def main():
start_time = time.time()
init_out_dir()
if args.clear_checkpoint:
clear_checkpoint()
last_step = get_last_checkpoint_step()
if last_step >= 0:
my_log('\nCheckpoint found: {}\n'.format(last_step))
else:
clear_log()
print_args()
if args.net == 'made':
net = MADE(**vars(args))
elif args.net == 'pixelcnn':
net = PixelCNN(**vars(args))
elif args.net == 'bernoulli':
net = BernoulliMixture(**vars(args))
else:
raise ValueError('Unknown net: {}'.format(args.net))
net.to(args.device)
my_log('{}\n'.format(net))
params = list(net.parameters())
params = list(filter(lambda p: p.requires_grad, params))
nparams = int(sum([np.prod(p.shape) for p in params]))
my_log('Total number of trainable parameters: {}'.format(nparams))
named_params = list(net.named_parameters())
if args.optimizer == 'sgd':
optimizer = torch.optim.SGD(params, lr=args.lr)
elif args.optimizer == 'sgdm':
optimizer = torch.optim.SGD(params, lr=args.lr, momentum=0.9)
elif args.optimizer == 'rmsprop':
optimizer = torch.optim.RMSprop(params, lr=args.lr, alpha=0.99)
elif args.optimizer == 'adam':
optimizer = torch.optim.Adam(params, lr=args.lr, betas=(0.9, 0.999))
elif args.optimizer == 'adam0.5':
optimizer = torch.optim.Adam(params, lr=args.lr, betas=(0.5, 0.999))
else:
raise ValueError('Unknown optimizer: {}'.format(args.optimizer))
if args.lr_schedule:
# 0.92**80 ~ 1e-3
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
optimizer, factor=0.92, patience=100, threshold=1e-4, min_lr=1e-6)
if last_step >= 0:
state = torch.load('{}_save/{}.state'.format(args.out_filename,
last_step))
ignore_param(state['net'], net)
net.load_state_dict(state['net'])
if state.get('optimizer'):
optimizer.load_state_dict(state['optimizer'])
if args.lr_schedule and state.get('scheduler'):
scheduler.load_state_dict(state['scheduler'])
init_time = time.time() - start_time
my_log('init_time = {:.3f}'.format(init_time))
my_log('Training...')
sample_time = 0
train_time = 0
start_time = time.time()
for step in range(last_step + 1, args.max_step + 1):
optimizer.zero_grad()
sample_start_time = time.time()
with torch.no_grad():
sample, x_hat = net.sample(args.batch_size)
assert not sample.requires_grad
assert not x_hat.requires_grad
sample_time += time.time() - sample_start_time
train_start_time = time.time()
log_prob = net.log_prob(sample)
# 0.998**9000 ~ 1e-8
beta = args.beta * (1 - args.beta_anneal**step)
with torch.no_grad():
energy = ising.energy(sample, args.ham, args.lattice,
args.boundary)
loss = log_prob + beta * energy
assert not energy.requires_grad
assert not loss.requires_grad
loss_reinforce = torch.mean((loss - loss.mean()) * log_prob)
loss_reinforce.backward()
if args.clip_grad:
nn.utils.clip_grad_norm_(params, args.clip_grad)
optimizer.step()
if args.lr_schedule:
scheduler.step(loss.mean())
train_time += time.time() - train_start_time
if args.print_step and step % args.print_step == 0:
free_energy_mean = loss.mean() / args.beta / args.L**2
free_energy_std = loss.std() / args.beta / args.L**2
entropy_mean = -log_prob.mean() / args.L**2
energy_mean = energy.mean() / args.L**2
mag = sample.mean(dim=0)
mag_mean = mag.mean()
mag_sqr_mean = (mag**2).mean()
if step > 0:
sample_time /= args.print_step
train_time /= args.print_step
used_time = time.time() - start_time
my_log(
'step = {}, F = {:.8g}, F_std = {:.8g}, S = {:.8g}, E = {:.8g}, M = {:.8g}, Q = {:.8g}, lr = {:.3g}, beta = {:.8g}, sample_time = {:.3f}, train_time = {:.3f}, used_time = {:.3f}'
.format(
step,
free_energy_mean.item(),
free_energy_std.item(),
entropy_mean.item(),
energy_mean.item(),
mag_mean.item(),
mag_sqr_mean.item(),
optimizer.param_groups[0]['lr'],
beta,
sample_time,
train_time,
used_time,
))
sample_time = 0
train_time = 0
if args.save_sample:
state = {
'sample': sample,
'x_hat': x_hat,
'log_prob': log_prob,
'energy': energy,
'loss': loss,
}
torch.save(state, '{}_save/{}.sample'.format(
args.out_filename, step))
if (args.out_filename and args.save_step
and step % args.save_step == 0):
state = {
'net': net.state_dict(),
'optimizer': optimizer.state_dict(),
}
if args.lr_schedule:
state['scheduler'] = scheduler.state_dict()
torch.save(state, '{}_save/{}.state'.format(
args.out_filename, step))
if (args.out_filename and args.visual_step
and step % args.visual_step == 0):
torchvision.utils.save_image(
sample,
'{}_img/{}.png'.format(args.out_filename, step),
nrow=int(sqrt(sample.shape[0])),
padding=0,
normalize=True)
if args.print_sample:
x_hat_np = x_hat.view(x_hat.shape[0], -1).cpu().numpy()
x_hat_std = np.std(x_hat_np, axis=0).reshape([args.L] * 2)
x_hat_cov = np.cov(x_hat_np.T)
x_hat_cov_diag = np.diag(x_hat_cov)
x_hat_corr = x_hat_cov / (
sqrt(x_hat_cov_diag[:, None] * x_hat_cov_diag[None, :]) +
args.epsilon)
x_hat_corr = np.tril(x_hat_corr, -1)
x_hat_corr = np.max(np.abs(x_hat_corr), axis=1)
x_hat_corr = x_hat_corr.reshape([args.L] * 2)
energy_np = energy.cpu().numpy()
energy_count = np.stack(
np.unique(energy_np, return_counts=True)).T
my_log(
'\nsample\n{}\nx_hat\n{}\nlog_prob\n{}\nenergy\n{}\nloss\n{}\nx_hat_std\n{}\nx_hat_corr\n{}\nenergy_count\n{}\n'
.format(
sample[:args.print_sample, 0],
x_hat[:args.print_sample, 0],
log_prob[:args.print_sample],
energy[:args.print_sample],
loss[:args.print_sample],
x_hat_std,
x_hat_corr,
energy_count,
))
if args.print_grad:
my_log('grad max_abs min_abs mean std')
for name, param in named_params:
if param.grad is not None:
grad = param.grad
grad_abs = torch.abs(grad)
my_log('{} {:.3g} {:.3g} {:.3g} {:.3g}'.format(
name,
torch.max(grad_abs).item(),
torch.min(grad_abs).item(),
torch.mean(grad).item(),
torch.std(grad).item(),
))
else:
my_log('{} None'.format(name))
my_log('')
data_train = data_train[:, init_scope]
data_pos = data_pos[:, init_scope]
data_neg = data_neg[:, init_scope]
xtr = torch.from_numpy(data_train).float().cuda()
xte = torch.from_numpy(data_pos).float().cuda()
xod = torch.from_numpy(data_neg).float().cuda()
# construct model and ship to GPU
hidden_list = list(map(int, args.hiddens.split(',')))
model = MADE(xtr.size(1),
hidden_list,
xtr.size(1) * 2,
num_masks=args.num_masks)
print("number of model parameters:",
sum([np.prod(p.size()) for p in model.parameters()]))
model.cuda()
# set up the optimizer
opt = torch.optim.Adam(model.parameters(),
args.learning_rate,
weight_decay=args.weight_decay)
scheduler = torch.optim.lr_scheduler.StepLR(opt, step_size=45, gamma=0.1)
# list to store loss
loss_tr = []
loss_te = []
loss_od = []
# start the training
for epoch in range(args.epoch):
scheduler.step(epoch)
# load the dataset
print("loading binarized mnist from", args.data_path)
mnist = np.load(args.data_path)
xtr, xte = mnist['train_data'], mnist['valid_data']
xtr = torch.from_numpy(xtr).cuda()
xte = torch.from_numpy(xte).cuda()
# construct model and ship to GPU
hidden_list = list(map(int, args.hiddens.split(',')))
model = MADE(xtr.size(1),
hidden_list,
xtr.size(1),
num_masks=args.num_masks)
print("number of model parameters:",
sum([np.prod(p.size()) for p in model.parameters()]))
model.cuda()
# set up the optimizer
opt = torch.optim.Adam(model.parameters(), 1e-3, weight_decay=1e-4)
scheduler = torch.optim.lr_scheduler.StepLR(opt, step_size=45, gamma=0.1)
# start the training
for epoch in range(100):
print("epoch %d" % (epoch, ))
scheduler.step(epoch)
run_epoch(
'test',
upto=5) # run only a few batches for approximate test accuracy
run_epoch('train')
if __name__ == "__main__":
# load the dataset from some path
mnist = np.load("binarized_mnist.npz")
x_train, x_test = mnist["train_data"], mnist["valid_data"]
x_train = torch.as_tensor(x_train).cuda()
x_test = torch.as_tensor(x_test).cuda()
hidden_list = [500]
resample_every = 20
model = MADE(x_train.size(1), hidden_list, x_train.size(1))
print(
"number of model parameters: {np.sum([np.prod(p.size()) for p in model.parameters()])}"
)
model.cuda()
opt = torch.optim.Adam(model.parameters(), lr=1e-3, weight_decay=1e-4)
scheduler = torch.optim.lr_scheduler.StepLR(opt, step_size=50, gamma=0.1)
# The training
for epoch in range(100):
print(f"Epoch {epoch}")
scheduler.step()
# get an estimate of the test loss
run_one_epoch("test", upto=5)
run_one_epoch("train")
print("Final test eval:")
run_one_epoch("test")
scope_list = np.arange(n_RV)
scope_temp = np.delete(scope_list, np.where(scope_list % 34 == 17))
init_scope = list(np.delete(scope_temp, np.where(scope_temp % 34 == 33)))
# modify data to remove 0 (imag) columns
data_train = data_train[:, init_scope]
data_pos = data_pos[:, init_scope]
data_neg = data_neg[:, init_scope]
xtr = torch.from_numpy(data_train).float().cuda()
xte = torch.from_numpy(data_pos).float().cuda()
xod = torch.from_numpy(data_neg).float().cuda()
# construct model and ship to GPU
hidden_list = list(map(int, args.hiddens.split(',')))
model = MADE(xtr.size(1), hidden_list, xtr.size(1) * 2, num_masks=args.num_masks)
print("number of model parameters:", sum([np.prod(p.size()) for p in model.parameters()]))
model.cuda()
# set up the optimizer
opt = torch.optim.Adam(model.parameters(), args.learning_rate, weight_decay=args.weight_decay)
scheduler = torch.optim.lr_scheduler.StepLR(opt, step_size=45, gamma=0.1)
# list to store loss
loss_tr = []
loss_te = []
loss_od = []
# start the training
for epoch in range(args.epoch):
scheduler.step(epoch)
loss_tr.append(run_epoch('train'))
loss_te.append(run_epoch('test')) # run validation, which is pos
def BuckyBall():
start_time = time.time()
init_out_dir()
print_args()
if args.ham == 'buckey':
ham = buckyball_2(args.beta)
# elif args.ham == 'sk':
# ham = SKModel(args.n, args.beta, args.device, seed=args.seed)
# elif args.ham == 'full':
# ham = FullModel()
# elif args.ham == 'buckey':
# ham = buckyball_2(args.beta)
else:
raise ValueError('Unknown ham: {}'.format(args.ham))
#ham.J.requires_grad = False
net = MADE(**vars(args))
net.to(args.device)
my_log('{}\n'.format(net))
params = list(net.parameters())
params = list(filter(lambda p: p.requires_grad, params))
nparams = int(sum([np.prod(p.shape) for p in params]))
my_log('Total number of trainable parameters: {}'.format(nparams))
if args.optimizer == 'sgd':
optimizer = torch.optim.SGD(params, lr=args.lr)
elif args.optimizer == 'sgdm':
optimizer = torch.optim.SGD(params, lr=args.lr, momentum=0.9)
elif args.optimizer == 'rmsprop':
optimizer = torch.optim.RMSprop(params, lr=args.lr, alpha=0.99)
elif args.optimizer == 'adam':
optimizer = torch.optim.Adam(params, lr=args.lr, betas=(0.9, 0.999))
elif args.optimizer == 'adam0.5':
optimizer = torch.optim.Adam(params, lr=args.lr, betas=(0.5, 0.999))
else:
raise ValueError('Unknown optimizer: {}'.format(args.optimizer))
init_time = time.time() - start_time
my_log('init_time = {:.3f}'.format(init_time))
my_log('Training...')
sample_time = 0
train_time = 0
start_time = time.time()
if args.beta_anneal_to < args.beta:
args.beta_anneal_to = args.beta
beta = args.beta
while beta <= args.beta_anneal_to:
for step in range(args.max_step):
optimizer.zero_grad()
sample_start_time = time.time()
with torch.no_grad():
sample, x_hat = net.sample(args.batch_size)
assert not sample.requires_grad
assert not x_hat.requires_grad
sample_time += time.time() - sample_start_time
train_start_time = time.time()
log_prob = net.log_prob(sample)
with torch.no_grad():
energy = ham.energy(sample)
loss = log_prob + beta * energy
assert not energy.requires_grad
assert not loss.requires_grad
loss_reinforce = torch.mean((loss - loss.mean()) * log_prob)
loss_reinforce.backward()
if args.clip_grad > 0:
# nn.utils.clip_grad_norm_(params, args.clip_grad)
parameters = list(filter(lambda p: p.grad is not None, params))
max_norm = float(args.clip_grad)
norm_type = 2
total_norm = 0
for p in parameters:
param_norm = p.grad.data.norm(norm_type)
total_norm += param_norm.item()**norm_type
total_norm = total_norm**(1 / norm_type)
clip_coef = max_norm / (total_norm + args.epsilon)
for p in parameters:
p.grad.data.mul_(clip_coef)
optimizer.step()
train_time += time.time() - train_start_time
if args.print_step and step % args.print_step == 0:
free_energy_mean = loss.mean() / beta / args.n
free_energy_std = loss.std() / beta / args.n
entropy_mean = -log_prob.mean() / args.n
energy_mean = energy.mean() / args.n
mag = sample.mean(dim=0)
mag_mean = mag.mean()
if step > 0:
sample_time /= args.print_step
train_time /= args.print_step
used_time = time.time() - start_time
my_log(
'beta = {:.3g}, # {}, F = {:.8g}, F_std = {:.8g}, S = {:.5g}, E = {:.5g}, M = {:.5g}, sample_time = {:.3f}, train_time = {:.3f}, used_time = {:.3f}'
.format(
beta,
step,
free_energy_mean.item(),
free_energy_std.item(),
entropy_mean.item(),
energy_mean.item(),
mag_mean.item(),
sample_time,
train_time,
used_time,
))
sample_time = 0
train_time = 0
with open(args.fname, 'a', newline='\n') as f:
f.write('{} {} {:.3g} {:.8g} {:.8g} {:.8g} {:.8g}\n'.format(
args.n,
args.seed,
beta,
free_energy_mean.item(),
free_energy_std.item(),
energy_mean.item(),
entropy_mean.item(),
))
if args.ham == 'hop':
ensure_dir(args.out_filename + '_sample/')
np.savetxt('{}_sample/sample{:.2f}.txt'.format(
args.out_filename, beta),
sample.cpu().numpy(),
delimiter=' ',
fmt='%d')
np.savetxt('{}_sample/log_prob{:.2f}.txt'.format(
args.out_filename, beta),
log_prob.cpu().detach().numpy(),
delimiter=' ',
fmt='%.5f')
beta += args.beta_inc
def train(train_data, test_data, image_shape):
""" Trains MADE model on binary image dataset.
Arguments:
train_data: A (n_train, H, W, 1) uint8 numpy array of binary images with values in {0, 1}
test_data: An (n_test, H, W, 1) uint8 numpy array of binary images with values in {0, 1}
image_shape: (H, W), height and width of the image
Returns:
- a (# of training iterations,) numpy array of train_losses evaluated every minibatch
- a (# of epochs + 1,) numpy array of test_losses evaluated once at initialization and after each epoch
- a numpy array of size (100, H, W, 1) of samples with values in {0, 1}
"""
use_cuda = True
device = torch.device('cuda') if use_cuda else None
train_data = torch.from_numpy(
train_data.reshape(
(train_data.shape[0],
train_data.shape[1] * train_data.shape[2]))).float().to(device)
test_data = torch.from_numpy(
test_data.reshape(
(test_data.shape[0],
test_data.shape[1] * test_data.shape[2]))).float().to(device)
def nll_loss(batch, output):
return F.binary_cross_entropy(torch.sigmoid(output), batch)
H, W = image_shape
input_dim = H * W
made = MADE(input_dim)
epochs = 10
lr = 0.005
batch_size = 32
train_loader = torch.utils.data.DataLoader(train_data,
batch_size=batch_size,
shuffle=True)
optimizer = torch.optim.Adam(made.parameters(), lr=lr)
init_test_loss = nll_loss(test_data, made(test_data))
train_losses = []
test_losses = [init_test_loss.item()]
# Training
for epoch in range(epochs):
for batch in train_loader:
optimizer.zero_grad()
output = made(batch)
loss = nll_loss(batch, output)
loss.backward()
optimizer.step()
train_losses.append(loss.item())
test_loss = nll_loss(test_data, made(test_data))
test_losses.append(test_loss.item())
print(f'{epoch + 1}/{epochs} epochs')
# Generate samples
made.eval()
samples = torch.zeros(size=(100, H * W)).to(device)
with torch.no_grad():
for i in range(H * W):
logits = made(samples)
probas = torch.sigmoid(logits)
pixel_i_samples = torch.bernoulli(probas[:, i])
samples[:, i] = pixel_i_samples
return np.array(train_losses), np.array(test_losses), samples.reshape(
(100, H, W, 1)).detach().cpu().numpy()
def run(split, upto=None):
torch.set_grad_enabled(split=='train')
model.train() if split == 'train' else model.eval()
nsamples = 1 if split == 'train' else xte
N, D = x.size()
B = 128
n_steps = N // B if upto is None else min(N//B, upto)
losses = []
for step in range(n_steps):
xb = Variable(x[step * B: step * B + B])
xbhat = torch.zeros_like(xb)
for s in range(nsamples):
if step % args.resample_every == 0 or split == 'test':
model.update_masks()
xbhat += model(xb)
xbhat /= nsamples
loss = F.binary_cross_entropy_with_logits(xbhat, xb, size_average=False) / B
lossf = loss.data.item()
losses.append(lossf)
if split == 'train':
opt.zero_grad()
loss.backward()
opt.step()
print("%s epoch avg loss: %f" %(split, np.mean(losses)))
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('-d', '--data-path', required=True, type=str, help="Path to binarized_mnist.npz")
parser.add_argument('-q', '--hiddens', type=str, default='500', help="Comma separated sizes for hidden layers, e.g. 500, or 500,500")
parser.add_argument('-n', '--num-masks', type=int, default=1, help="Number of orderings for order/connection-agnostic training")
parser.add_argument('-r', '--resample-every', type=int, default=20, help="For efficiency we can choose to resample orders/masks only once every this many steps")
parser.add_argument('-s', '--samples', type=int, default=1, help="How many samples of connectivity/masks to average logits over during inference")
args = parser.parse_args()
np.random_seed(42)
torch.manual_seed(42)
torch.cuda.manual_seed_all(42)
print("loading binarized mnist from", args.data_path)
mnist = np.load(args.data_path)
xtr, xte = mnist['train_data'], mnist['valid_data']
xtr = torch.from_numpy(xtr).cuda()
xte = torch.from_numpy(xte).cuda()
# construct model and ship to GPU
hidden_list = list(map(int, args.hiddens.split(',')))
model = MADE(xtr.size(1), hidden_list, xtr.size(1), num_masks=args.num_masks)
print("number of model parameters:",sum([np.prod(p.size()) for p in model.parameters()]))
model.cuda()
# set up the optimizer
opt = torch.optim.Adam(model.parameters(), 1e-3, weight_decay=1e-4)
scheduler = torch.optim.lr_scheduler.StepLR(opt, step_size=45, gamma=0.1)
# start the training
for epoch in range(100):
print("epoch %d" % (epoch, ))
scheduler.step(epoch)
run_epoch('test', upto=5) # run only a few batches for approximate test accuracy
run_epoch('train')
print("optimization done. full test set eval:")
run_epoch('test')