def main():
env = gym.make('CartPole-v0')
gen_idx = 0
nets = [
model.CartPoleNet(env.observation_space.shape[0], env.action_space.n)
for _ in range(POPULATION_SIZE)
]
population = [(net, evaluate(env, net)) for net in nets]
while True:
population.sort(key=lambda p: p[1], reverse=True)
rewards = [p[1] for p in population[:PARENTS_COUNT]]
reward_mean = np.mean(rewards)
reward_max = np.max(rewards)
reward_std = np.std(rewards)
print("%d: reward_mean=%.2f, reward_max=%.2f, reward_std=%.2f" %
(gen_idx, reward_mean, reward_max, reward_std))
if reward_mean > 199:
print("Solved in %d steps" % gen_idx)
break
prev_population = population
population = [population[0]]
for _ in range(POPULATION_SIZE - 1):
parent_idx = np.random.randint(0, PARENTS_COUNT)
parent = prev_population[parent_idx][0]
net = mutate_parent(parent)
fitness = utils.evaluate(env, net)
population.append((net, fitness))
gen_idx += 1
def test_evaluate_runs(predictions_file, out_folder):
df_preds_w_targets = pd.read_csv(predictions_file)
evaluate(df_preds_w_targets, out_folder=out_folder)
for val_w, val_d, val_r, val_t in val_loader:
net([val_w, val_d, val_r])
break
net.initialize(ctx = ctx, init = MyInit(), force_reinit = True)
# initialize a trainer to train model
trainer = gluon.Trainer(net.collect_params(), optimizer, {'learning_rate': learning_rate})
# initialize a SummaryWriter to write information into logs dir
sw = SummaryWriter(logdir = params_path, flush_secs = 5)
# compute validation loss before training
compute_val_loss(net, val_loader, loss_function, sw, 0)
# compute testing set MAE, RMSE, MAPE before training
evaluate(net, test_loader, true_value, num_of_vertices, sw, 0)
# train model
global_step = 1
for epoch in range(1, epochs + 1):
for train_w, train_d, train_r, train_t in train_loader:
start_time = time()
with autograd.record():
output = net([train_w, train_d, train_r])
l = loss_function(output, train_t)
l.backward()
trainer.step(train_t.shape[0])
training_loss = l.mean().asscalar()
recent=24,
K=3,
Kt=3)
net.to(device) #to cuda
optimizer = optim.Adam(net.parameters(),
lr=learning_rate,
weight_decay=wdecay)
scheduler = optim.lr_scheduler.ExponentialLR(optimizer, decay)
#scheduler = optim.lr_scheduler.MultiStepLR(optimizer, [20,30], gamma=0.7, last_epoch=-1)
#calculate origin loss in epoch 0
compute_val_loss(net, val_loader, loss_function, supports, device, epoch=0)
# compute testing set MAE, RMSE, MAPE before training
evaluate(net, test_loader, true_value, supports, device, epoch=0)
clip = 5
his_loss = []
train_time = []
for epoch in range(1, epochs + 1):
train_l = []
start_time_train = time()
for train_w, train_d, train_r, train_t in train_loader:
train_w = train_w.to(device)
train_d = train_d.to(device)
train_r = train_r.to(device)
train_t = train_t.to(device)
net.train() #train pattern
optimizer.zero_grad() #grad to 0
# initialize a trainer to train model
trainer = gluon.Trainer(net.collect_params(), optimizer,
{'learning_rate': learning_rate})
# initialize a SummaryWriter to write information into logs dir
sw = SummaryWriter(logdir=params_path, flush_secs=5)
# compute validation loss before training
compute_val_loss(net, val_loader, loss_function, sw, epoch=0, ctx=ctx)
# compute testing set MAE, RMSE, MAPE before training
evaluate(net,
test_loader,
true_value,
num_of_vertices,
sw,
epoch=0,
ctx=ctx)
# train model
global_step = 1
for epoch in range(1, epochs + 1):
for train_w, train_d, train_r, train_t in train_loader:
# running on single gpu
train_w = train_w.as_in_context(ctx)
train_d = train_d.as_in_context(ctx)
train_r = train_r.as_in_context(ctx)
train_t = train_t.as_in_context(ctx)
with autograd.record():
time: {batch_end_time - batch_start_time:.2f}'
)
print(
f's:[{epoch:d}, {i + 1:5d}] loss: {running_loss_s / group_num:.2f}, \
time: {batch_end_time - batch_start_time:.2f}'
)
print(
'--------------------------------------------------------------------'
)
running_loss = 0.0
running_loss_f = 0.0
running_loss_o = 0.0
running_loss_s = 0.0
batch_start_time = batch_end_time
epoch_end_time = time.perf_counter()
print(f'Epoch cost {epoch_end_time - epoch_start_time:.2f} seconds')
# probably not need to run this after every epoch:可能不需要在每一个时代之后都运行这个
with torch.no_grad():
# compute validation loss:计算验证损失
compute_val_loss(net, val_loader, loss_function, None, epoch,
device, all_data['stats']['stats'])
# testing:测试
evaluate(net, test_loader, true_value, num_of_vertices, None,
epoch, device, all_data['stats']['stats'])
end_time = time.perf_counter()
print(f'Total running time is {end_time - start_time:.2f} seconds.')
train_t = train_t.to(device)
outputs = net([train_w, train_d, train_r])
loss = loss_function(outputs, train_t) # loss is a tensor on the same device as outpus and train_t
loss.backward()
optimizer.step()
running_loss += loss.item() # type of running_loss is float, loss.item() is a float on CPU
if i % group_num == group_num - 1:
batch_end_time = time.perf_counter()
print(f'[{epoch:d}, {i + 1:5d}] loss: {running_loss / group_num:.2f}, \
time: {batch_end_time - batch_start_time:.2f}')
running_loss = 0.0
batch_start_time = batch_end_time
epoch_end_time = time.perf_counter()
print(f'Epoch cost {epoch_end_time - epoch_start_time:.2f} seconds')
# probably not need to run this after every epoch
with torch.no_grad():
# compute validation loss
compute_val_loss(net, val_loader, loss_function, None, epoch, device)
# testing
evaluate(net, test_loader, true_value, num_of_vertices, None, epoch, device)
end_time = time.perf_counter()
print(f'Total running time is {end_time - start_time:.2f} seconds.')
torch.cuda.get_device_name(0))
model = model.cuda() # move model to GPU if available
else:
print('[STATUS] Nvidia GPU not found, training using CPU')
criterion = nn.CrossEntropyLoss() # criterion
optimizer = torch.optim.Adam(
model.parameters(), lr=float(args['learning_rate'])
) # model optimizer (see torch.optim documentation to explore other optimizers)
print('[INFO] Training and evaluating the model...')
# train and evaluate the model
if args['file_name'] is None:
print(
'[WARNING] Trained model will not be saved, no path specified. Use -f or --file_name args to save the trained model'
)
else:
print('[INFO] Trained model will be saved as:', args['file_name'])
final_model, training_losses = train(int(args['epochs']),
model,
train_loader,
criterion,
optimizer,
use_cuda,
save_path=args['file_name'])
evaluate(model, test_loader, criterion, use_cuda)
print('[STATUS] Model training and evaluation completed')
print('Job completed!')
def main():
cudnn.benchmark = True
device = 'cuda' if torch.cuda.is_available() else 'cpu'
assert config.MISC.TEST_INTERVAL is not 0, 'Illegal setting: config.MISC.TEST_INTERVAL = 0!'
# set random seed
if config.MISC.RANDOM_SEED:
random.seed(config.MISC.RANDOM_SEED)
np.random.seed(config.MISC.RANDOM_SEED)
torch.manual_seed(config.MISC.RANDOM_SEED)
if torch.cuda.is_available():
torch.cuda.manual_seed_all(config.MISC.RANDOM_SEED)
train_transformer = Mytransforms.Compose([
Mytransforms.KeyAreaCrop(20),
Mytransforms.RandomRotate(40),
Mytransforms.TestResized(config.MODEL.IMG_SIZE),
Mytransforms.RandomHorizontalFlip()
])
test_transformer = Mytransforms.Compose([
Mytransforms.KeyAreaCrop(20),
Mytransforms.TestResized(config.MODEL.IMG_SIZE)
])
# train
source_dset = HandKptDataset(config.DATA.SOURCE.TRAIN.DIR, config.DATA.SOURCE.TRAIN.LBL_FILE,
stride=config.MODEL.HEATMAP_STRIDE, transformer=train_transformer)
# target_dset = HandKptDataset(config.DATA.TARGET.TRAIN.DIR, config.DATA.TARGET.TRAIN.LBL_FILE,
# stride=config.MODEL.HEATMAP_STRIDE, transformer=train_transformer)
source_val_dset = HandKptDataset(config.DATA.SOURCE.VAL.DIR, config.DATA.SOURCE.VAL.LBL_FILE,
stride=config.MODEL.HEATMAP_STRIDE, transformer=test_transformer)
target_val_dset = HandKptDataset(config.DATA.TARGET.VAL.DIR, config.DATA.TARGET.VAL.LBL_FILE,
stride=config.MODEL.HEATMAP_STRIDE, transformer=test_transformer)
# source only
train_loader = torch.utils.data.DataLoader(
source_dset,
batch_size=config.TRAIN.BATCH_SIZE, shuffle=True,
num_workers=config.MISC.WORKERS, pin_memory=True)
# val
source_val_loader = torch.utils.data.DataLoader(
source_val_dset,
batch_size=config.TRAIN.BATCH_SIZE, shuffle=False,
num_workers=config.MISC.WORKERS, pin_memory=True)
target_val_loader = torch.utils.data.DataLoader(
target_val_dset,
batch_size=config.TRAIN.BATCH_SIZE, shuffle=False,
num_workers=config.MISC.WORKERS, pin_memory=True)
logger = Logger(ckpt_path=os.path.join(config.DATA.CKPT_PATH, config.PROJ_NAME),
tsbd_path=os.path.join(config.DATA.VIZ_PATH, config.PROJ_NAME))
net = pose_resnet.get_pose_net(config).to(device)
optimizer = torch.optim.Adam(net.parameters(), config.TRAIN.BASE_LR,
weight_decay=config.TRAIN.WEIGHT_DECAY)
input_shape = (config.TRAIN.BATCH_SIZE, 3, config.MODEL.IMG_SIZE, config.MODEL.IMG_SIZE)
logger.add_graph(net, input_shape, device)
if len(config.MODEL.RESUME) > 0:
print("=> loading checkpoint '{}'".format(config.MODEL.RESUME))
resume_ckpt = torch.load(config.MODEL.RESUME)
net.load_state_dict(resume_ckpt['net'])
optimizer.load_state_dict(resume_ckpt['optim'])
config.TRAIN.START_ITERS = resume_ckpt['iter']
logger.global_step = resume_ckpt['iter']
logger.best_metric_val = resume_ckpt['best_metric_val']
net = torch.nn.DataParallel(net)
if config.EVALUATE:
pck05, pck2 = evaluate(net, target_val_loader, img_size=config.MODEL.IMG_SIZE, vis=True,
logger=logger, disp_interval=config.MISC.DISP_INTERVAL)
print("=> validate [email protected] = {}, [email protected] = {}".format(pck05 * 100, pck2 * 100))
return
criterion = nn.SmoothL1Loss(reduction='none').to(device)
total_progress_bar = tqdm.tqdm(desc='Train iter', ncols=80,
total=config.TRAIN.MAX_ITER,
initial=config.TRAIN.START_ITERS)
epoch = 0
while logger.global_step < config.TRAIN.MAX_ITER:
for (stu_inputs, stu_heatmap, _) in tqdm.tqdm(
train_loader, total=len(train_loader),
desc='Current epoch', ncols=80, leave=False):
stu_inputs = stu_inputs.to(device)
stu_heatmap = stu_heatmap.to(device)
stu_heats = net(stu_inputs)
loss = criterion(stu_heats, stu_heatmap).sum() / stu_inputs.size(0)
optimizer.zero_grad()
loss.backward()
optimizer.step()
# val
if logger.global_step % config.MISC.TEST_INTERVAL == 0:
pck05, pck2 = evaluate(net, source_val_loader, img_size=config.MODEL.IMG_SIZE, vis=True,
logger=logger, disp_interval=config.MISC.DISP_INTERVAL,
show_gt=(logger.global_step == 0), is_target=False)
logger.add_scalar('[email protected]', pck05 * 100)
logger.add_scalar('[email protected]', pck2 * 100)
pck05, pck2 = evaluate(net, target_val_loader, img_size=config.MODEL.IMG_SIZE, vis=True,
logger=logger, disp_interval=config.MISC.DISP_INTERVAL,
show_gt=(logger.global_step == 0), is_target=True)
logger.add_scalar('[email protected]', pck05 * 100)
logger.add_scalar('[email protected]', pck2 * 100)
logger.save_ckpt(state={
'net': net.module.state_dict(),
'optim': optimizer.state_dict(),
'iter': logger.global_step,
'best_metric_val': logger.best_metric_val,
}, cur_metric_val=pck05)
logger.step(1)
total_progress_bar.update(1)
# log
logger.add_scalar('regress_loss', loss.item())
epoch += 1
total_progress_bar.close()
