def create_file_list(
source_dir_root,
target_dir_root,
output_file,
compressions,
data_types,
samples_per_video,
min_sequence_length,
):
try:
# if file exists, we don't have to create it again
file_list = FileList.load(output_file)
except FileNotFoundError:
file_list = _create_file_list(
compressions,
data_types,
min_sequence_length,
output_file,
samples_per_video,
source_dir_root,
)
if target_dir_root:
file_list.copy_to(Path(target_dir_root))
file_list.save(output_file)
for split in [TRAIN_NAME, VAL_NAME, TEST_NAME]:
data_set = FileList.get_dataset_form_file(output_file, split)
logger.info(f"{split}-data-set: {data_set}")
def visualize():
net = SimpleAEL1Pretrained().eval()
file_list = FileList.load(
"/data/ssd1/file_lists/avspeech/resampled_and_avspeech_100_samples_consolidated.json"
)
dataset = file_list.get_dataset("train", sequence_length=1)
for class_id in range(6):
sample_images = _get_images_from_class(dataset,
class_id=class_id,
num_images=10)
sample_images = torch.stack(sample_images)
output_images = net(sample_images)[RECON_X]
diff = sample_images - output_images
x_12 = sample_images.view(-1, 3, 112, 112)
x_12_recon = output_images.contiguous().view(-1, 3, 112, 112)
diff_recon = diff.contiguous().view(-1, 3, 112, 112)
x_12 = torch.cat((x_12, x_12_recon, diff_recon),
dim=2) # this needs to stack the images differently
datapoints = make_grid(x_12, nrow=10, range=(-1, 1), normalize=True)
d = datapoints.detach().permute(1, 2, 0).numpy() * 255
d = cv2.cvtColor(d, cv2.COLOR_BGR2RGB)
cv2.imwrite(f"images_classes_{class_id}.png", d)
def change_root(file_list_path, new_root):
f = FileList.load(file_list_path)
print(f"Old root: {f.root}")
f.root = str(new_root)
print(f"New root: {f.root}")
save_filelist = (click.prompt("Is this root correct and save file list?",
type=str,
default="y").lower() == "y")
if save_filelist:
f.save(file_list_path)
print("Succesfully saved.")
else:
print("Aborted.")
def transform_dataset(size, source_file_list, target_dataset_folder, target_file_list):
f = FileList.load(source_file_list)
old_root = Path(f.root)
new_root = Path(target_dataset_folder)
new_root.mkdir(exist_ok=False)
transform = Compose(resized_crop(size))
for split in [TRAIN_NAME, VAL_NAME, TEST_NAME]:
Parallel(n_jobs=mp.cpu_count())(
delayed(lambda sample: resize(old_root, sample[0], new_root, transform))(
sample_
)
for sample_ in tqdm(f.samples[split])
)
f.root = str(new_root)
f.save(target_file_list)
def _create_file_list(
min_sequence_length: int,
output_file: str,
samples_per_video: int,
source_dir_root: str,
):
file_list = FileList(
root=source_dir_root,
classes=[AVSPEECH_NAME],
min_sequence_length=min_sequence_length,
)
source_dir_root = Path(source_dir_root)
# split between train and val
videos = sorted(source_dir_root.iterdir())
train = videos[:int(len(videos) * 0.9)]
val = videos[int(len(videos) * 0.9):]
for split, split_name in [(train, TRAIN_NAME), (val, VAL_NAME)]:
for video_folder in sorted(split):
images = sorted(video_folder.glob("*.png"))
filtered_images_idx = []
# find all frames that have at least min_sequence_length-1 preceeding
# frames
sequence_start = img_name_to_int(images[0])
last_idx = sequence_start
for list_idx, image in enumerate(images):
image_idx = img_name_to_int(image)
if last_idx + 1 != image_idx:
sequence_start = image_idx
elif image_idx - sequence_start >= min_sequence_length - 1:
filtered_images_idx.append(list_idx)
last_idx = image_idx
selected_frames = select_frames(len(filtered_images_idx),
samples_per_video)
sampled_images_idx = np.asarray(
filtered_images_idx)[selected_frames]
file_list.add_data_points(
path_list=images,
target_label=AVSPEECH_NAME,
split=split_name,
sampled_images_idx=sampled_images_idx,
)
file_list.save(output_file)
logger.info(f"{output_file} created.")
return file_list
def run_inference_for_video(audio_mode, folder, model, trainer):
image_file_list = folder / "image_file_list.json"
audio_file_list = folder / "audio_file_list.json"
f = FileList.load(str(image_file_list))
test_video_dataset = f.get_dataset(
"test",
image_transforms=model.resize_transform,
tensor_transforms=model.tensor_augmentation_transforms,
sequence_length=model.model.sequence_length,
audio_file_list=SimpleFileList.load(audio_file_list),
audio_mode=audio_mode,
)
loader = get_fixed_dataloader(
test_video_dataset,
model.hparams["batch_size"],
sampler=SequentialSampler,
num_workers=model.hparams["n_cpu"],
worker_init_fn=lambda worker_id: np.random.seed(worker_id),
)
model.test_dataloader = lambda: loader # this seems not to work at all
print(folder, len(loader))
trainer.test(model)
def change_class_order():
file_list = (
"/mnt/ssd1/sebastian/file_lists/c40/"
"youtube_Deepfakes_Face2Face_FaceSwap_NeuralTextures_c40_face_images_tracked_100_100_8.json"
)
# file_list = "/data/ssd1/file_lists/c40/tracked_resampled_faces.json"
f_new = FileList.load(file_list)
f_new.class_to_idx = {
"Deepfakes": 0,
"Face2Face": 1,
"FaceSwap": 2,
"NeuralTextures": 3,
"youtube": 4,
}
f_new.classes = [
"Deepfakes", "Face2Face", "FaceSwap", "NeuralTextures", "youtube"
]
for split in f_new.samples.values():
for item in split:
item[1] = (item[1] + 4) % 5
def do_error_stuff():
r = Five00BatchesResnet18().cuda().eval()
f = FileList.load("/data/ssd1/file_lists/imagenet/ssd_raw.json")
val_data = f.get_dataset(
"val", image_transforms=[transforms.Resize(256), transforms.CenterCrop(224)]
)
val_data_loader = DataLoader(
val_data, batch_size=256, shuffle=True, num_workers=2, pin_memory=True
)
with torch.no_grad():
acc = 0
num_items = 0
t_bar = tqdm(val_data_loader)
for batch in t_bar:
images, targets = batch
targets, images = targets.cuda(), images.cuda()
predictions = r.forward(images)
top1_error = r.calculate_accuracy(predictions, targets)
acc += top1_error
num_items += 1
t_bar.set_description(f"Curr acc: {acc/num_items}")
acc /= num_items
print(acc)
def _compute_statistics_of_path(path: str,
model,
batch_size,
dims,
cuda,
split="val"):
if path.endswith(".npz"):
f = np.load(path)
m, s = f["mu"][:], f["sigma"][:]
f.close()
else:
file_list = FileList.load(path)
root = Path(file_list.root)
idx = file_list.samples_idx[split]
files = np.array(
list(map(lambda x: root / x[0], file_list.samples[split])))
files = files[idx]
m, s = calculate_activation_statistics(files, model, batch_size, dims,
cuda)
return m, s
from forgery_detection.data.file_lists import FileList
resampled_file_list = FileList.load(
"/mnt/ssd1/sebastian/file_lists/c40/tracked_resampled_faces.json")
resampled_file_list.samples["train"] = resampled_file_list.samples["val"]
resampled_file_list.samples["val"] = resampled_file_list.samples["test"]
resampled_file_list.samples_idx["train"] = resampled_file_list.samples_idx[
"val"]
resampled_file_list.samples_idx["val"] = resampled_file_list.samples_idx[
"test"]
resampled_file_list.save(
"/mnt/ssd1/sebastian/file_lists/c40/tracked_resampled_faces_val_test_as_train_val.json"
)
import json
from pathlib import Path
import numpy as np
from forgery_detection.data.face_forensics.splits import TEST_NAME
from forgery_detection.data.face_forensics.splits import TRAIN_NAME
from forgery_detection.data.face_forensics.splits import VAL_NAME
from forgery_detection.data.file_lists import FileList
root_dir = Path("/data/hdd/dfdc/extracted_images")
with open(root_dir.parent / "all_metadata.json", "r") as f:
all_meta_data = json.load(f)
f = FileList(str(root_dir), classes=["FAKE", "REAL"], min_sequence_length=1)
train_data_numbers = list(range(5, 50))
val_data_numbers = list(range(5))
for train_data_number in train_data_numbers:
block = root_dir / f"extracted_images_{train_data_number}"
if block.exists():
for label in block.iterdir():
images = list(label.glob("*/*.png"))
f.add_data_points(images, label.name, "train",
np.arange(0, len(images)))
for val_data_number in val_data_numbers:
block = root_dir / f"extracted_images_{val_data_number}"
if block.exists():
def __init__(self, kwargs: Union[dict, Namespace]):
super(Supervised, self).__init__()
self.hparams = DictHolder(kwargs)
# load data-sets
self.file_list = FileList.load(self.hparams["data_dir"])
self.model: LightningModel = self.MODEL_DICT[self.hparams["model"]](
num_classes=len(self.file_list.classes))
if len(self.file_list.classes) != self.model.num_classes:
logger.error(
f"Classes of model ({self.model.num_classes}) != classes of dataset"
f" ({len(self.file_list.classes)})")
self.sampling_probs = self.hparams["sampling_probs"]
if self.sampling_probs:
self.sampling_probs = np.array(self.sampling_probs.split(" "),
dtype=float)
if self.sampling_probs is not None and len(
self.file_list.classes) != len(self.sampling_probs):
raise ValueError(
f"Classes of dataset ({len(self.file_list.classes)}) != classes of "
f"sampling probs ({len(self.sampling_probs)})!")
self.resize_transform = self._get_transforms(
self.hparams["resize_transforms"])
image_augmentation_transforms = self._get_transforms(
self.hparams["image_augmentation_transforms"])
self.tensor_augmentation_transforms = self._get_transforms(
self.hparams["tensor_augmentation_transforms"])
if self.hparams["audio_file"]:
self.audio_file_list = SimpleFileList.load(
self.hparams["audio_file"])
else:
self.audio_file_list = None
self.audio_mode = AudioMode[self.hparams["audio_mode"]]
self.train_data = self.file_list.get_dataset(
TRAIN_NAME,
image_transforms=self.resize_transform +
image_augmentation_transforms,
tensor_transforms=self.tensor_augmentation_transforms,
sequence_length=self.model.sequence_length,
audio_file_list=self.audio_file_list,
audio_mode=self.audio_mode,
should_align_faces=self.hparams["crop_faces"],
)
self.val_data = self.file_list.get_dataset(
VAL_NAME,
image_transforms=self.resize_transform,
tensor_transforms=self.tensor_augmentation_transforms,
sequence_length=self.model.sequence_length,
audio_file_list=self.audio_file_list,
audio_mode=self.audio_mode,
should_align_faces=self.hparams["crop_faces"],
)
# handle empty test_data better
self.test_data = self.file_list.get_dataset(
TEST_NAME,
image_transforms=self.resize_transform,
tensor_transforms=self.tensor_augmentation_transforms,
sequence_length=self.model.sequence_length,
audio_file_list=self.audio_file_list,
audio_mode=self.audio_mode,
should_align_faces=self.hparams["crop_faces"],
)
self.hparams.add_dataset_size(len(self.train_data), TRAIN_NAME)
self.hparams.add_dataset_size(len(self.val_data), VAL_NAME)
self.hparams.add_dataset_size(len(self.test_data), TEST_NAME)
if self.hparams["dont_balance_data"]:
self.sampler_cls = RandomSampler
else:
self.sampler_cls = BalancedSampler
self.system_mode = self.hparams.pop("mode")
if self.system_mode is SystemMode.TRAIN:
self.hparams.add_nb_trainable_params(self.model)
if self.hparams["class_weights"]:
labels, weights = calculate_class_weights(self.val_data)
self.hparams.add_class_weights(labels, weights)
self.class_weights = torch.tensor(weights, dtype=torch.float)
else:
self.class_weights = None
elif self.system_mode is SystemMode.TEST:
self.class_weights = None
elif self.system_mode is SystemMode.BENCHMARK:
pass
# hparams logging
self.decay = 0.95
self.acc = -1
self.loss = -1
logger.warning(f"{self.train_data.class_to_idx}")
self._optimizer = None
import os
from pathlib import Path
import numpy as np
from forgery_detection.data.face_forensics.splits import TEST_NAME
from forgery_detection.data.face_forensics.splits import TRAIN_NAME
from forgery_detection.data.face_forensics.splits import VAL_NAME
from forgery_detection.data.file_lists import FileList
file_list_a = FileList.load(
"/data/ssd1/file_lists/avspeech/100k_100_samples_consolidated.json"
)
file_list_b = FileList.load(
"/data/ssd1/file_lists/avspeech/avspeech_moria_20k_100_samples_consolidated.json"
)
a_root = Path(file_list_a.root)
b_root = Path(file_list_b.root)
common_path = os.path.commonpath([b_root, a_root])
a_relative_to_root = os.path.relpath(a_root, common_path)
b_relative_to_root = os.path.relpath(b_root, common_path)
print(common_path, a_relative_to_root, b_relative_to_root)
file_list_a.root = common_path
for split in file_list_a.samples.values():
for item in split:
import os
from pathlib import Path
import numpy as np
from forgery_detection.data.face_forensics.splits import TEST_NAME
from forgery_detection.data.face_forensics.splits import TRAIN_NAME
from forgery_detection.data.face_forensics.splits import VAL_NAME
from forgery_detection.data.file_lists import FileList
resampled_file_list = FileList.load(
"/data/ssd1/file_lists/c40/tracked_resampled_faces_224.json"
)
imagenet = FileList.load("/data/ssd1/file_lists/imagenet/ssd_256_center_crop_224.json")
resampled_root = Path(resampled_file_list.root)
imagenet_root = Path(imagenet.root)
common_path = os.path.commonpath([imagenet_root, resampled_root])
resampled_relative_to_root = os.path.relpath(resampled_root, common_path)
imagenet_relative_to_root = os.path.relpath(imagenet_root, common_path)
print("path stuff", common_path, resampled_relative_to_root, imagenet_relative_to_root)
resampled_file_list.class_to_idx = {
**imagenet.class_to_idx,
**dict(
map(
lambda x: (x[0], x[1] + len(imagenet.class_to_idx)),
resampled_file_list.class_to_idx.items(),
)
),
# flake8: noqa
#%%
from pathlib import Path
from forgery_detection.data.file_lists import FileList
from forgery_detection.data.file_lists import SimpleFileList
from forgery_detection.data.utils import resized_crop
f = FileList.load(
"/home/sebastian/data/file_lists/avspeech_crop_tests/aligned_faces.json")
a = f.get_dataset(
"test",
audio_file_list=SimpleFileList.load(
"/home/sebastian/data/file_lists/avspeech_crop_tests/mfcc_features.json"
),
sequence_length=8,
image_transforms=resized_crop(112),
)
path = Path("/home/ondyari/avspeech_test_formats/cropped_images_aligned_faces")
for p in sorted(path.iterdir()):
if p.is_dir():
try:
a.audio_file_list.files[p.name]
except KeyError:
print(f"{p.name} not in audio")
#%%
from forgery_detection.data.loading import get_fixed_dataloader
from forgery_detection.data.loading import BalancedSampler
from forgery_detection.data.file_lists import FileList
from forgery_detection.data.utils import resized_crop
import matplotlib
from cv2 import cv2
from torchvision.utils import make_grid
from forgery_detection.data.face_forensics.splits import TRAIN_NAME
from forgery_detection.data.file_lists import FileList
from forgery_detection.data.loading import get_fixed_dataloader
matplotlib.use("agg")
avspeech_filelist = FileList.load(
"/data/ssd1/file_lists/avspeech/avspeech_seti_20k_100_samples.json"
)
dataloader = get_fixed_dataloader(
avspeech_filelist.get_dataset(TRAIN_NAME, sequence_length=8), 4, num_workers=1
)
for batch in dataloader:
# generate image out of batch
x, target = batch
x = x.view(-1, 3, 112, 112)
im = make_grid(x, normalize=True, range=(-1, 1)).permute(1, 2, 0)
im = im.numpy() * 255
# torch converts images to bgr colour space
im = cv2.cvtColor(im, cv2.COLOR_BGR2RGB)
# save it for now
cv2.imwrite("/home/sebastian/loaded_iamge.png", im)
# plt.imshow(im)
# plt.show()
def _create_file_list(
compressions,
data_types,
min_sequence_length,
output_file,
samples_per_video,
source_dir_root,
):
file_list = FileList(
root=source_dir_root,
classes=FaceForensicsDataStructure.METHODS,
min_sequence_length=min_sequence_length,
)
# use faceforensicsdatastructure to iterate elegantly over the correct
# image folders
source_dir_data_structure = FaceForensicsDataStructure(
source_dir_root, compressions=compressions, data_types=data_types)
_min_sequence_length = _get_min_sequence_length(source_dir_data_structure)
if _min_sequence_length < samples_per_video:
logger.warning(
f"There is a sequence that is sequence that has less frames "
f"then you would like to sample: "
f"{_min_sequence_length}<{samples_per_video}")
for split, split_name in [(TRAIN, TRAIN_NAME), (VAL, VAL_NAME),
(TEST, TEST_NAME)]:
for source_sub_dir, target in zip(
source_dir_data_structure.get_subdirs(), file_list.classes):
for video_folder in sorted(source_sub_dir.iterdir()):
if video_folder.name.split("_")[0] in split:
images = sorted(video_folder.glob("*.png"))
filtered_images_idx = []
# find all frames that have at least min_sequence_length-1 preceeding
# frames
sequence_start = img_name_to_int(images[0])
last_idx = sequence_start
for list_idx, image in enumerate(images):
image_idx = img_name_to_int(image)
if last_idx + 1 != image_idx:
sequence_start = image_idx
elif image_idx - sequence_start >= min_sequence_length - 1:
filtered_images_idx.append(list_idx)
last_idx = image_idx
selected_frames = select_frames(len(filtered_images_idx),
samples_per_video)
sampled_images_idx = np.asarray(
filtered_images_idx)[selected_frames]
file_list.add_data_points(
path_list=images,
target_label=target,
split=split_name,
sampled_images_idx=sampled_images_idx,
)
file_list.save(output_file)
logger.info(f"{output_file} created.")
return file_list
import os
from pathlib import Path
import numpy as np
from forgery_detection.data.face_forensics.splits import TEST_NAME
from forgery_detection.data.face_forensics.splits import TRAIN_NAME
from forgery_detection.data.face_forensics.splits import VAL_NAME
from forgery_detection.data.file_lists import FileList
resampled_file_list = FileList.load(
"/data/ssd1/file_lists/c40/tracked_resampled_faces.json")
celeba = FileList.load(
"/data/ssd1/file_lists/avspeech/100k_100_samples_consolidated.json")
resampled_file_list.root = "/mnt/ssd2/sebastian/set/tracked_resampled_faces_112/"
resampled_root = Path(resampled_file_list.root)
celeba_root = Path(celeba.root)
common_path = os.path.commonpath([celeba_root, resampled_root])
resampled_relative_to_root = os.path.relpath(resampled_root, common_path)
celeba_relative_to_root = os.path.relpath(celeba_root, common_path)
print(common_path, resampled_relative_to_root, celeba_relative_to_root)
# change class idx values for resampled file list
# make youtube one value higher
resampled_file_list.class_to_idx["avspeech"] = 5
resampled_file_list.classes.append("avspeech")
# resampled_file_list.class_to_idx["youtube"] = 5
# resampled_file_list.classes.append("youtube")
# flake8: noqa
#%%
import torch
from forgery_detection.data.file_lists import FileList
from forgery_detection.data.utils import resized_crop
f = FileList.load(
"/mnt/ssd1/sebastian/file_lists/c40/"
"youtube_Deepfakes_Face2Face_FaceSwap_NeuralTextures_c40_face_images_tracked_100_100_8.json"
)
unnormalize = lambda x: x * torch.tensor([0.229, 0.224, 0.225]).unsqueeze(
1).unsqueeze(1) + torch.tensor([0.485, 0.456, 0.406]).unsqueeze(
1).unsqueeze(1)
d = f.get_dataset(
"train",
image_transforms=resized_crop(229),
tensor_transforms=[unnormalize],
should_align_faces=True,
)
class_idx_to_label = """0: unlabeled
1: person
2: bicycle
3: car
4: motorcycle
5: airplane
6: bus
7: train
8: truck
9: boat
from pathlib import Path
import numpy as np
from forgery_detection.data.file_lists import FileList
root_dir = Path("/mnt/ssd2/sebastian/set/avspeech_moria_112")
f = FileList(root_dir, ["avspeech"], 8)
#
# images = list((root_dir / "train19k").glob("*.jpg"))
# f.add_data_points(images, "celeba", "train", np.arange(0, len(images)))
#
# images = list((root_dir / "val").glob("*.jpg"))
# f.add_data_points(images, "celeba", "train", np.arange(0, len(images)))
#
# images = list((root_dir / "test").glob("*.jpg"))
# f.add_data_points(images, "celeba", "val", np.arange(0, len(images)))
#
# images = list((root_dir / "test").glob("*.jpg"))
# f.add_data_points(images, "celeba", "test", np.arange(0, len(images)))
videos = sorted(root_dir.iterdir())
train = videos[: int(len(videos) * 0.9)]
val = videos[int(len(videos) * 0.9) :]
samples_per_video = 100
for label in train:
images = sorted(label.glob("*.png"))
import os
from pathlib import Path
from forgery_detection.data.face_forensics.splits import TRAIN_NAME
from forgery_detection.data.file_lists import FileList
resampled_file_list = FileList.load(
"/data/ssd1/file_lists/c40/tracked_resampled_faces.json")
detection_file_list = FileList.load(
"/data/ssd1/file_lists/c40/detection_challenge_112.json")
resampled_root = Path(resampled_file_list.root)
detection_root = Path(detection_file_list.root)
common_path = os.path.commonpath([detection_root, resampled_root])
resampled_relative_to_root = os.path.relpath(resampled_root, common_path)
detection_relative_to_root = os.path.relpath(detection_root, common_path)
print(common_path, resampled_relative_to_root, detection_relative_to_root)
# change class idx values for resampled file list
# make youtube one value higher
resampled_file_list.class_to_idx = {"DeepFakeDetection": 0, "youtube": 1}
resampled_file_list.classes = ["DeepFakeDetection", "youtube"]
resampled_file_list.root = common_path
for split in resampled_file_list.samples.values():
for item in split:
if item[1] == 4:
item[1] = 1
else:
item[1] = 0
"NeuralTextures": 3,
"youtube": 4,
}
f_new.classes = [
"Deepfakes", "Face2Face", "FaceSwap", "NeuralTextures", "youtube"
]
for split in f_new.samples.values():
for item in split:
item[1] = (item[1] + 4) % 5
if __name__ == "__main__":
# change_class_order()
# 1 / 0
file_list = "/data/ssd1/file_lists/c40/trf_100_100_full_size_relative_bb_8_sl.json"
# file_list = "/data/ssd1/file_lists/c40/tracked_resampled_faces.json"
f = FileList.load(file_list)
# first_val_path = f.samples["val"][f.samples_idx["val"][0]]
val_loader = get_fixed_dataloader(
f.get_dataset("val", sequence_length=1, should_align_faces=True),
batch_size=1,
num_workers=1,
)
iter = val_loader.__iter__()
for i in tqdm.trange(1000):
next(iter)
# flake8: noqa
#%%
from pathlib import Path
from forgery_detection.data.file_lists import FileList
from forgery_detection.data.file_lists import SimpleFileList
# trained with /mnt/ssd1/sebastian/file_lists/c40/tracked_resampled_faces_yt_only_112_8_sequence_length.json
# and /data/hdd/audio_features/mfcc_features.npy -> /data/hdd/audio_features/mfcc_features_file_list.json
audio_features = SimpleFileList.load(
"/mnt/raid5/sebastian/audio_features/mfcc_features_file_list.json")
f = FileList.load(
"/mnt/ssd1/sebastian/file_lists/c40/tracked_resampled_faces_yt_only_112_8_sequence_length.json"
)
#%%
vids = sorted(set(
map(lambda x: x[0].split("/")[-2],
f.samples["train"])))[:10] # take first 10 videos in train set
images = {}
def _path_to_frame_nr(path):
return int(path.split("/")[-1].split(".")[0])
for vid in vids:
for path, _ in f.samples["train"]:
if vid in path.split("/")[-2]:
frame_nr = _path_to_frame_nr(path)
# flake8: noqa
#%%
from forgery_detection.data.file_lists import FileList
from forgery_detection.data.utils import resized_crop
f = FileList.load(
"/home/sebastian/data/file_lists/c40/trf_-1_-1_full_size_relative_bb_8_sl.json"
)
a = f.get_dataset("test",
audio_file_list=None,
sequence_length=8,
image_transforms=resized_crop(112))
#%%
f_2 = FileList.load(
"/home/sebastian/data/file_lists/c40/tracked_resampled_faces_all_112_8_sequence_length.json"
)
a_2 = f_2.get_dataset("test",
audio_file_list=None,
sequence_length=8,
image_transforms=[])
#%%
zero = a[((100000, 100000), None)][0]
zero_2 = a_2[((100001, 100001), None)][0]
#%%
print((zero == zero_2).all())