return model
if __name__ == '__main__':
input_size = 128
minibatch_size = 16
batch_size = 3200
checkpoint_path = 'checkpoints'
if not os.path.exists(checkpoint_path):
os.makedirs(checkpoint_path)
model = build_model()
model.load('road_model1-800000')
# Train
for i in range(num_iterations):
image_data, label_data = \
generate_data.generate_batch(
height=input_size,
width=input_size,
minibatch_size=batch_size)
X = image_data
Y = label_data[:,newaxis]
model.fit({'input': X}, {'target': Y},
n_epoch=1,
batch_size=minibatch_size,
snapshot_epoch=True, show_metric=True, run_id='road_model1')
if __name__ == '__main__':
input_size = 128
minibatch_size = 16
batch_size = 3200
checkpoint_path = 'checkpoints'
if not os.path.exists(checkpoint_path):
os.makedirs(checkpoint_path)
model = build_model()
model.load('road_model1-800000')
# Train
for i in range(num_iterations):
image_data, label_data = \
generate_data.generate_batch(
height=input_size,
width=input_size,
minibatch_size=batch_size)
X = image_data
Y = label_data[:, newaxis]
model.fit({'input': X}, {'target': Y},
n_epoch=1,
batch_size=minibatch_size,
snapshot_epoch=True,
show_metric=True,
run_id='road_model1')
samples_val[i]['xq_padded'].flatten()))
y_s = samples_val[i]['ys_padded'].flatten()
y_q = samples_val[i]['yq_padded'].flatten()
x_val[:len(x_b), i] = x_b
y_val[:len(y_s), i] = y_s
z_val[:len(y_q), i] = y_q
# start training
episode_start_time = time.time()
for episode in range(1, args.num_episodes + 1):
model.train()
# generate batch
x, y, z = generate_batch(args.batch_size, generate_episode_train,
tabu_episodes)
# update params
model.zero_grad()
output = model(x.cuda(), y.cuda(), src_mask=None, tgt_mask=None)
loss = criterion(output.view(-1, ntoken), z.view(-1).cuda())
optimizer.zero_grad()
loss.backward()
optimizer.step()
if episode % args.eval_interval == 0:
model.eval()
with torch.no_grad():
val_out = model(x_val.cuda(),