def alexnet():
if not valid_request(request, UPLOAD_FOLDER): return redirect(request.url)
# hard coded image as orig.png
model = AlexNet(pretrained=True)
model.eval()
image = Image.open(os.path.join(UPLOAD_FOLDER, 'orig.png'))
image = image.convert('RGB')
out = topk_to_rank_string(*model(image))
return render_template('/pages/alexnet.html', out=Markup(out[:-2]))
def main():
train_images = tf.data.TextLineDataset(
'/home/caique/datasets/caltech101/caltech101_train.txt'
)
train_labels = tf.data.TextLineDataset(
'/home/caique/datasets/caltech101/caltech101_train_labels.txt'
)
valid_images = tf.data.TextLineDataset(
'/home/caique/datasets/caltech101/caltech101_test.txt'
)
valid_labels = tf.data.TextLineDataset(
'/home/caique/datasets/caltech101/caltech101_test_labels.txt'
)
drop_data = tf.data.Dataset.zip((train_images, train_labels))
drop_data = drop_data.map(get_parser(False)).batch(1)
valid_data = tf.data.Dataset.zip((valid_images, valid_labels))
valid_data = valid_data.map(get_parser(False)).batch(120)
train_data = tf.data.Dataset.zip((train_images, train_labels))
train_data = train_data.map(get_parser(True)).shuffle(3030).batch(101)
session = tf.Session()
model = AlexNet(101)
saver = tf.train.Saver()
saver.restore(session, './variables/alexnet-caltech101-finetunned-2-2')
# dropped_filters = drop_filters(
# session, model, drop_data, valid_data, drop_total=500,
# drop_n=20
# )
# model.train(session, train_data, valid_data,
# epochs=40,
# lr=0.00001,
# # dropped_filters=dropped_filters,
# # train_layers=['fc8'],
# # weights_path='alexnet_weights.npy',
# variables_path='./variables/alexnet-caltech101-finetunned-33',
# model_name='alexnet-caltech101-finetunned-2')
session.run(model.iterator.make_initializer(valid_data))
print('final eval: {}'.format(model.eval(session)))
session.close()
optimizer.zero_grad()
outputs = net(signals)
loss = criterion(outputs, labels)
loss.backward()
optimizer.step()
running_loss += loss.item()
#if batch_index > 50:
# if batch_index % 100 == 0: # print every 2000 mini-batches
# print('[%d, %5d] loss: %.3f' %
# (epoch + 1, batch_index + 1, running_loss / (batch_index+1) ))
# running_loss = 0.0
print("Avg Loss Train: {}".format(running_loss / len(train_loader)))
# Validation
print("Epoch: {} - Val".format(epoch))
net.eval()
running_loss = 0.
with torch.no_grad():
for batch_index, (signals, labels) in enumerate(tqdm(val_loader)):
signals, labels = signals.to(device=device), labels.to(
device=device)
outputs = net(signals)
loss = criterion(outputs, labels)
running_loss += loss.item()
print("Avg Loss Val: {}".format(running_loss / len(val_loader)))
accumulated_train_loss.append(loss.item())
# zero gradients
optimizer.zero_grad()
# backward
loss.backward()
# Step to update optimizer params
optimizer.step()
iteration += 1
# Validation
# Set mode in inference mode
model.eval()
accumulated_val_loss = []
for batch_x, batch_y in val_generator:
# Forward
preds = model(batch_x)
# compute loss
loss = loss_func(preds, batch_y)
accumulated_val_loss.append(loss.item())
train_loss = sum(accumulated_train_loss) / len(accumulated_train_loss)
val_loss = sum(accumulated_val_loss) / len(accumulated_val_loss)
print("Epoch: {} -- -- train loss: {}, val loss: {}".format(
epoch, train_loss, val_loss))
tensorboard.add_scalar("Loss/Train", train_loss, epoch)
tensorboard.add_scalar("Loss/Validation", val_loss, epoch)