hugging-apps/echo-memory
0
1from .sd_unet import SDUNet, Attention, GEGLU2import torch3from einops import rearrange, repeat4 5 6class TemporalTransformerBlock(torch.nn.Module):7 8 def __init__(self, dim, num_attention_heads, attention_head_dim, max_position_embeddings=32):9 super().__init__()10 11 # 1. Self-Attn12 self.pe1 = torch.nn.Parameter(torch.zeros(1, max_position_embeddings, dim))13 self.norm1 = torch.nn.LayerNorm(dim, elementwise_affine=True)14 self.attn1 = Attention(q_dim=dim, num_heads=num_attention_heads, head_dim=attention_head_dim, bias_out=True)15 16 # 2. Cross-Attn17 self.pe2 = torch.nn.Parameter(torch.zeros(1, max_position_embeddings, dim))18 self.norm2 = torch.nn.LayerNorm(dim, elementwise_affine=True)19 self.attn2 = Attention(q_dim=dim, num_heads=num_attention_heads, head_dim=attention_head_dim, bias_out=True)20 21 # 3. Feed-forward22 self.norm3 = torch.nn.LayerNorm(dim, elementwise_affine=True)23 self.act_fn = GEGLU(dim, dim * 4)24 self.ff = torch.nn.Linear(dim * 4, dim)25 26 27 def forward(self, hidden_states, batch_size=1):28 29 # 1. Self-Attention30 norm_hidden_states = self.norm1(hidden_states)31 norm_hidden_states = rearrange(norm_hidden_states, "(b f) h c -> (b h) f c", b=batch_size)32 attn_output = self.attn1(norm_hidden_states + self.pe1[:, :norm_hidden_states.shape[1]])33 attn_output = rearrange(attn_output, "(b h) f c -> (b f) h c", b=batch_size)34 hidden_states = attn_output + hidden_states35 36 # 2. Cross-Attention37 norm_hidden_states = self.norm2(hidden_states)38 norm_hidden_states = rearrange(norm_hidden_states, "(b f) h c -> (b h) f c", b=batch_size)39 attn_output = self.attn2(norm_hidden_states + self.pe2[:, :norm_hidden_states.shape[1]])40 attn_output = rearrange(attn_output, "(b h) f c -> (b f) h c", b=batch_size)41 hidden_states = attn_output + hidden_states42 43 # 3. Feed-forward44 norm_hidden_states = self.norm3(hidden_states)45 ff_output = self.act_fn(norm_hidden_states)46 ff_output = self.ff(ff_output)47 hidden_states = ff_output + hidden_states48 49 return hidden_states50 51 52class TemporalBlock(torch.nn.Module):53 54 def __init__(self, num_attention_heads, attention_head_dim, in_channels, num_layers=1, norm_num_groups=32, eps=1e-5):55 super().__init__()56 inner_dim = num_attention_heads * attention_head_dim57 58 self.norm = torch.nn.GroupNorm(num_groups=norm_num_groups, num_channels=in_channels, eps=eps, affine=True)59 self.proj_in = torch.nn.Linear(in_channels, inner_dim)60 61 self.transformer_blocks = torch.nn.ModuleList([62 TemporalTransformerBlock(63 inner_dim,64 num_attention_heads,65 attention_head_dim66 )67 for d in range(num_layers)68 ])69 70 self.proj_out = torch.nn.Linear(inner_dim, in_channels)71 72 def forward(self, hidden_states, time_emb, text_emb, res_stack, batch_size=1):73 batch, _, height, width = hidden_states.shape74 residual = hidden_states75 76 hidden_states = self.norm(hidden_states)77 inner_dim = hidden_states.shape[1]78 hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * width, inner_dim)79 hidden_states = self.proj_in(hidden_states)80 81 for block in self.transformer_blocks:82 hidden_states = block(83 hidden_states,84 batch_size=batch_size85 )86 87 hidden_states = self.proj_out(hidden_states)88 hidden_states = hidden_states.reshape(batch, height, width, inner_dim).permute(0, 3, 1, 2).contiguous()89 hidden_states = hidden_states + residual90 91 return hidden_states, time_emb, text_emb, res_stack92 93 94class SDMotionModel(torch.nn.Module):95 def __init__(self):96 super().__init__()97 self.motion_modules = torch.nn.ModuleList([98 TemporalBlock(8, 40, 320, eps=1e-6),99 TemporalBlock(8, 40, 320, eps=1e-6),100 TemporalBlock(8, 80, 640, eps=1e-6),101 TemporalBlock(8, 80, 640, eps=1e-6),102 TemporalBlock(8, 160, 1280, eps=1e-6),103 TemporalBlock(8, 160, 1280, eps=1e-6),104 TemporalBlock(8, 160, 1280, eps=1e-6),105 TemporalBlock(8, 160, 1280, eps=1e-6),106 TemporalBlock(8, 160, 1280, eps=1e-6),107 TemporalBlock(8, 160, 1280, eps=1e-6),108 TemporalBlock(8, 160, 1280, eps=1e-6),109 TemporalBlock(8, 160, 1280, eps=1e-6),110 TemporalBlock(8, 160, 1280, eps=1e-6),111 TemporalBlock(8, 160, 1280, eps=1e-6),112 TemporalBlock(8, 160, 1280, eps=1e-6),113 TemporalBlock(8, 80, 640, eps=1e-6),114 TemporalBlock(8, 80, 640, eps=1e-6),115 TemporalBlock(8, 80, 640, eps=1e-6),116 TemporalBlock(8, 40, 320, eps=1e-6),117 TemporalBlock(8, 40, 320, eps=1e-6),118 TemporalBlock(8, 40, 320, eps=1e-6),119 ])120 self.call_block_id = {121 1: 0,122 4: 1,123 9: 2,124 12: 3,125 17: 4,126 20: 5,127 24: 6,128 26: 7,129 29: 8,130 32: 9,131 34: 10,132 36: 11,133 40: 12,134 43: 13,135 46: 14,136 50: 15,137 53: 16,138 56: 17,139 60: 18,140 63: 19,141 66: 20142 }143 144 def forward(self):145 pass146 147 @staticmethod148 def state_dict_converter():149 return SDMotionModelStateDictConverter()150 151 152class SDMotionModelStateDictConverter:153 def __init__(self):154 pass155 156 def from_diffusers(self, state_dict):157 rename_dict = {158 "norm": "norm",159 "proj_in": "proj_in",160 "transformer_blocks.0.attention_blocks.0.to_q": "transformer_blocks.0.attn1.to_q",161 "transformer_blocks.0.attention_blocks.0.to_k": "transformer_blocks.0.attn1.to_k",162 "transformer_blocks.0.attention_blocks.0.to_v": "transformer_blocks.0.attn1.to_v",163 "transformer_blocks.0.attention_blocks.0.to_out.0": "transformer_blocks.0.attn1.to_out",164 "transformer_blocks.0.attention_blocks.0.pos_encoder": "transformer_blocks.0.pe1",165 "transformer_blocks.0.attention_blocks.1.to_q": "transformer_blocks.0.attn2.to_q",166 "transformer_blocks.0.attention_blocks.1.to_k": "transformer_blocks.0.attn2.to_k",167 "transformer_blocks.0.attention_blocks.1.to_v": "transformer_blocks.0.attn2.to_v",168 "transformer_blocks.0.attention_blocks.1.to_out.0": "transformer_blocks.0.attn2.to_out",169 "transformer_blocks.0.attention_blocks.1.pos_encoder": "transformer_blocks.0.pe2",170 "transformer_blocks.0.norms.0": "transformer_blocks.0.norm1",171 "transformer_blocks.0.norms.1": "transformer_blocks.0.norm2",172 "transformer_blocks.0.ff.net.0.proj": "transformer_blocks.0.act_fn.proj",173 "transformer_blocks.0.ff.net.2": "transformer_blocks.0.ff",174 "transformer_blocks.0.ff_norm": "transformer_blocks.0.norm3",175 "proj_out": "proj_out",176 }177 name_list = sorted([i for i in state_dict if i.startswith("down_blocks.")])178 name_list += sorted([i for i in state_dict if i.startswith("mid_block.")])179 name_list += sorted([i for i in state_dict if i.startswith("up_blocks.")])180 state_dict_ = {}181 last_prefix, module_id = "", -1182 for name in name_list:183 names = name.split(".")184 prefix_index = names.index("temporal_transformer") + 1185 prefix = ".".join(names[:prefix_index])186 if prefix != last_prefix:187 last_prefix = prefix188 module_id += 1189 middle_name = ".".join(names[prefix_index:-1])190 suffix = names[-1]191 if "pos_encoder" in names:192 rename = ".".join(["motion_modules", str(module_id), rename_dict[middle_name]])193 else:194 rename = ".".join(["motion_modules", str(module_id), rename_dict[middle_name], suffix])195 state_dict_[rename] = state_dict[name]196 return state_dict_197 198 def from_civitai(self, state_dict):199 return self.from_diffusers(state_dict)200