Custom Dataloader and Accuracy being constant

[Mukil07]

Im trying to load a custom dataset which has multiple videos for the same sequence (6 videos), I need to load them together at once. Please find the custom dataloader i defined,

#!/usr/bin/env python3
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.

import os
import random

import numpy as np
import pandas

import slowfast.utils.logging as logging
import torch
import torch.utils.data
from slowfast.utils.env import pathmgr
from torchvision import transforms

from . import (
    decoder as decoder,
    transform as transform,
    utils as utils,
    video_container as container,
)
from .build import DATASET_REGISTRY
from .random_erasing import RandomErasing
from .transform import create_random_augment, MaskingGenerator, MaskingGenerator3D

logger = logging.get_logger(__name__)


@DATASET_REGISTRY.register()
class Kinetics_v2(torch.utils.data.Dataset):
    """
    Kinetics video loader. Construct the Kinetics video loader, then sample
    clips from the videos. For training and validation, a single clip is
    randomly sampled from every video with random cropping, scaling, and
    flipping. For testing, multiple clips are uniformaly sampled from every
    video with uniform cropping. For uniform cropping, we take the left, center,
    and right crop if the width is larger than height, or take top, center, and
    bottom crop if the height is larger than the width.
    """

    def __init__(self, cfg, mode, num_retries=100):
        """
        Construct the Kinetics video loader with a given csv file. The format of
        the csv file is:
        ```
        path_to_video_1 label_1
        path_to_video_2 label_2
        ...
        path_to_video_N label_N
        ```
        Args:
            cfg (CfgNode): configs.
            mode (string): Options includes `train`, `val`, or `test` mode.
                For the train and val mode, the data loader will take data
                from the train or val set, and sample one clip per video.
                For the test mode, the data loader will take data from test set,
                and sample multiple clips per video.
            num_retries (int): number of retries.
        """
        # Only support train, val, and test mode.
        assert mode in [
            "train",
            "val",
            "test",
        ], "Split '{}' not supported for Kinetics".format(mode)
        self.mode = mode
        self.cfg = cfg
        self.p_convert_gray = self.cfg.DATA.COLOR_RND_GRAYSCALE
        self.p_convert_dt = self.cfg.DATA.TIME_DIFF_PROB
        self._video_meta = {}
        self._num_retries = num_retries
        self._num_epoch = 0.0
        self._num_yielded = 0
        self.gaze = None
        self.ego = None
        self.skip_rows = self.cfg.DATA.SKIP_ROWS
        self.use_chunk_loading = (
            True
            if self.mode in ["train"] and self.cfg.DATA.LOADER_CHUNK_SIZE > 0
            else False
        )
        self.dummy_output = None
        # For training or validation mode, one single clip is sampled from every
        # video. For testing, NUM_ENSEMBLE_VIEWS clips are sampled from every
        # video. For every clip, NUM_SPATIAL_CROPS is cropped spatially from
        # the frames.
        if self.mode in ["train", "val"]:
            self._num_clips = 1
        elif self.mode in ["test"]:
            self._num_clips = cfg.TEST.NUM_ENSEMBLE_VIEWS * cfg.TEST.NUM_SPATIAL_CROPS

        logger.info("Constructing Kinetics {}...".format(mode))
        self._construct_loader()
        self.aug = False
        self.rand_erase = False
        self.use_temporal_gradient = False
        self.temporal_gradient_rate = 0.0
        self.cur_epoch = 0

        if self.mode == "train" and self.cfg.AUG.ENABLE:
            self.aug = True
            if self.cfg.AUG.RE_PROB > 0:
                self.rand_erase = True

    def _construct_loader(self):
        """
        Construct the video loader.
        """
    # import pdb;pdb.set_trace()
        path_to_file = os.path.join(
            self.cfg.DATA.PATH_TO_DATA_DIR, "{}.csv".format(self.mode)
        )
        assert pathmgr.exists(path_to_file), "{} dir not found".format(path_to_file)


        self._path_to_videos_1 = []
        self._path_to_videos_2 = []
        self._path_to_videos_3 = []
        self._path_to_videos_4 = []
        self._path_to_videos_5 = []
        self._path_to_videos_6 = []
        self.gaze = []
        self.ego = []
        self._labels = []
        self._spatial_temporal_idx = []
        self.cur_iter = 0
        self.chunk_epoch = 0
        self.epoch = 0.0
        self.skip_rows = self.cfg.DATA.SKIP_ROWS

        path_ = None

        #import pdb;pdb.set_trace()
        with pathmgr.open(path_to_file, "r") as f:
            if self.use_chunk_loading:
                rows = self._get_chunk(f, self.cfg.DATA.LOADER_CHUNK_SIZE)
            else:
                #import pdb;pdb.set_trace()
                rows = f.read().splitlines()
            for clip_idx, annotation in enumerate(rows):
                fetch_info = annotation.split(',',3)

                if len(fetch_info) == 4:
                    path, label, gaze, ego = fetch_info
                    #self.gaze = gaze
                    #self.ego = ego
                else:
                    import pdb;pdb.set_trace()
                    raise RuntimeError(
                        "Failed to parse video fetch {} info {} retries.".format(
                            path_to_file, fetch_info
                        )
                    )
                
                direc = "/scratch/xxx/dataset/test/"

                f1 = os.path.join(direc,'view0_common')
                f2 = os.path.join(direc,'view1_common')
                f3 = os.path.join(direc,'view2_common')
                f4 = os.path.join(direc,'view3_common')
                f5 = os.path.join(direc,'view4_common')
                f6 = os.path.join(direc,'view5_common')

                path1 = os.path.join(f1,path)
                path2 = os.path.join(f2,path)
                path3 = os.path.join(f3,path)
                path4 = os.path.join(f4,path)
                path5 = os.path.join(f5,path)
                path6 = os.path.join(f6,path)

                if not (os.path.exists(path1) and os.path.exists(path2) and os.path.exists(path3) and os.path.exists(path4) and os.path.exists(path5) and os.path.exists(path6)):
                    continue


                for idx in range(self._num_clips):
                    self._path_to_videos_1.append(
                        os.path.join(self.cfg.DATA.PATH_PREFIX, path1)
                    )
                    self._path_to_videos_2.append(
                        os.path.join(self.cfg.DATA.PATH_PREFIX, path2)
                    )
                    self._path_to_videos_3.append(
                        os.path.join(self.cfg.DATA.PATH_PREFIX, path3)
                    )
                    self._path_to_videos_4.append(
                        os.path.join(self.cfg.DATA.PATH_PREFIX, path4)
                    )
                    self._path_to_videos_5.append(
                        os.path.join(self.cfg.DATA.PATH_PREFIX, path5)
                    )
                    self._path_to_videos_6.append(
                        os.path.join(self.cfg.DATA.PATH_PREFIX, path6)
                    )               
                   
                    ego = list(map(int, ego.strip('[]').split()))
                    ego = [torch.tensor(i,dtype=torch.int8) for i in ego]
                    gaze = torch.tensor(int(gaze),dtype=torch.int8)
                    self.gaze.append(gaze)        
                    self.ego.append(ego)          

                    self._labels.append(int(label))
                    self._spatial_temporal_idx.append(idx)
                    self._video_meta[clip_idx * self._num_clips + idx] = {}
                   # self._path_to_videos_3 = 
        assert (
            len(self._path_to_videos_1) > 0
        ), "Failed to load Kinetics split {} from {}".format(
            self._split_idx, path_to_file
        )
        logger.info(
            "Constructing kinetics dataloader (size: {} skip_rows {}) from {} ".format(
                len(self._path_to_videos_1), self.skip_rows, path_to_file
            )
        )

    def _set_epoch_num(self, epoch):
        self.epoch = epoch

    def _get_chunk(self, path_to_file, chunksize):
        try:
            for chunk in pandas.read_csv(
                path_to_file,
                chunksize=self.cfg.DATA.LOADER_CHUNK_SIZE,
                skiprows=self.skip_rows,
            ):
                break
        except Exception:
            self.skip_rows = 0
            return self._get_chunk(path_to_file, chunksize)
        else:
            return pandas.array(chunk.values.flatten(), dtype="string")

    def __getitem__(self, index):
        """
        Given the video index, return the list of frames, label, and video
        index if the video can be fetched and decoded successfully, otherwise
        repeatly find a random video that can be decoded as a replacement.
        Args:
            index (int): the video index provided by the pytorch sampler.
        Returns:
            frames (tensor): the frames of sampled from the video. The dimension
                is `channel` x `num frames` x `height` x `width`.
            label (int): the label of the current video.
            index (int): if the video provided by pytorch sampler can be
                decoded, then return the index of the video. If not, return the
                index of the video replacement that can be decoded.
        """
        short_cycle_idx = None
        # When short cycle is used, input index is a tupple.
        if isinstance(index, tuple):
            index, self._num_yielded = index
            if self.cfg.MULTIGRID.SHORT_CYCLE:
                index, short_cycle_idx = index
        if self.dummy_output is not None:
            return self.dummy_output
        if self.mode in ["train", "val"]:
            # -1 indicates random sampling.
            temporal_sample_index = -1
            spatial_sample_index = -1
            min_scale = self.cfg.DATA.TRAIN_JITTER_SCALES[0]
            max_scale = self.cfg.DATA.TRAIN_JITTER_SCALES[1]
            crop_size = self.cfg.DATA.TRAIN_CROP_SIZE
            if short_cycle_idx in [0, 1]:
                crop_size = int(
                    round(
                        self.cfg.MULTIGRID.SHORT_CYCLE_FACTORS[short_cycle_idx]
                        * self.cfg.MULTIGRID.DEFAULT_S
                    )
                )
            if self.cfg.MULTIGRID.DEFAULT_S > 0:
                # Decreasing the scale is equivalent to using a larger "span"
                # in a sampling grid.
                min_scale = int(
                    round(float(min_scale) * crop_size / self.cfg.MULTIGRID.DEFAULT_S)
                )
        elif self.mode in ["test"]:
            temporal_sample_index = (
                self._spatial_temporal_idx[index] // self.cfg.TEST.NUM_SPATIAL_CROPS
            )
            # spatial_sample_index is in [0, 1, 2]. Corresponding to left,
            # center, or right if width is larger than height, and top, middle,
            # or bottom if height is larger than width.
            spatial_sample_index = (
                (self._spatial_temporal_idx[index] % self.cfg.TEST.NUM_SPATIAL_CROPS)
                if self.cfg.TEST.NUM_SPATIAL_CROPS > 1
                else 1
            )
            min_scale, max_scale, crop_size = (
                [self.cfg.DATA.TEST_CROP_SIZE] * 3
                if self.cfg.TEST.NUM_SPATIAL_CROPS > 1
                else [self.cfg.DATA.TRAIN_JITTER_SCALES[0]] * 2
                + [self.cfg.DATA.TEST_CROP_SIZE]
            )
            # The testing is deterministic and no jitter should be performed.
            # min_scale, max_scale, and crop_size are expect to be the same.
            assert len({min_scale, max_scale}) == 1
        else:
            raise NotImplementedError("Does not support {} mode".format(self.mode))
        num_decode = (
            self.cfg.DATA.TRAIN_CROP_NUM_TEMPORAL if self.mode in ["train"] else 1
        )
        min_scale, max_scale, crop_size = [min_scale], [max_scale], [crop_size]
        if len(min_scale) < num_decode:
            min_scale += [self.cfg.DATA.TRAIN_JITTER_SCALES[0]] * (
                num_decode - len(min_scale)
            )
            max_scale += [self.cfg.DATA.TRAIN_JITTER_SCALES[1]] * (
                num_decode - len(max_scale)
            )
            crop_size += (
                [self.cfg.MULTIGRID.DEFAULT_S] * (num_decode - len(crop_size))
                if self.cfg.MULTIGRID.LONG_CYCLE or self.cfg.MULTIGRID.SHORT_CYCLE
                else [self.cfg.DATA.TRAIN_CROP_SIZE] * (num_decode - len(crop_size))
            )
            assert self.mode in ["train", "val"]
        # Try to decode and sample a clip from a video. If the video can not be
        # decoded, repeatly find a random video replacement that can be decoded.
        clean = True
        
        while clean:
            #while retry:
            #for i_try in range(self._num_retries):
            FRAMES=[]
            path_list = [self._path_to_videos_1, self._path_to_videos_2, self._path_to_videos_3, self._path_to_videos_4, self._path_to_videos_5, self._path_to_videos_6]

            video_container = None
            
            for path_ in path_list:
                ################################## LOADING ######################################################
                if isinstance(index, list):
                    
                    index=index[0]
                try:
                    video_container = container.get_video_container(
                        path_[index],
                        self.cfg.DATA_LOADER.ENABLE_MULTI_THREAD_DECODE,
                        self.cfg.DATA.DECODING_BACKEND,
                    )
                except Exception as e:
                    import pdb;pdb.set_trace()
                    logger.info(
                        "Failed to load video from {} with error {}".format(
                            path_[index], e
                        )
                    )
                    # if self.mode not in ["test"]:
                    #     # let's try another one
                    #     #index = random.randint(0, len(path_) - 1)
                    #continue  # Select a random video if the current video was not able to access.
                    index = random.randint(0, len(path_) - 1)
                    break

                if video_container is None:
                    index = random.randint(0, len(path_) - 1)
                    print("print video container is empty", path_[index])
                    break
                ################################## LOADING ######################################################
                frames_decoded, time_idx_decoded = (
                    [None] * num_decode,
                    [None] * num_decode,
                )

                # for i in range(num_decode):
                num_frames = [self.cfg.DATA.NUM_FRAMES]
                sampling_rate = utils.get_random_sampling_rate(
                    self.cfg.MULTIGRID.LONG_CYCLE_SAMPLING_RATE,
                    self.cfg.DATA.SAMPLING_RATE,
                )
                sampling_rate = [sampling_rate]
                if len(num_frames) < num_decode:
                    num_frames.extend(
                        [num_frames[-1] for i in range(num_decode - len(num_frames))]
                    )
                    # base case where keys have same frame-rate as query
                    sampling_rate.extend(
                        [sampling_rate[-1] for i in range(num_decode - len(sampling_rate))]
                    )
                elif len(num_frames) > num_decode:
                    num_frames = num_frames[:num_decode]
                    sampling_rate = sampling_rate[:num_decode]

                if self.mode in ["train"]:
                    assert len(min_scale) == len(max_scale) == len(crop_size) == num_decode

                target_fps = self.cfg.DATA.TARGET_FPS
                if self.cfg.DATA.TRAIN_JITTER_FPS > 0.0 and self.mode in ["train"]:
                    target_fps += random.uniform(0.0, self.cfg.DATA.TRAIN_JITTER_FPS)
                
                ################################### DECODING #################################################

                try:# Decode video. Meta info is used to perform selective decoding.
                    frames, time_idx, tdiff = decoder.decode(
                        video_container,
                        sampling_rate,
                        num_frames,
                        temporal_sample_index,
                        self.cfg.TEST.NUM_ENSEMBLE_VIEWS,
                        video_meta=(
                            self._video_meta[index] if len(self._video_meta) < 5e6 else {}
                        ),  # do not cache on huge datasets
                        target_fps=target_fps,
                        backend=self.cfg.DATA.DECODING_BACKEND,
                        use_offset=self.cfg.DATA.USE_OFFSET_SAMPLING,
                        max_spatial_scale=(
                            min_scale[0] if all(x == min_scale[0] for x in min_scale) else 0
                        ),  # if self.mode in ["test"] else 0,
                        time_diff_prob=self.p_convert_dt if self.mode in ["train"] else 0.0,
                        temporally_rnd_clips=True,
                        min_delta=self.cfg.CONTRASTIVE.DELTA_CLIPS_MIN,
                        max_delta=self.cfg.CONTRASTIVE.DELTA_CLIPS_MAX,
                    )
                except:
                    #index = random.randint(0, len(path_) - 1)
                    index = random.randint(0, len(path_) - 1)
                    print("couldnt decode the video", path_[index])
                    break
                    #import pdb;pdb.set_trace()
                frames_decoded = frames
                time_idx_decoded = time_idx

                # If decoding failed (wrong format, video is too short, and etc),
                # select another video.
                if frames_decoded is None or None in frames_decoded:
                    print("cant decode the video",path_[index])
                    break

                ################################### DECODING #################################################

                #import pdb;pdb.set_trace()   
                gaze = self.gaze[index]
                ego = self.ego[index]
                
                num_aug = (
                    self.cfg.DATA.TRAIN_CROP_NUM_SPATIAL * self.cfg.AUG.NUM_SAMPLE
                    if self.mode in ["train"]
                    else 1
                )
                num_out = num_aug * num_decode
                f_out, time_idx_out = [None] * num_out, [None] * num_out
                idx = -1
                label = self._labels[index]

                for i in range(num_decode):
                    for _ in range(num_aug):
                        idx += 1
                        f_out[idx] = frames_decoded[i].clone()
                        time_idx_out[idx] = time_idx_decoded[i, :]

                        f_out[idx] = f_out[idx].float()
                        f_out[idx] = f_out[idx] / 255.0

                        if self.mode in ["train"] and self.cfg.DATA.SSL_COLOR_JITTER:
                            f_out[idx] = transform.color_jitter_video_ssl(
                                f_out[idx],
                                bri_con_sat=self.cfg.DATA.SSL_COLOR_BRI_CON_SAT,
                                hue=self.cfg.DATA.SSL_COLOR_HUE,
                                p_convert_gray=self.p_convert_gray,
                                moco_v2_aug=self.cfg.DATA.SSL_MOCOV2_AUG,
                                gaussan_sigma_min=self.cfg.DATA.SSL_BLUR_SIGMA_MIN,
                                gaussan_sigma_max=self.cfg.DATA.SSL_BLUR_SIGMA_MAX,
                            )

                        if self.aug and self.cfg.AUG.AA_TYPE:
                            aug_transform = create_random_augment(
                                input_size=(f_out[idx].size(1), f_out[idx].size(2)),
                                auto_augment=self.cfg.AUG.AA_TYPE,
                                interpolation=self.cfg.AUG.INTERPOLATION,
                            )
                            # T H W C -> T C H W.
                            f_out[idx] = f_out[idx].permute(0, 3, 1, 2)
                            list_img = self._frame_to_list_img(f_out[idx])
                            list_img = aug_transform(list_img)
                            f_out[idx] = self._list_img_to_frames(list_img)
                            f_out[idx] = f_out[idx].permute(0, 2, 3, 1)

                        # Perform color normalization.
                        f_out[idx] = utils.tensor_normalize(
                            f_out[idx], self.cfg.DATA.MEAN, self.cfg.DATA.STD
                        )

                        # T H W C -> C T H W.
                        f_out[idx] = f_out[idx].permute(3, 0, 1, 2)

                        scl, asp = (
                            self.cfg.DATA.TRAIN_JITTER_SCALES_RELATIVE,
                            self.cfg.DATA.TRAIN_JITTER_ASPECT_RELATIVE,
                        )
                        relative_scales = (
                            None if (self.mode not in ["train"] or len(scl) == 0) else scl
                        )
                        relative_aspect = (
                            None if (self.mode not in ["train"] or len(asp) == 0) else asp
                        )
                        f_out[idx] = utils.spatial_sampling(
                            f_out[idx],
                            spatial_idx=spatial_sample_index,
                            min_scale=min_scale[i],
                            max_scale=max_scale[i],
                            crop_size=crop_size[i],
                            #random_horizontal_flip=self.cfg.DATA.RANDOM_FLIP,
                            random_horizontal_flip=False,
                            inverse_uniform_sampling=self.cfg.DATA.INV_UNIFORM_SAMPLE,
                            aspect_ratio=relative_aspect,
                            scale=relative_scales,
                            motion_shift=(
                                self.cfg.DATA.TRAIN_JITTER_MOTION_SHIFT
                                if self.mode in ["train"]
                                else False
                            ),
                        )

                        if self.rand_erase:
                            erase_transform = RandomErasing(
                                self.cfg.AUG.RE_PROB,
                                mode=self.cfg.AUG.RE_MODE,
                                max_count=self.cfg.AUG.RE_COUNT,
                                num_splits=self.cfg.AUG.RE_COUNT,
                                device="cpu",
                            )
                            f_out[idx] = erase_transform(
                                f_out[idx].permute(1, 0, 2, 3)
                            ).permute(1, 0, 2, 3)

                        f_out[idx] = utils.pack_pathway_output(self.cfg, f_out[idx])
                        if self.cfg.AUG.GEN_MASK_LOADER:
                            mask = self._gen_mask()
                            f_out[idx] = f_out[idx] + [torch.Tensor(), mask]
                frames = f_out[0] if num_out == 1 else f_out
                time_idx = np.array(time_idx_out)
                if (
                    num_aug * num_decode > 1
                    and not self.cfg.MODEL.MODEL_NAME == "ContrastiveModel"
                ):
                    label = [label] * num_aug * num_decode
                    gaze = [gaze] * num_aug * num_decode
                    ego = [ego] * num_aug * num_decode
                    index = [index] * num_aug * num_decode
                if self.cfg.DATA.DUMMY_LOAD:
                    if self.dummy_output is None:
                        self.dummy_output = (frames, label, index, time_idx, {})
                #import pdb;pdb.set_trace()
                FRAMES.append(frames)
            if len(FRAMES) == 6:
               # import pdb;pdb.set_trace()
                return FRAMES, gaze, torch.FloatTensor(ego), label, index, time_idx, {}

            else:
                index = random.randint(0, len(path_) - 1)
                print("Something is fishy")

    def _gen_mask(self):
        if self.cfg.AUG.MASK_TUBE:
            num_masking_patches = round(
                np.prod(self.cfg.AUG.MASK_WINDOW_SIZE) * self.cfg.AUG.MASK_RATIO
            )
            min_mask = num_masking_patches // 5
            masked_position_generator = MaskingGenerator(
                mask_window_size=self.cfg.AUG.MASK_WINDOW_SIZE,
                num_masking_patches=num_masking_patches,
                max_num_patches=None,
                min_num_patches=min_mask,
            )
            mask = masked_position_generator()
            mask = np.tile(mask, (8, 1, 1))
        elif self.cfg.AUG.MASK_FRAMES:
            mask = np.zeros(shape=self.cfg.AUG.MASK_WINDOW_SIZE, dtype=int)
            n_mask = round(self.cfg.AUG.MASK_WINDOW_SIZE[0] * self.cfg.AUG.MASK_RATIO)
            mask_t_ind = random.sample(
                range(0, self.cfg.AUG.MASK_WINDOW_SIZE[0]), n_mask
            )
            mask[mask_t_ind, :, :] += 1
        else:
            num_masking_patches = round(
                np.prod(self.cfg.AUG.MASK_WINDOW_SIZE) * self.cfg.AUG.MASK_RATIO
            )
            max_mask = np.prod(self.cfg.AUG.MASK_WINDOW_SIZE[1:])
            min_mask = max_mask // 5
            masked_position_generator = MaskingGenerator3D(
                mask_window_size=self.cfg.AUG.MASK_WINDOW_SIZE,
                num_masking_patches=num_masking_patches,
                max_num_patches=max_mask,
                min_num_patches=min_mask,
            )
            mask = masked_position_generator()
        return mask

    def _frame_to_list_img(self, frames):
        img_list = [transforms.ToPILImage()(frames[i]) for i in range(frames.size(0))]
        return img_list

    def _list_img_to_frames(self, img_list):
        img_list = [transforms.ToTensor()(img) for img in img_list]
        return torch.stack(img_list)

    def __len__(self):
        """
        Returns:
            (int): the number of videos in the dataset.
        """
        return self.num_videos

    @property
    def num_videos(self):
        """
        Returns:
            (int): the number of videos in the dataset.
        """
        return len(self._path_to_videos_1)

After loading, I used MViTv2_S model to train them. I have 250 videos in training set and 50 in test set. During training, the loss converges, but the confusion matrix says it always predicts class 0.

Please help me navigate through this issue, thanks !!