# Copyright 2021 Zilliz. 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.
import pathlib

from torch import nn
from timm.models.vision_transformer import resize_pos_embed
from towhee.types.image_utils import to_pil


class ClipCaptionReward(NNOperator):
    """
    BLIP multi-modal embedding operator
    """
    def __init__(self, model_name: str):
        super().__init__()
        sys.path.append(str(Path(__file__).parent))
        from utils import opts
        import clip
        opt = opts.parse_opt(parse=False, cfg=cfg)
        path = pathlib.Path(__file__).parent
        dict_json = json.load(open("{}/data/cocotalk.json".format(path)))
        ix_to_word = dict_json["ix_to_word"]
        self.device = "cuda" if torch.cuda.is_available() else "cpu"

        clip_model, clip_transform = clip.load("RN50", jit=False, device=self.device)
        self.clip_model = clip_model
        self.clip_transform = clip_transform

        vocab_size = len(ix_to_word)
        seq_length = 1
        opt.vocab_size = vocab_size
        opt.seq_length = seq_length
        opt.batch_size = 1
        opt.vocab = ix_to_word

        num_patches = 196  # 600 * 1000 // 32 // 32

        pos_embed = nn.Parameter(
            torch.zeros(
                1,
                num_patches + 1,
                clip_model.visual.attnpool.positional_embedding.shape[-1],
                device=self.device,
            ),
        )
        pos_embed.weight = resize_pos_embed(
            clip_model.visual.attnpool.positional_embedding.unsqueeze(0), pos_embed
        )
        self.clip_model.visual.attnpool.positional_embedding = pos_embed

        self.model = TransformerModel(opt)
        self.image_mean  = (
            torch.Tensor([0.48145466, 0.4578275, 0.40821073])
            .to(self.device)
            .reshape(3, 1, 1)
        )
        self.image_std = (
            torch.Tensor([0.26862954, 0.26130258, 0.27577711])
            .to(self.device)
            .reshape(3, 1, 1)
        )

    @arg(1, to_image_color('RGB'))
    def inference_single_data(self, data):
        text = self._inference_from_image(data)
        return text

   @arg(1, to_image_color('RGB'))
    def _inference_from_image(self, img):
        img = to_pil(img) 
        img = self._preprocess(img)
        self._inference_from_image(img)
        img -= self.image_mean
        img /= self.image_std
        tmp_att, tmp_fc = self.clip_model.encode_image(img)
        tmp_att = tmp_att[0].permute(1, 2, 0)
    
        att_feat = tmp_att

        return att_feat

    def __call__(self, data):
        if not isinstance(data, list):
            data = [data]
        else:
            data = data
        for single_data in data:
            result = self.inference_single_data(single_data)
            results.append(result)
        if len(data) == 1:
            return results[0]
        else:
            return results