dataset_recognition.py

# -*- coding: utf-8 -*-

""" Implements a dataset object which allows to read representations from LMDB datasets."""
import lmdb
import numpy as np
import pandas as pd
from torch.utils import data
from tqdm import tqdm


class SequenceDataset(data.Dataset):
    def __init__(self, path_to_lmdb,
                 path_to_csv,
                 label_type='action',
                 img_tmpl="frame_{:010d}.jpg",
                 challenge=False,
                 fps=30,
                 args=None):
        """
            Inputs:
                path_to_lmdb: path to the folder containing the LMDB dataset
                path_to_csv: path to training/validation csv
                label_type: which label to return (verb, noun, or action)
                img_tmpl: image template to load the features
                challenge: allows to load csvs containing only time-stamp for the challenge
                fps: framerate
        """

        # read the csv file
        if challenge:
            self.annotations = pd.read_csv(path_to_csv, header=None, names=['video', 'start', 'end'])
        else:
            self.annotations = pd.read_csv(path_to_csv, header=None,
                                           names=['video', 'start', 'end', 'verb', 'noun', 'action'])

        self.challenge = challenge
        self.path_to_lmdb = path_to_lmdb
        self.fps = fps
        self.label_type = label_type
        self.img_tmpl = img_tmpl

        self.recent_sec1 = args.recent_sec1
        self.recent_sec2 = args.recent_sec2
        self.recent_sec3 = args.recent_sec3
        self.recent_sec4 = args.recent_sec4
        self.recent_dim = args.recent_dim

        self.spanning_sec = args.spanning_sec
        self.span_dim1 = args.span_dim1
        self.span_dim2 = args.span_dim2
        self.span_dim3 = args.span_dim3

        self.feat_dim = args.video_feat_dim

        self.debug_on = args.debug_on

        # initialize some lists
        self.ids = []  # action ids
        self.discarded_ids = []  # list of ids discarded (e.g., if there were
        # no enough frames before the beginning of the action
        self.discarded_labels = [] # list of labels discarded (e.g., if there 
        # were no enough frames before the beginning of the action
        self.recent_frames = []  # recent past
        self.spanning_frames = []  # spanning past
        self.labels = []  # labels of each action

        # populate them
        self.__populate_lists()

        # if a list to datasets has been provided, load all of them
        if isinstance(self.path_to_lmdb, list):
            self.env = [lmdb.open(l_m, readonly=True, lock=False) for l_m in self.path_to_lmdb]
        else:
            # otherwise, just load the single LMDB dataset
            self.env = lmdb.open(self.path_to_lmdb, readonly=True, lock=False)

    def __populate_lists(self):
        count_debug = 0
        """ Samples a sequence for each action and populates the lists. """
        for _, a in tqdm(self.annotations.iterrows(), 'Populating Dataset', total=len(self.annotations)):
            count_debug += 1
            if self.debug_on:
                if count_debug > 10:
                    break

            # sample frames before the beginning of the action
            recent_f, spanning_f = self.__get_snippet_features(a.start, a.end, a.video)

            # check if there were enough frames before the beginning of the action
            # if the smaller frame is at least 1, the sequence is valid
            if spanning_f is not None and recent_f is not None:
                self.spanning_frames.append(spanning_f)
                self.recent_frames.append(recent_f)
                self.ids.append(a.name)

                # handle whether a list of labels is required (e.g., [verb, noun]), rather than a single action
                if isinstance(self.label_type, list):
                    if self.challenge:  # if sampling for the challenge, there are no labels, just add -1
                        self.labels.append(-1)
                    else:
                        # otherwise get the required labels
                        self.labels.append(a[self.label_type].values.astype(int))
                else:  # single label version
                    if self.challenge:
                        self.labels.append(-1)
                    else:
                        self.labels.append(a[self.label_type])
            else:
                # if the sequence is invalid, do nothing, but add the id to the discarded_ids list
                self.discarded_ids.append(a.name)
                if isinstance(self.label_type, list):
                    if self.challenge: # if sampling for the challenge, there are no labels, just add -1
                        self.discarded_labels.append(-1)
                    else:
                        # otherwise get the required labels
                        self.discarded_labels.append(a[self.label_type].values.astype(int))
                else: #single label version
                    if self.challenge:
                        self.discarded_labels.append(-1)
                    else:
                        self.discarded_labels.append(a[self.label_type])

    def __get_snippet_features(self, point_start, point_end, video):

        # Spanning snippets
        start_spanning = max(point_start - (self.spanning_sec * self.fps), 0)
        end_spanning = point_end + (self.spanning_sec * self.fps)

        select_spanning_frames1 = np.linspace(start_spanning, end_spanning, self.span_dim1 + 1, dtype=int)
        select_spanning_frames2 = np.linspace(start_spanning, end_spanning, self.span_dim2 + 1, dtype=int)
        select_spanning_frames3 = np.linspace(start_spanning, end_spanning, self.span_dim3 + 1, dtype=int)

        spanning_past = [self.__get_frames_from_indices(video, select_spanning_frames1),
                         self.__get_frames_from_indices(video, select_spanning_frames2),
                         self.__get_frames_from_indices(video, select_spanning_frames3)]

        # Recent snippets
        start_recent1 = int(max(point_start - (self.recent_sec1 * self.fps), 0))
        end_recent1 = int(point_end + (self.recent_sec1 * self.fps))
        start_recent2 = int(max(point_start - (self.recent_sec2 * self.fps), 0))
        end_recent2 = int(point_end + (self.recent_sec2 * self.fps))
        start_recent3 = int(max(point_start - (self.recent_sec3 * self.fps), 0))
        end_recent3 = int(point_end + (self.recent_sec3 * self.fps))
        start_recent4 = int(max(point_start - (self.recent_sec4 * self.fps), 0))
        end_recent4 = int(point_end + (self.recent_sec4 * self.fps))

        select_recent_frames1 = np.linspace(start_recent1, end_recent1, self.recent_dim + 1, dtype=int)
        select_recent_frames2 = np.linspace(start_recent2, end_recent2, self.recent_dim + 1, dtype=int)
        select_recent_frames3 = np.linspace(start_recent3, end_recent3, self.recent_dim + 1, dtype=int)
        select_recent_frames4 = np.linspace(start_recent4, end_recent4, self.recent_dim + 1, dtype=int)

        recent_past = [self.__get_frames_from_indices(video, select_recent_frames1),
                       self.__get_frames_from_indices(video, select_recent_frames2),
                       self.__get_frames_from_indices(video, select_recent_frames3),
                       self.__get_frames_from_indices(video, select_recent_frames4)]

        return recent_past, spanning_past

    def __get_frames_from_indices(self, video, indices):
        list_data = []
        for kkl in range(len(indices) - 1):
            cur_start = np.floor(indices[kkl]).astype('int')
            cur_end = np.floor(indices[kkl + 1]).astype('int')
            list_frames = list(range(cur_start, cur_end + 1))
            list_data.append(self.__get_frames(list_frames, video))
        return list_data

    def __get_frames(self, frames, video):
        """ format file names using the image template """
        frames = np.array(list(map(lambda x: video + "_" + self.img_tmpl.format(x), frames)))
        return frames

    def __len__(self):
        return len(self.ids)

    def __getitem__(self, index):
        """ sample a given sequence """

        # get spanning and recent frames
        spanning_frames = self.spanning_frames[index]
        recent_frames = self.recent_frames[index]

        # return a dictionary containing the id of the current sequence
        # this is useful to produce the jsons for the challenge
        out = {'id': self.ids[index]}

        # read representations for spanning and recent frames
        out['recent_features'], out['spanning_features'] = read_data(recent_frames, spanning_frames, self.env,
                                                                     self.feat_dim)

        # get the label of the current sequence
        label = self.labels[index]
        out['label'] = label

        return out


def read_representations(recent_frames, spanning_frames, env, feat_dim):
    """ Reads a set of representations, given their frame names and an LMDB environment."""

    recent_features1 = []
    recent_features2 = []
    recent_features3 = []
    recent_features4 = []
    spanning_features1 = []
    spanning_features2 = []
    spanning_features3 = []
    for e in env:
        spanning_features1.append(get_max_pooled_features(e, spanning_frames[0], feat_dim))
        spanning_features2.append(get_max_pooled_features(e, spanning_frames[1], feat_dim))
        spanning_features3.append(get_max_pooled_features(e, spanning_frames[2], feat_dim))

        recent_features1.append(get_max_pooled_features(e, recent_frames[0], feat_dim))
        recent_features2.append(get_max_pooled_features(e, recent_frames[1], feat_dim))
        recent_features3.append(get_max_pooled_features(e, recent_frames[2], feat_dim))
        recent_features4.append(get_max_pooled_features(e, recent_frames[3], feat_dim))

    spanning_features1 = np.concatenate(spanning_features1, axis=-1)
    spanning_features2 = np.concatenate(spanning_features2, axis=-1)
    spanning_features3 = np.concatenate(spanning_features3, axis=-1)

    recent_features1 = np.concatenate(recent_features1, axis=-1)
    recent_features2 = np.concatenate(recent_features2, axis=-1)
    recent_features3 = np.concatenate(recent_features3, axis=-1)
    recent_features4 = np.concatenate(recent_features4, axis=-1)

    spanning_snippet_features = [spanning_features1, spanning_features2, spanning_features3]
    recent_snippet_features = [recent_features1, recent_features2, recent_features3, recent_features4]

    return recent_snippet_features, spanning_snippet_features


def get_max_pooled_features(env, frame_names, feat_dim):
    list_features = []
    missing_features = []

    for kkl in range(len(frame_names)):
        with env.begin() as e:
            pool_list = []
            for name in frame_names[kkl]:
                dd = e.get(name.strip().encode('utf-8'))
                if dd is None:
                    continue
                data_curr = np.frombuffer(dd, 'float32')  # convert to numpy array
                feat_dim = data_curr.shape[0]
                pool_list.append(data_curr)

            if len(pool_list) == 0:  # Missing frames indices
                missing_features.append(kkl)
                list_features.append(np.zeros(feat_dim, dtype='float32'))
            else:
                max_pool = np.max(np.array(pool_list), 0)
                list_features.append(max_pool.squeeze())

    for index in missing_features[::-1]: 
        list_features[index] = list_features[0]
        
    list_features = np.stack(list_features)
    return list_features


def read_data(recent_frames, spanning_frames, env, feat_dim):
    """A wrapper form read_representations to handle loading from more environments.
    This is used for multimodal data loading (e.g., RGB + Flow)"""

    # if env is a list
    if isinstance(env, list):
        # read the representations from all environments
        return read_representations(recent_frames, spanning_frames, env, feat_dim)
    else:
        # otherwise, just read the representations
        env = [env]
        return read_representations(recent_frames, spanning_frames, env, feat_dim)