declare-lab/tango2

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""" PyTorch - Flax general utilities."""16import re17 18import jax.numpy as jnp19from flax.traverse_util import flatten_dict, unflatten_dict20from jax.random import PRNGKey21 22from ..utils import logging23 24 25logger = logging.get_logger(__name__)26 27 28def rename_key(key):29    regex = r"\w+[.]\d+"30    pats = re.findall(regex, key)31    for pat in pats:32        key = key.replace(pat, "_".join(pat.split(".")))33    return key34 35 36#####################37# PyTorch => Flax #38#####################39 40 41# Adapted from https://github.com/huggingface/transformers/blob/c603c80f46881ae18b2ca50770ef65fa4033eacd/src/transformers/modeling_flax_pytorch_utils.py#L6942# and https://github.com/patil-suraj/stable-diffusion-jax/blob/main/stable_diffusion_jax/convert_diffusers_to_jax.py43def rename_key_and_reshape_tensor(pt_tuple_key, pt_tensor, random_flax_state_dict):44    """Rename PT weight names to corresponding Flax weight names and reshape tensor if necessary"""45 46    # conv norm or layer norm47    renamed_pt_tuple_key = pt_tuple_key[:-1] + ("scale",)48    if (49        any("norm" in str_ for str_ in pt_tuple_key)50        and (pt_tuple_key[-1] == "bias")51        and (pt_tuple_key[:-1] + ("bias",) not in random_flax_state_dict)52        and (pt_tuple_key[:-1] + ("scale",) in random_flax_state_dict)53    ):54        renamed_pt_tuple_key = pt_tuple_key[:-1] + ("scale",)55        return renamed_pt_tuple_key, pt_tensor56    elif pt_tuple_key[-1] in ["weight", "gamma"] and pt_tuple_key[:-1] + ("scale",) in random_flax_state_dict:57        renamed_pt_tuple_key = pt_tuple_key[:-1] + ("scale",)58        return renamed_pt_tuple_key, pt_tensor59 60    # embedding61    if pt_tuple_key[-1] == "weight" and pt_tuple_key[:-1] + ("embedding",) in random_flax_state_dict:62        pt_tuple_key = pt_tuple_key[:-1] + ("embedding",)63        return renamed_pt_tuple_key, pt_tensor64 65    # conv layer66    renamed_pt_tuple_key = pt_tuple_key[:-1] + ("kernel",)67    if pt_tuple_key[-1] == "weight" and pt_tensor.ndim == 4:68        pt_tensor = pt_tensor.transpose(2, 3, 1, 0)69        return renamed_pt_tuple_key, pt_tensor70 71    # linear layer72    renamed_pt_tuple_key = pt_tuple_key[:-1] + ("kernel",)73    if pt_tuple_key[-1] == "weight":74        pt_tensor = pt_tensor.T75        return renamed_pt_tuple_key, pt_tensor76 77    # old PyTorch layer norm weight78    renamed_pt_tuple_key = pt_tuple_key[:-1] + ("weight",)79    if pt_tuple_key[-1] == "gamma":80        return renamed_pt_tuple_key, pt_tensor81 82    # old PyTorch layer norm bias83    renamed_pt_tuple_key = pt_tuple_key[:-1] + ("bias",)84    if pt_tuple_key[-1] == "beta":85        return renamed_pt_tuple_key, pt_tensor86 87    return pt_tuple_key, pt_tensor88 89 90def convert_pytorch_state_dict_to_flax(pt_state_dict, flax_model, init_key=42):91    # Step 1: Convert pytorch tensor to numpy92    pt_state_dict = {k: v.numpy() for k, v in pt_state_dict.items()}93 94    # Step 2: Since the model is stateless, get random Flax params95    random_flax_params = flax_model.init_weights(PRNGKey(init_key))96 97    random_flax_state_dict = flatten_dict(random_flax_params)98    flax_state_dict = {}99 100    # Need to change some parameters name to match Flax names101    for pt_key, pt_tensor in pt_state_dict.items():102        renamed_pt_key = rename_key(pt_key)103        pt_tuple_key = tuple(renamed_pt_key.split("."))104 105        # Correctly rename weight parameters106        flax_key, flax_tensor = rename_key_and_reshape_tensor(pt_tuple_key, pt_tensor, random_flax_state_dict)107 108        if flax_key in random_flax_state_dict:109            if flax_tensor.shape != random_flax_state_dict[flax_key].shape:110                raise ValueError(111                    f"PyTorch checkpoint seems to be incorrect. Weight {pt_key} was expected to be of shape "112                    f"{random_flax_state_dict[flax_key].shape}, but is {flax_tensor.shape}."113                )114 115        # also add unexpected weight so that warning is thrown116        flax_state_dict[flax_key] = jnp.asarray(flax_tensor)117 118    return unflatten_dict(flax_state_dict)119