tsi-org/tango

1import argparse2import json3 4import torch5 6from diffusers import AutoencoderKL, DDPMPipeline, DDPMScheduler, UNet2DModel, VQModel7 8 9def shave_segments(path, n_shave_prefix_segments=1):10    """11    Removes segments. Positive values shave the first segments, negative shave the last segments.12    """13    if n_shave_prefix_segments >= 0:14        return ".".join(path.split(".")[n_shave_prefix_segments:])15    else:16        return ".".join(path.split(".")[:n_shave_prefix_segments])17 18 19def renew_resnet_paths(old_list, n_shave_prefix_segments=0):20    mapping = []21    for old_item in old_list:22        new_item = old_item23        new_item = new_item.replace("block.", "resnets.")24        new_item = new_item.replace("conv_shorcut", "conv1")25        new_item = new_item.replace("in_shortcut", "conv_shortcut")26        new_item = new_item.replace("temb_proj", "time_emb_proj")27 28        new_item = shave_segments(new_item, n_shave_prefix_segments=n_shave_prefix_segments)29 30        mapping.append({"old": old_item, "new": new_item})31 32    return mapping33 34 35def renew_attention_paths(old_list, n_shave_prefix_segments=0, in_mid=False):36    mapping = []37    for old_item in old_list:38        new_item = old_item39 40        # In `model.mid`, the layer is called `attn`.41        if not in_mid:42            new_item = new_item.replace("attn", "attentions")43        new_item = new_item.replace(".k.", ".key.")44        new_item = new_item.replace(".v.", ".value.")45        new_item = new_item.replace(".q.", ".query.")46 47        new_item = new_item.replace("proj_out", "proj_attn")48        new_item = new_item.replace("norm", "group_norm")49 50        new_item = shave_segments(new_item, n_shave_prefix_segments=n_shave_prefix_segments)51        mapping.append({"old": old_item, "new": new_item})52 53    return mapping54 55 56def assign_to_checkpoint(57    paths, checkpoint, old_checkpoint, attention_paths_to_split=None, additional_replacements=None, config=None58):59    assert isinstance(paths, list), "Paths should be a list of dicts containing 'old' and 'new' keys."60 61    if attention_paths_to_split is not None:62        if config is None:63            raise ValueError("Please specify the config if setting 'attention_paths_to_split' to 'True'.")64 65        for path, path_map in attention_paths_to_split.items():66            old_tensor = old_checkpoint[path]67            channels = old_tensor.shape[0] // 368 69            target_shape = (-1, channels) if len(old_tensor.shape) == 3 else (-1)70 71            num_heads = old_tensor.shape[0] // config.get("num_head_channels", 1) // 372 73            old_tensor = old_tensor.reshape((num_heads, 3 * channels // num_heads) + old_tensor.shape[1:])74            query, key, value = old_tensor.split(channels // num_heads, dim=1)75 76            checkpoint[path_map["query"]] = query.reshape(target_shape).squeeze()77            checkpoint[path_map["key"]] = key.reshape(target_shape).squeeze()78            checkpoint[path_map["value"]] = value.reshape(target_shape).squeeze()79 80    for path in paths:81        new_path = path["new"]82 83        if attention_paths_to_split is not None and new_path in attention_paths_to_split:84            continue85 86        new_path = new_path.replace("down.", "down_blocks.")87        new_path = new_path.replace("up.", "up_blocks.")88 89        if additional_replacements is not None:90            for replacement in additional_replacements:91                new_path = new_path.replace(replacement["old"], replacement["new"])92 93        if "attentions" in new_path:94            checkpoint[new_path] = old_checkpoint[path["old"]].squeeze()95        else:96            checkpoint[new_path] = old_checkpoint[path["old"]]97 98 99def convert_ddpm_checkpoint(checkpoint, config):100    """101    Takes a state dict and a config, and returns a converted checkpoint.102    """103    new_checkpoint = {}104 105    new_checkpoint["time_embedding.linear_1.weight"] = checkpoint["temb.dense.0.weight"]106    new_checkpoint["time_embedding.linear_1.bias"] = checkpoint["temb.dense.0.bias"]107    new_checkpoint["time_embedding.linear_2.weight"] = checkpoint["temb.dense.1.weight"]108    new_checkpoint["time_embedding.linear_2.bias"] = checkpoint["temb.dense.1.bias"]109 110    new_checkpoint["conv_norm_out.weight"] = checkpoint["norm_out.weight"]111    new_checkpoint["conv_norm_out.bias"] = checkpoint["norm_out.bias"]112 113    new_checkpoint["conv_in.weight"] = checkpoint["conv_in.weight"]114    new_checkpoint["conv_in.bias"] = checkpoint["conv_in.bias"]115    new_checkpoint["conv_out.weight"] = checkpoint["conv_out.weight"]116    new_checkpoint["conv_out.bias"] = checkpoint["conv_out.bias"]117 118    num_down_blocks = len({".".join(layer.split(".")[:2]) for layer in checkpoint if "down" in layer})119    down_blocks = {120        layer_id: [key for key in checkpoint if f"down.{layer_id}" in key] for layer_id in range(num_down_blocks)121    }122 123    num_up_blocks = len({".".join(layer.split(".")[:2]) for layer in checkpoint if "up" in layer})124    up_blocks = {layer_id: [key for key in checkpoint if f"up.{layer_id}" in key] for layer_id in range(num_up_blocks)}125 126    for i in range(num_down_blocks):127        block_id = (i - 1) // (config["layers_per_block"] + 1)128 129        if any("downsample" in layer for layer in down_blocks[i]):130            new_checkpoint[f"down_blocks.{i}.downsamplers.0.conv.weight"] = checkpoint[131                f"down.{i}.downsample.op.weight"132            ]133            new_checkpoint[f"down_blocks.{i}.downsamplers.0.conv.bias"] = checkpoint[f"down.{i}.downsample.op.bias"]134        #            new_checkpoint[f'down_blocks.{i}.downsamplers.0.op.weight'] = checkpoint[f'down.{i}.downsample.conv.weight']135        #            new_checkpoint[f'down_blocks.{i}.downsamplers.0.op.bias'] = checkpoint[f'down.{i}.downsample.conv.bias']136 137        if any("block" in layer for layer in down_blocks[i]):138            num_blocks = len(139                {".".join(shave_segments(layer, 2).split(".")[:2]) for layer in down_blocks[i] if "block" in layer}140            )141            blocks = {142                layer_id: [key for key in down_blocks[i] if f"block.{layer_id}" in key]143                for layer_id in range(num_blocks)144            }145 146            if num_blocks > 0:147                for j in range(config["layers_per_block"]):148                    paths = renew_resnet_paths(blocks[j])149                    assign_to_checkpoint(paths, new_checkpoint, checkpoint)150 151        if any("attn" in layer for layer in down_blocks[i]):152            num_attn = len(153                {".".join(shave_segments(layer, 2).split(".")[:2]) for layer in down_blocks[i] if "attn" in layer}154            )155            attns = {156                layer_id: [key for key in down_blocks[i] if f"attn.{layer_id}" in key]157                for layer_id in range(num_blocks)158            }159 160            if num_attn > 0:161                for j in range(config["layers_per_block"]):162                    paths = renew_attention_paths(attns[j])163                    assign_to_checkpoint(paths, new_checkpoint, checkpoint, config=config)164 165    mid_block_1_layers = [key for key in checkpoint if "mid.block_1" in key]166    mid_block_2_layers = [key for key in checkpoint if "mid.block_2" in key]167    mid_attn_1_layers = [key for key in checkpoint if "mid.attn_1" in key]168 169    # Mid new 2170    paths = renew_resnet_paths(mid_block_1_layers)171    assign_to_checkpoint(172        paths,173        new_checkpoint,174        checkpoint,175        additional_replacements=[{"old": "mid.", "new": "mid_new_2."}, {"old": "block_1", "new": "resnets.0"}],176    )177 178    paths = renew_resnet_paths(mid_block_2_layers)179    assign_to_checkpoint(180        paths,181        new_checkpoint,182        checkpoint,183        additional_replacements=[{"old": "mid.", "new": "mid_new_2."}, {"old": "block_2", "new": "resnets.1"}],184    )185 186    paths = renew_attention_paths(mid_attn_1_layers, in_mid=True)187    assign_to_checkpoint(188        paths,189        new_checkpoint,190        checkpoint,191        additional_replacements=[{"old": "mid.", "new": "mid_new_2."}, {"old": "attn_1", "new": "attentions.0"}],192    )193 194    for i in range(num_up_blocks):195        block_id = num_up_blocks - 1 - i196 197        if any("upsample" in layer for layer in up_blocks[i]):198            new_checkpoint[f"up_blocks.{block_id}.upsamplers.0.conv.weight"] = checkpoint[199                f"up.{i}.upsample.conv.weight"200            ]201            new_checkpoint[f"up_blocks.{block_id}.upsamplers.0.conv.bias"] = checkpoint[f"up.{i}.upsample.conv.bias"]202 203        if any("block" in layer for layer in up_blocks[i]):204            num_blocks = len(205                {".".join(shave_segments(layer, 2).split(".")[:2]) for layer in up_blocks[i] if "block" in layer}206            )207            blocks = {208                layer_id: [key for key in up_blocks[i] if f"block.{layer_id}" in key] for layer_id in range(num_blocks)209            }210 211            if num_blocks > 0:212                for j in range(config["layers_per_block"] + 1):213                    replace_indices = {"old": f"up_blocks.{i}", "new": f"up_blocks.{block_id}"}214                    paths = renew_resnet_paths(blocks[j])215                    assign_to_checkpoint(paths, new_checkpoint, checkpoint, additional_replacements=[replace_indices])216 217        if any("attn" in layer for layer in up_blocks[i]):218            num_attn = len(219                {".".join(shave_segments(layer, 2).split(".")[:2]) for layer in up_blocks[i] if "attn" in layer}220            )221            attns = {222                layer_id: [key for key in up_blocks[i] if f"attn.{layer_id}" in key] for layer_id in range(num_blocks)223            }224 225            if num_attn > 0:226                for j in range(config["layers_per_block"] + 1):227                    replace_indices = {"old": f"up_blocks.{i}", "new": f"up_blocks.{block_id}"}228                    paths = renew_attention_paths(attns[j])229                    assign_to_checkpoint(paths, new_checkpoint, checkpoint, additional_replacements=[replace_indices])230 231    new_checkpoint = {k.replace("mid_new_2", "mid_block"): v for k, v in new_checkpoint.items()}232    return new_checkpoint233 234 235def convert_vq_autoenc_checkpoint(checkpoint, config):236    """237    Takes a state dict and a config, and returns a converted checkpoint.238    """239    new_checkpoint = {}240 241    new_checkpoint["encoder.conv_norm_out.weight"] = checkpoint["encoder.norm_out.weight"]242    new_checkpoint["encoder.conv_norm_out.bias"] = checkpoint["encoder.norm_out.bias"]243 244    new_checkpoint["encoder.conv_in.weight"] = checkpoint["encoder.conv_in.weight"]245    new_checkpoint["encoder.conv_in.bias"] = checkpoint["encoder.conv_in.bias"]246    new_checkpoint["encoder.conv_out.weight"] = checkpoint["encoder.conv_out.weight"]247    new_checkpoint["encoder.conv_out.bias"] = checkpoint["encoder.conv_out.bias"]248 249    new_checkpoint["decoder.conv_norm_out.weight"] = checkpoint["decoder.norm_out.weight"]250    new_checkpoint["decoder.conv_norm_out.bias"] = checkpoint["decoder.norm_out.bias"]251 252    new_checkpoint["decoder.conv_in.weight"] = checkpoint["decoder.conv_in.weight"]253    new_checkpoint["decoder.conv_in.bias"] = checkpoint["decoder.conv_in.bias"]254    new_checkpoint["decoder.conv_out.weight"] = checkpoint["decoder.conv_out.weight"]255    new_checkpoint["decoder.conv_out.bias"] = checkpoint["decoder.conv_out.bias"]256 257    num_down_blocks = len({".".join(layer.split(".")[:3]) for layer in checkpoint if "down" in layer})258    down_blocks = {259        layer_id: [key for key in checkpoint if f"down.{layer_id}" in key] for layer_id in range(num_down_blocks)260    }261 262    num_up_blocks = len({".".join(layer.split(".")[:3]) for layer in checkpoint if "up" in layer})263    up_blocks = {layer_id: [key for key in checkpoint if f"up.{layer_id}" in key] for layer_id in range(num_up_blocks)}264 265    for i in range(num_down_blocks):266        block_id = (i - 1) // (config["layers_per_block"] + 1)267 268        if any("downsample" in layer for layer in down_blocks[i]):269            new_checkpoint[f"encoder.down_blocks.{i}.downsamplers.0.conv.weight"] = checkpoint[270                f"encoder.down.{i}.downsample.conv.weight"271            ]272            new_checkpoint[f"encoder.down_blocks.{i}.downsamplers.0.conv.bias"] = checkpoint[273                f"encoder.down.{i}.downsample.conv.bias"274            ]275 276        if any("block" in layer for layer in down_blocks[i]):277            num_blocks = len(278                {".".join(shave_segments(layer, 3).split(".")[:3]) for layer in down_blocks[i] if "block" in layer}279            )280            blocks = {281                layer_id: [key for key in down_blocks[i] if f"block.{layer_id}" in key]282                for layer_id in range(num_blocks)283            }284 285            if num_blocks > 0:286                for j in range(config["layers_per_block"]):287                    paths = renew_resnet_paths(blocks[j])288                    assign_to_checkpoint(paths, new_checkpoint, checkpoint)289 290        if any("attn" in layer for layer in down_blocks[i]):291            num_attn = len(292                {".".join(shave_segments(layer, 3).split(".")[:3]) for layer in down_blocks[i] if "attn" in layer}293            )294            attns = {295                layer_id: [key for key in down_blocks[i] if f"attn.{layer_id}" in key]296                for layer_id in range(num_blocks)297            }298 299            if num_attn > 0:300                for j in range(config["layers_per_block"]):301                    paths = renew_attention_paths(attns[j])302                    assign_to_checkpoint(paths, new_checkpoint, checkpoint, config=config)303 304    mid_block_1_layers = [key for key in checkpoint if "mid.block_1" in key]305    mid_block_2_layers = [key for key in checkpoint if "mid.block_2" in key]306    mid_attn_1_layers = [key for key in checkpoint if "mid.attn_1" in key]307 308    # Mid new 2309    paths = renew_resnet_paths(mid_block_1_layers)310    assign_to_checkpoint(311        paths,312        new_checkpoint,313        checkpoint,314        additional_replacements=[{"old": "mid.", "new": "mid_new_2."}, {"old": "block_1", "new": "resnets.0"}],315    )316 317    paths = renew_resnet_paths(mid_block_2_layers)318    assign_to_checkpoint(319        paths,320        new_checkpoint,321        checkpoint,322        additional_replacements=[{"old": "mid.", "new": "mid_new_2."}, {"old": "block_2", "new": "resnets.1"}],323    )324 325    paths = renew_attention_paths(mid_attn_1_layers, in_mid=True)326    assign_to_checkpoint(327        paths,328        new_checkpoint,329        checkpoint,330        additional_replacements=[{"old": "mid.", "new": "mid_new_2."}, {"old": "attn_1", "new": "attentions.0"}],331    )332 333    for i in range(num_up_blocks):334        block_id = num_up_blocks - 1 - i335 336        if any("upsample" in layer for layer in up_blocks[i]):337            new_checkpoint[f"decoder.up_blocks.{block_id}.upsamplers.0.conv.weight"] = checkpoint[338                f"decoder.up.{i}.upsample.conv.weight"339            ]340            new_checkpoint[f"decoder.up_blocks.{block_id}.upsamplers.0.conv.bias"] = checkpoint[341                f"decoder.up.{i}.upsample.conv.bias"342            ]343 344        if any("block" in layer for layer in up_blocks[i]):345            num_blocks = len(346                {".".join(shave_segments(layer, 3).split(".")[:3]) for layer in up_blocks[i] if "block" in layer}347            )348            blocks = {349                layer_id: [key for key in up_blocks[i] if f"block.{layer_id}" in key] for layer_id in range(num_blocks)350            }351 352            if num_blocks > 0:353                for j in range(config["layers_per_block"] + 1):354                    replace_indices = {"old": f"up_blocks.{i}", "new": f"up_blocks.{block_id}"}355                    paths = renew_resnet_paths(blocks[j])356                    assign_to_checkpoint(paths, new_checkpoint, checkpoint, additional_replacements=[replace_indices])357 358        if any("attn" in layer for layer in up_blocks[i]):359            num_attn = len(360                {".".join(shave_segments(layer, 3).split(".")[:3]) for layer in up_blocks[i] if "attn" in layer}361            )362            attns = {363                layer_id: [key for key in up_blocks[i] if f"attn.{layer_id}" in key] for layer_id in range(num_blocks)364            }365 366            if num_attn > 0:367                for j in range(config["layers_per_block"] + 1):368                    replace_indices = {"old": f"up_blocks.{i}", "new": f"up_blocks.{block_id}"}369                    paths = renew_attention_paths(attns[j])370                    assign_to_checkpoint(paths, new_checkpoint, checkpoint, additional_replacements=[replace_indices])371 372    new_checkpoint = {k.replace("mid_new_2", "mid_block"): v for k, v in new_checkpoint.items()}373    new_checkpoint["quant_conv.weight"] = checkpoint["quant_conv.weight"]374    new_checkpoint["quant_conv.bias"] = checkpoint["quant_conv.bias"]375    if "quantize.embedding.weight" in checkpoint:376        new_checkpoint["quantize.embedding.weight"] = checkpoint["quantize.embedding.weight"]377    new_checkpoint["post_quant_conv.weight"] = checkpoint["post_quant_conv.weight"]378    new_checkpoint["post_quant_conv.bias"] = checkpoint["post_quant_conv.bias"]379 380    return new_checkpoint381 382 383if __name__ == "__main__":384    parser = argparse.ArgumentParser()385 386    parser.add_argument(387        "--checkpoint_path", default=None, type=str, required=True, help="Path to the checkpoint to convert."388    )389 390    parser.add_argument(391        "--config_file",392        default=None,393        type=str,394        required=True,395        help="The config json file corresponding to the architecture.",396    )397 398    parser.add_argument("--dump_path", default=None, type=str, required=True, help="Path to the output model.")399 400    args = parser.parse_args()401    checkpoint = torch.load(args.checkpoint_path)402 403    with open(args.config_file) as f:404        config = json.loads(f.read())405 406    # unet case407    key_prefix_set = {key.split(".")[0] for key in checkpoint.keys()}408    if "encoder" in key_prefix_set and "decoder" in key_prefix_set:409        converted_checkpoint = convert_vq_autoenc_checkpoint(checkpoint, config)410    else:411        converted_checkpoint = convert_ddpm_checkpoint(checkpoint, config)412 413    if "ddpm" in config:414        del config["ddpm"]415 416    if config["_class_name"] == "VQModel":417        model = VQModel(**config)418        model.load_state_dict(converted_checkpoint)419        model.save_pretrained(args.dump_path)420    elif config["_class_name"] == "AutoencoderKL":421        model = AutoencoderKL(**config)422        model.load_state_dict(converted_checkpoint)423        model.save_pretrained(args.dump_path)424    else:425        model = UNet2DModel(**config)426        model.load_state_dict(converted_checkpoint)427 428        scheduler = DDPMScheduler.from_config("/".join(args.checkpoint_path.split("/")[:-1]))429 430        pipe = DDPMPipeline(unet=model, scheduler=scheduler)431        pipe.save_pretrained(args.dump_path)432