def main():
with timer('load data'):
df = pd.read_csv(TEST_PATH)
df[['ID', 'Image', 'Diagnosis']] = df['ID'].str.split('_', expand=True)
df = df[['Image', 'Diagnosis', 'Label']]
df.drop_duplicates(inplace=True)
df = df.pivot(index='Image', columns='Diagnosis',
values='Label').reset_index()
df['Image'] = 'ID_' + df['Image']
df = df[["Image"]]
ids = df["Image"].values
gc.collect()
with timer('preprocessing'):
test_augmentation = Compose([
CenterCrop(512 - 50, 512 - 50, p=1.0),
Resize(img_size, img_size, p=1)
])
test_dataset = RSNADatasetTest(df,
img_size,
IMAGE_PATH,
id_colname=ID_COLUMNS,
transforms=test_augmentation,
black_crop=False,
subdural_window=True,
conc_type="concat_all",
conc_type2="concat_prepost",
n_tta=N_TTA)
test_loader = DataLoader(test_dataset,
batch_size=batch_size,
shuffle=False,
num_workers=16,
pin_memory=True)
del df, test_dataset
gc.collect()
with timer('create model'):
model = CnnModel(num_classes=N_CLASSES,
encoder="se_resnext50_32x4d",
pretrained="imagenet",
pool_type="avg")
model.load_state_dict(torch.load(model_path))
model.to(device)
model = torch.nn.DataParallel(model)
with timer('predict'):
pred = predict(model, test_loader, device, n_tta=N_TTA)
pred = np.clip(pred, 1e-6, 1 - 1e-6)
with timer('sub'):
sub = pd.DataFrame(pred, columns=TARGET_COLUMNS)
sub["ID"] = ids
sub = sub.set_index("ID")
sub = sub.unstack().reset_index()
sub["ID"] = sub["ID"] + "_" + sub["level_0"]
sub = sub.rename(columns={0: "Label"})
sub = sub.drop("level_0", axis=1)
LOGGER.info(sub.head())
sub.to_csv("../output/{}_sub_st2.csv".format(EXP_ID), index=False)
def main():
with timer('load data'):
df = pd.read_csv(TRAIN_PATH)[:10]
df = df[df.Image != "ID_6431af929"].reset_index(drop=True)
df.loc[df.pre_SOPInstanceUID == "ID_6431af929",
"pre1_SOPInstanceUID"] = df.loc[df.pre_SOPInstanceUID ==
"ID_6431af929", "Image"]
df.loc[df.post_SOPInstanceUID == "ID_6431af929",
"post1_SOPInstanceUID"] = df.loc[df.post_SOPInstanceUID ==
"ID_6431af929", "Image"]
df = df[["Image", "pre1_SOPInstanceUID", "post1_SOPInstanceUID"]]
ids = df["Image"].values
gc.collect()
with timer('preprocessing'):
test_augmentation = Compose([
CenterCrop(512 - 50, 512 - 50, p=1.0),
Resize(img_size, img_size, p=1)
])
test_dataset = RSNADatasetTest(df,
img_size,
IMAGE_PATH,
id_colname=ID_COLUMNS,
transforms=test_augmentation,
black_crop=False,
three_window=True,
rescaling=False,
n_tta=N_TTA)
test_loader = DataLoader(test_dataset,
batch_size=batch_size,
shuffle=False,
num_workers=16,
pin_memory=True)
del df, test_dataset
gc.collect()
with timer('create model'):
model = CnnModel(num_classes=N_CLASSES,
encoder="se_resnext50_32x4d",
pretrained="imagenet",
pool_type="avg")
model.load_state_dict(torch.load(model_path))
model.to(device)
model = torch.nn.DataParallel(model)
with timer('predict'):
pred = predict(model, test_loader, device, n_tta=N_TTA)
pred = np.clip(pred, 1e-6, 1 - 1e-6)
with timer('sub'):
sub = pd.DataFrame(pred, columns=TARGET_COLUMNS)
sub["ID"] = ids
sub = sub.set_index("ID")
sub = sub.unstack().reset_index()
sub["ID"] = sub["ID"] + "_" + sub["level_0"]
sub = sub.rename(columns={0: "Label"})
sub = sub.drop("level_0", axis=1)
LOGGER.info(sub.head())
sub.to_csv("../output/{}_train.csv".format(EXP_ID), index=False)
def main():
with timer('load data'):
df = pd.read_csv(TRAIN_PATH)
df = df[df.Image != "ID_6431af929"].reset_index(drop=True)
df.loc[df.pre_SOPInstanceUID=="ID_6431af929", "pre1_SOPInstanceUID"] = df.loc[
df.pre_SOPInstanceUID=="ID_6431af929", "Image"]
df.loc[df.post_SOPInstanceUID == "ID_6431af929", "post1_SOPInstanceUID"] = df.loc[
df.post_SOPInstanceUID == "ID_6431af929", "Image"]
df.loc[df.prepre_SOPInstanceUID == "ID_6431af929", "pre2_SOPInstanceUID"] = df.loc[
df.prepre_SOPInstanceUID == "ID_6431af929", "pre1_SOPInstanceUID"]
df.loc[df.postpost_SOPInstanceUID == "ID_6431af929", "post2_SOPInstanceUID"] = df.loc[
df.postpost_SOPInstanceUID == "ID_6431af929", "post1_SOPInstanceUID"]
y = df[TARGET_COLUMNS].values
df = df[["Image", "pre1_SOPInstanceUID", "post1_SOPInstanceUID", "pre2_SOPInstanceUID", "post2_SOPInstanceUID"]]
gc.collect()
with timer('preprocessing'):
train_augmentation = Compose([
CenterCrop(512 - 50, 512 - 50, p=1.0),
HorizontalFlip(p=0.5),
OneOf([
ElasticTransform(p=0.5, alpha=120, sigma=120 * 0.05, alpha_affine=120 * 0.03),
GridDistortion(p=0.5),
OpticalDistortion(p=1, distort_limit=2, shift_limit=0.5)
], p=0.5),
Rotate(limit=30, border_mode=0, p=0.7),
Resize(img_size, img_size, p=1)
])
train_dataset = RSNADataset(df, y, img_size, IMAGE_PATH, id_colname=ID_COLUMNS,
transforms=train_augmentation, black_crop=False, subdural_window=True, user_window=2)
train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True, num_workers=8, pin_memory=True)
del df, train_dataset
gc.collect()
with timer('create model'):
model = CnnModel(num_classes=N_CLASSES, encoder="se_resnext50_32x4d", pretrained="imagenet", pool_type="avg")
if model_path is not None:
model.load_state_dict(torch.load(model_path))
model.to(device)
criterion = torch.nn.BCEWithLogitsLoss(weight=torch.FloatTensor([2, 1, 1, 1, 1, 1]).cuda())
optimizer = torch.optim.Adam(model.parameters(), lr=1e-4, eps=1e-4)
model = torch.nn.DataParallel(model)
with timer('train'):
for epoch in range(1, epochs + 1):
if epoch == 5:
for param_group in optimizer.param_groups:
param_group['lr'] = param_group['lr'] * 0.1
seed_torch(SEED + epoch)
LOGGER.info("Starting {} epoch...".format(epoch))
tr_loss = train_one_epoch(model, train_loader, criterion, optimizer, device)
LOGGER.info('Mean train loss: {}'.format(round(tr_loss, 5)))
torch.save(model.module.state_dict(), 'models/{}_ep{}.pth'.format(EXP_ID, epoch))
def main():
with timer('load data'):
df = pd.read_csv(TEST_PATH)
df["post_SOPInstanceUID"] = df["post_SOPInstanceUID"].fillna(
df["SOPInstanceUID"])
df["pre_SOPInstanceUID"] = df["pre_SOPInstanceUID"].fillna(
df["SOPInstanceUID"])
df = df[["Image", "pre_SOPInstanceUID", "post_SOPInstanceUID"]]
ids = df["Image"].values
pre_ids = df["pre_SOPInstanceUID"].values
pos_ids = df["post_SOPInstanceUID"].values
gc.collect()
with timer('preprocessing'):
test_augmentation = Compose([
CenterCrop(512 - 50, 512 - 50, p=1.0),
Resize(img_size, img_size, p=1)
])
test_dataset = RSNADatasetTest(df,
img_size,
IMAGE_PATH,
id_colname=ID_COLUMNS,
transforms=test_augmentation,
black_crop=False,
subdural_window=True,
n_tta=N_TTA)
test_loader = DataLoader(test_dataset,
batch_size=batch_size,
shuffle=False,
num_workers=16,
pin_memory=True)
del df, test_dataset
gc.collect()
with timer('create model'):
model = CnnModel(num_classes=N_CLASSES,
encoder="se_resnext50_32x4d",
pretrained="imagenet",
pool_type="avg")
model.load_state_dict(torch.load(model_path))
model.to(device)
model = torch.nn.DataParallel(model)
with timer('predict'):
pred = predict(model, test_loader, device, n_tta=N_TTA)
pred = np.clip(pred, 1e-6, 1 - 1e-6)
with timer('sub'):
sub = pd.DataFrame(pred, columns=TARGET_COLUMNS)
sub["ID"] = ids
sub["PRE_ID"] = pre_ids
sub["POST_ID"] = pos_ids
sub = postprocess_multitarget(sub)
LOGGER.info(sub.head())
sub.to_csv("../output/{}_sub_st2.csv".format(EXP_ID), index=False)
def main():
with timer('load data'):
path = glob.glob("../input_ext/*/*/*/*.dcm")
df = pd.DataFrame({"Image": path})
df = df[["Image"]]
ids = df["Image"].values
gc.collect()
with timer('preprocessing'):
test_augmentation = Compose([
CenterCrop(512 - 50, 512 - 50, p=1.0),
Resize(img_size, img_size, p=1)
])
test_dataset = RSNADatasetTest(df,
img_size,
IMAGE_PATH,
id_colname=ID_COLUMNS,
transforms=test_augmentation,
black_crop=False,
subdural_window=True,
n_tta=N_TTA,
img_type="",
external=True)
test_loader = DataLoader(test_dataset,
batch_size=batch_size,
shuffle=False,
num_workers=0,
pin_memory=True)
del df, test_dataset
gc.collect()
with timer('create model'):
model = CnnModel(num_classes=N_CLASSES,
encoder="se_resnext50_32x4d",
pretrained="imagenet",
pool_type="avg")
model.load_state_dict(torch.load(model_path))
model.to(device)
model = torch.nn.DataParallel(model)
with timer('predict'):
pred, is_dicoms = predict_external(model,
test_loader,
device,
n_tta=N_TTA)
pred = np.clip(pred, 1e-6, 1 - 1e-6)
with timer('sub'):
sub = pd.DataFrame(pred, columns=TARGET_COLUMNS)
sub["is_dicom"] = is_dicoms.reshape(-1)
sub["Image"] = ids.reshape(-1)
LOGGER.info(sub.head())
sub.to_csv("../input_ext/{}_externalv2.csv".format(EXP_ID),
index=False)
def main():
with timer('load data'):
df = pd.read_csv(TRAIN_PATH)
df["loc_x"] = df["loc_x"] / 100
df["loc_y"] = df["loc_y"] / 100
y = df[TARGET_COLUMNS].values
df = df[[ID_COLUMNS]]
gc.collect()
with timer("split data"):
folds = StratifiedKFold(n_splits=5, shuffle=True, random_state=0).split(df, y)
for n_fold, (train_index, val_index) in enumerate(folds):
train_df = df.loc[train_index]
val_df = df.loc[val_index]
y_train = y[train_index]
y_val = y[val_index]
if n_fold == fold_id:
break
with timer('preprocessing'):
train_augmentation = Compose([
HorizontalFlip(p=0.5),
OneOf([
ElasticTransform(p=0.5, alpha=120, sigma=120 * 0.05, alpha_affine=120 * 0.03),
GridDistortion(p=0.5),
OpticalDistortion(p=1, distort_limit=2, shift_limit=0.5)
], p=0.5),
RandomBrightnessContrast(p=0.5),
ShiftScaleRotate(rotate_limit=20, p=0.5),
Resize(img_size, img_size, p=1)
])
val_augmentation = Compose([
Resize(img_size, img_size, p=1)
])
train_dataset = KDDataset(train_df, y_train, img_size, IMAGE_PATH, id_colname=ID_COLUMNS,
transforms=train_augmentation)
train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True, num_workers=2, pin_memory=True)
val_dataset = KDDataset(val_df, y_val, img_size, IMAGE_PATH, id_colname=ID_COLUMNS,
transforms=val_augmentation)
val_loader = DataLoader(val_dataset, batch_size=batch_size, shuffle=False, num_workers=2, pin_memory=True)
del df, train_dataset, val_dataset
gc.collect()
with timer('create model'):
model = CnnModel(num_classes=N_CLASSES, encoder="se_resnext50_32x4d",
pretrained="../input/pytorch-pretrained-models/se_resnext50_32x4d-a260b3a4.pth",
pool_type="avg")
if model_path is not None:
model.load_state_dict(torch.load(model_path))
model.to(device)
criterion = torch.nn.BCEWithLogitsLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=1e-4, eps=1e-4)
# model, optimizer = amp.initialize(model, optimizer, opt_level="O1", verbosity=0)
with timer('train'):
best_score = 0
for epoch in range(1, epochs + 1):
seed_torch(SEED + epoch)
if epoch == epochs - 3:
for param_group in optimizer.param_groups:
param_group['lr'] = param_group['lr'] * 0.1
LOGGER.info("Starting {} epoch...".format(epoch))
tr_loss = train_one_epoch(model, train_loader, criterion, optimizer, device, N_CLASSES)
LOGGER.info('Mean train loss: {}'.format(round(tr_loss, 5)))
y_pred, target, val_loss = validate(model, val_loader, criterion, device, N_CLASSES)
score = roc_auc_score(target, y_pred)
LOGGER.info('Mean val loss: {}'.format(round(val_loss, 5)))
LOGGER.info('val score: {}'.format(round(score, 5)))
if score > best_score:
best_score = score
np.save("y_pred.npy", y_pred)
torch.save(model.state_dict(), save_path)
np.save("target.npy", target)
with timer('predict'):
test_df = pd.read_csv(TEST_PATH)
test_ids = test_df["id"].values
test_augmentation = Compose([
Resize(img_size, img_size, p=1)
])
test_dataset = KDDatasetTest(test_df, img_size, TEST_IMAGE_PATH, id_colname=ID_COLUMNS,
transforms=test_augmentation, n_tta=2)
test_loader = DataLoader(test_dataset, batch_size=batch_size, shuffle=False, num_workers=2, pin_memory=True)
model.load_state_dict(torch.load(save_path))
pred = predict(model, test_loader, device, N_CLASSES, n_tta=2)
print(pred.shape)
results = pd.DataFrame({"id": test_ids,
"is_star": pred.reshape(-1)})
results.to_csv("results.csv", index=False)