# Copyright (c) 2023, NVIDIA CORPORATION.  All rights reserved.
#
# 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.

from typing import List, Optional, Union

import torch
import torch.nn as nn
from omegaconf import ListConfig

from nemo.collections.multimodal.parts.stable_diffusion.utils import append_dims, instantiate_from_config
from nemo.collections.multimodal.parts.utils import randn_like


class StandardDiffusionLoss(nn.Module):
    def __init__(
        self,
        sigma_sampler,
        type="l2",
        offset_noise_level=0.0,
        batch2model_keys: Optional[Union[str, List[str], ListConfig]] = None,
    ):
        super().__init__()

        assert type in ["l2", "l1", "lpips"]

        self.sigma_sampler = sigma_sampler

        self.type = type
        self.offset_noise_level = offset_noise_level

        if type == "lpips":
            self.lpips = LPIPS().eval()

        if not batch2model_keys:
            batch2model_keys = []

        if isinstance(batch2model_keys, str):
            batch2model_keys = [batch2model_keys]

        self.batch2model_keys = set(batch2model_keys)

    def __call__(self, network, denoiser, conditioner, input, batch, rng=None):
        cond = conditioner(batch)
        additional_model_inputs = {key: batch[key] for key in self.batch2model_keys.intersection(batch)}

        rand = torch.randint(0, self.sigma_sampler.num_idx, (input.shape[0],), generator=rng, device=rng.device).to(
            self.sigma_sampler.sigmas.device
        )
        sigmas = self.sigma_sampler(input.shape[0], rand=rand).to(input.device)
        noise = randn_like(input, generator=rng)
        if self.offset_noise_level > 0.0:
            noise = noise + self.offset_noise_level * append_dims(
                torch.randn(input.shape[0], device=input.device, generator=rng), input.ndim
            )
        noised_input = input + noise * append_dims(sigmas, input.ndim)
        model_output = denoiser(network, noised_input, sigmas, cond, **additional_model_inputs)
        w = append_dims(denoiser.w(sigmas), input.ndim)
        return self.get_loss(model_output, input, w)

    def get_loss(self, model_output, target, w):
        if self.type == "l2":
            return torch.mean((w * (model_output - target) ** 2).reshape(target.shape[0], -1), 1)
        elif self.type == "l1":
            return torch.mean((w * (model_output - target).abs()).reshape(target.shape[0], -1), 1)