super-gradients/src/super_gradients/training/utils/checkpoint_utils.py

import os
import tempfile
import pkg_resources

import torch

from super_gradients.common.abstractions.abstract_logger import get_logger
from super_gradients.common import explicit_params_validation, ADNNModelRepositoryDataInterfaces
from super_gradients.training.pretrained_models import MODEL_URLS
from super_gradients.common.environment import environment_config

try:
    from torch.hub import download_url_to_file, load_state_dict_from_url
except (ModuleNotFoundError, ImportError, NameError):
    from torch.hub import _download_url_to_file as download_url_to_file


logger = get_logger(__name__)


def get_checkpoints_dir_path(experiment_name: str, ckpt_root_dir: str = None):
    """Creating the checkpoint directory of a given experiment.
    :param experiment_name:     Name of the experiment.
    :param ckpt_root_dir:       Local root directory path where all experiment logging directories will
                                reside. When none is give, it is assumed that pkg_resources.resource_filename('checkpoints', "")
                                exists and will be used.
    :return:                    checkpoints_dir_path
    """
    if ckpt_root_dir:
        return os.path.join(ckpt_root_dir, experiment_name)
    elif os.path.exists(environment_config.PKG_CHECKPOINTS_DIR):
        return os.path.join(environment_config.PKG_CHECKPOINTS_DIR, experiment_name)
    else:
        raise ValueError("Illegal checkpoints directory: pass ckpt_root_dir that exists, or add 'checkpoints' to resources.")


def get_ckpt_local_path(source_ckpt_folder_name: str, experiment_name: str, ckpt_name: str, external_checkpoint_path: str):
    """
    Gets the local path to the checkpoint file, which will be:
        - By default: YOUR_REPO_ROOT/super_gradients/checkpoints/experiment_name.
        - if the checkpoint file is remotely located:
            when overwrite_local_checkpoint=True then it will be saved in a temporary path which will be returned,
            otherwise it will be downloaded to YOUR_REPO_ROOT/super_gradients/checkpoints/experiment_name and overwrite
            YOUR_REPO_ROOT/super_gradients/checkpoints/experiment_name/ckpt_name if such file exists.
        - external_checkpoint_path when external_checkpoint_path != None

    @param source_ckpt_folder_name: The folder where the checkpoint is saved. When set to None- uses the experiment_name.
    @param experiment_name: experiment name attr in trainer
    @param ckpt_name: checkpoint filename
    @param external_checkpoint_path: full path to checkpoint file (that might be located outside of super_gradients/checkpoints directory)
    @return:
    """
    if external_checkpoint_path:
        return external_checkpoint_path
    else:
        checkpoints_folder_name = source_ckpt_folder_name or experiment_name
        checkpoints_dir_path = get_checkpoints_dir_path(checkpoints_folder_name)
        return os.path.join(checkpoints_dir_path, ckpt_name)


def adaptive_load_state_dict(net: torch.nn.Module, state_dict: dict, strict: str):
    """
    Adaptively loads state_dict to net, by adapting the state_dict to net's layer names first.

    @param net: (nn.Module) to load state_dict to
    @param state_dict: (dict) Chekpoint state_dict
    @param strict: (str) key matching strictness
    @return:
    """
    try:
        net.load_state_dict(state_dict["net"] if "net" in state_dict.keys() else state_dict, strict=strict)
    except (RuntimeError, ValueError, KeyError) as ex:
        if strict == "no_key_matching":
            adapted_state_dict = adapt_state_dict_to_fit_model_layer_names(net.state_dict(), state_dict)
            net.load_state_dict(adapted_state_dict["net"], strict=True)
        else:
            raise_informative_runtime_error(net.state_dict(), state_dict, ex)


@explicit_params_validation(validation_type="None")
def copy_ckpt_to_local_folder(
    local_ckpt_destination_dir: str,
    ckpt_filename: str,
    remote_ckpt_source_dir: str = None,
    path_src: str = "local",
    overwrite_local_ckpt: bool = False,
    load_weights_only: bool = False,
):
    """
    Copy the checkpoint from any supported source to a local destination path
        :param local_ckpt_destination_dir:  destination where the checkpoint will be saved to
        :param ckpt_filename:         ckpt_best.pth Or ckpt_latest.pth
        :param remote_ckpt_source_dir:       Name of the source checkpoint to be loaded (S3 Model\full URL)
        :param path_src:              S3 / url
        :param overwrite_local_ckpt:  determines if checkpoint will be saved in destination dir or in a temp folder

        :return: Path to checkpoint
    """
    ckpt_file_full_local_path = None

    # IF NOT DEFINED - IT IS SET TO THE TARGET's FOLDER NAME
    remote_ckpt_source_dir = local_ckpt_destination_dir if remote_ckpt_source_dir is None else remote_ckpt_source_dir

    if not overwrite_local_ckpt:
        # CREATE A TEMP FOLDER TO SAVE THE CHECKPOINT TO
        download_ckpt_destination_dir = tempfile.gettempdir()
        print(
            "PLEASE NOTICE - YOU ARE IMPORTING A REMOTE CHECKPOINT WITH overwrite_local_checkpoint = False "
            "-> IT WILL BE REDIRECTED TO A TEMP FOLDER AND DELETED ON MACHINE RESTART"
        )
    else:
        # SAVE THE CHECKPOINT TO MODEL's FOLDER
        download_ckpt_destination_dir = pkg_resources.resource_filename("checkpoints", local_ckpt_destination_dir)

    if path_src.startswith("s3"):
        model_checkpoints_data_interface = ADNNModelRepositoryDataInterfaces(data_connection_location=path_src)
        # DOWNLOAD THE FILE FROM S3 TO THE DESTINATION FOLDER
        ckpt_file_full_local_path = model_checkpoints_data_interface.load_remote_checkpoints_file(
            ckpt_source_remote_dir=remote_ckpt_source_dir,
            ckpt_destination_local_dir=download_ckpt_destination_dir,
            ckpt_file_name=ckpt_filename,
            overwrite_local_checkpoints_file=overwrite_local_ckpt,
        )

        if not load_weights_only:
            # COPY LOG FILES FROM THE REMOTE DIRECTORY TO THE LOCAL ONE ONLY IF LOADING THE CURRENT MODELs CKPT
            model_checkpoints_data_interface.load_all_remote_log_files(
                model_name=remote_ckpt_source_dir, model_checkpoint_local_dir=download_ckpt_destination_dir
            )

    if path_src == "url":
        ckpt_file_full_local_path = download_ckpt_destination_dir + os.path.sep + ckpt_filename
        # DOWNLOAD THE FILE FROM URL TO THE DESTINATION FOLDER
        download_url_to_file(remote_ckpt_source_dir, ckpt_file_full_local_path, progress=True)

    return ckpt_file_full_local_path


def read_ckpt_state_dict(ckpt_path: str, device="cpu"):
    if not os.path.exists(ckpt_path):
        raise FileNotFoundError(f"Incorrect Checkpoint path: {ckpt_path} (This should be an absolute path)")

    if device == "cuda":
        state_dict = torch.load(ckpt_path)

    else:
        state_dict = torch.load(ckpt_path, map_location=lambda storage, loc: storage)
    return state_dict


def adapt_state_dict_to_fit_model_layer_names(model_state_dict: dict, source_ckpt: dict, exclude: list = [], solver: callable = None):
    """
    Given a model state dict and source checkpoints, the method tries to correct the keys in the model_state_dict to fit
    the ckpt in order to properly load the weights into the model. If unsuccessful - returns None
        :param model_state_dict:               the model state_dict
        :param source_ckpt:                         checkpoint dict
        :param exclude                  optional list for excluded layers
        :param solver:                  callable with signature (ckpt_key, ckpt_val, model_key, model_val)
                                        that returns a desired weight for ckpt_val.
        :return: renamed checkpoint dict (if possible)
    """
    if "net" in source_ckpt.keys():
        source_ckpt = source_ckpt["net"]
    model_state_dict_excluded = {k: v for k, v in model_state_dict.items() if not any(x in k for x in exclude)}
    new_ckpt_dict = {}
    for (ckpt_key, ckpt_val), (model_key, model_val) in zip(source_ckpt.items(), model_state_dict_excluded.items()):
        if solver is not None:
            ckpt_val = solver(ckpt_key, ckpt_val, model_key, model_val)
        if ckpt_val.shape != model_val.shape:
            raise ValueError(f"ckpt layer {ckpt_key} with shape {ckpt_val.shape} does not match {model_key}" f" with shape {model_val.shape} in the model")
        new_ckpt_dict[model_key] = ckpt_val
    return {"net": new_ckpt_dict}


def raise_informative_runtime_error(state_dict, checkpoint, exception_msg):
    """
    Given a model state dict and source checkpoints, the method calls "adapt_state_dict_to_fit_model_layer_names"
    and enhances the exception_msg if loading the checkpoint_dict via the conversion method is possible
    """
    try:
        new_ckpt_dict = adapt_state_dict_to_fit_model_layer_names(state_dict, checkpoint)
        temp_file = tempfile.NamedTemporaryFile().name + ".pt"
        torch.save(new_ckpt_dict, temp_file)
        exception_msg = (
            f"\n{'=' * 200}\n{str(exception_msg)} \nconvert ckpt via the utils.adapt_state_dict_to_fit_"
            f"model_layer_names method\na converted checkpoint file was saved in the path {temp_file}\n{'=' * 200}"
        )
    except ValueError as ex:  # IN CASE adapt_state_dict_to_fit_model_layer_names WAS UNSUCCESSFUL
        exception_msg = f"\n{'=' * 200} \nThe checkpoint and model shapes do no fit, e.g.: {ex}\n{'=' * 200}"
    finally:
        raise RuntimeError(exception_msg)


def load_checkpoint_to_model(
    ckpt_local_path: str, load_backbone: bool, net: torch.nn.Module, strict: str, load_weights_only: bool, load_ema_as_net: bool = False
):
    """
    Loads the state dict in ckpt_local_path to net and returns the checkpoint's state dict.

    @param load_ema_as_net: Will load the EMA inside the checkpoint file to the network when set
    @param ckpt_local_path: local path to the checkpoint file
    @param load_backbone: whether to load the checkpoint as a backbone
    @param net: network to load the checkpoint to
    @param strict:
    @param load_weights_only:
    @return:
    """
    if ckpt_local_path is None or not os.path.exists(ckpt_local_path):
        error_msg = "Error - loading Model Checkpoint: Path {} does not exist".format(ckpt_local_path)
        raise RuntimeError(error_msg)

    if load_backbone and not hasattr(net, "backbone"):
        raise ValueError("No backbone attribute in net - Can't load backbone weights")

    # LOAD THE LOCAL CHECKPOINT PATH INTO A state_dict OBJECT
    checkpoint = read_ckpt_state_dict(ckpt_path=ckpt_local_path)

    if load_ema_as_net:
        if "ema_net" not in checkpoint.keys():
            raise ValueError("Can't load ema network- no EMA network stored in checkpoint file")
        else:
            checkpoint["net"] = checkpoint["ema_net"]

    # LOAD THE CHECKPOINTS WEIGHTS TO THE MODEL
    if load_backbone:
        adaptive_load_state_dict(net.backbone, checkpoint, strict)
    else:
        adaptive_load_state_dict(net, checkpoint, strict)

    message_suffix = " checkpoint." if not load_ema_as_net else " EMA checkpoint."
    message_model = "model" if not load_backbone else "model's backbone"
    logger.info("Successfully loaded " + message_model + " weights from " + ckpt_local_path + message_suffix)

    if load_weights_only or load_backbone:
        # DISCARD ALL THE DATA STORED IN CHECKPOINT OTHER THAN THE WEIGHTS
        [checkpoint.pop(key) for key in list(checkpoint.keys()) if key != "net"]

    return checkpoint


class MissingPretrainedWeightsException(Exception):
    """Exception raised by unsupported pretrianed model.

    Attributes:
        message -- explanation of the error
    """

    def __init__(self, desc):
        self.message = "Missing pretrained wights: " + desc
        super().__init__(self.message)


def _yolox_ckpt_solver(ckpt_key, ckpt_val, model_key, model_val):
    """
    Helper method for reshaping old pretrained checkpoint's focus weights to 6x6 conv weights.
    """

    if (
        ckpt_val.shape != model_val.shape
        and ckpt_key == "module._backbone._modules_list.0.conv.conv.weight"
        and model_key == "_backbone._modules_list.0.conv.weight"
    ):
        model_val.data[:, :, ::2, ::2] = ckpt_val.data[:, :3]
        model_val.data[:, :, 1::2, ::2] = ckpt_val.data[:, 3:6]
        model_val.data[:, :, ::2, 1::2] = ckpt_val.data[:, 6:9]
        model_val.data[:, :, 1::2, 1::2] = ckpt_val.data[:, 9:12]
        replacement = model_val
    else:
        replacement = ckpt_val

    return replacement


def load_pretrained_weights(model: torch.nn.Module, architecture: str, pretrained_weights: str):

    """
    Loads pretrained weights from the MODEL_URLS dictionary to model
    @param architecture: name of the model's architecture
    @param model: model to load pretrinaed weights for
    @param pretrained_weights: name for the pretrianed weights (i.e imagenet)
    @return: None
    """
    model_url_key = architecture + "_" + str(pretrained_weights)
    if model_url_key not in MODEL_URLS.keys():
        raise MissingPretrainedWeightsException(model_url_key)

    url = MODEL_URLS[model_url_key]
    unique_filename = url.split("https://deci-pretrained-models.s3.amazonaws.com/")[1].replace("/", "_").replace(" ", "_")
    map_location = torch.device("cpu")
    pretrained_state_dict = load_state_dict_from_url(url=url, map_location=map_location, file_name=unique_filename)
    _load_weights(architecture, model, pretrained_state_dict)


def _load_weights(architecture, model, pretrained_state_dict):
    if "ema_net" in pretrained_state_dict.keys():
        pretrained_state_dict["net"] = pretrained_state_dict["ema_net"]
    solver = _yolox_ckpt_solver if "yolox" in architecture else None
    adapted_pretrained_state_dict = adapt_state_dict_to_fit_model_layer_names(
        model_state_dict=model.state_dict(), source_ckpt=pretrained_state_dict, solver=solver
    )
    model.load_state_dict(adapted_pretrained_state_dict["net"], strict=False)


def load_pretrained_weights_local(model: torch.nn.Module, architecture: str, pretrained_weights: str):

    """
    Loads pretrained weights from the MODEL_URLS dictionary to model
    @param architecture: name of the model's architecture
    @param model: model to load pretrinaed weights for
    @param pretrained_weights: path tp pretrained weights
    @return: None
    """

    map_location = torch.device("cpu")

    pretrained_state_dict = torch.load(pretrained_weights, map_location=map_location)
    _load_weights(architecture, model, pretrained_state_dict)