def test_train():
dataset_df = load_dataset_df('/store/kaggle/tgs')
train_df, val_df = train_test_split(dataset_df)
output_dir = '/store/tmp'
train(
model_path=f"{output_dir}/model.pt",
train_dataset=TgsSaltDataset(train_df),
val_dataset=TgsSaltDataset(val_df),
epochs=1000,
batch_size=32,
feature_size=32,
patience=5,
base_size=5,
log_dir=f'{config["TENSORBORAD_LOG_DIR"]}/test',
)
def test_dataset():
dataset_df = load_dataset_df('/store/kaggle/tgs')
dataset = TgsSaltDataset(dataset_df)
scores = []
for i in range(12):
sample = dataset[i]
score = iou(sample['mask'].numpy(), sample['mask'].numpy())
scores.append(score)
assert np.round(np.mean(scores), decimals=3) == 1.0
def test_flip():
writer = SummaryWriter(f'{config["TENSORBORAD_LOG_DIR"]}/test')
dataset_df = load_dataset_df('/store/kaggle/tgs')
dataset = TgsSaltDataset(dataset_df)
writer.add_image(
f"flip",
vutils.make_grid(
pipe(range(8), map(lambda x: dataset[12]),
map(lambda x: [x['image'], x['mask']]), concat, list)),
)
def test_predict(csv_fn, is_train):
dataset_df = load_dataset_df('/store/kaggle/tgs', csv_fn).sample(10)
dataset = TgsSaltDataset(dataset_df, is_train=is_train)
model = UNet()
model_paths = ['/store/tmp/model.pt']
torch.save(model, model_paths[0])
predicted_df = predict(model_paths=model_paths,
log_dir=f'{config["TENSORBORAD_LOG_DIR"]}/test',
dataset=dataset,
log_interval=1)
print(predicted_df)
assert len(predicted_df) == len(dataset_df)
def test_pool():
dataset_df = load_dataset_df(
'/store/kaggle/tgs',
'train.csv',
)
dataset = TgsSaltDataset(dataset_df, has_y=True)
dataloader = DataLoader(dataset, batch_size=8, shuffle=True)
sample = pipe(dataloader, first)['mask']
interpolated = F.max_pool2d(sample, kernel_size=101)
writer = SummaryWriter(f'{config["TENSORBORAD_LOG_DIR"]}/test')
dataset_df = load_dataset_df('/store/kaggle/tgs')
writer.add_image(f"mask", vutils.make_grid([*sample]), 0)
writer.add_image(f"pooled", vutils.make_grid([*interpolated]), 0)
def test_add_noise():
dataset_df = load_dataset_df(
'/store/kaggle/tgs',
'train.csv',
)
dataset = TgsSaltDataset(dataset_df)
dataloader = DataLoader(dataset, batch_size=2, shuffle=True)
sample = pipe(dataloader, first)['image']
noised = add_noise(sample)
writer = SummaryWriter(f'{config["TENSORBORAD_LOG_DIR"]}/test')
dataset_df = load_dataset_df('/store/kaggle/tgs')
writer.add_image(
f"add_noise",
vutils.make_grid([*sample, *noised]),
)
def test_erase():
dataset_df = load_dataset_df(
'/store/kaggle/tgs',
'train.csv',
)
dataset = TgsSaltDataset(dataset_df, has_y=True)
dataloader = DataLoader(dataset, batch_size=8, shuffle=True)
sample = pipe(dataloader, first)['image']
random_erase = RandomErasing()
noised = add_noise(sample, erase_num=5, erase_p=1)
writer = SummaryWriter(f'{config["TENSORBORAD_LOG_DIR"]}/test')
dataset_df = load_dataset_df('/store/kaggle/tgs')
writer.add_image(
f"random_erase",
vutils.make_grid([*sample, *noised]),
)
def test_dataset():
dataset_df = load_dataset_df('/store/kaggle/tgs')
dataset = TgsSaltDataset(dataset_df)
assert len(dataset) == 4000
def test_dense_crf():
dataset_df = load_dataset_df('/store/kaggle/tgs').dropna().head(5)
dataset = TgsSaltDataset(dataset_df)
sample = dataset[0]
print(sample['image'].shape)
aa = dense_crf(sample['image'].view(-1, 101, 101), sample['mask'].view(-1, 101, 101))