def predict_on_eval_using_fold(fold, train_eval_data):
_, eval_indices = indices_for_fold(fold, train_eval_data)
eval_dataset = TrainEvalDataset(train_eval_data.iloc[eval_indices],
transform=eval_transform)
eval_data_loader = torch.utils.data.DataLoader(
eval_dataset,
batch_size=config.batch_size,
num_workers=args.workers,
worker_init_fn=worker_init_fn)
model = Model(config.model, NUM_CLASSES)
model = model.to(DEVICE)
model.load_state_dict(
torch.load(
os.path.join(args.experiment_path, 'model_{}.pth'.format(fold))))
model.eval()
with torch.no_grad():
fold_labels = []
fold_logits = []
fold_exps = []
fold_ids = []
for images, feats, exps, labels, ids in tqdm(
eval_data_loader, desc='fold {} evaluation'.format(fold)):
images, feats, labels = images.to(DEVICE), feats.to(
DEVICE), labels.to(DEVICE)
logits = model(images, feats)
fold_labels.append(labels)
fold_logits.append(logits)
fold_exps.extend(exps)
fold_ids.extend(ids)
fold_labels = torch.cat(fold_labels, 0)
fold_logits = torch.cat(fold_logits, 0)
tmp = train_eval_data.iloc[eval_indices].copy()
temp, _, _ = find_temp_global(input=fold_logits,
target=fold_labels,
exps=fold_exps)
classes = assign_classes(probs=(fold_logits *
temp).softmax(1).data.cpu().numpy(),
exps=fold_exps)
print('{:.2f}'.format((tmp['sirna'] == classes).mean()))
tmp['sirna'] = classes
tmp.to_csv(os.path.join(args.experiment_path,
'eval_{}.csv'.format(fold)),
index=False)
return fold_labels, fold_logits, fold_exps, fold_ids
def predict_on_eval_using_fold(fold, train_eval_data):
_, eval_indices = indices_for_fold(fold, train_eval_data)
eval_dataset = TrainEvalDataset(train_eval_data.iloc[eval_indices],
transform=eval_transform)
eval_data_loader = torch.utils.data.DataLoader(
eval_dataset,
batch_size=config.batch_size // 4,
num_workers=args.workers,
worker_init_fn=worker_init_fn)
model = Model(config.model, NUM_CLASSES)
model = model.to(DEVICE)
model.load_state_dict(
torch.load(
os.path.join(args.experiment_path, 'model_{}.pth'.format(fold))))
model.eval()
with torch.no_grad():
fold_labels = []
fold_probs = []
fold_exps = []
fold_ids = []
for images, feats, exps, labels, ids in tqdm(
eval_data_loader, desc='fold {} evaluation'.format(fold)):
images, feats, labels = images.to(DEVICE), feats.to(
DEVICE), labels.to(DEVICE)
b, n, c, h, w = images.size()
assert n == 2 * NUM_TTA
images = images.view(b * n, c, h, w)
feats = feats.view(b, 1, 2).repeat(1, n, 1).view(b * n, 2)
logits = model(images, feats)
logits = logits.view(b, n, NUM_CLASSES)
probs = softmax(logits)
fold_labels.append(labels)
fold_probs.append(probs)
fold_exps.extend(exps)
fold_ids.extend(ids)
fold_labels = torch.cat(fold_labels, 0)
fold_probs = torch.cat(fold_probs, 0)
fold_plates = train_eval_data.iloc[eval_indices]['plate'].values
return fold_labels, fold_probs, fold_exps, fold_plates, fold_ids
def predict_on_eval_using_fold(fold, train_eval_data):
_, eval_indices = indices_for_fold(fold, train_eval_data)
eval_dataset = TrainEvalDataset(train_eval_data.iloc[eval_indices], transform=eval_transform)
eval_data_loader = torch.utils.data.DataLoader(
eval_dataset,
batch_size=config.batch_size,
num_workers=args.workers,
worker_init_fn=worker_init_fn)
model = Model(config.model, NUM_CLASSES)
model = model.to(DEVICE)
model.load_state_dict(torch.load(os.path.join(args.experiment_path, 'model_{}.pth'.format(fold))))
model.eval()
with torch.no_grad():
fold_labels = []
fold_logits = []
fold_exps = []
fold_ids = []
for images, feats, exps, labels, ids in tqdm(eval_data_loader, desc='fold {} evaluation'.format(fold)):
images, feats, labels = images.to(DEVICE), feats.to(DEVICE), labels.to(DEVICE)
b, n, c, h, w = images.size()
images = images.view(b * n, c, h, w)
feats = feats.view(b, 1, 2).repeat(1, n, 1).view(b * n, 2)
logits = model(images, feats)
logits = logits.view(b, n, NUM_CLASSES)
fold_labels.append(labels)
fold_logits.append(logits)
fold_exps.extend(exps)
fold_ids.extend(ids)
fold_labels = torch.cat(fold_labels, 0)
fold_logits = torch.cat(fold_logits, 0)
fold_plates = train_eval_data.iloc[eval_indices]['plate'].values
temp, _, _ = find_temp_global(input=fold_logits, target=fold_labels, exps=fold_exps)
classes = assign_classes(probs=to_prob(fold_logits, temp).data.cpu().numpy(), exps=fold_exps)
fold_logits = refine_scores(
fold_logits, classes, exps=fold_exps, plates=fold_plates, value=float('-inf'))
return fold_labels, fold_logits, fold_exps, fold_ids
def predict_on_test_using_fold(fold, test_data):
test_dataset = TestDataset(test_data, transform=test_transform)
test_data_loader = torch.utils.data.DataLoader(
test_dataset,
batch_size=config.batch_size // 2,
num_workers=args.workers,
worker_init_fn=worker_init_fn)
model = Model(config.model, NUM_CLASSES)
model = model.to(DEVICE)
model.load_state_dict(
torch.load(
os.path.join(args.experiment_path, 'model_{}.pth'.format(fold))))
model.eval()
with torch.no_grad():
fold_logits = []
fold_exps = []
fold_plates = []
fold_ids = []
for images, feats, exps, plates, ids in tqdm(
test_data_loader, desc='fold {} inference'.format(fold)):
images, feats = images.to(DEVICE), feats.to(DEVICE)
b, n, c, h, w = images.size()
images = images.view(b * n, c, h, w)
feats = feats.view(b, 1, 2).repeat(1, n, 1).view(b * n, 2)
logits = model(images, feats)
logits = logits.view(b, n, NUM_CLASSES)
fold_logits.append(logits)
fold_exps.extend(exps)
fold_plates.extend(plates)
fold_ids.extend(ids)
fold_logits = torch.cat(fold_logits, 0)
torch.save((fold_logits, fold_exps, fold_ids),
'./test_{}.pth'.format(fold))
return fold_logits, fold_exps, fold_plates, fold_ids
def compute_features_using_fold(fold, data):
dataset = TestDataset(data, transform=test_transform)
data_loader = torch.utils.data.DataLoader(dataset,
batch_size=config.batch_size //
2,
num_workers=args.workers,
worker_init_fn=worker_init_fn)
model = Model(config.model, NUM_CLASSES, return_features=True)
model = model.to(DEVICE)
model.load_state_dict(
torch.load(
os.path.join(args.experiment_path, 'model_{}.pth'.format(fold))))
model.eval()
with torch.no_grad():
fold_embs = []
fold_exps = []
fold_ids = []
for images, feats, exps, ids in tqdm(
data_loader, desc='fold {} inference'.format(fold)):
images, feats = images.to(DEVICE), feats.to(DEVICE)
b, n, c, h, w = images.size()
images = images.view(b * n, c, h, w)
feats = feats.view(b, 1, 2).repeat(1, n, 1).view(b * n, 2)
_, embds = model(images, feats)
embds = embds.view(b, n, embds.size(1))
fold_embs.append(embds)
fold_exps.extend(exps)
fold_ids.extend(ids)
fold_embs = torch.cat(fold_embs, 0)
return fold_embs, fold_exps, fold_ids
