tsi-org/tango
0
1# coding=utf-82# Copyright 2023 The HuggingFace Inc. team.3#4# Licensed under the Apache License, Version 2.0 (the "License");5# you may not use this file except in compliance with the License.6# You may obtain a copy of the License at7#8# http://www.apache.org/licenses/LICENSE-2.09#10# Unless required by applicable law or agreed to in writing, software11# distributed under the License is distributed on an "AS IS" BASIS,12# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.13# See the License for the specific language governing permissions and14# limitations under the License.15""" Conversion script for the LDM checkpoints. """16 17import argparse18import json19 20import torch21 22from diffusers import DDPMScheduler, LDMPipeline, UNet2DModel, VQModel23 24 25def shave_segments(path, n_shave_prefix_segments=1):26 """27 Removes segments. Positive values shave the first segments, negative shave the last segments.28 """29 if n_shave_prefix_segments >= 0:30 return ".".join(path.split(".")[n_shave_prefix_segments:])31 else:32 return ".".join(path.split(".")[:n_shave_prefix_segments])33 34 35def renew_resnet_paths(old_list, n_shave_prefix_segments=0):36 """37 Updates paths inside resnets to the new naming scheme (local renaming)38 """39 mapping = []40 for old_item in old_list:41 new_item = old_item.replace("in_layers.0", "norm1")42 new_item = new_item.replace("in_layers.2", "conv1")43 44 new_item = new_item.replace("out_layers.0", "norm2")45 new_item = new_item.replace("out_layers.3", "conv2")46 47 new_item = new_item.replace("emb_layers.1", "time_emb_proj")48 new_item = new_item.replace("skip_connection", "conv_shortcut")49 50 new_item = shave_segments(new_item, n_shave_prefix_segments=n_shave_prefix_segments)51 52 mapping.append({"old": old_item, "new": new_item})53 54 return mapping55 56 57def renew_attention_paths(old_list, n_shave_prefix_segments=0):58 """59 Updates paths inside attentions to the new naming scheme (local renaming)60 """61 mapping = []62 for old_item in old_list:63 new_item = old_item64 65 new_item = new_item.replace("norm.weight", "group_norm.weight")66 new_item = new_item.replace("norm.bias", "group_norm.bias")67 68 new_item = new_item.replace("proj_out.weight", "proj_attn.weight")69 new_item = new_item.replace("proj_out.bias", "proj_attn.bias")70 71 new_item = shave_segments(new_item, n_shave_prefix_segments=n_shave_prefix_segments)72 73 mapping.append({"old": old_item, "new": new_item})74 75 return mapping76 77 78def assign_to_checkpoint(79 paths, checkpoint, old_checkpoint, attention_paths_to_split=None, additional_replacements=None, config=None80):81 """82 This does the final conversion step: take locally converted weights and apply a global renaming83 to them. It splits attention layers, and takes into account additional replacements84 that may arise.85 86 Assigns the weights to the new checkpoint.87 """88 assert isinstance(paths, list), "Paths should be a list of dicts containing 'old' and 'new' keys."89 90 # Splits the attention layers into three variables.91 if attention_paths_to_split is not None:92 for path, path_map in attention_paths_to_split.items():93 old_tensor = old_checkpoint[path]94 channels = old_tensor.shape[0] // 395 96 target_shape = (-1, channels) if len(old_tensor.shape) == 3 else (-1)97 98 num_heads = old_tensor.shape[0] // config["num_head_channels"] // 399 100 old_tensor = old_tensor.reshape((num_heads, 3 * channels // num_heads) + old_tensor.shape[1:])101 query, key, value = old_tensor.split(channels // num_heads, dim=1)102 103 checkpoint[path_map["query"]] = query.reshape(target_shape)104 checkpoint[path_map["key"]] = key.reshape(target_shape)105 checkpoint[path_map["value"]] = value.reshape(target_shape)106 107 for path in paths:108 new_path = path["new"]109 110 # These have already been assigned111 if attention_paths_to_split is not None and new_path in attention_paths_to_split:112 continue113 114 # Global renaming happens here115 new_path = new_path.replace("middle_block.0", "mid_block.resnets.0")116 new_path = new_path.replace("middle_block.1", "mid_block.attentions.0")117 new_path = new_path.replace("middle_block.2", "mid_block.resnets.1")118 119 if additional_replacements is not None:120 for replacement in additional_replacements:121 new_path = new_path.replace(replacement["old"], replacement["new"])122 123 # proj_attn.weight has to be converted from conv 1D to linear124 if "proj_attn.weight" in new_path:125 checkpoint[new_path] = old_checkpoint[path["old"]][:, :, 0]126 else:127 checkpoint[new_path] = old_checkpoint[path["old"]]128 129 130def convert_ldm_checkpoint(checkpoint, config):131 """132 Takes a state dict and a config, and returns a converted checkpoint.133 """134 new_checkpoint = {}135 136 new_checkpoint["time_embedding.linear_1.weight"] = checkpoint["time_embed.0.weight"]137 new_checkpoint["time_embedding.linear_1.bias"] = checkpoint["time_embed.0.bias"]138 new_checkpoint["time_embedding.linear_2.weight"] = checkpoint["time_embed.2.weight"]139 new_checkpoint["time_embedding.linear_2.bias"] = checkpoint["time_embed.2.bias"]140 141 new_checkpoint["conv_in.weight"] = checkpoint["input_blocks.0.0.weight"]142 new_checkpoint["conv_in.bias"] = checkpoint["input_blocks.0.0.bias"]143 144 new_checkpoint["conv_norm_out.weight"] = checkpoint["out.0.weight"]145 new_checkpoint["conv_norm_out.bias"] = checkpoint["out.0.bias"]146 new_checkpoint["conv_out.weight"] = checkpoint["out.2.weight"]147 new_checkpoint["conv_out.bias"] = checkpoint["out.2.bias"]148 149 # Retrieves the keys for the input blocks only150 num_input_blocks = len({".".join(layer.split(".")[:2]) for layer in checkpoint if "input_blocks" in layer})151 input_blocks = {152 layer_id: [key for key in checkpoint if f"input_blocks.{layer_id}" in key]153 for layer_id in range(num_input_blocks)154 }155 156 # Retrieves the keys for the middle blocks only157 num_middle_blocks = len({".".join(layer.split(".")[:2]) for layer in checkpoint if "middle_block" in layer})158 middle_blocks = {159 layer_id: [key for key in checkpoint if f"middle_block.{layer_id}" in key]160 for layer_id in range(num_middle_blocks)161 }162 163 # Retrieves the keys for the output blocks only164 num_output_blocks = len({".".join(layer.split(".")[:2]) for layer in checkpoint if "output_blocks" in layer})165 output_blocks = {166 layer_id: [key for key in checkpoint if f"output_blocks.{layer_id}" in key]167 for layer_id in range(num_output_blocks)168 }169 170 for i in range(1, num_input_blocks):171 block_id = (i - 1) // (config["num_res_blocks"] + 1)172 layer_in_block_id = (i - 1) % (config["num_res_blocks"] + 1)173 174 resnets = [key for key in input_blocks[i] if f"input_blocks.{i}.0" in key]175 attentions = [key for key in input_blocks[i] if f"input_blocks.{i}.1" in key]176 177 if f"input_blocks.{i}.0.op.weight" in checkpoint:178 new_checkpoint[f"down_blocks.{block_id}.downsamplers.0.conv.weight"] = checkpoint[179 f"input_blocks.{i}.0.op.weight"180 ]181 new_checkpoint[f"down_blocks.{block_id}.downsamplers.0.conv.bias"] = checkpoint[182 f"input_blocks.{i}.0.op.bias"183 ]184 continue185 186 paths = renew_resnet_paths(resnets)187 meta_path = {"old": f"input_blocks.{i}.0", "new": f"down_blocks.{block_id}.resnets.{layer_in_block_id}"}188 resnet_op = {"old": "resnets.2.op", "new": "downsamplers.0.op"}189 assign_to_checkpoint(190 paths, new_checkpoint, checkpoint, additional_replacements=[meta_path, resnet_op], config=config191 )192 193 if len(attentions):194 paths = renew_attention_paths(attentions)195 meta_path = {196 "old": f"input_blocks.{i}.1",197 "new": f"down_blocks.{block_id}.attentions.{layer_in_block_id}",198 }199 to_split = {200 f"input_blocks.{i}.1.qkv.bias": {201 "key": f"down_blocks.{block_id}.attentions.{layer_in_block_id}.key.bias",202 "query": f"down_blocks.{block_id}.attentions.{layer_in_block_id}.query.bias",203 "value": f"down_blocks.{block_id}.attentions.{layer_in_block_id}.value.bias",204 },205 f"input_blocks.{i}.1.qkv.weight": {206 "key": f"down_blocks.{block_id}.attentions.{layer_in_block_id}.key.weight",207 "query": f"down_blocks.{block_id}.attentions.{layer_in_block_id}.query.weight",208 "value": f"down_blocks.{block_id}.attentions.{layer_in_block_id}.value.weight",209 },210 }211 assign_to_checkpoint(212 paths,213 new_checkpoint,214 checkpoint,215 additional_replacements=[meta_path],216 attention_paths_to_split=to_split,217 config=config,218 )219 220 resnet_0 = middle_blocks[0]221 attentions = middle_blocks[1]222 resnet_1 = middle_blocks[2]223 224 resnet_0_paths = renew_resnet_paths(resnet_0)225 assign_to_checkpoint(resnet_0_paths, new_checkpoint, checkpoint, config=config)226 227 resnet_1_paths = renew_resnet_paths(resnet_1)228 assign_to_checkpoint(resnet_1_paths, new_checkpoint, checkpoint, config=config)229 230 attentions_paths = renew_attention_paths(attentions)231 to_split = {232 "middle_block.1.qkv.bias": {233 "key": "mid_block.attentions.0.key.bias",234 "query": "mid_block.attentions.0.query.bias",235 "value": "mid_block.attentions.0.value.bias",236 },237 "middle_block.1.qkv.weight": {238 "key": "mid_block.attentions.0.key.weight",239 "query": "mid_block.attentions.0.query.weight",240 "value": "mid_block.attentions.0.value.weight",241 },242 }243 assign_to_checkpoint(244 attentions_paths, new_checkpoint, checkpoint, attention_paths_to_split=to_split, config=config245 )246 247 for i in range(num_output_blocks):248 block_id = i // (config["num_res_blocks"] + 1)249 layer_in_block_id = i % (config["num_res_blocks"] + 1)250 output_block_layers = [shave_segments(name, 2) for name in output_blocks[i]]251 output_block_list = {}252 253 for layer in output_block_layers:254 layer_id, layer_name = layer.split(".")[0], shave_segments(layer, 1)255 if layer_id in output_block_list:256 output_block_list[layer_id].append(layer_name)257 else:258 output_block_list[layer_id] = [layer_name]259 260 if len(output_block_list) > 1:261 resnets = [key for key in output_blocks[i] if f"output_blocks.{i}.0" in key]262 attentions = [key for key in output_blocks[i] if f"output_blocks.{i}.1" in key]263 264 resnet_0_paths = renew_resnet_paths(resnets)265 paths = renew_resnet_paths(resnets)266 267 meta_path = {"old": f"output_blocks.{i}.0", "new": f"up_blocks.{block_id}.resnets.{layer_in_block_id}"}268 assign_to_checkpoint(paths, new_checkpoint, checkpoint, additional_replacements=[meta_path], config=config)269 270 if ["conv.weight", "conv.bias"] in output_block_list.values():271 index = list(output_block_list.values()).index(["conv.weight", "conv.bias"])272 new_checkpoint[f"up_blocks.{block_id}.upsamplers.0.conv.weight"] = checkpoint[273 f"output_blocks.{i}.{index}.conv.weight"274 ]275 new_checkpoint[f"up_blocks.{block_id}.upsamplers.0.conv.bias"] = checkpoint[276 f"output_blocks.{i}.{index}.conv.bias"277 ]278 279 # Clear attentions as they have been attributed above.280 if len(attentions) == 2:281 attentions = []282 283 if len(attentions):284 paths = renew_attention_paths(attentions)285 meta_path = {286 "old": f"output_blocks.{i}.1",287 "new": f"up_blocks.{block_id}.attentions.{layer_in_block_id}",288 }289 to_split = {290 f"output_blocks.{i}.1.qkv.bias": {291 "key": f"up_blocks.{block_id}.attentions.{layer_in_block_id}.key.bias",292 "query": f"up_blocks.{block_id}.attentions.{layer_in_block_id}.query.bias",293 "value": f"up_blocks.{block_id}.attentions.{layer_in_block_id}.value.bias",294 },295 f"output_blocks.{i}.1.qkv.weight": {296 "key": f"up_blocks.{block_id}.attentions.{layer_in_block_id}.key.weight",297 "query": f"up_blocks.{block_id}.attentions.{layer_in_block_id}.query.weight",298 "value": f"up_blocks.{block_id}.attentions.{layer_in_block_id}.value.weight",299 },300 }301 assign_to_checkpoint(302 paths,303 new_checkpoint,304 checkpoint,305 additional_replacements=[meta_path],306 attention_paths_to_split=to_split if any("qkv" in key for key in attentions) else None,307 config=config,308 )309 else:310 resnet_0_paths = renew_resnet_paths(output_block_layers, n_shave_prefix_segments=1)311 for path in resnet_0_paths:312 old_path = ".".join(["output_blocks", str(i), path["old"]])313 new_path = ".".join(["up_blocks", str(block_id), "resnets", str(layer_in_block_id), path["new"]])314 315 new_checkpoint[new_path] = checkpoint[old_path]316 317 return new_checkpoint318 319 320if __name__ == "__main__":321 parser = argparse.ArgumentParser()322 323 parser.add_argument(324 "--checkpoint_path", default=None, type=str, required=True, help="Path to the checkpoint to convert."325 )326 327 parser.add_argument(328 "--config_file",329 default=None,330 type=str,331 required=True,332 help="The config json file corresponding to the architecture.",333 )334 335 parser.add_argument("--dump_path", default=None, type=str, required=True, help="Path to the output model.")336 337 args = parser.parse_args()338 339 checkpoint = torch.load(args.checkpoint_path)340 341 with open(args.config_file) as f:342 config = json.loads(f.read())343 344 converted_checkpoint = convert_ldm_checkpoint(checkpoint, config)345 346 if "ldm" in config:347 del config["ldm"]348 349 model = UNet2DModel(**config)350 model.load_state_dict(converted_checkpoint)351 352 try:353 scheduler = DDPMScheduler.from_config("/".join(args.checkpoint_path.split("/")[:-1]))354 vqvae = VQModel.from_pretrained("/".join(args.checkpoint_path.split("/")[:-1]))355 356 pipe = LDMPipeline(unet=model, scheduler=scheduler, vae=vqvae)357 pipe.save_pretrained(args.dump_path)358 except: # noqa: E722359 model.save_pretrained(args.dump_path)360