def validate(model, criterion, valid_loader, validation_size, batch_size, iter_size):
model.eval()
losses = []
accuracies = []
batches_count = validation_size // batch_size
valid_loader = islice(valid_loader, batches_count)
for i, (inputs, targets) in tqdm.tqdm(enumerate(valid_loader), total=batches_count, desc="validation"):
inputs = variable(inputs, volatile=True)
targets = variable(targets)
targets = long_tensor(targets)
inputs0_chunks = inputs[0].chunk(iter_size)
inputs1_chunks = inputs[1].chunk(iter_size)
targets_chunks = targets.chunk(iter_size)
loss = 0
acc = 0
for input1, input2, target in zip(inputs0_chunks, inputs1_chunks, targets_chunks):
outputs = model(input1, input2)
loss_iter = criterion(outputs, target)
loss_iter /= batch_size
loss += loss_iter.data[0]
acc_iter = accuracy(outputs, target)[0]
acc_iter /= iter_size
acc += acc_iter.data[0]
losses.append(loss)
accuracies.append(acc)
valid_loss = np.mean(losses)
valid_acc = np.mean(accuracies)
print('Valid loss: {:.4f}, acc: {:.4f}'.format(valid_loss, valid_acc))
return {'valid_loss': valid_loss, 'valid_acc': valid_acc}
def main(architecture, folds, tta):
test_dataset = InternValidDataset(transform=test_augm())
labels = None
for fold in folds:
model = get_model(num_classes=test_dataset.num_classes,
architecture=architecture)
state = torch.load('../results/{}/best-model_{}.pt'.format(
architecture, fold))
model.load_state_dict(state['model'])
model.eval()
labels = []
with open('../results/{}/{}_valid_prob.csv'.format(architecture, fold),
"w") as f:
for idx in tqdm.tqdm(range(len(test_dataset))):
best_conf = 0
best_pred = None
for rot in range(4):
test_dataset.rot = rot
in1 = []
in2 = []
for _ in range(tta):
x = test_dataset[idx][0]
in1.append(x[0])
in2.append(x[1])
in1 = variable(torch.stack(in1))
in2 = variable(torch.stack(in2))
pred = model(in1, in2).data.cpu().numpy()
pred = np.array([softmax(x) for x in pred])
pred = np.sum(pred, axis=0) / len(pred)
if np.max(pred) > best_conf:
best_conf = np.max(pred)
best_pred = pred
labels.append(test_dataset[idx][1])
probas = ','.join([str(x) for x in best_pred])
f.write('{}\n'.format(probas))
dfs = [
pd.read_csv('../results/{}/{}_valid_prob.csv'.format(architecture, i),
header=None) for i in folds
]
classes = [
'HTC-1-M7', 'LG-Nexus-5x', 'Motorola-Droid-Maxx', 'Motorola-Nexus-6',
'Motorola-X', 'Samsung-Galaxy-Note3', 'Samsung-Galaxy-S4',
'Sony-NEX-7', 'iPhone-4s', 'iPhone-6'
]
for df in dfs:
df.columns = classes
df = dfs[0].copy()
for i in np.arange(1, len(folds)):
df[classes] += dfs[i][classes]
df[classes] /= len(folds)
matched = 0
for i in np.arange(len(test_dataset)):
pred = df[classes].iloc[i].values.argmax()
real = labels[i]
if pred == real:
matched += 1
print('accuracy = {}'.format(matched / len(test_dataset)))
def predict(self, architecture, fold, tta=5, mode='submit', name="sub"):
test_dataset = TestDataset(transform=test_augm())
model = get_model(num_classes=test_dataset.num_classes, architecture=architecture)
state = torch.load('../results/{}/best-model_{}.pt'.format(architecture, fold))
model.load_state_dict(state['model'])
model.eval()
if mode == 'submit':
with open('../results/{}/{}_{}.csv'.format(architecture, name, fold), "w") as f:
f.write("fname,camera\n")
for idx in tqdm.tqdm(range(len(test_dataset))):
images = torch.stack([test_dataset[idx][0] for _ in range(tta)])
images = variable(images)
pred = model(images).data.cpu().numpy()
pred = np.sum(pred, axis=0)
fname = test_dataset[idx][1]
label = np.argmax(pred, 0)
camera_model = test_dataset.inverse_dict[label]
f.write('{},{}\n'.format(fname, camera_model))
else:
def softmax(x):
"""Compute softmax values for each sets of scores in x."""
e_x = np.exp(x - np.max(x))
return e_x / e_x.sum(axis=0)
with open('../results/{}/{}_{}_prob.csv'.format(architecture, name, fold), "w") as f:
for idx in tqdm.tqdm(range(len(test_dataset))):
best_conf = 0
best_pred = None
for rot in range(4):
test_dataset.rot = rot
in1 = []
in2 = []
for _ in range(tta):
x = test_dataset[idx][0]
in1.append(x[0])
in2.append(x[1])
in1 = variable(torch.stack(in1))
in2 = variable(torch.stack(in2))
fname = test_dataset[idx][1]
pred = model(in1, in2).data.cpu().numpy()
pred = np.array([softmax(x) for x in pred])
pred = np.sum(pred, axis=0) / len(pred)
if np.max(pred) > best_conf:
best_conf = np.max(pred)
best_pred = pred
probas = ','.join([str(x) for x in best_pred])
f.write('{},{}\n'.format(fname, probas))
def predict_validation(self, architecture, fold, tta, batch_size):
n_classes = 5270
model = get_model(num_classes=n_classes, architecture=architecture)
state = torch.load(f"../results/{architecture}/best-model_{fold}.pt")
model.load_state_dict(state['model'])
test_augm = valid_augm()
label_map = pd.read_csv("../data/labels_map.csv")
label_map.index = label_map['label_id']
loader = get_valid_loader(fold, batch_size, test_augm)
with open(f"../results/{architecture}/validation_{fold}.csv",
"w") as f:
f.write("_id,category_id\n")
for images, product_ids in tqdm.tqdm(loader):
images = variable(images)
preds = model(images).data.cpu().numpy()
for pred, product_id in zip(preds, product_ids):
label = np.argmax(pred, 0)
cat_id = label_map.ix[label]['category_id']
f.write(f"{product_id},{cat_id}\n")
def predict(self, architecture, fold, tta, batch_size, name="sub"):
print("Start predicting with following params:",
f"architecture = {architecture}", f"fold = {fold}",
f"tta = {tta}")
n_classes = 5270
model = get_model(num_classes=n_classes, architecture=architecture)
state = torch.load(f"../results/{architecture}/best-model_{fold}.pt")
model.load_state_dict(state['model'])
test_augm = valid_augm()
label_map = pd.read_csv("../data/labels_map.csv")
label_map.index = label_map['label_id']
test_dataset = TestDataset(transform=test_augm)
with open(f"../results/{architecture}/{name}_{fold}.csv", "w") as f:
f.write("_id,category_id\n")
for idx in tqdm.tqdm(range(len(test_dataset))):
images = torch.stack(
[test_dataset[idx][0] for i in range(tta)])
images = variable(images)
pred = model(images).data.cpu().numpy()
pred = sum(pred)
product_id = test_dataset[idx][1]
label = np.argmax(pred, 0)
cat_id = label_map.ix[label]['category_id']
f.write(f"{product_id},{cat_id}\n")
def _train(self,
args,
model,
criterion,
*,
train_loader,
valid_loader,
validation_size,
patience=2):
lr = self.lr
n_epochs = args['n_epochs']
optimizer = self._init_optimizer()
self._init_files()
self._init_model()
report_each = 10
valid_losses = []
lr_reset_epoch = self.epoch
batch_size = args['batch_size']
iter_size = args['iter_size']
for epoch in range(self.epoch, n_epochs + 1):
model.train()
random.seed()
tq = tqdm.tqdm(total=(args['epoch_size'] or
len(train_loader) * batch_size))
tq.set_description('Epoch {}, lr {}'.format(epoch, lr))
losses = []
tl = train_loader
epoch_loss = 0
if args['epoch_size']:
tl = islice(tl, args['epoch_size'] // batch_size)
try:
mean_loss = 0
batches_count = 0
for i, (inputs, targets) in enumerate(tl):
batches_count += 1
inputs, targets = variable(inputs), variable(targets)
targets = long_tensor(targets)
inputs_0_chunks = inputs[0].chunk(iter_size)
inputs_1_chunks = inputs[1].chunk(iter_size)
targets_chunks = targets.chunk(iter_size)
optimizer.zero_grad()
iter_loss = 0
for input1, input2, target in zip(inputs_0_chunks, inputs_1_chunks, targets_chunks):
outputs = model(input1, input2)
loss = criterion(outputs, target)
loss /= batch_size
iter_loss += loss.data[0]
loss.backward()
optimizer.step()
self.step += 1
tq.update(batch_size)
epoch_loss += iter_loss
losses.append(iter_loss)
mean_loss = np.mean(losses[-report_each:])
tq.set_postfix(loss='{:.3f}'.format(mean_loss))
if i and i % report_each == 0:
self._write_event(loss=mean_loss)
epoch_loss /= batches_count
self._write_event(loss=mean_loss)
tq.close()
self._save_model(epoch + 1)
valid_metrics = validate(model, criterion, valid_loader, validation_size, batch_size, iter_size)
self._write_event(**valid_metrics)
valid_loss = valid_metrics['valid_loss']
valid_losses.append(valid_loss)
if valid_loss < self.best_valid_loss:
print("Best validation loss improved from {} to {}".format(self.best_valid_loss, valid_loss))
self.best_valid_loss = valid_loss
shutil.copy(str(self.model_path), str(self.best_model_path))
elif patience and epoch - lr_reset_epoch > patience and min(
valid_losses[-patience:]) > self.best_valid_loss:
lr /= 10
if lr < 1e-8:
exit(0)
lr_reset_epoch = epoch
optimizer = self._init_optimizer()
except KeyboardInterrupt:
tq.close()
print('Ctrl+C, saving snapshot')
self._save_model(epoch)
print('done.')
break
return
