# from https://github.com/pytorch/fairseq/blob/master/scripts/average_checkpoints.py
import collections
import torch


def average_checkpoints(inputs):
    params_dict = collections.OrderedDict()
    params_keys = None
    new_state = None
    num_models = len(inputs)
    for fpath in inputs:
        with open(fpath, "rb") as f:
            state = torch.load(
                f,
                map_location=(
                    lambda s, _: torch.serialization.default_restore_location(s, "cpu")
                ),
            )
        # Copies over the settings from the first checkpoint
        if new_state is None:
            new_state = state
        model_params = state["model"]
        model_params_keys = list(model_params.keys())
        if params_keys is None:
            params_keys = model_params_keys
        elif params_keys != model_params_keys:
            raise KeyError(
                "For checkpoint {}, expected list of params: {}, "
                "but found: {}".format(f, params_keys, model_params_keys)
            )
        for k in params_keys:
            p = model_params[k]
            if isinstance(p, torch.HalfTensor):
                p = p.float()
            if k not in params_dict:
                params_dict[k] = p.clone()
                # NOTE: clone() is needed in case of p is a shared parameter
            else:
                params_dict[k] += p
    averaged_params = collections.OrderedDict()
    for k, v in params_dict.items():
        averaged_params[k] = v
        if averaged_params[k].is_floating_point():
            averaged_params[k].div_(num_models)
        else:
            averaged_params[k] //= num_models
    new_state["model"] = averaged_params
    return new_state


def avg(epochs, filename):
    paths = ["model_{}.pth".format(i) for i in epochs]
    weights = average_checkpoints(paths)
    torch.save(weights, filename.format(len(epochs)))

avg([549, 528, 408], "model_best{}avg.pth")