pytorchvideo/pytorchvideo.py

import logging
import os
import json
from pathlib import Path
from typing import List, Union, Iterable, Callable

import torch
from torch import nn
import numpy

from towhee import register
from towhee.types import VideoFrame
from towhee.operator.base import NNOperator
from towhee.models.utils.video_transforms import transform_video

log = logging.getLogger()


@register(output_schema=['labels', 'scores', 'features'])
class PytorchVideo(NNOperator):
    """
    Generate a list of class labels given a video input data.
    Default labels are from [Kinetics400 Dataset](https://deepmind.com/research/open-source/kinetics).

    Args:
        model_name (`str`):
            The pretrained model name from torch hub.
            Supported model names:
            - c2d_r50
            - i3d_r50
            - slow_r50
            - slowfast_r50
            - slowfast_r101
            - x3d_xs
            - x3d_s
            - x3d_m
            - mvit_base_16x4
            - mvit_base_32x3
        skip_preprocess (`str`):
            Flag to skip video transforms.
        classmap (`dict=None`):
            The dictionary maps classes to integers.
        topk (`int=5`):
            The number of classification labels to be returned (ordered by possibility from high to low).
    """

    def __init__(
            self,
            model_name: str = 'x3d_xs',
            framework: str = 'pytorch',
            skip_preprocess: bool = False,
            classmap: dict = None,
            topk: int = 5,
    ) -> None:
        super().__init__(framework=framework)
        self.model_name = model_name
        self.skip_preprocess = skip_preprocess
        self.topk = topk
        if classmap is None:
            class_file = os.path.join(str(Path(__file__).parent), 'kinetics_400.json')
            with open(class_file, 'r') as f:
                kinetics_classes = json.load(f)
            self.classmap = {}
            for k, v in kinetics_classes.items():
                self.classmap[v] = str(k).replace('"', '')
        else:
            self.classmap = classmap
        self.device = 'cuda' if torch.cuda.is_available() else 'cpu'
        # self.device = 'cpu'

        self.model = torch.hub.load('facebookresearch/pytorchvideo', model=model_name, pretrained=True)
        self.model.eval()
        self.model.to(self.device)

    def __call__(self, frames: List[VideoFrame]):
        """
        Args:
            frames (`List[VideoFrame]`):
                Video frames in towhee.types.video_frame.VideoFrame.

        Returns:
            labels, scores:
                A tuple of lists (labels, scores).
            video embedding:
                A video embedding in numpy.ndarray.
        """
        # Convert list of towhee.types.Image to numpy.ndarray in float32
        video = numpy.stack([img.astype(numpy.float32) / 255. for img in frames], axis=0)
        assert len(video.shape) == 4
        video = video.transpose(3, 0, 1, 2)  # twhc -> ctwh

        if self.skip_preprocess:
            data = transform_video(
                video=video,
                model_name=self.model_name,
                num_frames=None
            )
        else:
            data = transform_video(
                video=video,
                model_name=self.model_name
            )
        if self.model_name.startswith('slowfast'):
            inputs = [data[0].to(self.device)[None, ...], data[1].to(self.device)[None, ...]]
        else:
            inputs = data.to(self.device)[None, ...]

        feats, outs = self.new_forward(inputs)
        features = feats.to('cpu').squeeze(0).detach().numpy()

        post_act = torch.nn.Softmax(dim=1)
        preds = post_act(outs)
        pred_scores, pred_classes = preds.topk(k=self.topk)
        labels = [self.classmap[int(i)] for i in pred_classes[0]]
        scores = [round(float(x), 5) for x in pred_scores[0]]
        return labels, scores, features

    def new_forward(self, x: Union[torch.Tensor, list]):
        """
        Generate embeddings returned by the second last hidden layer.

        Args:
            x (`Union[torch.Tensor, list]`):
                tensor or list of input video after transforms

        Returns:
            Tensor of layer outputs.
        """
        blocks = list(self.model.children())
        if len(blocks) == 1:
            blocks = blocks[0]
        if self.model_name.startswith('x3d'):
            sub_blocks = list(blocks[-1].children())
            extractor = FeatureExtractor(self.model, sub_blocks, layer=0)
        elif self.model_name.startswith('mvit'):
            sub_blocks = list(blocks[-1].children())
            extractor = FeatureExtractor(self.model, sub_blocks, layer=0)
        else:
            extractor = FeatureExtractor(self.model, blocks, layer=-2)
        features, outs = extractor(x)
        if features.dim() == 5:
            global_pool = nn.AdaptiveAvgPool3d(1)
            features = global_pool(features)
        return features.flatten(), outs

    def get_model_name(self):
        full_list = [
            'c2d_r50',
            'i3d_r50',
            'slow_r50',
            'slowfast_r50',
            'slowfast_r101',
            'x3d_xs',
            'x3d_s',
            'x3d_m',
            'mvit_base_16x4',
            'mvit_base_32x3'
        ]
        full_list.sort()
        return full_list


class FeatureExtractor(nn.Module):
    def __init__(self, model: nn.Module, blocks: List[nn.Module], layer: int):
        super().__init__()
        self.model = model
        self.features = None

        target_layer = blocks[layer]
        self.handler = target_layer.register_forward_hook(self.save_outputs_hook())

    def save_outputs_hook(self) -> Callable:
        def fn(_, __, output):
            self.features = output
        return fn

    def forward(self, x):
        outs = self.model(x)
        self.handler.remove()
        return self.features, outs