def main():
# Initialize variables
device = "/gpu:0" if Constants.USE_GPU else "/cpu:0"
learning_rate_input = tf.placeholder("float")
initial_learning_rate = log_uniform(Constants.ALPHA.LOW,
Constants.ALPHA.HIGH,
Constants.ALPHA.LOG_RATE)
# Initialize network
global_network, grad_applier = init_network(device, learning_rate_input)
# Load our saver class
saver = Saver()
# Initialize threads
Data.a3c_threads = create_a3c_threads(global_network, learning_rate_input,
initial_learning_rate, grad_applier,
saver.data, device)
# Initialize TF
session = init_tf()
summary = init_tf_summary(session)
# Initialize / Load existing checkpoint
saver.load(session)
# Install Ctrl+C signal handler
def signal_handler(sig, frame):
print('CTRL+C was pressed, attempting to stop and save.')
Data.stop_requested = True
signal.signal(signal.SIGINT, signal_handler)
# Create and trigger the threads
threads = run_system_threads(session, Data, summary, saver)
print('Press Ctrl+C to stop')
# Finish and save results
finish(threads)
saver.save(session)
device = torch.device(
"cuda:{}".format(gpu) if torch.cuda.is_available() else "cpu")
model = SegNet(sigma=opt.sigma).to(device)
optimizer = optim.Adam(model.parameters(), lr=1e-4)
# Recover weights, if required
if opt.recover:
ckpt_file = os.path.join(model_dir, opt.reco_type + '_weights.pth')
ckpt = torch.load(ckpt_file, map_location=device)
model.load_state_dict(ckpt['model_state_dict'])
epoch = ckpt['iter_nb'] + 1
print('Model recovered from {}.'.format(ckpt_file))
if 'optimizer_state_dict' in ckpt:
optimizer.load_state_dict(ckpt['optimizer_state_dict'])
print('Optimizer recovered from {}.'.format(ckpt_file))
saver.load()
else:
epoch = 0
# Metrics
metrics = [
'SEMANTIC_LOSS', 'IOU_0', 'IOU_1', 'IOU_2', 'IOU_3', 'IOU_4', 'IOU_5',
'IOU_6', 'MEAN_IOU', 'PIX_ACC', 'DEPTH_LOSS', 'ABS_ERR', 'REL_ERR'
]
saver.add_metrics(metrics)
# Create datasets
dataset_path = opt.dataroot
city_train_set = Cityscapes(root=dataset_path, train=True)
city_test_set = Cityscapes(root=dataset_path, train=False)
if __name__ == '__main__':
map = Map(133, 33, '#')
start = Start(133, 33)
finish = GameOver(133, 33)
paconator = Paco(1, 15, 133, 33)
getKey = GetKey()
save = Saver()
pigeons = []
lasers = []
bones = []
play_game = False
game_over = False
ammo = 20
points = 0
x = save.load()
record = int(x)
WindowSize()
# ====== START SCREEN ======
while True:
key = getKey()
for y in range(11):
ClearConsole()
if y < 10:
for x in range(87):
map.set_point(start.screen[y][x].x, start.screen[y][x].y,
start.screen[y][x].character)
map.set_point(start.screen[y + 12][x].x,
start.screen[y + 12][x].y,
start.screen[y + 12][x].character)
#saver manages data and tf model saving
saver = Saver(data)
#testing env for playing in the end
env_test = Environment(gym.make(Constants.ENV),
Agent(master_network, Constants.EPS_START,
Constants.EPS_STOP, Constants.EPS_STEPS),
summary,
saver,
render=True,
eps_start=0.,
eps_end=0.)
#load last tf checkpoint
saver.load(master_network.session)
#dont train if replaymode is set
if not Constants.REPLAY_MODE:
#initialize threads
envs = [
Environment(
gym.make(Constants.ENV),
Agent(master_network, Constants.EPS_START, Constants.EPS_STOP,
Constants.EPS_STEPS), summary, saver)
for i in range(Constants.THREADS - 1)
]
#create cvshow thread
envs.append(
Environment(
class Experiment:
def __init__(self, opt):
assert opt.save_dir is not None, "No directory given for saving the model."
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = opt.gpu
self.opt = opt
self.device = torch.device("cpu" if opt.gpu == "" or
not torch.cuda.is_available() else "cuda")
self.train_set = Chembl("train",
kekulize=opt.kekulize,
random_order=opt.random_order)
self.valid_set = Chembl("valid",
kekulize=opt.kekulize,
random_order=opt.random_order)
self.test_set = Chembl("test",
kekulize=opt.kekulize,
random_order=opt.random_order)
self.train_loader = torch.utils.data.DataLoader(
self.train_set,
batch_size=opt.batch_size,
shuffle=True,
collate_fn=Chembl.collate,
num_workers=1,
pin_memory=True)
self.valid_loader = torch.utils.data.DataLoader(
self.valid_set,
batch_size=256,
shuffle=False,
collate_fn=Chembl.collate,
num_workers=1,
pin_memory=True)
self.test_loader = torch.utils.data.DataLoader(
self.test_set,
batch_size=256,
shuffle=False,
collate_fn=Chembl.collate,
num_workers=1,
pin_memory=True)
self.model = GraphGen(self.train_set.n_node_types(),
self.train_set.n_edge_types(),
opt.state_size,
n_max_nodes=30,
n_max_edges=2 * self.train_set.get_max_bonds(),
dropout=opt.dropout)
self.model = self.model.to(self.device)
if opt.optimizer == "adam":
self.optimizer = torch.optim.Adam(self.model.parameters(),
lr=opt.lr,
weight_decay=opt.wd,
eps=1e-10)
elif opt.optimizer == "sgd":
self.optimizer = torch.optim.SGD(self.model.parameters(),
lr=opt.lr,
momentum=0.99,
nesterov=True)
else:
assert False, "Invalid optimizer: %s" % opt.optimizer
self.loss_plot = Plot2D("loss", 10, xlabel="iter", ylabel="loss")
self.valid_loss_plot = Plot2D("Validation loss",
1,
xlabel="iter",
ylabel="loss")
self.percent_valid = Plot2D("Valid molecules",
1,
xlabel="iter",
ylabel="%")
self.mol_images = Image("Molecules")
self.iteration = 0
self.epoch = 0
self.best_loss = float("inf")
self.best_loss_iteration = 0
self.best_loss_epoch = 0
self.test_loss = None
self.patience = 5
self.lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(self.optimizer, opt.lr_milestones, opt.lr_gamma) \
if opt.lr_milestones else None
self.saver = Saver(self, os.path.join(opt.save_dir, "save"))
self.saver.load()
def _move_to_device(self, d):
if isinstance(d, tuple):
return tuple(self._move_to_device(list(d)))
elif isinstance(d, list):
return [self._move_to_device(v) for v in d]
elif isinstance(d, dict):
return {k: self._move_to_device(v) for k, v in d}
elif torch.is_tensor(d):
return d.to(self.device)
else:
return d
def state_dict(self):
return {
"model":
self.model.state_dict(),
"optimizer":
self.optimizer.state_dict(),
"loss_plot":
self.loss_plot.state_dict(),
"valid_loss_plot":
self.valid_loss_plot.state_dict(),
"percent_valid":
self.percent_valid.state_dict(),
"iteration":
self.iteration,
"epoch":
self.epoch,
"best_loss":
self.best_loss,
"best_loss_iteration":
self.best_loss_iteration,
"best_loss_epoch":
self.best_loss_epoch,
"test_loss":
self.test_loss,
"lr_scheduler":
self.lr_scheduler.state_dict()
if self.lr_scheduler is not None else None
}
def load_state_dict(self, state_dict):
self.model.load_state_dict(state_dict["model"])
self.optimizer.load_state_dict(state_dict["optimizer"])
self.valid_loss_plot.load_state_dict(state_dict["valid_loss_plot"])
self.loss_plot.load_state_dict(state_dict["loss_plot"])
self.percent_valid.load_state_dict(state_dict["percent_valid"])
self.iteration = state_dict["iteration"]
self.epoch = state_dict["epoch"]
self.best_loss = state_dict["best_loss"]
self.best_loss_iteration = state_dict["best_loss_iteration"]
self.best_loss_epoch = state_dict["best_loss_epoch"]
self.test_loss = state_dict["test_loss"]
if self.lr_scheduler is not None:
s = state_dict.get("lr_scheduler")
if s is not None:
self.lr_scheduler.load_state_dict(s)
def test(self, loader=None):
self.model.eval()
loss_sum = 0
cnt = 0
with torch.no_grad():
for d in tqdm(loader):
d = self._move_to_device(d)
_, loss = self.model(d)
cnt += d[0].shape[0]
loss_sum += loss.item() * d[0].shape[0]
loss = loss_sum / cnt
return loss
def load_the_best(self):
self.saver.load(self.best_loss_iteration)
def do_final_test(self):
self.load_the_best()
self.test_loss = self.test(self.test_loader)
print("----------------------------------------------")
print("Training done.")
print(" Validation loss:", self.best_loss)
print(" Test loss:", self.test_loss)
print(" Epoch:", self.best_loss_epoch)
print(" Iteration:", self.best_loss_iteration)
self.saver.save("best")
def display_generated(self):
self.model.eval()
graphs = []
with torch.no_grad():
for i in range(1):
g = self.model.generate(32, self.device)
g = g.get_final_graph()
graphs.append(g)
self.model.train()
img = self.train_set.draw_molecules(graphs)
self.mol_images.draw(img)
def train(self):
running = True
while running:
print("Epoch %d" % self.epoch)
self.model.train()
for d in tqdm(self.train_loader):
d = self._move_to_device(d)
g, loss = self.model(d)
assert torch.isfinite(loss), "Loss is %s" % loss.item()
self.loss_plot.add_point(self.iteration, loss.item())
self.optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(self.model.parameters(), 0.25)
self.optimizer.step()
self.iteration += 1
if self.iteration % 30 == 0:
self.display_generated()
# Do a validation step
validation_loss = self.test(self.valid_loader)
self.valid_loss_plot.add_point(self.iteration, validation_loss)
g = self.generate(10000)
self.percent_valid.add_point(self.iteration, g["ratio_ok"] * 100)
# Early stopping
if validation_loss <= self.best_loss:
self.best_loss = validation_loss
self.best_loss_iteration = self.iteration
self.best_loss_epoch = self.epoch
elif (self.epoch - self.best_loss_epoch) > self.patience:
running = not opt.early_stop
if self.lr_scheduler is not None:
self.lr_scheduler.step()
# Save the model
self.saver.save(self.iteration)
self.epoch += 1
def train_done(self):
return self.test_loss is not None
def generate(self, n_test=100000):
v = self.train_set.start_verification()
bsize = 200
self.model.eval()
with torch.no_grad():
for i in tqdm(range(n_test // bsize)):
res = self.model.generate(bsize, self.device)
self.train_set.verify(v, res.get_final_graph())
gc.collect()
return self.train_set.get_verification_results(v)
shuffle=True)
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model = YOLO()
if torch.cuda.device_count() > 1:
print("Let's use", torch.cuda.device_count(), "GPUs!")
model = nn.DataParallel(model)
model.to(device)
criterion = YOLO_loss()
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
yolo_saver = Saver(model,
join(save_root, r'yolo', r'save'),
"yolo",
max_to_keep=20)
loaded_index = yolo_saver.load()
for epoch in range(num_epochs):
for i, (images, labels) in enumerate(data_loader):
images = images.cuda()
labels = labels.cuda()
# Forward pass
outputs = model(images)
loss = criterion(outputs, labels)
# Backward and optimize
optimizer.zero_grad()
loss.backward()
optimizer.step()
ModelClass = load_class(config.model._class_)
net: nn.Module = ModelClass(
**{k: v
for k, v in config.model.items() if k != '_class_'})
net.to(config.opts.device)
OptimizerClass = load_class(config.optimizer._class_)
optimizer: optim.Optimizer = OptimizerClass(
params=net.parameters(),
**{k: v
for k, v in config.optimizer.items() if k != '_class_'})
if config.training.restore:
train_state = saver.load(model=net,
optimizer=optimizer,
device=config.training.device)
else:
train_state = Munch(epochs=0, samples=0)
if config.opts.log:
with open(folder_run / 'config.yml', mode='w') as f:
f.write(config.toYAML())
logger.add_text('Config',
textwrap.indent(config.toYAML(), ' '),
global_step=train_state.samples)
def make_dataloader(dataset, shuffle) -> data.DataLoader:
return data.DataLoader(dataset,
batch_size=config.training.batch_size,