scripts/convert_i2vgen_to_diffusers.py

# coding=utf-8
# Copyright 2024 The HuggingFace Inc. team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Conversion script for the LDM checkpoints."""

import argparse

import torch
from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection

from diffusers import DDIMScheduler, I2VGenXLPipeline, I2VGenXLUNet, StableDiffusionPipeline


CLIP_ID = "laion/CLIP-ViT-H-14-laion2B-s32B-b79K"


def assign_to_checkpoint(
    paths, checkpoint, old_checkpoint, attention_paths_to_split=None, additional_replacements=None, config=None
):
    """
    This does the final conversion step: take locally converted weights and apply a global renaming to them. It splits
    attention layers, and takes into account additional replacements that may arise.

    Assigns the weights to the new checkpoint.
    """
    assert isinstance(paths, list), "Paths should be a list of dicts containing 'old' and 'new' keys."

    # Splits the attention layers into three variables.
    if attention_paths_to_split is not None:
        for path, path_map in attention_paths_to_split.items():
            old_tensor = old_checkpoint[path]
            channels = old_tensor.shape[0] // 3

            target_shape = (-1, channels) if len(old_tensor.shape) == 3 else (-1)

            num_heads = old_tensor.shape[0] // config["num_head_channels"] // 3

            old_tensor = old_tensor.reshape((num_heads, 3 * channels // num_heads) + old_tensor.shape[1:])
            query, key, value = old_tensor.split(channels // num_heads, dim=1)

            checkpoint[path_map["query"]] = query.reshape(target_shape)
            checkpoint[path_map["key"]] = key.reshape(target_shape)
            checkpoint[path_map["value"]] = value.reshape(target_shape)

    for path in paths:
        new_path = path["new"]

        # These have already been assigned
        if attention_paths_to_split is not None and new_path in attention_paths_to_split:
            continue

        if additional_replacements is not None:
            for replacement in additional_replacements:
                new_path = new_path.replace(replacement["old"], replacement["new"])

        # proj_attn.weight has to be converted from conv 1D to linear
        weight = old_checkpoint[path["old"]]
        names = ["proj_attn.weight"]
        names_2 = ["proj_out.weight", "proj_in.weight"]
        if any(k in new_path for k in names):
            checkpoint[new_path] = weight[:, :, 0]
        elif any(k in new_path for k in names_2) and len(weight.shape) > 2 and ".attentions." not in new_path:
            checkpoint[new_path] = weight[:, :, 0]
        else:
            checkpoint[new_path] = weight


def renew_attention_paths(old_list, n_shave_prefix_segments=0):
    """
    Updates paths inside attentions to the new naming scheme (local renaming)
    """
    mapping = []
    for old_item in old_list:
        new_item = old_item
        mapping.append({"old": old_item, "new": new_item})

    return mapping


def shave_segments(path, n_shave_prefix_segments=1):
    """
    Removes segments. Positive values shave the first segments, negative shave the last segments.
    """
    if n_shave_prefix_segments >= 0:
        return ".".join(path.split(".")[n_shave_prefix_segments:])
    else:
        return ".".join(path.split(".")[:n_shave_prefix_segments])


def renew_temp_conv_paths(old_list, n_shave_prefix_segments=0):
    """
    Updates paths inside resnets to the new naming scheme (local renaming)
    """
    mapping = []
    for old_item in old_list:
        mapping.append({"old": old_item, "new": old_item})

    return mapping


def renew_resnet_paths(old_list, n_shave_prefix_segments=0):
    """
    Updates paths inside resnets to the new naming scheme (local renaming)
    """
    mapping = []
    for old_item in old_list:
        new_item = old_item.replace("in_layers.0", "norm1")
        new_item = new_item.replace("in_layers.2", "conv1")

        new_item = new_item.replace("out_layers.0", "norm2")
        new_item = new_item.replace("out_layers.3", "conv2")

        new_item = new_item.replace("emb_layers.1", "time_emb_proj")
        new_item = new_item.replace("skip_connection", "conv_shortcut")

        new_item = shave_segments(new_item, n_shave_prefix_segments=n_shave_prefix_segments)

        if "temopral_conv" not in old_item:
            mapping.append({"old": old_item, "new": new_item})

    return mapping


def convert_ldm_unet_checkpoint(checkpoint, config, path=None, extract_ema=False):
    """
    Takes a state dict and a config, and returns a converted checkpoint.
    """

    # extract state_dict for UNet
    unet_state_dict = {}
    keys = list(checkpoint.keys())

    unet_key = "model.diffusion_model."

    # at least a 100 parameters have to start with `model_ema` in order for the checkpoint to be EMA
    if sum(k.startswith("model_ema") for k in keys) > 100 and extract_ema:
        print(f"Checkpoint {path} has both EMA and non-EMA weights.")
        print(
            "In this conversion only the EMA weights are extracted. If you want to instead extract the non-EMA"
            " weights (useful to continue fine-tuning), please make sure to remove the `--extract_ema` flag."
        )
        for key in keys:
            if key.startswith("model.diffusion_model"):
                flat_ema_key = "model_ema." + "".join(key.split(".")[1:])
                unet_state_dict[key.replace(unet_key, "")] = checkpoint.pop(flat_ema_key)
    else:
        if sum(k.startswith("model_ema") for k in keys) > 100:
            print(
                "In this conversion only the non-EMA weights are extracted. If you want to instead extract the EMA"
                " weights (usually better for inference), please make sure to add the `--extract_ema` flag."
            )

        for key in keys:
            unet_state_dict[key.replace(unet_key, "")] = checkpoint.pop(key)

    new_checkpoint = {}

    new_checkpoint["time_embedding.linear_1.weight"] = unet_state_dict["time_embed.0.weight"]
    new_checkpoint["time_embedding.linear_1.bias"] = unet_state_dict["time_embed.0.bias"]
    new_checkpoint["time_embedding.linear_2.weight"] = unet_state_dict["time_embed.2.weight"]
    new_checkpoint["time_embedding.linear_2.bias"] = unet_state_dict["time_embed.2.bias"]

    additional_embedding_substrings = [
        "local_image_concat",
        "context_embedding",
        "local_image_embedding",
        "fps_embedding",
    ]
    for k in unet_state_dict:
        if any(substring in k for substring in additional_embedding_substrings):
            diffusers_key = k.replace("local_image_concat", "image_latents_proj_in").replace(
                "local_image_embedding", "image_latents_context_embedding"
            )
            new_checkpoint[diffusers_key] = unet_state_dict[k]

    # temporal encoder.
    new_checkpoint["image_latents_temporal_encoder.norm1.weight"] = unet_state_dict[
        "local_temporal_encoder.layers.0.0.norm.weight"
    ]
    new_checkpoint["image_latents_temporal_encoder.norm1.bias"] = unet_state_dict[
        "local_temporal_encoder.layers.0.0.norm.bias"
    ]

    # attention
    qkv = unet_state_dict["local_temporal_encoder.layers.0.0.fn.to_qkv.weight"]
    q, k, v = torch.chunk(qkv, 3, dim=0)
    new_checkpoint["image_latents_temporal_encoder.attn1.to_q.weight"] = q
    new_checkpoint["image_latents_temporal_encoder.attn1.to_k.weight"] = k
    new_checkpoint["image_latents_temporal_encoder.attn1.to_v.weight"] = v
    new_checkpoint["image_latents_temporal_encoder.attn1.to_out.0.weight"] = unet_state_dict[
        "local_temporal_encoder.layers.0.0.fn.to_out.0.weight"
    ]
    new_checkpoint["image_latents_temporal_encoder.attn1.to_out.0.bias"] = unet_state_dict[
        "local_temporal_encoder.layers.0.0.fn.to_out.0.bias"
    ]

    # feedforward
    new_checkpoint["image_latents_temporal_encoder.ff.net.0.proj.weight"] = unet_state_dict[
        "local_temporal_encoder.layers.0.1.net.0.0.weight"
    ]
    new_checkpoint["image_latents_temporal_encoder.ff.net.0.proj.bias"] = unet_state_dict[
        "local_temporal_encoder.layers.0.1.net.0.0.bias"
    ]
    new_checkpoint["image_latents_temporal_encoder.ff.net.2.weight"] = unet_state_dict[
        "local_temporal_encoder.layers.0.1.net.2.weight"
    ]
    new_checkpoint["image_latents_temporal_encoder.ff.net.2.bias"] = unet_state_dict[
        "local_temporal_encoder.layers.0.1.net.2.bias"
    ]

    if "class_embed_type" in config:
        if config["class_embed_type"] is None:
            # No parameters to port
            ...
        elif config["class_embed_type"] == "timestep" or config["class_embed_type"] == "projection":
            new_checkpoint["class_embedding.linear_1.weight"] = unet_state_dict["label_emb.0.0.weight"]
            new_checkpoint["class_embedding.linear_1.bias"] = unet_state_dict["label_emb.0.0.bias"]
            new_checkpoint["class_embedding.linear_2.weight"] = unet_state_dict["label_emb.0.2.weight"]
            new_checkpoint["class_embedding.linear_2.bias"] = unet_state_dict["label_emb.0.2.bias"]
        else:
            raise NotImplementedError(f"Not implemented `class_embed_type`: {config['class_embed_type']}")

    new_checkpoint["conv_in.weight"] = unet_state_dict["input_blocks.0.0.weight"]
    new_checkpoint["conv_in.bias"] = unet_state_dict["input_blocks.0.0.bias"]

    first_temp_attention = [v for v in unet_state_dict if v.startswith("input_blocks.0.1")]
    paths = renew_attention_paths(first_temp_attention)
    meta_path = {"old": "input_blocks.0.1", "new": "transformer_in"}
    assign_to_checkpoint(paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config)

    new_checkpoint["conv_norm_out.weight"] = unet_state_dict["out.0.weight"]
    new_checkpoint["conv_norm_out.bias"] = unet_state_dict["out.0.bias"]
    new_checkpoint["conv_out.weight"] = unet_state_dict["out.2.weight"]
    new_checkpoint["conv_out.bias"] = unet_state_dict["out.2.bias"]

    # Retrieves the keys for the input blocks only
    num_input_blocks = len({".".join(layer.split(".")[:2]) for layer in unet_state_dict if "input_blocks" in layer})
    input_blocks = {
        layer_id: [key for key in unet_state_dict if f"input_blocks.{layer_id}" in key]
        for layer_id in range(num_input_blocks)
    }

    # Retrieves the keys for the middle blocks only
    num_middle_blocks = len({".".join(layer.split(".")[:2]) for layer in unet_state_dict if "middle_block" in layer})
    middle_blocks = {
        layer_id: [key for key in unet_state_dict if f"middle_block.{layer_id}" in key]
        for layer_id in range(num_middle_blocks)
    }

    # Retrieves the keys for the output blocks only
    num_output_blocks = len({".".join(layer.split(".")[:2]) for layer in unet_state_dict if "output_blocks" in layer})
    output_blocks = {
        layer_id: [key for key in unet_state_dict if f"output_blocks.{layer_id}" in key]
        for layer_id in range(num_output_blocks)
    }

    for i in range(1, num_input_blocks):
        block_id = (i - 1) // (config["layers_per_block"] + 1)
        layer_in_block_id = (i - 1) % (config["layers_per_block"] + 1)

        resnets = [
            key for key in input_blocks[i] if f"input_blocks.{i}.0" in key and f"input_blocks.{i}.0.op" not in key
        ]
        attentions = [key for key in input_blocks[i] if f"input_blocks.{i}.1" in key]
        temp_attentions = [key for key in input_blocks[i] if f"input_blocks.{i}.2" in key]

        if f"input_blocks.{i}.op.weight" in unet_state_dict:
            new_checkpoint[f"down_blocks.{block_id}.downsamplers.0.conv.weight"] = unet_state_dict.pop(
                f"input_blocks.{i}.op.weight"
            )
            new_checkpoint[f"down_blocks.{block_id}.downsamplers.0.conv.bias"] = unet_state_dict.pop(
                f"input_blocks.{i}.op.bias"
            )

        paths = renew_resnet_paths(resnets)
        meta_path = {"old": f"input_blocks.{i}.0", "new": f"down_blocks.{block_id}.resnets.{layer_in_block_id}"}
        assign_to_checkpoint(
            paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config
        )

        temporal_convs = [key for key in resnets if "temopral_conv" in key]
        paths = renew_temp_conv_paths(temporal_convs)
        meta_path = {
            "old": f"input_blocks.{i}.0.temopral_conv",
            "new": f"down_blocks.{block_id}.temp_convs.{layer_in_block_id}",
        }
        assign_to_checkpoint(
            paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config
        )

        if len(attentions):
            paths = renew_attention_paths(attentions)
            meta_path = {"old": f"input_blocks.{i}.1", "new": f"down_blocks.{block_id}.attentions.{layer_in_block_id}"}
            assign_to_checkpoint(
                paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config
            )

        if len(temp_attentions):
            paths = renew_attention_paths(temp_attentions)
            meta_path = {
                "old": f"input_blocks.{i}.2",
                "new": f"down_blocks.{block_id}.temp_attentions.{layer_in_block_id}",
            }
            assign_to_checkpoint(
                paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config
            )

    resnet_0 = middle_blocks[0]
    temporal_convs_0 = [key for key in resnet_0 if "temopral_conv" in key]
    attentions = middle_blocks[1]
    temp_attentions = middle_blocks[2]
    resnet_1 = middle_blocks[3]
    temporal_convs_1 = [key for key in resnet_1 if "temopral_conv" in key]

    resnet_0_paths = renew_resnet_paths(resnet_0)
    meta_path = {"old": "middle_block.0", "new": "mid_block.resnets.0"}
    assign_to_checkpoint(
        resnet_0_paths, new_checkpoint, unet_state_dict, config=config, additional_replacements=[meta_path]
    )

    temp_conv_0_paths = renew_temp_conv_paths(temporal_convs_0)
    meta_path = {"old": "middle_block.0.temopral_conv", "new": "mid_block.temp_convs.0"}
    assign_to_checkpoint(
        temp_conv_0_paths, new_checkpoint, unet_state_dict, config=config, additional_replacements=[meta_path]
    )

    resnet_1_paths = renew_resnet_paths(resnet_1)
    meta_path = {"old": "middle_block.3", "new": "mid_block.resnets.1"}
    assign_to_checkpoint(
        resnet_1_paths, new_checkpoint, unet_state_dict, config=config, additional_replacements=[meta_path]
    )

    temp_conv_1_paths = renew_temp_conv_paths(temporal_convs_1)
    meta_path = {"old": "middle_block.3.temopral_conv", "new": "mid_block.temp_convs.1"}
    assign_to_checkpoint(
        temp_conv_1_paths, new_checkpoint, unet_state_dict, config=config, additional_replacements=[meta_path]
    )

    attentions_paths = renew_attention_paths(attentions)
    meta_path = {"old": "middle_block.1", "new": "mid_block.attentions.0"}
    assign_to_checkpoint(
        attentions_paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config
    )

    temp_attentions_paths = renew_attention_paths(temp_attentions)
    meta_path = {"old": "middle_block.2", "new": "mid_block.temp_attentions.0"}
    assign_to_checkpoint(
        temp_attentions_paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config
    )

    for i in range(num_output_blocks):
        block_id = i // (config["layers_per_block"] + 1)
        layer_in_block_id = i % (config["layers_per_block"] + 1)
        output_block_layers = [shave_segments(name, 2) for name in output_blocks[i]]
        output_block_list = {}

        for layer in output_block_layers:
            layer_id, layer_name = layer.split(".")[0], shave_segments(layer, 1)
            if layer_id in output_block_list:
                output_block_list[layer_id].append(layer_name)
            else:
                output_block_list[layer_id] = [layer_name]

        if len(output_block_list) > 1:
            resnets = [key for key in output_blocks[i] if f"output_blocks.{i}.0" in key]
            attentions = [key for key in output_blocks[i] if f"output_blocks.{i}.1" in key]
            temp_attentions = [key for key in output_blocks[i] if f"output_blocks.{i}.2" in key]

            resnet_0_paths = renew_resnet_paths(resnets)
            paths = renew_resnet_paths(resnets)

            meta_path = {"old": f"output_blocks.{i}.0", "new": f"up_blocks.{block_id}.resnets.{layer_in_block_id}"}
            assign_to_checkpoint(
                paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config
            )

            temporal_convs = [key for key in resnets if "temopral_conv" in key]
            paths = renew_temp_conv_paths(temporal_convs)
            meta_path = {
                "old": f"output_blocks.{i}.0.temopral_conv",
                "new": f"up_blocks.{block_id}.temp_convs.{layer_in_block_id}",
            }
            assign_to_checkpoint(
                paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config
            )

            output_block_list = {k: sorted(v) for k, v in output_block_list.items()}
            if ["conv.bias", "conv.weight"] in output_block_list.values():
                index = list(output_block_list.values()).index(["conv.bias", "conv.weight"])
                new_checkpoint[f"up_blocks.{block_id}.upsamplers.0.conv.weight"] = unet_state_dict[
                    f"output_blocks.{i}.{index}.conv.weight"
                ]
                new_checkpoint[f"up_blocks.{block_id}.upsamplers.0.conv.bias"] = unet_state_dict[
                    f"output_blocks.{i}.{index}.conv.bias"
                ]

                # Clear attentions as they have been attributed above.
                if len(attentions) == 2:
                    attentions = []

            if len(attentions):
                paths = renew_attention_paths(attentions)
                meta_path = {
                    "old": f"output_blocks.{i}.1",
                    "new": f"up_blocks.{block_id}.attentions.{layer_in_block_id}",
                }
                assign_to_checkpoint(
                    paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config
                )

            if len(temp_attentions):
                paths = renew_attention_paths(temp_attentions)
                meta_path = {
                    "old": f"output_blocks.{i}.2",
                    "new": f"up_blocks.{block_id}.temp_attentions.{layer_in_block_id}",
                }
                assign_to_checkpoint(
                    paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config
                )
        else:
            resnet_0_paths = renew_resnet_paths(output_block_layers, n_shave_prefix_segments=1)
            for path in resnet_0_paths:
                old_path = ".".join(["output_blocks", str(i), path["old"]])
                new_path = ".".join(["up_blocks", str(block_id), "resnets", str(layer_in_block_id), path["new"]])
                new_checkpoint[new_path] = unet_state_dict[old_path]

            temopral_conv_paths = [l for l in output_block_layers if "temopral_conv" in l]
            for path in temopral_conv_paths:
                pruned_path = path.split("temopral_conv.")[-1]
                old_path = ".".join(["output_blocks", str(i), str(block_id), "temopral_conv", pruned_path])
                new_path = ".".join(["up_blocks", str(block_id), "temp_convs", str(layer_in_block_id), pruned_path])
                new_checkpoint[new_path] = unet_state_dict[old_path]

    return new_checkpoint


if __name__ == "__main__":
    parser = argparse.ArgumentParser()

    parser.add_argument(
        "--unet_checkpoint_path", default=None, type=str, required=True, help="Path to the checkpoint to convert."
    )
    parser.add_argument("--dump_path", default=None, type=str, required=True, help="Path to the output model.")
    parser.add_argument("--push_to_hub", action="store_true")
    args = parser.parse_args()

    # UNet
    unet_checkpoint = torch.load(args.unet_checkpoint_path, map_location="cpu")
    unet_checkpoint = unet_checkpoint["state_dict"]
    unet = I2VGenXLUNet(sample_size=32)

    converted_ckpt = convert_ldm_unet_checkpoint(unet_checkpoint, unet.config)

    diff_0 = set(unet.state_dict().keys()) - set(converted_ckpt.keys())
    diff_1 = set(converted_ckpt.keys()) - set(unet.state_dict().keys())

    assert len(diff_0) == len(diff_1) == 0, "Converted weights don't match"

    unet.load_state_dict(converted_ckpt, strict=True)

    # vae
    temp_pipe = StableDiffusionPipeline.from_single_file(
        "https://huggingface.co/ali-vilab/i2vgen-xl/blob/main/models/v2-1_512-ema-pruned.ckpt"
    )
    vae = temp_pipe.vae
    del temp_pipe

    # text encoder and tokenizer
    text_encoder = CLIPTextModel.from_pretrained(CLIP_ID)
    tokenizer = CLIPTokenizer.from_pretrained(CLIP_ID)

    # image encoder and feature extractor
    image_encoder = CLIPVisionModelWithProjection.from_pretrained(CLIP_ID)
    feature_extractor = CLIPImageProcessor.from_pretrained(CLIP_ID)

    # scheduler
    # https://github.com/ali-vilab/i2vgen-xl/blob/main/configs/i2vgen_xl_train.yaml
    scheduler = DDIMScheduler(
        beta_schedule="squaredcos_cap_v2",
        rescale_betas_zero_snr=True,
        set_alpha_to_one=True,
        clip_sample=False,
        steps_offset=1,
        timestep_spacing="leading",
        prediction_type="v_prediction",
    )

    # final
    pipeline = I2VGenXLPipeline(
        unet=unet,
        vae=vae,
        image_encoder=image_encoder,
        feature_extractor=feature_extractor,
        text_encoder=text_encoder,
        tokenizer=tokenizer,
        scheduler=scheduler,
    )

    pipeline.save_pretrained(args.dump_path, push_to_hub=args.push_to_hub)