def eval_cavia(args,
model,
task_family,
num_updates,
n_tasks=100,
return_gradnorm=False,
encoder=False,
p_encoder=False):
global temp
# get the task family
input_range = task_family.get_input_range().to(args.device)
# logging
losses = []
gradnorms = []
# --- inner loop ---
for t in range(n_tasks):
# sample a task
target_function, ty = task_family.sample_task(True)
# reset context parameters
model.reset_context_params()
# get data for current task
curr_inputs = task_family.sample_inputs(
args.k_shot_eval, args.use_ordered_pixels).to(args.device)
curr_targets = target_function(curr_inputs)
train_inputs = torch.cat([curr_inputs, curr_targets], dim=1)
a = encoder(train_inputs)
#embedding,_ = torch.max(a,dim=0)
embedding = torch.mean(a, dim=0)
#model.set_context_params(embedding)
logits = p_encoder(embedding)
logits = logits.reshape([latent_dim, categorical_dim])
y = gumbel_softmax(logits, temp, hard=True)
y = y[:, 1]
# ------------ update on current task ------------
for _ in range(1, num_updates + 1):
# forward pass
curr_outputs = model(curr_inputs)
# compute loss for current task
task_loss = F.mse_loss(curr_outputs, curr_targets)
# compute gradient wrt context params
task_gradients = \
torch.autograd.grad(task_loss, model.context_params, create_graph=not args.first_order)[0]
# update context params
if args.first_order:
model.context_params = model.context_params - args.lr_inner * task_gradients.detach(
) * y
else:
model.context_params = model.context_params - args.lr_inner * task_gradients * y
# keep track of gradient norms
gradnorms.append(task_gradients[0].norm().item())
# ------------ logging ------------
# compute true loss on entire input range
model.eval()
losses.append(
F.mse_loss(model(input_range),
target_function(input_range)).detach().item())
model.train()
losses_mean = np.mean(losses)
losses_conf = st.t.interval(0.95,
len(losses) - 1,
loc=losses_mean,
scale=st.sem(losses))
if not return_gradnorm:
return losses_mean, np.mean(np.abs(losses_conf - losses_mean))
else:
return losses_mean, np.mean(np.abs(losses_conf -
losses_mean)), np.mean(gradnorms)
def run(args, log_interval=5000, rerun=False):
global temp
assert not args.maml
# see if we already ran this experiment
code_root = os.path.dirname(os.path.realpath(__file__))
if not os.path.isdir('{}/{}_result_files/'.format(code_root, args.task)):
os.mkdir('{}/{}_result_files/'.format(code_root, args.task))
path = '{}/{}_result_files/'.format(
code_root, args.task) + utils.get_path_from_args(args)
if os.path.exists(path + '.pkl') and not rerun:
return utils.load_obj(path)
start_time = time.time()
utils.set_seed(args.seed)
# --- initialise everything ---
# get the task family
if args.task == 'sine':
task_family_train = tasks_sine.RegressionTasksSinusoidal()
task_family_valid = tasks_sine.RegressionTasksSinusoidal()
task_family_test = tasks_sine.RegressionTasksSinusoidal()
elif args.task == 'celeba':
task_family_train = tasks_celebA.CelebADataset('train',
device=args.device)
task_family_valid = tasks_celebA.CelebADataset('valid',
device=args.device)
task_family_test = tasks_celebA.CelebADataset('test',
device=args.device)
elif args.task == 'multi':
task_family_train = multi()
task_family_valid = multi()
task_family_test = multi()
else:
raise NotImplementedError
# initialise network
model = CaviaModel(n_in=task_family_train.num_inputs,
n_out=task_family_train.num_outputs,
num_context_params=args.num_context_params,
n_hidden=args.num_hidden_layers,
device=args.device).to(args.device)
# intitialise meta-optimiser
# (only on shared params - context parameters are *not* registered parameters of the model)
meta_optimiser = optim.Adam(model.parameters(), args.lr_meta)
encoder = pool_encoder().to(args.device)
encoder_optimiser = optim.Adam(encoder.parameters(), lr=1e-3)
decoder = pool_decoder().to(args.device)
decoder_optimiser = optim.Adam(decoder.parameters(), lr=1e-3)
#encoder.load_state_dict(torch.load('./model/encoder'))
p_encoder = place().to(args.device)
p_optimiser = optim.Adam(p_encoder.parameters(), lr=1e-3)
# initialise loggers
logger = Logger()
logger.best_valid_model = copy.deepcopy(model)
# --- main training loop ---
for i_iter in range(args.n_iter):
# initialise meta-gradient
meta_gradient = [0 for _ in range(len(model.state_dict()))]
place_gradient = [0 for _ in range(len(p_encoder.state_dict()))]
encoder_gradient = [0 for _ in range(len(encoder.state_dict()))]
#print(meta_gradient)
# sample tasks
target_functions, ty = task_family_train.sample_tasks(
args.tasks_per_metaupdate, True)
# --- inner loop ---
for t in range(args.tasks_per_metaupdate):
# reset private network weights
model.reset_context_params()
# get data for current task
x = task_family_train.sample_inputs(
args.k_meta_train, args.use_ordered_pixels).to(args.device)
y = target_functions[t](x)
train_inputs = torch.cat([x, y], dim=1)
a = encoder(train_inputs)
#embedding,_ = torch.max(a,dim=0)
embedding = torch.mean(a, dim=0)
logits = p_encoder(embedding)
logits = logits.reshape([latent_dim, categorical_dim])
y = gumbel_softmax(logits, temp, hard=True)
y = y[:, 1]
#print(temp)
#model.set_context_params(embedding)
#print(model.context_params)
for _ in range(args.num_inner_updates):
# forward through model
train_outputs = model(x)
# get targets
train_targets = target_functions[t](x)
# ------------ update on current task ------------
# compute loss for current task
task_loss = F.mse_loss(train_outputs, train_targets)
# compute gradient wrt context params
task_gradients = \
torch.autograd.grad(task_loss, model.context_params, create_graph=not args.first_order)[0]
# update context params (this will set up the computation graph correctly)
model.context_params = model.context_params - args.lr_inner * task_gradients * y
#print(model.context_params)
# ------------ compute meta-gradient on test loss of current task ------------
# get test data
test_inputs = task_family_train.sample_inputs(
args.k_meta_test, args.use_ordered_pixels).to(args.device)
# get outputs after update
test_outputs = model(test_inputs)
# get the correct targets
test_targets = target_functions[t](test_inputs)
# compute loss after updating context (will backprop through inner loop)
loss_meta = F.mse_loss(test_outputs, test_targets)
#print(torch.norm(y,1)/1000)
#loss_meta += torch.norm(y,1)/700
qy = F.softmax(logits, dim=-1)
log_ratio = torch.log(qy * categorical_dim + 1e-20)
KLD = torch.sum(qy * log_ratio, dim=-1).mean() / 5
# print(KLD)
loss_meta += KLD
# compute gradient + save for current task
task_grad = torch.autograd.grad(loss_meta,
model.parameters(),
retain_graph=True)
for i in range(len(task_grad)):
# clip the gradient
meta_gradient[i] += task_grad[i].detach().clamp_(-10, 10)
task_grad_place = torch.autograd.grad(loss_meta,
p_encoder.parameters(),
retain_graph=True)
for i in range(len(task_grad_place)):
# clip the gradient
place_gradient[i] += task_grad_place[i].detach().clamp_(
-10, 10)
task_grad_encoder = torch.autograd.grad(loss_meta,
encoder.parameters())
for i in range(len(task_grad_encoder)):
# clip the gradient
encoder_gradient[i] += task_grad_encoder[i].detach().clamp_(
-10, 10)
# ------------ meta update ------------
# assign meta-gradient
for i, param in enumerate(model.parameters()):
param.grad = meta_gradient[i] / args.tasks_per_metaupdate
meta_optimiser.step()
# do update step on shared model
for i, param in enumerate(p_encoder.parameters()):
param.grad = place_gradient[i] / args.tasks_per_metaupdate
p_optimiser.step()
for i, param in enumerate(encoder.parameters()):
param.grad = encoder_gradient[i] / args.tasks_per_metaupdate
encoder_optimiser.step()
# reset context params
model.reset_context_params()
if i_iter % 350 == 1:
temp = np.maximum(temp * np.exp(-ANNEAL_RATE * i_iter), 0.5)
print(temp)
# ------------ logging ------------
if i_iter % log_interval == 0:
# evaluate on training set
loss_mean, loss_conf = eval_cavia(
args,
copy.deepcopy(model),
task_family=task_family_train,
num_updates=args.num_inner_updates,
encoder=encoder,
p_encoder=p_encoder)
logger.train_loss.append(loss_mean)
logger.train_conf.append(loss_conf)
# evaluate on test set
loss_mean, loss_conf = eval_cavia(
args,
copy.deepcopy(model),
task_family=task_family_valid,
num_updates=args.num_inner_updates,
encoder=encoder,
p_encoder=p_encoder)
logger.valid_loss.append(loss_mean)
logger.valid_conf.append(loss_conf)
# evaluate on validation set
if i_iter % log_interval == 0:
loss_mean, loss_conf = eval_cavia(
args,
copy.deepcopy(model),
task_family=task_family_test,
num_updates=args.num_inner_updates,
encoder=encoder,
p_encoder=p_encoder)
logger.test_loss.append(loss_mean)
logger.test_conf.append(loss_conf)
# save logging results
utils.save_obj(logger, path)
# save best model
if logger.valid_loss[-1] == np.min(logger.valid_loss):
print('saving best model at iter', i_iter)
logger.best_valid_model = copy.deepcopy(model)
logger.best_encoder_valid_model = copy.deepcopy(encoder)
logger.best_place_valid_model = copy.deepcopy(p_encoder)
if i_iter % (4 * log_interval) == 0:
print('saving model at iter', i_iter)
logger.valid_model.append(copy.deepcopy(model))
logger.encoder_valid_model.append(copy.deepcopy(encoder))
logger.place_valid_model.append(copy.deepcopy(p_encoder))
# visualise results
if args.task == 'celeba':
task_family_train.visualise(
task_family_train, task_family_test,
copy.deepcopy(logger.best_valid_model), args, i_iter)
# print current results
logger.print_info(i_iter, start_time)
start_time = time.time()
return logger