "norm1.bias",281 "conv1.bias": "conv1.bias",282 "norm2.bias": "norm2.bias",283 "conv2.bias": "conv2.bias",284 "conv_shortcut.bias": "conv_shortcut.bias",285 "mix_factor": "mix_factor",286 }287 288 state_dict_ = {}289 for name in static_rename_dict:290 state_dict_[static_rename_dict[name] + ".weight"] = state_dict[name + ".weight"]291 state_dict_[static_rename_dict[name] + ".bias"] = state_dict[name + ".bias"]292 for prefix_name in prefix_rename_dict:293 for suffix_name in suffix_rename_dict:294 name = prefix_name + "." + suffix_name295 name_ = prefix_rename_dict[prefix_name] + "." + suffix_rename_dict[suffix_name]296 if name in state_dict:297 state_dict_[name_] = state_dict[name]298 299 return state_dict_300 301 302 def from_civitai(self, state_dict):303 rename_dict = {304 "first_stage_model.decoder.conv_in.bias": "conv_in.bias",305 "first_stage_model.decoder.conv_in.weight": "conv_in.weight",306 "first_stage_model.decoder.conv_out.bias": "conv_out.bias",307 "first_stage_model.decoder.conv_out.time_mix_conv.bias": "time_conv_out.bias",308 "first_stage_model.decoder.conv_out.time_mix_conv.weight": "time_conv_out.weight",309 "first_stage_model.decoder.conv_out.weight": "conv_out.weight",310 "first_stage_model.decoder.mid.attn_1.k.bias": "blocks.2.transformer_blocks.0.to_k.bias",311 "first_stage_model.decoder.mid.attn_1.k.weight": "blocks.2.transformer_blocks.0.to_k.weight",312 "first_stage_model.decoder.mid.attn_1.norm.bias": "blocks.2.norm.bias",313 "first_stage_model.decoder.mid.attn_1.norm.weight": "blocks.2.norm.weight",314 "first_stage_model.decoder.mid.attn_1.proj_out.bias": "blocks.2.transformer_blocks.0.to_out.bias",315 "first_stage_model.decoder.mid.attn_1.proj_out.weight": "blocks.2.transformer_blocks.0.to_out.weight",316 "first_stage_model.decoder.mid.attn_1.q.bias": "blocks.2.transformer_blocks.0.to_q.bias",317 "first_stage_model.decoder.mid.attn_1.q.weight": "blocks.2.transformer_blocks.0.to_q.weight",318 "first_stage_model.decoder.mid.attn_1.v.bias": "blocks.2.transformer_blocks.0.to_v.bias",319 "first_stage_model.decoder.mid.attn_1.v.weight": "blocks.2.transformer_blocks.0.to_v.weight",320 "first_stage_model.decoder.mid.block_1.conv1.bias": "blocks.0.conv1.bias",321 "first_stage_model.decoder.mid.block_1.conv1.weight": "blocks.0.conv1.weight",322 "first_stage_model.decoder.mid.block_1.conv2.bias": "blocks.0.conv2.bias",323 "first_stage_model.decoder.mid.block_1.conv2.weight": "blocks.0.conv2.weight",324 "first_stage_model.decoder.mid.block_1.mix_factor": "blocks.1.mix_factor",325 "first_stage_model.decoder.mid.block_1.norm1.bias": "blocks.0.norm1.bias",326 "first_stage_model.decoder.mid.block_1.norm1.weight": "blocks.0.norm1.weight",327 "first_stage_model.decoder.mid.block_1.norm2.bias": "blocks.0.norm2.bias",328 "first_stage_model.decoder.mid.block_1.norm2.weight": "blocks.0.norm2.weight",329 "first_stage_model.decoder.mid.block_1.time_stack.in_layers.0.bias": "blocks.1.norm1.bias",330 "first_stage_model.decoder.mid.block_1.time_stack.in_layers.0.weight": "blocks.1.norm1.weight",331 "first_stage_model.decoder.mid.block_1.time_stack.in_layers.2.bias": "blocks.1.conv1.bias",332 "first_stage_model.decoder.mid.block_1.time_stack.in_layers.2.weight": "blocks.1.conv1.weight",333 "first_stage_model.decoder.mid.block_1.time_stack.out_layers.0.bias": "blocks.1.norm2.bias",334 "first_stage_model.decoder.mid.block_1.time_stack.out_layers.0.weight": "blocks.1.norm2.weight",335 "first_stage_model.decoder.mid.block_1.time_stack.out_layers.3.bias": "blocks.1.conv2.bias",336 "first_stage_model.decoder.mid.block_1.time_stack.out_layers.3.weight": "blocks.1.conv2.weight",337 "first_stage_model.decoder.mid.block_2.conv1.bias": "blocks.3.conv1.bias",338 "first_stage_model.decoder.mid.block_2.conv1.weight": "blocks.3.conv1.weight",339 "first_stage_model.decoder.mid.block_2.conv2.bias": "blocks.3.conv2.bias",340 "first_stage_model.decoder.mid.block_2.conv2.weight": "blocks.3.conv2.weight",341 "first_stage_model.decoder.mid.block_2.mix_factor": "blocks.4.mix_factor",342 "first_stage_model.decoder.mid.block_2.norm1.bias": "blocks.3.norm1.bias",343 "first_stage_model.decoder.mid.block_2.norm1.weight": "blocks.3.norm1.weight",344 "first_stage_model.decoder.mid.block_2.norm2.bias": "blocks.3.norm2.bias",345 "first_stage_model.decoder.mid.block_2.norm2.weight": "blocks.3.norm2.weight",346 "first_stage_model.decoder.mid.block_2.time_stack.in_layers.0.bias": "blocks.4.norm1.bias",347 "first_stage_model.decoder.mid.block_2.time_stack.in_layers.0.weight": "blocks.4.norm1.weight",348 "first_stage_model.decoder.mid.block_2.time_stack.in_layers.2.bias": "blocks.4.conv1.bias",349 "first_stage_model.decoder.mid.block_2.time_stack.in_layers.2.weight": "blocks.4.conv1.weight",350 "first_stage_model.decoder.mid.block_2.time_stack.out_layers.0.bias": "blocks.4.norm2.bias",351 "first_stage_model.decoder.mid.block_2.time_stack.out_layers.0.weight": "blocks.4.norm2.weight",352 "first_stage_model.decoder.mid.block_2.time_stack.out_layers.3.bias": "blocks.4.conv2.bias",353 "first_stage_model.decoder.mid.block_2.time_stack.out_layers.3.weight": "blocks.4.conv2.weight",354 "first_stage_model.decoder.norm_out.bias": "conv_norm_out.bias",355 "first_stage_model.decoder.norm_out.weight": "conv_norm_out.weight",356 "first_stage_model.decoder.up.0.block.0.conv1.bias": "blocks.26.conv1.bias",357 "first_stage_model.decoder.up.0.block.0.conv1.weight": "blocks.26.conv1.weight",358 "first_stage_model.decoder.up.0.block.0.conv2.bias": "blocks.26.conv2.bias",359 "first_stage_model.decoder.up.0.block.0.conv2.weight": "blocks.26.conv2.weight",360 "first_stage_model.decoder.up.0.block.0.mix_factor": "blocks.27.mix_factor",361 "first_stage_model.decoder.up.0.block.0.nin_shortcut.bias": "blocks.26.conv_shortcut.bias",362 "first_stage_model.decoder.up.0.block.0.nin_shortcut.weight": "blocks.26.conv_shortcut.weight",363 "first_stage_model.decoder.up.0.block.0.norm1.bias": "blocks.26.norm1.bias",364 "first_stage_model.decoder.up.0.block.0.norm1.weight": "blocks.26.norm1.weight",365 "first_stage_model.decoder.up.0.block.0.norm2.bias": "blocks.26.norm2.bias",366 "first_stage_model.decoder.up.0.block.0.norm2.weight": "blocks.26.norm2.weight",367 "first_stage_model.decoder.up.0.block.0.time_stack.in_layers.0.bias": "blocks.27.norm1.bias",368 "first_stage_model.decoder.up.0.block.0.time_stack.in_layers.0.weight": "blocks.27.norm1.weight",369 "first_stage_model.decoder.up.0.block.0.time_stack.in_layers.2.bias": "blocks.27.conv1.bias",370 "first_stage_model.decoder.up.0.block.0.time_stack.in_layers.2.weight": "blocks.27.conv1.weight",371 "first_stage_model.decoder.up.0.block.0.time_stack.out_layers.0.bias": "blocks.27.norm2.bias",372 "first_stage_model.decoder.up.0.block.0.time_stack.out_layers.0.weight": "blocks.27.norm2.weight",373 "first_stage_model.decoder.up.0.block.0.time_stack.out_layers.3.bias": "blocks.27.conv2.bias",374 "first_stage_model.decoder.up.0.block.0.time_stack.out_layers.3.weight": "blocks.27.conv2.weight",375 "first_stage_model.decoder.up.0.block.1.conv1.bias": "blocks.28.conv1.bias",376 "first_stage_model.decoder.up.0.block.1.conv1.weight": "blocks.28.conv1.weight",377 "first_stage_model.decoder.up.0.block.1.conv2.bias": "blocks.28.conv2.bias",378 "first_stage_model.decoder.up.0.block.1.conv2.weight": "blocks.28.conv2.weight",379 "first_stage_model.decoder.up.0.block.1.mix_factor": "blocks.29.mix_factor",380 "first_stage_model.decoder.up.0.block.1.norm1.bias": "blocks.28.norm1.bias",381 "first_stage_model.decoder.up.0.block.1.norm1.weight": "blocks.28.norm1.weight",382 "first_stage_model.decoder.up.0.block.1.norm2.bias": "blocks.28.norm2.bias",383 "first_stage_model.decoder.up.0.block.1.norm2.weight": "blocks.28.norm2.weight",384 "first_stage_model.decoder.up.0.block.1.time_stack.in_layers.0.bias": "blocks.29.norm1.bias",385 "first_stage_model.decoder.up.0.block.1.time_stack.in_layers.0.weight": "blocks.29.norm1.weight",386 "first_stage_model.decoder.up.0.block.1.time_stack.in_layers.2.bias": "blocks.29.conv1.bias",387 "first_stage_model.decoder.up.0.block.1.time_stack.in_layers.2.weight": "blocks.29.conv1.weight",388 "first_stage_model.decoder.up.0.block.1.time_stack.out_layers.0.bias": "blocks.29.norm2.bias",389 "first_stage_model.decoder.up.0.block.1.time_stack.out_layers.0.weight": "blocks.29.norm2.weight",390 "first_stage_model.decoder.up.0.block.1.time_stack.out_layers.3.bias": "blocks.29.conv2.bias",391 "first_stage_model.decoder.up.0.block.1.time_stack.out_layers.3.weight": "blocks.29.conv2.weight",392 "first_stage_model.decoder.up.0.block.2.conv1.bias": "blocks.30.conv1.bias",393 "first_stage_model.decoder.up.0.block.2.conv1.weight": "blocks.30.conv1.weight",394 "first_stage_model.decoder.up.0.block.2.conv2.bias": "blocks.30.conv2.bias",395 "first_stage_model.decoder.up.0.block.2.conv2.weight": "blocks.30.conv2.weight",396 "first_stage_model.decoder.up.0.block.2.mix_factor": "blocks.31.mix_factor",397 "first_stage_model.decoder.up.0.block.2.norm1.bias": "blocks.30.norm1.bias",398 "first_stage_model.decoder.up.0.block.2.norm1.weight": "blocks.30.norm1.weight",399 "first_stage_model.decoder.up.0.block.2.norm2.bias": "blocks.30.norm2.bias",400 "first_stage_model.decoder.up.0.block.2.norm2.weight": "blocks.30.norm2.weight",401 "first_stage_model.decoder.up.0.block.2.time_stack.in_layers.0.bias": "blocks.31.norm1.bias",402 "first_stage_model.decoder.up.0.block.2.time_stack.in_layers.0.weight": "blocks.31.norm1.weight",403 "first_stage_model.decoder.up.0.block.2.time_stack.in_layers.2.bias": "blocks.31.conv1.bias",404 "first_stage_model.decoder.up.0.block.2.time_stack.in_layers.2.weight": "blocks.31.conv1.weight",405 "first_stage_model.decoder.up.0.block.2.time_stack.out_layers.0.bias": "blocks.31.norm2.bias",406 "first_stage_model.decoder.up.0.block.2.time_stack.out_layers.0.weight": "blocks.31.norm2.weight",407 "first_stage_model.decoder.up.0.block.2.time_stack.out_layers.3.bias": "blocks.31.conv2.bias",408 "first_stage_model.decoder.up.0.block.2.time_stack.out_layers.3.weight": "blocks.31.conv2.weight",409 "first_stage_model.decoder.up.1.block.0.conv1.bias": "blocks.19.conv1.bias",410 "first_stage_model.decoder.up.1.block.0.conv1.weight": "blocks.19.conv1.weight",411 "first_stage_model.decoder.up.1.block.0.conv2.bias": "blocks.19.conv2.bias",412 "first_stage_model.decoder.up.1.block.0.conv2.weight": "blocks.19.conv2.weight",413 "first_stage_model.decoder.up.1.block.0.mix_factor": "blocks.20.mix_factor",414 "first_stage_model.decoder.up.1.block.0.nin_shortcut.bias": "blocks.19.conv_shortcut.bias",415 "first_stage_model.decoder.up.1.block.0.nin_shortcut.weight": "blocks.19.conv_shortcut.weight",416 "first_stage_model.decoder.up.1.block.0.norm1.bias": "blocks.19.norm1.bias",417 "first_stage_model.decoder.up.1.block.0.norm1.weight": "blocks.19.norm1.weight",418 "first_stage_model.decoder.up.1.block.0.norm2.bias": "blocks.19.norm2.bias",419 "first_stage_model.decoder.up.1.block.0.norm2.weight": "blocks.19.norm2.weight",420 "first_stage_model.decoder.up.1.block.0.time_stack.in_layers.0.bias": "blocks.20.norm1.bias",421 "first_stage_model.decoder.up.1.block.0.time_stack.in_layers.0.weight": "blocks.20.norm1.weight",422 "first_stage_model.decoder.up.1.block.0.time_stack.in_layers.2.bias": "blocks.20.conv1.bias",423 "first_stage_model.decoder.up.1.block.0.time_stack.in_layers.2.weight": "blocks.20.conv1.weight",424 "first_stage_model.decoder.up.1.block.0.time_stack.out_layers.0.bias": "blocks.20.norm2.bias",425 "first_stage_model.decoder.up.1.block.0.time_stack.out_layers.0.weight": "blocks.20.norm2.weight",426 "first_stage_model.decoder.up.1.block.0.time_stack.out_layers.3.bias": "blocks.20.conv2.bias",427 "first_stage_model.decoder.up.1.block.0.time_stack.out_layers.3.weight": "blocks.20.conv2.weight",428 "first_stage_model.decoder.up.1.block.1.conv1.bias": "blocks.21.conv1.bias",429 "first_stage_model.decoder.up.1.block.1.conv1.weight": "blocks.21.conv1.weight",430 "first_stage_model.decoder.up.1.block.1.conv2.bias": "blocks.21.conv2.bias",431 "first_stage_model.decoder.up.1.block.1.conv2.weight": "blocks.21.conv2.weight",432 "first_stage_model.decoder.up.1.block.1.mix_factor": "blocks.22.mix_factor",433 "first_stage_model.decoder.up.1.block.1.norm1.bias": "blocks.21.norm1.bias",434 "first_stage_model.decoder.up.1.block.1.norm1.weight": "blocks.21.norm1.weight",435 "first_stage_model.decoder.up.1.block.1.norm2.bias": "blocks.21.norm2.bias",436 "first_stage_model.decoder.up.1.block.1.norm2.weight": "blocks.21.norm2.weight",437 "first_stage_model.decoder.up.1.block.1.time_stack.in_layers.0.bias": "blocks.22.norm1.bias",438 "first_stage_model.decoder.up.1.block.1.time_stack.in_layers.0.weight": "blocks.22.norm1.weight",439 "first_stage_model.decoder.up.1.block.1.time_stack.in_layers.2.bias": "blocks.22.conv1.bias",440 "first_stage_model.decoder.up.1.block.1.time_stack.in_layers.2.weight": "blocks.22.conv1.weight",441 "first_stage_model.decoder.up.1.block.1.time_stack.out_layers.0.bias": "blocks.22.norm2.bias",442 "first_stage_model.decoder.up.1.block.1.time_stack.out_layers.0.weight": "blocks.22.norm2.weight",443 "first_stage_model.decoder.up.1.block.1.time_stack.out_layers.3.bias": "blocks.22.conv2.bias",444 "first_stage_model.decoder.up.1.block.1.time_stack.out_layers.3.weight": "blocks.22.conv2.weight",445 "first_stage_model.decoder.up.1.block.2.conv1.bias": "blocks.23.conv1.bias",446 "first_stage_model.decoder.up.1.block.2.conv1.weight": "blocks.23.conv1.weight",447 "first_stage_model.decoder.up.1.block.2.conv2.bias": "blocks.23.conv2.bias",448 "first_stage_model.decoder.up.1.block.2.conv2.weight": "blocks.23.conv2.weight",449 "first_stage_model.decoder.up.1.block.2.mix_factor": "blocks.24.mix_factor",450 "first_stage_model.decoder.up.1.block.2.norm1.bias": "blocks.23.norm1.bias",451 "first_stage_model.decoder.up.1.block.2.norm1.weight": "blocks.23.norm1.weight",452 "first_stage_model.decoder.up.1.block.2.norm2.bias": "blocks.23.norm2.bias",453 "first_stage_model.decoder.up.1.block.2.norm2.weight": "blocks.23.norm2.weight",454 "first_stage_model.decoder.up.1.block.2.time_stack.in_layers.0.bias": "blocks.24.norm1.bias",455 "first_stage_model.decoder.up.1.block.2.time_stack.in_layers.0.weight": "blocks.24.norm1.weight",456 "first_stage_model.decoder.up.1.block.2.time_stack.in_layers.2.bias": "blocks.24.conv1.bias",457 "first_stage_model.decoder.up.1.block.2.time_stack.in_layers.2.weight": "blocks.24.conv1.weight",458 "first_stage_model.decoder.up.1.block.2.time_stack.out_layers.0.bias": "blocks.24.norm2.bias",459 "first_stage_model.decoder.up.1.block.2.time_stack.out_layers.0.weight": "blocks.24.norm2.weight",460 "first_stage_model.decoder.up.1.block.2.time_stack.out_layers.3.bias": "blocks.24.conv2.bias",461 "first_stage_model.decoder.up.1.block.2.time_stack.out_layers.3.weight": "blocks.24.conv2.weight",462 "first_stage_model.decoder.up.1.upsample.conv.bias": "blocks.25.conv.bias",463 "first_stage_model.decoder.up.1.upsample.conv.weight": "blocks.25.conv.weight",464 "first_stage_model.decoder.up.2.block.0.conv1.bias": "blocks.12.conv1.bias",465 "first_stage_model.decoder.up.2.block.0.conv1.weight": "blocks.12.conv1.weight",466 "first_stage_model.decoder.up.2.block.0.conv2.bias": "blocks.12.conv2.bias",467 "first_stage_model.decoder.up.2.block.0.conv2.weight": "blocks.12.conv2.weight",468 "first_stage_model.decoder.up.2.block.0.mix_factor": "blocks.13.mix_factor",469 "first_stage_model.decoder.up.2.block.0.norm1.bias": "blocks.12.norm1.bias",470 "first_stage_model.decoder.up.2.block.0.norm1.weight": "blocks.12.norm1.weight",471 "first_stage_model.decoder.up.2.block.0.norm2.bias": "blocks.12.norm2.bias",472 "first_stage_model.decoder.up.2.block.0.norm2.weight": "blocks.12.norm2.weight",473 "first_stage_model.decoder.up.2.block.0.time_stack.in_layers.0.bias": "blocks.13.norm1.bias",474 "first_stage_model.decoder.up.2.block.0.time_stack.in_layers.0.weight": "blocks.13.norm1.weight",475 "first_stage_model.decoder.up.2.block.0.time_stack.in_layers.2.bias": "blocks.13.conv1.bias",476 "first_stage_model.decoder.up.2.block.0.time_stack.in_layers.2.weight": "blocks.13.conv1.weight",477 "first_stage_model.decoder.up.2.block.0.time_stack.out_layers.0.bias": "blocks.13.norm2.bias",478 "first_stage_model.decoder.up.2.block.0.time_stack.out_layers.0.weight": "blocks.13.norm2.weight",479 "first_stage_model.decoder.up.2.block.0.time_stack.out_layers.3.bias": "blocks.13.conv2.bias",480 "first_stage_model.decoder.up.2.block.0.time_stack.out_layers.3.weight": "blocks.13.conv2.weight",481 "first_stage_model.decoder.up.2.block.1.conv1.bias": "blocks.14.conv1.bias",482 "first_stage_model.decoder.up.2.block.1.conv1.weight": "blocks.14.conv1.weight",483 "first_stage_model.decoder.up.2.block.1.conv2.bias": "blocks.14.conv2.bias",484 "first_stage_model.decoder.up.2.block.1.conv2.weight": "blocks.14.conv2.weight",485 "first_stage_model.decoder.up.2.block.1.mix_factor": "blocks.15.mix_factor",486 "first_stage_model.decoder.up.2.block.1.norm1.bias": "blocks.14.norm1.bias",487 "first_stage_model.decoder.up.2.block.1.norm1.weight": "blocks.14.norm1.weight",488 "first_stage_model.decoder.up.2.block.1.norm2.bias": "blocks.14.norm2.bias",489 "first_stage_model.decoder.up.2.block.1.norm2.weight": "blocks.14.norm2.weight",490 "first_stage_model.decoder.up.2.block.1.time_stack.in_layers.0.bias": "blocks.15.norm1.bias",491 "first_stage_model.decoder.up.2.block.1.time_stack.in_layers.0.weight": "blocks.15.norm1.weight",492 "first_stage_model.decoder.up.2.block.1.time_stack.in_layers.2.bias": "blocks.15.conv1.bias",493 "first_stage_model.decoder.up.2.block.1.time_stack.in_layers.2.weight": "blocks.15.conv1.weight",494 "first_stage_model.decoder.up.2.block.1.time_stack.out_layers.0.bias": "blocks.15.norm2.bias",495 "first_stage_model.decoder.up.2.block.1.time_stack.out_layers.0.weight": "blocks.15.norm2.weight",496 "first_stage_model.decoder.up.2.block.1.time_stack.out_layers.3.bias": "blocks.15.conv2.bias",497 "first_stage_model.decoder.up.2.block.1.time_stack.out_layers.3.weight": "blocks.15.conv2.weight",498 "first_stage_model.decoder.up.2.block.2.conv1.bias": "blocks.16.conv1.bias",499 "first_stage_model.decoder.up.2.block.2.conv1.weight": "blocks.16.conv1.weight",500 "first_stage_model.decoder.up.2.block.2.conv2.bias": "blocks.16.conv2.bias",501 "first_stage_model.decoder.up.2.block.2.conv2.weight": "blocks.16.conv2.weight",502 "first_stage_model.decoder.up.2.block.2.mix_factor": "blocks.17.mix_factor",503 "first_stage_model.decoder.up.2.block.2.norm1.bias": "blocks.16.norm1.bias",504 "first_stage_model.decoder.up.2.block.2.norm1.weight": "blocks.16.norm1.weight",505 "first_stage_model.decoder.up.2.block.2.norm2.bias": "blocks.16.norm2.bias",506 "first_stage_model.decoder.up.2.block.2.norm2.weight": "blocks.16.norm2.weight",507 "first_stage_model.decoder.up.2.block.2.time_stack.in_layers.0.bias": "blocks.17.norm1.bias",508 "first_stage_model.decoder.up.2.block.2.time_stack.in_layers.0.weight": "blocks.17.norm1.weight",509 "first_stage_model.decoder.up.2.block.2.time_stack.in_layers.2.bias": "blocks.17.conv1.bias",510 "first_stage_model.decoder.up.2.block.2.time_stack.in_layers.2.weight": "blocks.17.conv1.weight",511 "first_stage_model.decoder.up.2.block.2.time_stack.out_layers.0.bias": "blocks.17.norm2.bias",512 "first_stage_model.decoder.up.2.block.2.time_stack.out_layers.0.weight": "blocks.17.norm2.weight",513 "first_stage_model.decoder.up.2.block.2.time_stack.out_layers.3.bias": "blocks.17.conv2.bias",514 "first_stage_model.decoder.up.2.block.2.time_stack.out_layers.3.weight": "blocks.17.conv2.weight",515 "first_stage_model.decoder.up.2.upsample.conv.bias": "blocks.18.conv.bias",516 "first_stage_model.decoder.up.2.upsample.conv.weight": "blocks.18.conv.weight",517 "first_stage_model.decoder.up.3.block.0.conv1.bias": "blocks.5.conv1.bias",518 "first_stage_model.decoder.up.3.block.0.conv1.weight": "blocks.5.conv1.weight",519 "first_stage_model.decoder.up.3.block.0.conv2.bias": "blocks.5.conv2.bias",520 "first_stage_model.decoder.up.3.block.0.conv2.weight": "blocks.5.conv2.weight",521 "first_stage_model.decoder.up.3.block.0.mix_factor": "blocks.6.mix_factor",522 "first_stage_model.decoder.up.3.block.0.norm1.bias": "blocks.5.norm1.bias",523 "first_stage_model.decoder.up.3.block.0.norm1.weight": "blocks.5.norm1.weight",524 "first_stage_model.decoder.up.3.block.0.norm2.bias": "blocks.5.norm2.bias",525 "first_stage_model.decoder.up.3.block.0.norm2.weight": "blocks.5.norm2.weight",526 "first_stage_model.decoder.up.3.block.0.time_stack.in_layers.0.bias": "blocks.6.norm1.bias",527 "first_stage_model.decoder.up.3.block.0.time_stack.in_layers.0.weight": "blocks.6.norm1.weight",528 "first_stage_model.decoder.up.3.block.0.time_stack.in_layers.2.bias": "blocks.6.conv1.bias",529 "first_stage_model.decoder.up.3.block.0.time_stack.in_layers.2.weight": "blocks.6.conv1.weight",530 "first_stage_model.decoder.up.3.block.0.time_stack.out_layers.0.bias": "blocks.6.norm2.bias",531 "first_stage_model.decoder.up.3.block.0.time_stack.out_layers.0.weight": "blocks.6.norm2.weight",532 "first_stage_model.decoder.up.3.block.0.time_stack.out_layers.3.bias": "blocks.6.conv2.bias",533 "first_stage_model.decoder.up.3.block.0.time_stack.out_layers.3.weight": "blocks.6.conv2.weight",534 "first_stage_model.decoder.up.3.block.1.conv1.bias": "blocks.7.conv1.bias",535 "first_stage_model.decoder.up.3.block.1.conv1.weight": "blocks.7.conv1.weight",536 "first_stage_model.decoder.up.3.block.1.conv2.bias": "blocks.7.conv2.bias",537 "first_stage_model.decoder.up.3.block.1.conv2.weight": "blocks.7.conv2.weight",538 "first_stage_model.decoder.up.3.block.1.mix_factor": "blocks.8.mix_factor",539 "first_stage_model.decoder.up.3.block.1.norm1.bias": "blocks.7.norm1.bias",540 "first_stage_model.decoder.up.3.block.1.norm1.weight": "blocks.7.norm1.weight",541 "first_stage_model.decoder.up.3.block.1.norm2.bias": "blocks.7.norm2.bias",542 "first_stage_model.decoder.up.3.block.1.norm2.weight": "blocks.7.norm2.weight",543 "first_stage_model.decoder.up.3.block.1.time_stack.in_layers.0.bias": "blocks.8.norm1.bias",544 "first_stage_model.decoder.up.3.block.1.time_stack.in_layers.0.weight": "blocks.8.norm1.weight",545 "first_stage_model.decoder.up.3.block.1.time_stack.in_layers.2.bias": "blocks.8.conv1.bias",546 "first_stage_model.decoder.up.3.block.1.time_stack.in_layers.2.weight": "blocks.8.conv1.weight",547 "first_stage_model.decoder.up.3.block.1.time_stack.out_layers.0.bias": "blocks.8.norm2.bias",548 "first_stage_model.decoder.up.3.block.1.time_stack.out_layers.0.weight": "blocks.8.norm2.weight",549 "first_stage_model.decoder.up.3.block.1.time_stack.out_layers.3.bias": "blocks.8.conv2.bias",550 "first_stage_model.decoder.up.3.block.1.time_stack.out_layers.3.weight": "blocks.8.conv2.weight",551 "first_stage_model.decoder.up.3.block.2.conv1.bias": "blocks.9.conv1.bias",552 "first_stage_model.decoder.up.3.block.2.conv1.weight": "blocks.9.conv1.weight",553 "first_stage_model.decoder.up.3.block.2.conv2.bias": "blocks.9.conv2.bias",554 "first_stage_model.decoder.up.3.block.2.conv2.weight": "blocks.9.conv2.weight",555 "first_stage_model.decoder.up.3.block.2.mix_factor": "blocks.10.mix_factor",556 "first_stage_model.decoder.up.3.block.2.norm1.bias": "blocks.9.norm1.bias",557 "first_stage_model.decoder.up.3.block.2.norm1.weight": "blocks.9.norm1.weight",558 "first_stage_model.decoder.up.3.block.2.norm2.bias": "blocks.9.norm2.bias",559 "first_stage_model.decoder.up.3.block.2.norm2.weight": "blocks.9.norm2.weight",560 "first_stage_model.decoder.up.3.block.2.time_stack.in_layers.0.bias": "blocks.10.norm1.bias",561 "first_stage_model.decoder.up.3.block.2.time_stack.in_layers.0.weight": "blocks.10.norm1.weight",562 "first_stage_model.decoder.up.3.block.2.time_stack.in_layers.2.bias": "blocks.10.conv1.bias",563 "first_stage_model.decoder.up.3.block.2.time_stack.in_layers.2.weight": "blocks.10.conv1.weight",564 "first_stage_model.decoder.up.3.block.2.time_stack.out_layers.0.bias": "blocks.10.norm2.bias",565 "first_stage_model.decoder.up.3.block.2.time_stack.out_layers.0.weight": "blocks.10.norm2.weight",566 "first_stage_model.decoder.up.3.block.2.time_stack.out_layers.3.bias": "blocks.10.conv2.bias",567 "first_stage_model.decoder.up.3.block.2.time_stack.out_layers.3.weight": "blocks.10.conv2.weight",568 "first_stage_model.decoder.up.3.upsample.conv.bias": "blocks.11.conv.bias",569 "first_stage_model.decoder.up.3.upsample.conv.weight": "blocks.11.conv.weight",570 }571 state_dict_ = {}572 for name in state_dict:573 if name in rename_dict:574 param = state_dict[name]575 if "blocks.2.transformer_blocks.0" in rename_dict[name]:576 param = param.squeeze()577 state_dict_[rename_dict[name]] = param578 return state_dict_579