def maml_supervised(model,
Task,
n_epochs,
task_batch_n,
npts,
k_shot,
n_gradient_steps,
test_fn=None,
**_):
"""
supervised MAML. Task need to implement .proper and .sample methods, where proper is the full,
dense set of data from which samples are drawn.
:param model:
:param Task:
:param n_epochs:
:param task_batch_n:
:param npts: the total number of samples for the sinusoidal task
:param k_shot:
:param n_gradient_steps:
:param _:
:return:
"""
import playground.maml.maml_torch.archive.paper_metrics as metrics
device = t.device('cuda' if t.cuda.is_available() else 'cpu')
model.to(device)
ps = list(model.parameters())
# for ep_ind in trange(n_epochs, desc='Epochs', ncols=50, leave=False):
M.tic('epoch')
for ep_ind in trange(n_epochs):
M.split('epoch', silent=True)
meta_grads = defaultdict(lambda: 0)
theta = copy.deepcopy(model.state_dict())
tasks = [Task(npts=npts) for _ in range(task_batch_n)]
for task_ind, task in enumerate(tasks): # sample a new problem
# todo: this part is highly-parallelizable
if task_ind != 0:
model.load_state_dict(theta)
task_grads = defaultdict(deque)
proper = t.tensor(task.proper()).to(device)
samples = t.tensor(task.samples(k_shot)).to(device)
for grad_ind in range(n_gradient_steps):
# done: ready to be repackaged
loss, _ = metrics.comp_loss(*samples, model)
model.zero_grad()
# back-propagate once, retain graph.
loss.backward(t.ones(1).to(device), retain_graph=True)
# done: need to use gradient descent, plus creating a meta graph.
U, grad_outputs = [], []
for p in model.parameters():
U.append(p - G.alpha * p.grad) # meta update
grad_outputs.append(t.ones(1).to(device).expand_as(p))
# t.autograd.grad returns sum of gradient between all U and all grad_outputs
# note: this is the row sum of \partial theta_prime \partial theta, which is a matrix.
dU = t.autograd.grad(outputs=U,
grad_outputs=grad_outputs,
inputs=model.parameters())
# Now update the param.figs
for p, updated_p, du in zip(ps, U, dU):
p.data = updated_p.data # these are leaf notes, so we can directly manipulate the data attribute.
task_grads[p].append(du)
# note: evaluate the 1-grad loss
if G.test_interval and ep_ind % G.test_interval and grad_ind + 1 in G.test_grad_steps:
with t.no_grad():
if test_fn is not None:
test_fn(grad_ind + 1,
model,
task=task,
epoch=ep_ind)
_loss, _ = metrics.comp_loss(*proper, model)
logger.log_keyvalue(
ep_ind,
key=f"{grad_ind + 1:d}-grad-loss-{task_ind:02d}",
value=_loss.item(),
silent=True)
# compute Loss_theta_prime
samples = t.tensor(task.samples(k_shot)).to(
device) # sample from this problem
loss, _ = metrics.comp_loss(*samples, model)
model.zero_grad()
loss.backward()
for i, grad in enumerate(
model.gradients()): # Now accumulate the gradient
p = ps[i]
task_grads[p].append(grad)
meta_grads[p] += t.prod(t.cat(list(
map(lambda d: d.unsqueeze(dim=-1), task_grads[p])),
dim=-1),
dim=-1)
# theta_prime = copy.deepcopy(model.state_dict())
model.load_state_dict(theta)
for p in ps:
p.grad = t.tensor(
(meta_grads[p] / task_batch_n).detach()).to(device)
model.meta_step(lr=G.beta)
with t.no_grad():
if test_fn is not None:
test_fn(0, model, epoch=ep_ind)
_loss, _ = metrics.comp_loss(*proper, model)
# it is very easy to use the wrong loss here.
logger.log_keyvalue(ep_ind, '0-grad-loss', _loss.item(), silent=True)
# save model weight
if G.save_interval and ep_ind % G.save_interval == 0:
logger.log_module(ep_ind, **{type(model).__name__: model})
def maml(model=None, test_fn=None):
from playground.maml.maml_torch.tasks import Sine
model = model or (StandardMLP(1, 1) if G.debug else FunctionalMLP(1, 1))
meta_optimizer = t.optim.Adam(model.parameters(), lr=G.beta)
mse = t.nn.MSELoss()
M.tic('start')
M.tic('epoch')
for ep_ind in range(G.n_epochs):
dt = M.split('epoch', silent=True)
dt_ = M.toc('start', silent=True)
print(f"epoch {ep_ind} @ {dt:.4f}sec/ep, {dt_:.1f} sec from start")
original_ps = OrderedDict(model.named_parameters())
tasks = [Sine() for _ in range(G.task_batch_n)]
for task_ind, task in enumerate(tasks):
if task_ind != 0:
model.params.update(original_ps)
if G.test_mode:
_gradient = original_ps['bias_var'].grad
if task_ind == 0:
assert _gradient is None or _gradient.sum().item(
) == 0, f"{_gradient} is not zero or None, epoch {ep_ind}."
else:
assert _gradient.sum().item(
) != 0, f"{_gradient} should be non-zero"
assert (
original_ps['bias_var'] == model.params['bias_var']
).all().item() == 1, 'the two parameters should be the same'
xs, labels = t.DoubleTensor(task.samples(G.k_shot))
_silent = task_ind != 0
for grad_ind in range(G.n_gradient_steps):
if hasattr(model,
"is_autoregressive") and model.is_autoregressive:
h0 = model.h0_init()
ys, ht = model(xs.view(G.k_shot, 1, 1), h0,
labels.view(G.k_shot, 1, 1))
ys = ys.squeeze(-1) # ys:Size[5, batch_n: 1, 1]
elif hasattr(model, "is_recurrent") and model.is_recurrent:
h0 = model.h0_init()
ys, ht = model(xs.view(G.k_shot, 1, 1), h0)
ys = ys.squeeze(-1) # ys:Size[5, batch_n: 1, 1]
else:
ys = model(xs.unsqueeze(-1))
ht = None
loss = mse(ys, labels.unsqueeze(-1))
logger.log_keyvalue(ep_ind,
f"{task_ind}-grad-{grad_ind}-loss",
loss.item(),
silent=_silent)
if callable(test_fn) and \
ep_ind % G.test_interval == 0 and grad_ind in G.test_grad_steps:
test_fn(model,
task=task,
task_id=task_ind,
epoch=ep_ind,
grad_step=grad_ind,
silent=_silent,
h0=ht)
dps = t.autograd.grad(loss,
model.parameters(),
create_graph=True,
retain_graph=True)
# 1. update parameters, use updated theta'.
# 2. run forward, get direct gradient to update the network
for (name, p), dp in zip(model.named_parameters(), dps):
model.params[name] = p - G.alpha * dp
grad_ind = G.n_gradient_steps
# meta gradient
if hasattr(model, "is_autoregressive") and model.is_autoregressive:
h0 = model.h0_init()
ys, ht = model(xs.view(G.k_shot, 1, 1), h0,
labels.view(G.k_shot, 1, 1))
ys = ys.squeeze(-1) # ys:Size[5, batch_n: 1, 1]
elif hasattr(model, "is_recurrent") and model.is_recurrent:
h0 = model.h0_init()
ys, ht = model(xs.view(G.k_shot, 1, 1), h0)
ys = ys.squeeze(-1) # ys:Size[5, batch_n: 1, 1]
else:
ys = model(xs.unsqueeze(-1))
ht = None
loss = mse(ys, labels.unsqueeze(-1))
logger.log_keyvalue(ep_ind,
f"{task_ind}-grad-{grad_ind}-loss",
loss.item(),
silent=_silent)
if callable(test_fn) and \
ep_ind % G.test_interval == 0 and grad_ind in G.test_grad_steps:
test_fn(model,
task=task,
task_id=task_ind,
epoch=ep_ind,
grad_step=grad_ind,
silent=_silent,
h0=ht)
meta_dps = t.autograd.grad(loss, original_ps.values())
with t.no_grad():
for (name, p), meta_dp in zip(original_ps.items(), meta_dps):
p.grad = (0 if p.grad is None else p.grad) + meta_dp
# normalize the gradient.
with t.no_grad():
for (name, p) in original_ps.items():
p.grad /= G.task_batch_n
model.params.update(original_ps)
meta_optimizer.step()
meta_optimizer.zero_grad()
if G.save_interval and ep_ind % G.save_interval == 0:
logger.log_module(ep_ind, **{type(model).__name__: model})
logger.flush()
def train():
from moleskin import moleskin as M
M.tic('Full Run')
if G.model == "lenet":
model = Conv2d()
elif G.model == 'mlp':
model = Mlp()
else:
raise NotImplementedError('only lenet and mlp are allowed')
model.train()
print(model)
G.log_prefix = f"mnist_{type(model).__name__}"
logger.configure(log_directory=G.log_dir, prefix=G.log_prefix)
logger.log_params(G=vars(G), Model=dict(architecture=str(model)))
from torchvision import datasets, transforms
trans = transforms.Compose(
[transforms.ToTensor(),
transforms.Normalize((0.5, ), (1.0, ))])
train_set = datasets.MNIST(root=G.data_dir,
train=True,
transform=trans,
download=True)
test_set = datasets.MNIST(root=G.data_dir,
train=False,
transform=trans,
download=True)
train_loader = torch.utils.data.DataLoader(dataset=train_set,
batch_size=G.batch_size,
shuffle=True)
test_loader = torch.utils.data.DataLoader(dataset=test_set,
batch_size=G.batch_size,
shuffle=False)
celoss = nn.CrossEntropyLoss()
adam = optim.SGD(model.parameters(), lr=G.learning_rate, momentum=0.9)
for epoch in range(G.n_epochs):
for it, (x, target) in enumerate(train_loader):
adam.zero_grad()
ys = model(x)
loss = celoss(ys, target)
loss.backward()
adam.step()
if it % G.test_interval == 0:
with h.Eval(model), torch.no_grad():
accuracy = h.Average()
for x, label in test_loader:
acc = h.cast(
h.one_hot_to_int(model(x).detach()) == label,
float).sum() / len(x)
accuracy.add(acc.detach().numpy())
logger.log(float(epoch) + it / len(train_loader),
accuracy=accuracy.value)
M.split("epoch")
# logger.log(epoch, it=it, loss=loss.detach().numpy())
M.toc('Full Run')
def reptile(model=None, test_fn=None):
from playground.maml.maml_torch.tasks import Sine
model = model or FunctionalMLP(1, 1)
meta_optimizer = t.optim.Adam(model.parameters(), lr=G.beta)
mse = t.nn.MSELoss()
M.tic('epoch')
for ep_ind in range(G.n_epochs):
M.split('epoch')
original_ps = OrderedDict(model.named_parameters())
tasks = [Sine() for _ in range(G.task_batch_n)]
for task_ind, task in enumerate(tasks):
if task_ind != 0:
model.params.update(original_ps)
xs, labels = t.DoubleTensor(task.samples(G.k_shot))
_silent = task_ind != 0
for grad_ind in range(G.n_gradient_steps):
if hasattr(model,
"is_autoregressive") and model.is_autoregressive:
h0 = model.h0_init()
ys, ht = model(xs.view(G.k_shot, 1, 1), h0,
labels.view(G.k_shot, 1, 1))
ys = ys.squeeze(-1) # ys:Size(5, batch_n:1, 1).
elif hasattr(model, "is_recurrent") and model.is_recurrent:
h0 = model.h0_init()
ys, ht = model(xs.view(G.k_shot, 1, 1), h0)
ys = ys.squeeze(-1) # ys:Size(5, batch_n:1, 1).
else:
ys = model(xs.unsqueeze(-1))
ht = None
loss = mse(ys, labels.unsqueeze(-1))
with t.no_grad():
logger.log_keyvalue(ep_ind,
f"{task_ind}-grad-{grad_ind}-loss",
loss.item(),
silent=_silent)
if callable(
test_fn
) and ep_ind % G.test_interval == 0 and grad_ind in G.test_grad_steps:
test_fn(model,
task,
task_id=task_ind,
epoch=ep_ind,
grad_step=grad_ind,
h0=ht,
silent=_silent)
dps = t.autograd.grad(loss, model.parameters())
with t.no_grad():
for (name, p), dp in zip(model.named_parameters(), dps):
model.params[name] = p - G.alpha * dp
model.params[name].requires_grad = True
grad_ind = G.n_gradient_steps
with t.no_grad():
# domain adaptation
if hasattr(model,
"is_autoregressive") and model.is_autoregressive:
h0 = model.h0_init()
ys, ht = model(xs.view(G.k_shot, 1, 1), h0,
labels.view(G.k_shot, 1, 1))
ys = ys.squeeze(-1) # ys:Size(5, batch_n:1, 1).
elif hasattr(model, "is_recurrent") and model.is_recurrent:
h0 = model.h0_init()
ys, ht = model(xs.view(G.k_shot, 1, 1), h0)
ys = ys.squeeze(-1) # ys:Size(5, batch_n:1, 1).
else:
ys = model(xs.unsqueeze(-1))
ht = None
loss = mse(ys, labels.unsqueeze(-1))
logger.log_keyvalue(ep_ind,
f"{task_ind}-grad-{grad_ind}-loss",
loss.item(),
silent=_silent)
if callable(test_fn) and \
ep_ind % G.test_interval == 0 and grad_ind in G.test_grad_steps:
test_fn(model,
task,
task_id=task_ind,
epoch=ep_ind,
grad_step=grad_ind,
h0=ht,
silent=_silent)
# Compute REPTILE 1st-order gradient
with t.no_grad():
for name, p in original_ps.items():
# let's do the division at the end.
p.grad = (0 if p.grad is None else p.grad) + (
p - model.params[name]) # / G.task_batch_n
with t.no_grad():
for name, p in original_ps.items():
# let's do the division at the end.
p.grad /= G.task_batch_n
model.params.update(original_ps)
meta_optimizer.step()
meta_optimizer.zero_grad()
if G.save_interval and ep_ind % G.save_interval == 0:
logger.log_module(ep_ind, **{type(model).__name__: model})
logger.flush()
def train():
from linear_schedule import Linear
ledger = defaultdict(lambda: MovingAverage(Reporting.reward_average))
M.config(file=os.path.join(RUN.log_directory, RUN.log_file))
M.diff()
with U.make_session(
RUN.num_cpu), Logger(RUN.log_directory) as logger, contextify(
gym.make(G.env_name)) as env:
env = ScaledFloatFrame(wrap_dqn(env))
if G.seed is not None:
env.seed(G.seed)
logger.log_params(G=vars(G), RUN=vars(RUN), Reporting=vars(Reporting))
inputs = TrainInputs(action_space=env.action_space,
observation_space=env.observation_space)
trainer = QTrainer(inputs=inputs,
action_space=env.action_space,
observation_space=env.observation_space)
if G.prioritized_replay:
replay_buffer = PrioritizedReplayBuffer(size=G.buffer_size,
alpha=G.alpha)
else:
replay_buffer = ReplayBuffer(size=G.buffer_size)
class schedules:
# note: it is important to have this start from the begining.
eps = Linear(G.n_timesteps * G.exploration_fraction, 1,
G.final_eps)
if G.prioritized_replay:
beta = Linear(G.n_timesteps - G.learning_start, G.beta_start,
G.beta_end)
U.initialize()
trainer.update_target()
x = np.array(env.reset())
ep_ind = 0
M.tic('episode')
for t_step in range(G.n_timesteps):
# schedules
eps = 0 if G.param_noise else schedules.eps[t_step]
if G.prioritized_replay:
beta = schedules.beta[t_step - G.learning_start]
x0 = x
M.tic('sample', silent=True)
(action, *_), action_q, q = trainer.runner.act([x], eps)
x, rew, done, info = env.step(action)
ledger['action_q_value'].append(action_q.max())
ledger['action_q_value/mean'].append(action_q.mean())
ledger['action_q_value/var'].append(action_q.var())
ledger['q_value'].append(q.max())
ledger['q_value/mean'].append(q.mean())
ledger['q_value/var'].append(q.var())
ledger['timing/sample'].append(M.toc('sample', silent=True))
# note: adding sample to the buffer is identical between the prioritized and the standard replay strategy.
replay_buffer.add(s0=x0,
action=action,
reward=rew,
s1=x,
done=float(done))
logger.log(
t_step, {
'q_value': ledger['q_value'].latest,
'q_value/mean': ledger['q_value/mean'].latest,
'q_value/var': ledger['q_value/var'].latest,
'q_value/action': ledger['action_q_value'].latest,
'q_value/action/mean':
ledger['action_q_value/mean'].latest,
'q_value/action/var': ledger['action_q_value/var'].latest
},
action=action,
eps=eps,
silent=True)
if G.prioritized_replay:
logger.log(t_step, beta=beta, silent=True)
if done:
ledger['timing/episode'].append(M.split('episode',
silent=True))
ep_ind += 1
x = np.array(env.reset())
ledger['rewards'].append(info['total_reward'])
silent = (ep_ind % Reporting.print_interval != 0)
logger.log(t_step,
timestep=t_step,
episode=green(ep_ind),
total_reward=ledger['rewards'].latest,
episode_length=info['timesteps'],
silent=silent)
logger.log(t_step, {
'total_reward/mean':
yellow(ledger['rewards'].mean, lambda v: f"{v:.1f}"),
'total_reward/max':
yellow(ledger['rewards'].max, lambda v: f"{v:.1f}"),
"time_spent_exploring":
default(eps, percent),
"timing/episode":
green(ledger['timing/episode'].latest, sec),
"timing/episode/mean":
green(ledger['timing/episode'].mean, sec),
},
silent=silent)
try:
logger.log(t_step, {
"timing/sample":
default(ledger['timing/sample'].latest, sec),
"timing/sample/mean":
default(ledger['timing/sample'].mean, sec),
"timing/train":
default(ledger['timing/train'].latest, sec),
"timing/train/mean":
green(ledger['timing/train'].mean, sec),
"timing/log_histogram":
default(ledger['timing/log_histogram'].latest, sec),
"timing/log_histogram/mean":
default(ledger['timing/log_histogram'].mean, sec)
},
silent=silent)
if G.prioritized_replay:
logger.log(t_step, {
"timing/update_priorities":
default(ledger['timing/update_priorities'].latest,
sec),
"timing/update_priorities/mean":
default(ledger['timing/update_priorities'].mean,
sec)
},
silent=silent)
except Exception as e:
pass
if G.prioritized_replay:
logger.log(
t_step,
{"replay_beta": default(beta, lambda v: f"{v:.2f}")},
silent=silent)
# note: learn here.
if t_step >= G.learning_start and t_step % G.learn_interval == 0:
if G.prioritized_replay:
experiences, weights, indices = replay_buffer.sample(
G.replay_batch_size, beta)
logger.log_histogram(t_step, weights=weights)
else:
experiences, weights = replay_buffer.sample(
G.replay_batch_size), None
M.tic('train', silent=True)
x0s, actions, rewards, x1s, dones = zip(*experiences)
td_error_val, loss_val = trainer.train(s0s=x0s,
actions=actions,
rewards=rewards,
s1s=x1s,
dones=dones,
sample_weights=weights)
ledger['timing/train'].append(M.toc('train', silent=True))
M.tic('log_histogram', silent=True)
logger.log_histogram(t_step, td_error=td_error_val)
ledger['timing/log_histogram'].append(
M.toc('log_histogram', silent=True))
if G.prioritized_replay:
M.tic('update_priorities', silent=True)
new_priorities = np.abs(td_error_val) + eps
replay_buffer.update_priorities(indices, new_priorities)
ledger['timing/update_priorities'].append(
M.toc('update_priorities', silent=True))
if t_step % G.target_network_update_interval == 0:
trainer.update_target()
if t_step % Reporting.checkpoint_interval == 0:
U.save_state(os.path.join(RUN.log_directory, RUN.checkpoint))