def test(args):
# Setup Dataloader
data_json = json.load(open('config.json'))
data_path = data_json[args.dataset]['data_path']
t_loader = SaltLoader(data_path, split="test")
test_df=t_loader.test_df
test_loader = data.DataLoader(t_loader, batch_size=args.batch_size, num_workers=8)
# load Model
if args.arch=='unet':
model = Unet(start_fm=16)
else:
model=Unet_upsample(start_fm=16)
model_path = data_json[args.model]['model_path']
model.load_state_dict(torch.load(model_path)['model_state'])
model.cuda()
total = sum([param.nelement() for param in model.parameters()])
print('Number of params: %.2fM' % (total / 1e6))
#test
pred_list=[]
for images in test_loader:
images = Variable(images.cuda())
y_preds = model(images)
y_preds_shaped = y_preds.reshape(-1, args.img_size_target, args.img_size_target)
for idx in range(args.batch_size):
y_pred = y_preds_shaped[idx]
pred = torch.sigmoid(y_pred)
pred = pred.cpu().data.numpy()
pred_ori = resize(pred, (args.img_size_ori, args.img_size_ori), mode='constant', preserve_range=True)
pred_list.append(pred_ori)
#submit the test image predictions.
threshold_best=args.threshold
pred_dict = {idx: RLenc(np.round(pred_list[i] > threshold_best)) for i, idx in
enumerate(tqdm_notebook(test_df.index.values))}
sub = pd.DataFrame.from_dict(pred_dict, orient='index')
sub.index.names = ['id']
sub.columns = ['rle_mask']
sub.to_csv('./results/{}_submission.csv'.format(args.model))
print("The submission.csv saved in ./results")
def test(args):
# Setup Data
data_json = json.load(open('config.json'))
x = Variable(torch.randn(32, 1, 128, 128))
x = x.cuda()
# load Model
if args.arch == 'unet':
model = Unet(start_fm=16)
else:
model = Unet_upsample(start_fm=16)
model_path = data_json[args.model]['model_path']
model.load_state_dict(torch.load(model_path)['model_state'])
model.cuda()
total = sum([param.nelement() for param in model.parameters()])
print('Number of params: %.2fM' % (total / 1e6))
#visualize
y = model(x)
g = make_dot(y)
g.render('k')
def main():
videos, audios = get_data("v1", 1)
print("Data Loaded")
# if os.path.exists('./Unet.pt'):
# unet = torch.load('./Unet.pt')
# frame_discriminator = torch.load()
unet = Unet(debug=False)
frame_discriminator = FrameDiscriminator()
sequence_discriminator = SequenceDiscriminator()
if cuda:
print('SAHI JA RHA.......')
unet = unet.cuda()
frame_discriminator = frame_discriminator.cuda()
sequence_discriminator = sequence_discriminator.cuda()
# if torch.cuda.device_count() > 1:
# print("Using ", torch.cuda.device_count(), " GPUs!")
# unet = nn.DataParallel(unet)
# frame_discriminator = nn.DataParallel(frame_discriminator)
# sequence_discriminator = nn.DataParallel(sequence_discriminator)
train(audios, videos, unet, frame_discriminator, sequence_discriminator)
def validate(args):
# Setup Dataloader
data_json = json.load(open('config.json'))
data_path = data_json[args.dataset]['data_path']
v_loader = SaltLoader(data_path, split='val')
train_df = v_loader.train_df
val_loader = data.DataLoader(v_loader,
batch_size=args.batch_size,
num_workers=8)
# load Model
if args.arch == 'unet':
model = Unet(start_fm=16)
else:
model = Unet_upsample(start_fm=16)
model_path = data_json[args.model]['model_path']
model.load_state_dict(torch.load(model_path)['model_state'])
model.cuda()
total = sum([param.nelement() for param in model.parameters()])
print('Number of params: %.2fM' % (total / 1e6))
#validate
pred_list = []
for images, masks in val_loader:
images = Variable(images.cuda())
y_preds = model(images)
# print(y_preds.shape)
y_preds_shaped = y_preds.reshape(-1, args.img_size_target,
args.img_size_target)
for idx in range(args.batch_size):
y_pred = y_preds_shaped[idx]
pred = torch.sigmoid(y_pred)
pred = pred.cpu().data.numpy()
pred_ori = resize(pred, (args.img_size_ori, args.img_size_ori),
mode='constant',
preserve_range=True)
pred_list.append(pred_ori)
preds_valid = np.array(pred_list)
y_valid_ori = np.array(
[train_df.loc[idx].masks for idx in v_loader.ids_valid])
#jaccard score
accuracies_best = 0.0
for threshold in np.linspace(0, 1, 11):
ious = []
for y_pred, mask in zip(preds_valid, y_valid_ori):
prediction = (y_pred > threshold).astype(int)
iou = jaccard_similarity_score(mask.flatten(),
prediction.flatten())
ious.append(iou)
accuracies = [
np.mean(ious > iou_threshold)
for iou_threshold in np.linspace(0.5, 0.95, 10)
]
if accuracies_best < np.mean(accuracies):
accuracies_best = np.mean(accuracies)
threshold_best = threshold
print('Threshold: %.1f, Metric: %.3f' %
(threshold, np.mean(accuracies)))
print("jaccard score gets threshold_best=", threshold_best)
#other score way
thresholds = np.linspace(0, 1, 50)
ious = np.array([
iou_metric_batch(y_valid_ori, np.int32(preds_valid > threshold))
for threshold in tqdm_notebook(thresholds)
])
#don't understand
threshold_best_index = np.argmax(ious[9:-10]) + 9
iou_best = ious[threshold_best_index]
threshold_best = thresholds[threshold_best_index]
print("other way gets iou_best=", iou_best, "threshold_best=",
threshold_best)
def train(args):
# Setup Dataloader
data_json = json.load(open('config.json'))
data_path=data_json[args.dataset]['data_path']
t_loader = SaltLoader(data_path, img_size_ori=args.img_size_ori,img_size_target=args.img_size_target)
v_loader = SaltLoader(data_path, split='val',img_size_ori=args.img_size_ori, img_size_target=args.img_size_target)
train_loader = data.DataLoader(t_loader, batch_size=args.batch_size, num_workers=8, shuffle=True)
val_loader = data.DataLoader(v_loader, batch_size=args.batch_size, num_workers=8)
# Setup Model
if args.arch=='unet':
model = Unet(start_fm=16)
else:
model=Unet_upsample(start_fm=16)
print(model)
total = sum([param.nelement() for param in model.parameters()])
print('Number of params: %.2fM' % (total / 1e6))
model.cuda()
# Check if model has custom optimizer / loss
optimizer = torch.optim.Adam(model.parameters(), lr=args.l_rate)
loss_fn= nn.BCEWithLogitsLoss()
best_loss=100
mean_train_losses = []
mean_val_losses = []
for epoch in range(args.n_epoch):
train_losses = []
val_losses = []
for images, masks in train_loader:
images = Variable(images.cuda())
masks = Variable(masks.cuda())
outputs = model(images)
loss = loss_fn(outputs, masks)
train_losses.append(loss.data)
optimizer.zero_grad()
loss.backward()
optimizer.step()
for images, masks in val_loader:
images = Variable(images.cuda())
masks = Variable(masks.cuda())
outputs = model(images)
loss = loss_fn(outputs, masks)
val_losses.append(loss.data)
mean_train_losses.append(np.mean(train_losses))
mean_val_losses.append(np.mean(val_losses))
if np.mean(val_losses) < best_loss:
best_loss = np.mean(val_losses)
state = {'epoch': epoch + 1,
'model_state': model.state_dict(),
'optimizer_state': optimizer.state_dict(), }
torch.save(state, "./saved_models/{}_{}_best_model.pkl".format(args.arch, args.dataset))
# Print Loss
print('Epoch: {}. Train Loss: {}. Val Loss: {}'.format(epoch + 1, np.mean(train_losses), np.mean(val_losses)))
state = {'model_state': model.state_dict(),
'optimizer_state': optimizer.state_dict(), }
torch.save(state, "./saved_models/{}_{}_final_model.pkl".format(args.arch, args.dataset))
print("saved two models in ./saved_models")