generate_train_split.py

import Loader
import numpy as np
import os

from sklearn.model_selection import train_test_split

TEST_FRAC = 0.2
RANDOM_SEED = 42

loader = Loader.GridLoader()
# we'll leave out speaker 34 for testing
video_paths = loader.load_video_paths(
    verbose=True, fetch_all_paths=False,
    excluded_speakers=[34]
)

# speaker-sentence pairs where lip extraction had problems
bad_lip_pairs = {
    (1, 'pbio7a'), (1, 'bwwuzn'), (7, 'bbir1s'), (1, 'prii9a'),
    (2, 'pbwxzs'), (7, 'bbir2p'), (7, 'lbad1s'), (17, 'lbib9a'),
    (1, 'bbizzn'), (7, 'lgws5s'), (3, 'lgbz9s'), (1, 'lrarzn'),
    (3, 'pbiu6n'), (1, 'pbwx1s'), (3, 'bgit2n'), (3, 'lbij5s'),
    (3, 'bramzn'), (1, 'lgbf8n'), (7, 'lrwe6p'), (1, 'brwg8p'),
    (1, 'sbbh4p'), (3, 'swiu2n'), (7, 'lwak8p'), (7, 'sbatzn'),
    (2, 'pwbd6s'), (7, 'pwii6n'), (9, 'bwaf6n'), (3, 'pgwy7s'),
    (7, 'lwws1a'), (1, 'sran9s'), (7, 'bgam9s'), (3, 'bgbn9a'),
    (3, 'prwq3a'), (7, 'sgio2p'), (4, 'lwiy3n'), (3, 'lbij7a'),
    (1, 'brwa4p'), (2, 'pbib7p'), (3, 'lrbr3s')
}

new_video_paths = []
for video_path in video_paths:
    sentence = os.path.basename(video_path)
    sentence, _ = os.path.splitext(sentence)
    speaker_name = os.path.basename(os.path.dirname(video_path))
    speaker_no = int(speaker_name[1:])
    cache_key = (speaker_no, sentence)
    # print(cache_key)

    if cache_key in bad_lip_pairs:
        print('SKIPPING', video_path)
        continue

    new_video_paths.append(video_path)


video_paths = new_video_paths
train_paths, validate_paths, _, _ = train_test_split(
    video_paths, video_paths,
    test_size=TEST_FRAC, random_state=RANDOM_SEED
)


def get_speakers(filepaths):
    return set([os.path.basename(os.path.dirname(x)) for x in filepaths])


train_paths = sorted(train_paths)
validate_paths = sorted(validate_paths)

print(f'ALL_SPEAKERS {get_speakers(video_paths)}')
print(f'TRAIN_PATHS: {len(train_paths)}')
print(f'TRAIN_SPEAKERS: {get_speakers(train_paths)}')
print(f'VALIDATE_PATHS: {len(validate_paths)}')
print(f'VALIDATE_SPEAKERS: {get_speakers(validate_paths)}')

open('data/unseen_train.txt', 'w').write('\n'.join(train_paths))
open('data/unseen_val.txt', 'w').write('\n'.join(validate_paths))