def composition():
data_dir = pathlib.Path('train_val')
data_anno_raw = pd.read_csv('train_val/keys.csv')
data_anno = pd.DataFrame({
'id':
data_anno_raw['id'].values,
'category': [
TRAFFIC_LABELS_TO_NUM[label]
for label in data_anno_raw['category'].values
]
})
orig = MyDataset(data_dir, data_anno, transform=None)
str_transform = "Composition"
path = f'aug_pics/{str_transform}.png'
data = MyDataset(data_dir, data_anno)
fig, axes = plt.subplots(8, 2, figsize=(4, 16))
for row in axes:
for ax in row:
ax.set_axis_off()
for i in range(8):
tens, _ = orig.__getitem__(i)
img = tensor2img(tens)
tens, _ = data.__getitem__(i)
img_aug = tensor2img(tens)
axes[i, 0].imshow(img)
axes[i, 1].imshow(img_aug)
fig.suptitle(str_transform)
fig.savefig(path)
def one_by_one():
data_dir = 'train_val/pic'
data_anno = pd.read_csv('train_val/keys.csv')
orig = MyDataset(data_dir, data_anno)
for transform in tqdm(alb_transforms):
str_transform = str(transform)
str_transform = str_transform[:str_transform.find('(')]
path = f'aug_pics/{str_transform}.png'
if not pathlib.Path(path).exists():
data = MyDataset(data_dir,
data_anno,
transform=AlbuWrapper(transform))
fig, axes = plt.subplots(4, 2, figsize=(4, 8))
for row in axes:
for ax in row:
ax.set_axis_off()
for i in range(4):
tens, _ = orig.__getitem__(i)
img = tensor2img(tens)
tens, _ = data.__getitem__(i)
img_aug = tensor2img(tens)
axes[i, 0].imshow(img)
axes[i, 1].imshow(img_aug)
fig.suptitle(str_transform)
fig.savefig(path)