modelscope/DiffSynth-Painter

1import torch2from .attention import Attention3 4 5class CLIPEncoderLayer(torch.nn.Module):6    def __init__(self, embed_dim, intermediate_size, num_heads=12, head_dim=64, use_quick_gelu=True):7        super().__init__()8        self.attn = Attention(q_dim=embed_dim, num_heads=num_heads, head_dim=head_dim, bias_q=True, bias_kv=True, bias_out=True)9        self.layer_norm1 = torch.nn.LayerNorm(embed_dim)10        self.layer_norm2 = torch.nn.LayerNorm(embed_dim)11        self.fc1 = torch.nn.Linear(embed_dim, intermediate_size)12        self.fc2 = torch.nn.Linear(intermediate_size, embed_dim)13 14        self.use_quick_gelu = use_quick_gelu15 16    def quickGELU(self, x):17        return x * torch.sigmoid(1.702 * x)18    19    def forward(self, hidden_states, attn_mask=None):20        residual = hidden_states21 22        hidden_states = self.layer_norm1(hidden_states)23        hidden_states = self.attn(hidden_states, attn_mask=attn_mask)24        hidden_states = residual + hidden_states25 26        residual = hidden_states27        hidden_states = self.layer_norm2(hidden_states)28        hidden_states = self.fc1(hidden_states)29        if self.use_quick_gelu:30            hidden_states = self.quickGELU(hidden_states)31        else:32            hidden_states = torch.nn.functional.gelu(hidden_states)33        hidden_states = self.fc2(hidden_states)34        hidden_states = residual + hidden_states35 36        return hidden_states37    38 39class SDTextEncoder(torch.nn.Module):40    def __init__(self, embed_dim=768, vocab_size=49408, max_position_embeddings=77, num_encoder_layers=12, encoder_intermediate_size=3072):41        super().__init__()42 43        # token_embedding44        self.token_embedding = torch.nn.Embedding(vocab_size, embed_dim)45 46        # position_embeds (This is a fixed tensor)47        self.position_embeds = torch.nn.Parameter(torch.zeros(1, max_position_embeddings, embed_dim))48 49        # encoders50        self.encoders = torch.nn.ModuleList([CLIPEncoderLayer(embed_dim, encoder_intermediate_size) for _ in range(num_encoder_layers)])51 52        # attn_mask53        self.attn_mask = self.attention_mask(max_position_embeddings)54 55        # final_layer_norm56        self.final_layer_norm = torch.nn.LayerNorm(embed_dim)57 58    def attention_mask(self, length):59        mask = torch.empty(length, length)60        mask.fill_(float("-inf"))61        mask.triu_(1)62        return mask63 64    def forward(self, input_ids, clip_skip=1):65        embeds = self.token_embedding(input_ids) + self.position_embeds66        attn_mask = self.attn_mask.to(device=embeds.device, dtype=embeds.dtype)67        for encoder_id, encoder in enumerate(self.encoders):68            embeds = encoder(embeds, attn_mask=attn_mask)69            if encoder_id + clip_skip == len(self.encoders):70                break71        embeds = self.final_layer_norm(embeds)72        return embeds73    74    @staticmethod75    def state_dict_converter():76        return SDTextEncoderStateDictConverter()77 78 79class SDTextEncoderStateDictConverter:80    def __init__(self):81        pass82 83    def from_diffusers(self, state_dict):84        rename_dict = {85            "text_model.embeddings.token_embedding.weight": "token_embedding.weight",86            "text_model.embeddings.position_embedding.weight": "position_embeds",87            "text_model.final_layer_norm.weight": "final_layer_norm.weight",88            "text_model.final_layer_norm.bias": "final_layer_norm.bias"89        }90        attn_rename_dict = {91            "self_attn.q_proj": "attn.to_q",92            "self_attn.k_proj": "attn.to_k",93            "self_attn.v_proj": "attn.to_v",94            "self_attn.out_proj": "attn.to_out",95            "layer_norm1": "layer_norm1",96            "layer_norm2": "layer_norm2",97            "mlp.fc1": "fc1",98            "mlp.fc2": "fc2",99        }100        state_dict_ = {}101        for name in state_dict:102            if name in rename_dict:103                param = state_dict[name]104                if name == "text_model.embeddings.position_embedding.weight":105                    param = param.reshape((1, param.shape[0], param.shape[1]))106                state_dict_[rename_dict[name]] = param107            elif name.startswith("text_model.encoder.layers."):108                param = state_dict[name]109                names = name.split(".")110                layer_id, layer_type, tail = names[3], ".".join(names[4:-1]), names[-1]111                name_ = ".".join(["encoders", layer_id, attn_rename_dict[layer_type], tail])112                state_dict_[name_] = param113        return state_dict_114    115    def from_civitai(self, state_dict):116        rename_dict = {117            "cond_stage_model.transformer.text_model.embeddings.token_embedding.weight": "token_embedding.weight",118            "cond_stage_model.transformer.text_model.encoder.layers.0.layer_norm1.bias": "encoders.0.layer_norm1.bias",119            "cond_stage_model.transformer.text_model.encoder.layers.0.layer_norm1.weight": "encoders.0.layer_norm1.weight",120            "cond_stage_model.transformer.text_model.encoder.layers.0.layer_norm2.bias": "encoders.0.layer_norm2.bias",121            "cond_stage_model.transformer.text_model.encoder.layers.0.layer_norm2.weight": "encoders.0.layer_norm2.weight",122            "cond_stage_model.transformer.text_model.encoder.layers.0.mlp.fc1.bias": "encoders.0.fc1.bias",123            "cond_stage_model.transformer.text_model.encoder.layers.0.mlp.fc1.weight": "encoders.0.fc1.weight",124            "cond_stage_model.transformer.text_model.encoder.layers.0.mlp.fc2.bias": "encoders.0.fc2.bias",125            "cond_stage_model.transformer.text_model.encoder.layers.0.mlp.fc2.weight": "encoders.0.fc2.weight",126            "cond_stage_model.transformer.text_model.encoder.layers.0.self_attn.k_proj.bias": "encoders.0.attn.to_k.bias",127            "cond_stage_model.transformer.text_model.encoder.layers.0.self_attn.k_proj.weight": "encoders.0.attn.to_k.weight",128            "cond_stage_model.transformer.text_model.encoder.layers.0.self_attn.out_proj.bias": "encoders.0.attn.to_out.bias",129            "cond_stage_model.transformer.text_model.encoder.layers.0.self_attn.out_proj.weight": "encoders.0.attn.to_out.weight",130            "cond_stage_model.transformer.text_model.encoder.layers.0.self_attn.q_proj.bias": "encoders.0.attn.to_q.bias",131            "cond_stage_model.transformer.text_model.encoder.layers.0.self_attn.q_proj.weight": "encoders.0.attn.to_q.weight",132            "cond_stage_model.transformer.text_model.encoder.layers.0.self_attn.v_proj.bias": "encoders.0.attn.to_v.bias",133            "cond_stage_model.transformer.text_model.encoder.layers.0.self_attn.v_proj.weight": "encoders.0.attn.to_v.weight",134            "cond_stage_model.transformer.text_model.encoder.layers.1.layer_norm1.bias": "encoders.1.layer_norm1.bias",135            "cond_stage_model.transformer.text_model.encoder.layers.1.layer_norm1.weight": "encoders.1.layer_norm1.weight",136            "cond_stage_model.transformer.text_model.encoder.layers.1.layer_norm2.bias": "encoders.1.layer_norm2.bias",137            "cond_stage_model.transformer.text_model.encoder.layers.1.layer_norm2.weight": "encoders.1.layer_norm2.weight",138            "cond_stage_model.transformer.text_model.encoder.layers.1.mlp.fc1.bias": "encoders.1.fc1.bias",139            "cond_stage_model.transformer.text_model.encoder.layers.1.mlp.fc1.weight": "encoders.1.fc1.weight",140            "cond_stage_model.transformer.text_model.encoder.layers.1.mlp.fc2.bias": "encoders.1.fc2.bias",141            "cond_stage_model.transformer.text_model.encoder.layers.1.mlp.fc2.weight": "encoders.1.fc2.weight",142            "cond_stage_model.transformer.text_model.encoder.layers.1.self_attn.k_proj.bias": "encoders.1.attn.to_k.bias",143            "cond_stage_model.transformer.text_model.encoder.layers.1.self_attn.k_proj.weight": "encoders.1.attn.to_k.weight",144            "cond_stage_model.transformer.text_model.encoder.layers.1.self_attn.out_proj.bias": "encoders.1.attn.to_out.bias",145            "cond_stage_model.transformer.text_model.encoder.layers.1.self_attn.out_proj.weight": "encoders.1.attn.to_out.weight",146            "cond_stage_model.transformer.text_model.encoder.layers.1.self_attn.q_proj.bias": "encoders.1.attn.to_q.bias",147            "cond_stage_model.transformer.text_model.encoder.layers.1.self_attn.q_proj.weight": "encoders.1.attn.to_q.weight",148            "cond_stage_model.transformer.text_model.encoder.layers.1.self_attn.v_proj.bias": "encoders.1.attn.to_v.bias",149            "cond_stage_model.transformer.text_model.encoder.layers.1.self_attn.v_proj.weight": "encoders.1.attn.to_v.weight",150            "cond_stage_model.transformer.text_model.encoder.layers.10.layer_norm1.bias": "encoders.10.layer_norm1.bias",151            "cond_stage_model.transformer.text_model.encoder.layers.10.layer_norm1.weight": "encoders.10.layer_norm1.weight",152            "cond_stage_model.transformer.text_model.encoder.layers.10.layer_norm2.bias": "encoders.10.layer_norm2.bias",153            "cond_stage_model.transformer.text_model.encoder.layers.10.layer_norm2.weight": "encoders.10.layer_norm2.weight",154            "cond_stage_model.transformer.text_model.encoder.layers.10.mlp.fc1.bias": "encoders.10.fc1.bias",155            "cond_stage_model.transformer.text_model.encoder.layers.10.mlp.fc1.weight": "encoders.10.fc1.weight",156            "cond_stage_model.transformer.text_model.encoder.layers.10.mlp.fc2.bias": "encoders.10.fc2.bias",157            "cond_stage_model.transformer.text_model.encoder.layers.10.mlp.fc2.weight": "encoders.10.fc2.weight",158            "cond_stage_model.transformer.text_model.encoder.layers.10.self_attn.k_proj.bias": "encoders.10.attn.to_k.bias",159            "cond_stage_model.transformer.text_model.encoder.layers.10.self_attn.k_proj.weight": "encoders.10.attn.to_k.weight",160            "cond_stage_model.transformer.text_model.encoder.layers.10.self_attn.out_proj.bias": "encoders.10.attn.to_out.bias",161            "cond_stage_model.transformer.text_model.encoder.layers.10.self_attn.out_proj.weight": "encoders.10.attn.to_out.weight",        162            "cond_stage_model.transformer.text_model.encoder.layers.10.self_attn.q_proj.bias": "encoders.10.attn.to_q.bias",163            "cond_stage_model.transformer.text_model.encoder.layers.10.self_attn.q_proj.weight": "encoders.10.attn.to_q.weight",164            "cond_stage_model.transformer.text_model.encoder.layers.10.self_attn.v_proj.bias": "encoders.10.attn.to_v.bias",165            "cond_stage_model.transformer.text_model.encoder.layers.10.self_attn.v_proj.weight": "encoders.10.attn.to_v.weight",166            "cond_stage_model.transformer.text_model.encoder.layers.11.layer_norm1.bias": "encoders.11.layer_norm1.bias",167            "cond_stage_model.transformer.text_model.encoder.layers.11.layer_norm1.weight": "encoders.11.layer_norm1.weight",168            "cond_stage_model.transformer.text_model.encoder.layers.11.layer_norm2.bias": "encoders.11.layer_norm2.bias",169            "cond_stage_model.transformer.text_model.encoder.layers.11.layer_norm2.weight": "encoders.11.layer_norm2.weight",170            "cond_stage_model.transformer.text_model.encoder.layers.11.mlp.fc1.bias": "encoders.11.fc1.bias",171            "cond_stage_model.transformer.text_model.encoder.layers.11.mlp.fc1.weight": "encoders.11.fc1.weight",172            "cond_stage_model.transformer.text_model.encoder.layers.11.mlp.fc2.bias": "encoders.11.fc2.bias",173            "cond_stage_model.transformer.text_model.encoder.layers.11.mlp.fc2.weight": "encoders.11.fc2.weight",174            "cond_stage_model.transformer.text_model.encoder.layers.11.self_attn.k_proj.bias": "encoders.11.attn.to_k.bias",175            "cond_stage_model.transformer.text_model.encoder.layers.11.self_attn.k_proj.weight": "encoders.11.attn.to_k.weight",176            "cond_stage_model.transformer.text_model.encoder.layers.11.self_attn.out_proj.bias": "encoders.11.attn.to_out.bias",177            "cond_stage_model.transformer.text_model.encoder.layers.11.self_attn.out_proj.weight": "encoders.11.attn.to_out.weight",178            "cond_stage_model.transformer.text_model.encoder.layers.11.self_attn.q_proj.bias": "encoders.11.attn.to_q.bias",179            "cond_stage_model.transformer.text_model.encoder.layers.11.self_attn.q_proj.weight": "encoders.11.attn.to_q.weight",180            "cond_stage_model.transformer.text_model.encoder.layers.11.self_attn.v_proj.bias": "encoders.11.attn.to_v.bias",181            "cond_stage_model.transformer.text_model.encoder.layers.11.self_attn.v_proj.weight": "encoders.11.attn.to_v.weight",182            "cond_stage_model.transformer.text_model.encoder.layers.2.layer_norm1.bias": "encoders.2.layer_norm1.bias",183            "cond_stage_model.transformer.text_model.encoder.layers.2.layer_norm1.weight": "encoders.2.layer_norm1.weight",184            "cond_stage_model.transformer.text_model.encoder.layers.2.layer_norm2.bias": "encoders.2.layer_norm2.bias",185            "cond_stage_model.transformer.text_model.encoder.layers.2.layer_norm2.weight": "encoders.2.layer_norm2.weight",186            "cond_stage_model.transformer.text_model.encoder.layers.2.mlp.fc1.bias": "encoders.2.fc1.bias",187            "cond_stage_model.transformer.text_model.encoder.layers.2.mlp.fc1.weight": "encoders.2.fc1.weight",188            "cond_stage_model.transformer.text_model.encoder.layers.2.mlp.fc2.bias": "encoders.2.fc2.bias",189            "cond_stage_model.transformer.text_model.encoder.layers.2.mlp.fc2.weight": "encoders.2.fc2.weight",190            "cond_stage_model.transformer.text_model.encoder.layers.2.self_attn.k_proj.bias": "encoders.2.attn.to_k.bias",191            "cond_stage_model.transformer.text_model.encoder.layers.2.self_attn.k_proj.weight": "encoders.2.attn.to_k.weight",192            "cond_stage_model.transformer.text_model.encoder.layers.2.self_attn.out_proj.bias": "encoders.2.attn.to_out.bias",193            "cond_stage_model.transformer.text_model.encoder.layers.2.self_attn.out_proj.weight": "encoders.2.attn.to_out.weight",194            "cond_stage_model.transformer.text_model.encoder.layers.2.self_attn.q_proj.bias": "encoders.2.attn.to_q.bias",195            "cond_stage_model.transformer.text_model.encoder.layers.2.self_attn.q_proj.weight": "encoders.2.attn.to_q.weight",196            "cond_stage_model.transformer.text_model.encoder.layers.2.self_attn.v_proj.bias": "encoders.2.attn.to_v.bias",197            "cond_stage_model.transformer.text_model.encoder.layers.2.self_attn.v_proj.weight": "encoders.2.attn.to_v.weight",198            "cond_stage_model.transformer.text_model.encoder.layers.3.layer_norm1.bias": "encoders.3.layer_norm1.bias",199            "cond_stage_model.transformer.text_model.encoder.layers.3.layer_norm1.weight": "encoders.3.layer_norm1.weight",200            "cond_stage_model.transformer.text_model.encoder.layers.3.layer_norm2.bias": "encoders.3.layer_norm2.bias",201            "cond_stage_model.transformer.text_model.encoder.layers.3.layer_norm2.weight": "encoders.3.layer_norm2.weight",202            "cond_stage_model.transformer.text_model.encoder.layers.3.mlp.fc1.bias": "encoders.3.fc1.bias",203            "cond_stage_model.transformer.text_model.encoder.layers.3.mlp.fc1.weight": "encoders.3.fc1.weight",204            "cond_stage_model.transformer.text_model.encoder.layers.3.mlp.fc2.bias": "encoders.3.fc2.bias",205            "cond_stage_model.transformer.text_model.encoder.layers.3.mlp.fc2.weight": "encoders.3.fc2.weight",206            "cond_stage_model.transformer.text_model.encoder.layers.3.self_attn.k_proj.bias": "encoders.3.attn.to_k.bias",207            "cond_stage_model.transformer.text_model.encoder.layers.3.self_attn.k_proj.weight": "encoders.3.attn.to_k.weight",208            "cond_stage_model.transformer.text_model.encoder.layers.3.self_attn.out_proj.bias": "encoders.3.attn.to_out.bias",209            "cond_stage_model.transformer.text_model.encoder.layers.3.self_attn.out_proj.weight": "encoders.3.attn.to_out.weight",210            "cond_stage_model.transformer.text_model.encoder.layers.3.self_attn.q_proj.bias": "encoders.3.attn.to_q.bias",211            "cond_stage_model.transformer.text_model.encoder.layers.3.self_attn.q_proj.weight": "encoders.3.attn.to_q.weight",212            "cond_stage_model.transformer.text_model.encoder.layers.3.self_attn.v_proj.bias": "encoders.3.attn.to_v.bias",213            "cond_stage_model.transformer.text_model.encoder.layers.3.self_attn.v_proj.weight": "encoders.3.attn.to_v.weight",214            "cond_stage_model.transformer.text_model.encoder.layers.4.layer_norm1.bias": "encoders.4.layer_norm1.bias",215            "cond_stage_model.transformer.text_model.encoder.layers.4.layer_norm1.weight": "encoders.4.layer_norm1.weight",216            "cond_stage_model.transformer.text_model.encoder.layers.4.layer_norm2.bias": "encoders.4.layer_norm2.bias",217            "cond_stage_model.transformer.text_model.encoder.layers.4.layer_norm2.weight": "encoders.4.layer_norm2.weight",218            "cond_stage_model.transformer.text_model.encoder.layers.4.mlp.fc1.bias": "encoders.4.fc1.bias",219            "cond_stage_model.transformer.text_model.encoder.layers.4.mlp.fc1.weight": "encoders.4.fc1.weight",220            "cond_stage_model.transformer.text_model.encoder.layers.4.mlp.fc2.bias": "encoders.4.fc2.bias",221            "cond_stage_model.transformer.text_model.encoder.layers.4.mlp.fc2.weight": "encoders.4.fc2.weight",222            "cond_stage_model.transformer.text_model.encoder.layers.4.self_attn.k_proj.bias": "encoders.4.attn.to_k.bias",223            "cond_stage_model.transformer.text_model.encoder.layers.4.self_attn.k_proj.weight": "encoders.4.attn.to_k.weight",224            "cond_stage_model.transformer.text_model.encoder.layers.4.self_attn.out_proj.bias": "encoders.4.attn.to_out.bias",225            "cond_stage_model.transformer.text_model.encoder.layers.4.self_attn.out_proj.weight": "encoders.4.attn.to_out.weight",226            "cond_stage_model.transformer.text_model.encoder.layers.4.self_attn.q_proj.bias": "encoders.4.attn.to_q.bias",227            "cond_stage_model.transformer.text_model.encoder.layers.4.self_attn.q_proj.weight": "encoders.4.attn.to_q.weight",228            "cond_stage_model.transformer.text_model.encoder.layers.4.self_attn.v_proj.bias": "encoders.4.attn.to_v.bias",229            "cond_stage_model.transformer.text_model.encoder.layers.4.self_attn.v_proj.weight": "encoders.4.attn.to_v.weight",230            "cond_stage_model.transformer.text_model.encoder.layers.5.layer_norm1.bias": "encoders.5.layer_norm1.bias",231            "cond_stage_model.transformer.text_model.encoder.layers.5.layer_norm1.weight": "encoders.5.layer_norm1.weight",232            "cond_stage_model.transformer.text_model.encoder.layers.5.layer_norm2.bias": "encoders.5.layer_norm2.bias",233            "cond_stage_model.transformer.text_model.encoder.layers.5.layer_norm2.weight": "encoders.5.layer_norm2.weight",234            "cond_stage_model.transformer.text_model.encoder.layers.5.mlp.fc1.bias": "encoders.5.fc1.bias",235            "cond_stage_model.transformer.text_model.encoder.layers.5.mlp.fc1.weight": "encoders.5.fc1.weight",236            "cond_stage_model.transformer.text_model.encoder.layers.5.mlp.fc2.bias": "encoders.5.fc2.bias",237            "cond_stage_model.transformer.text_model.encoder.layers.5.mlp.fc2.weight": "encoders.5.fc2.weight",238            "cond_stage_model.transformer.text_model.encoder.layers.5.self_attn.k_proj.bias": "encoders.5.attn.to_k.bias",239            "cond_stage_model.transformer.text_model.encoder.layers.5.self_attn.k_proj.weight": "encoders.5.attn.to_k.weight",240            "cond_stage_model.transformer.text_model.encoder.layers.5.self_attn.out_proj.bias": "encoders.5.attn.to_out.bias",241            "cond_stage_model.transformer.text_model.encoder.layers.5.self_attn.out_proj.weight": "encoders.5.attn.to_out.weight",242            "cond_stage_model.transformer.text_model.encoder.layers.5.self_attn.q_proj.bias": "encoders.5.attn.to_q.bias",243            "cond_stage_model.transformer.text_model.encoder.layers.5.self_attn.q_proj.weight": "encoders.5.attn.to_q.weight",244            "cond_stage_model.transformer.text_model.encoder.layers.5.self_attn.v_proj.bias": "encoders.5.attn.to_v.bias",245            "cond_stage_model.transformer.text_model.encoder.layers.5.self_attn.v_proj.weight": "encoders.5.attn.to_v.weight",246            "cond_stage_model.transformer.text_model.encoder.layers.6.layer_norm1.bias": "encoders.6.layer_norm1.bias",247            "cond_stage_model.transformer.text_model.encoder.layers.6.layer_norm1.weight": "encoders.6.layer_norm1.weight",248            "cond_stage_model.transformer.text_model.encoder.layers.6.layer_norm2.bias": "encoders.6.layer_norm2.bias",249            "cond_stage_model.transformer.text_model.encoder.layers.6.layer_norm2.weight": "encoders.6.layer_norm2.weight",250            "cond_stage_model.transformer.text_model.encoder.layers.6.mlp.fc1.bias": "encoders.6.fc1.bias",251            "cond_stage_model.transformer.text_model.encoder.layers.6.mlp.fc1.weight": "encoders.6.fc1.weight",252            "cond_stage_model.transformer.text_model.encoder.layers.6.mlp.fc2.bias": "encoders.6.fc2.bias",253            "cond_stage_model.transformer.text_model.encoder.layers.6.mlp.fc2.weight": "encoders.6.fc2.weight",254            "cond_stage_model.transformer.text_model.encoder.layers.6.self_attn.k_proj.bias": "encoders.6.attn.to_k.bias",255            "cond_stage_model.transformer.text_model.encoder.layers.6.self_attn.k_proj.weight": "encoders.6.attn.to_k.weight",256            "cond_stage_model.transformer.text_model.encoder.layers.6.self_attn.out_proj.bias": "encoders.6.attn.to_out.bias",257            "cond_stage_model.transformer.text_model.encoder.layers.6.self_attn.out_proj.weight": "encoders.6.attn.to_out.weight",258            "cond_stage_model.transformer.text_model.encoder.layers.6.self_attn.q_proj.bias": "encoders.6.attn.to_q.bias",259            "cond_stage_model.transformer.text_model.encoder.layers.6.self_attn.q_proj.weight": "encoders.6.attn.to_q.weight",260            "cond_stage_model.transformer.text_model.encoder.layers.6.self_attn.v_proj.bias": "encoders.6.attn.to_v.bias",261            "cond_stage_model.transformer.text_model.encoder.layers.6.self_attn.v_proj.weight": "encoders.6.attn.to_v.weight",262            "cond_stage_model.transformer.text_model.encoder.layers.7.layer_norm1.bias": "encoders.7.layer_norm1.bias",263            "cond_stage_model.transformer.text_model.encoder.layers.7.layer_norm1.weight": "encoders.7.layer_norm1.weight",264            "cond_stage_model.transformer.text_model.encoder.layers.7.layer_norm2.bias": "encoders.7.layer_norm2.bias",265            "cond_stage_model.transformer.text_model.encoder.layers.7.layer_norm2.weight": "encoders.7.layer_norm2.weight",266            "cond_stage_model.transformer.text_model.encoder.layers.7.mlp.fc1.bias": "encoders.7.fc1.bias",267            "cond_stage_model.transformer.text_model.encoder.layers.7.mlp.fc1.weight": "encoders.7.fc1.weight",268            "cond_stage_model.transformer.text_model.encoder.layers.7.mlp.fc2.bias": "encoders.7.fc2.bias",269            "cond_stage_model.transformer.text_model.encoder.layers.7.mlp.fc2.weight": "encoders.7.fc2.weight",270            "cond_stage_model.transformer.text_model.encoder.layers.7.self_attn.k_proj.bias": "encoders.7.attn.to_k.bias",271            "cond_stage_model.transformer.text_model.encoder.layers.7.self_attn.k_proj.weight": "encoders.7.attn.to_k.weight",272            "cond_stage_model.transformer.text_model.encoder.layers.7.self_attn.out_proj.bias": "encoders.7.attn.to_out.bias",273            "cond_stage_model.transformer.text_model.encoder.layers.7.self_attn.out_proj.weight": "encoders.7.attn.to_out.weight",274            "cond_stage_model.transformer.text_model.encoder.layers.7.self_attn.q_proj.bias": "encoders.7.attn.to_q.bias",275            "cond_stage_model.transformer.text_model.encoder.layers.7.self_attn.q_proj.weight": "encoders.7.attn.to_q.weight",276            "cond_stage_model.transformer.text_model.encoder.layers.7.self_attn.v_proj.bias": "encoders.7.attn.to_v.bias",277            "cond_stage_model.transformer.text_model.encoder.layers.7.self_attn.v_proj.weight": "encoders.7.attn.to_v.weight",278            "cond_stage_model.transformer.text_model.encoder.layers.8.layer_norm1.bias": "encoders.8.layer_norm1.bias",279            "cond_stage_model.transformer.text_model.encoder.layers.8.layer_norm1.weight": "encoders.8.layer_norm1.weight",280            "cond_stage_model.transformer.text_model.encoder.layers.8.layer_norm2.bias": "encoders.8.layer_norm2.bias",281            "cond_stage_model.transformer.text_model.encoder.layers.8.layer_norm2.weight": "encoders.8.layer_norm2.weight",282            "cond_stage_model.transformer.text_model.encoder.layers.8.mlp.fc1.bias": "encoders.8.fc1.bias",283            "cond_stage_model.transformer.text_model.encoder.layers.8.mlp.fc1.weight": "encoders.8.fc1.weight",284            "cond_stage_model.transformer.text_model.encoder.layers.8.mlp.fc2.bias": "encoders.8.fc2.bias",285            "cond_stage_model.transformer.text_model.encoder.layers.8.mlp.fc2.weight": "encoders.8.fc2.weight",286            "cond_stage_model.transformer.text_model.encoder.layers.8.self_attn.k_proj.bias": "encoders.8.attn.to_k.bias",287            "cond_stage_model.transformer.text_model.encoder.layers.8.self_attn.k_proj.weight": "encoders.8.attn.to_k.weight",288            "cond_stage_model.transformer.text_model.encoder.layers.8.self_attn.out_proj.bias": "encoders.8.attn.to_out.bias",289            "cond_stage_model.transformer.text_model.encoder.layers.8.self_attn.out_proj.weight": "encoders.8.attn.to_out.weight",290            "cond_stage_model.transformer.text_model.encoder.layers.8.self_attn.q_proj.bias": "encoders.8.attn.to_q.bias",291            "cond_stage_model.transformer.text_model.encoder.layers.8.self_attn.q_proj.weight": "encoders.8.attn.to_q.weight",292            "cond_stage_model.transformer.text_model.encoder.layers.8.self_attn.v_proj.bias": "encoders.8.attn.to_v.bias",293            "cond_stage_model.transformer.text_model.encoder.layers.8.self_attn.v_proj.weight": "encoders.8.attn.to_v.weight",294            "cond_stage_model.transformer.text_model.encoder.layers.9.layer_norm1.bias": "encoders.9.layer_norm1.bias",295            "cond_stage_model.transformer.text_model.encoder.layers.9.layer_norm1.weight": "encoders.9.layer_norm1.weight",296            "cond_stage_model.transformer.text_model.encoder.layers.9.layer_norm2.bias": "encoders.9.layer_norm2.bias",297            "cond_stage_model.transformer.text_model.encoder.layers.9.layer_norm2.weight": "encoders.9.layer_norm2.weight",298            "cond_stage_model.transformer.text_model.encoder.layers.9.mlp.fc1.bias": "encoders.9.fc1.bias",299            "cond_stage_model.transformer.text_model.encoder.layers.9.mlp.fc1.weight": "encoders.9.fc1.weight",300            "cond_stage_model.transformer.text_model.encoder.layers.9.mlp.fc2.bias": "encoders.9.fc2.bias",301            "cond_stage_model.transformer.text_model.encoder.layers.9.mlp.fc2.weight": "encoders.9.fc2.weight",302            "cond_stage_model.transformer.text_model.encoder.layers.9.self_attn.k_proj.bias": "encoders.9.attn.to_k.bias",303            "cond_stage_model.transformer.text_model.encoder.layers.9.self_attn.k_proj.weight": "encoders.9.attn.to_k.weight",304            "cond_stage_model.transformer.text_model.encoder.layers.9.self_attn.out_proj.bias": "encoders.9.attn.to_out.bias",305            "cond_stage_model.transformer.text_model.encoder.layers.9.self_attn.out_proj.weight": "encoders.9.attn.to_out.weight",306            "cond_stage_model.transformer.text_model.encoder.layers.9.self_attn.q_proj.bias": "encoders.9.attn.to_q.bias",307            "cond_stage_model.transformer.text_model.encoder.layers.9.self_attn.q_proj.weight": "encoders.9.attn.to_q.weight",308            "cond_stage_model.transformer.text_model.encoder.layers.9.self_attn.v_proj.bias": "encoders.9.attn.to_v.bias",309            "cond_stage_model.transformer.text_model.encoder.layers.9.self_attn.v_proj.weight": "encoders.9.attn.to_v.weight",310            "cond_stage_model.transformer.text_model.final_layer_norm.bias": "final_layer_norm.bias",311            "cond_stage_model.transformer.text_model.final_layer_norm.weight": "final_layer_norm.weight",312            "cond_stage_model.transformer.text_model.embeddings.position_embedding.weight": "position_embeds"313        }314        state_dict_ = {}315        for name in state_dict:316            if name in rename_dict:317                param = state_dict[name]318                if name == "cond_stage_model.transformer.text_model.embeddings.position_embedding.weight":319                    param = param.reshape((1, param.shape[0], param.shape[1]))320                state_dict_[rename_dict[name]] = param321        return state_dict_322