def model_analyzer(optim,
mode,
model_train,
params,
model_cls,
set_size,
data,
iter,
optim_it,
analyze_mask=False,
sample_mask=False,
draw_loss=False):
text_dir = 'analyze_model/analyze_mask'
result_dir = 'analyze_model/drawloss'
sample_dir = 'analyze_model/mask_compare'
mask_dict = ResultDict()
iter_interval = 10
sample_num = 10000
is_mask_dict = False
if mode == 'test' and analyze_mask:
analyzing_mask(optim.mask_gen, set_size, mode, iter, iter_interval,
text_dir)
if mode == 'test' and sample_mask:
mask_result = sampling_mask(optim.mask_gen, set_size, model_train,
params, sample_num, mode, iter,
iter_interval, sample_dir)
if mask_result is not None:
mask_dict.append(mask_result)
is_mask_dict = True
if mode == 'test' and draw_loss:
plot_loss(model_cls=model_cls,
model=model_train,
params=params,
input_data=data['in_train'].load(),
dataset=data['in_train'],
feature_gen=optim.feature_gen,
mask_gen=optim.mask_gen,
step_gen=optim.step_gen,
scale_way=None,
xmin=-2.0,
xmax=0.5,
num_x=20,
mode=mode,
iteration=iter,
iter_interval=iter_interval,
loss_dir=result_dir)
if sample_mask and is_mask_dict:
plot_mask_result(mask_dict, sample_dir)
return
def meta_optimize(self,
cfg,
meta_optim,
data,
model_cls,
writer=None,
mode='train'):
assert mode in ['train', 'valid', 'test']
self.set_mode(mode)
# mask_mode = ['no_mask', 'structured', 'unstructured'][2]
# data_mode = ['in_train', 'in_test'][0]
############################################################################
analyze_model = False
analyze_surface = False
############################################################################
result_dict = ResultDict()
unroll_losses = 0
walltime = Walltime()
test_kld = torch.tensor(0.)
params = C(model_cls()).params
self.feature_gen.new()
self.step_gen.new()
sparse_r = {} # sparsity
iter_pbar = tqdm(range(1, cfg['iter_' + mode] + 1), 'Inner_loop')
for iter in iter_pbar:
debug_1 = sigint.is_active(iter == 1 or iter % 10 == 0)
debug_2 = sigstp.is_active()
with WalltimeChecker(walltime):
model_train = C(model_cls(params.detach()))
data_train = data['in_train'].load()
train_nll, train_acc = model_train(*data_train)
train_nll.backward()
grad = model_train.params.grad.detach()
# g = model_train.params.grad.flat.detach()
# w = model_train.params.flat.detach()
# step & mask genration
feature, v_sqrt = self.feature_gen(grad.flat.detach())
# step = self.step_gen(feature, v_sqrt, debug=debug_1)
# step = params.new_from_flat(step[0])
size = params.size().unflat()
if cfg.mask_mode == 'structured':
mask = self.mask_gen(feature, size, debug=debug_1)
mask = ParamsFlattener(mask)
mask_layout = mask.expand_as(params)
params = params + grad.detach() * mask_layout * (
-cfg.inner_lr)
elif cfg.mask_mode == 'unstructured':
mask_flat = self.mask_gen.unstructured(feature, size)
mask = params.new_from_flat(mask_flat)
params = params + grad.detach() * mask * (-cfg.inner_lr)
# update = params.new_from_flat(params.flat + grad.flat.detach() * mask * (-cfg.lr))
# params = params + update
elif cfg.mask_mode == 'no_mask':
params = params + grad.detach() * (-cfg.inner_lr)
else:
raise Exception('Unknown setting!')
# import pdb; pdb.set_trace()
# step_masked = step * mask_layout
# params = params + step_masked
with WalltimeChecker(walltime if mode == 'train' else None):
model_test = C(model_cls(params))
if cfg.data_mode == 'in_train':
data_test = data_train
elif cfg.data_mode == 'in_test':
data_test = data['in_test'].load()
test_nll, test_acc = utils.isnan(*model_test(*data_test))
if debug_2: pdb.set_trace()
if mode == 'train':
unroll_losses += test_nll # + test_kld
if iter % cfg.unroll == 0:
meta_optim.zero_grad()
unroll_losses.backward()
nn.utils.clip_grad_value_(self.parameters(), 0.01)
meta_optim.step()
unroll_losses = 0
with WalltimeChecker(walltime):
if not mode == 'train' or iter % cfg.unroll == 0:
params = params.detach_()
##########################################################################
if analyze_model:
analyzers.model_analyzer(self,
mode,
model_train,
params,
model_cls,
mask.tsize(0),
data,
iter,
optim_it,
analyze_mask=True,
sample_mask=True,
draw_loss=False)
if analyze_surface:
analyzers.surface_analyzer(params, best_mask, step, writer,
iter)
##########################################################################
result = dict(
train_nll=train_nll.tolist(),
test_nll=test_nll.tolist(),
train_acc=train_acc.tolist(),
test_acc=test_acc.tolist(),
test_kld=test_kld.tolist(),
walltime=walltime.time,
)
if not cfg.mask_mode == 'no_mask':
result.update(
**mask.sparsity(overall=True),
**mask.sparsity(overall=False),
)
result_dict.append(result)
log_pbar(result, iter_pbar)
return result_dict, params
def meta_optimize(self,
meta_optim,
data,
model_cls,
optim_it,
unroll,
out_mul,
k_obsrv=0,
no_mask=False,
writer=None,
mode='train'):
assert mode in ['train', 'valid', 'test']
self.set_mode(mode)
############################################################################
analyze_model = False
analyze_surface = False
############################################################################
result_dict = ResultDict()
unroll_losses = 0
walltime = Walltime()
test_kld = torch.tensor(0.)
params = C(model_cls()).params
self.feature_gen.new()
self.step_gen.new()
sparse_r = {} # sparsity
iter_pbar = tqdm(range(1, optim_it + 1), 'Inner_loop')
for iter in iter_pbar:
debug_1 = sigint.is_active(iter == 1 or iter % 10 == 0)
debug_2 = sigstp.is_active()
with WalltimeChecker(walltime):
model_train = C(model_cls(params.detach()))
data_ = data['in_train'].load()
train_nll, train_acc = model_train(*data_)
train_nll.backward()
g = model_train.params.grad.flat.detach()
w = model_train.params.flat.detach()
# step & mask genration
feature, v_sqrt = self.feature_gen(g)
step = self.step_gen(feature, v_sqrt, debug=debug_1)
step = params.new_from_flat(step[0])
size = params.size().unflat()
if no_mask:
params = params + step
else:
kld = self.mask_gen(feature, size, debug=debug_1)
test_kld = kld / data['in_test'].full_size / unroll
## kl annealing function 'linear' / 'logistic' / None
test_kld2 = test_kld * kl_anneal_function(
anneal_function=None, step=iter, k=0.0025, x0=optim_it)
mask = self.mask_gen.sample_mask()
mask = ParamsFlattener(mask)
mask_layout = mask.expand_as(params)
import pdb
pdb.set_trace()
step_masked = step * mask_layout
params = params + step_masked
with WalltimeChecker(walltime if mode == 'train' else None):
model_test = C(model_cls(params))
test_nll, test_acc = utils.isnan(*model_test(
*data['in_test'].load()))
if debug_2: pdb.set_trace()
if mode == 'train':
unroll_losses += test_nll # + test_kld
if iter % unroll == 0:
meta_optim.zero_grad()
unroll_losses.backward()
nn.utils.clip_grad_value_(self.parameters(), 0.01)
meta_optim.step()
unroll_losses = 0
with WalltimeChecker(walltime):
if not mode == 'train' or iter % unroll == 0:
params = params.detach_()
##########################################################################
if analyze_model:
analyzers.model_analyzer(self,
mode,
model_train,
params,
model_cls,
mask.tsize(0),
data,
iter,
optim_it,
analyze_mask=True,
sample_mask=True,
draw_loss=False)
if analyze_surface:
analyzers.surface_analyzer(params, best_mask, step, writer,
iter)
##########################################################################
result = dict(
train_nll=train_nll.tolist(),
test_nll=test_nll.tolist(),
train_acc=train_acc.tolist(),
test_acc=test_acc.tolist(),
test_kld=test_kld.tolist(),
walltime=walltime.time,
)
if no_mask is False:
result.update(
**mask.sparsity(overall=True),
**mask.sparsity(overall=False),
)
result_dict.append(result)
log_pbar(result, iter_pbar)
return result_dict, params
def meta_optimize(self,
meta_optimizer,
data_cls,
model_cls,
optim_it,
unroll,
out_mul,
should_train=True):
data = data_cls(training=should_train)
model = C(model_cls())
g_states = C(
OptimizerStates.initial_zeros(
(self.n_layers, len(model.params), self.hidden_sz)))
if should_train:
self.train()
else:
self.eval()
result_dict = ResultDict()
model_dt = None
grad_losses = 0
train_with_lazy_tf = True
eval_test_grad_loss = True
grad_prev = C(torch.zeros(model.params.size().flat()))
update_prev = C(torch.zeros(model.params.size().flat()))
#grad_real_prev = None
grad_real_cur = None
grad_real_prev = C(torch.zeros(model.params.size().flat()))
iter_pbar = tqdm(range(1, optim_it + 1), 'optim_iteration')
iter_watch = utils.StopWatch('optim_iteration')
"""grad_real_prev:
teacher-forcing current input of RNN.
required during all of training steps + every 'k' test steps.
grad_real_cur:
MSE loss target for current output of RNN.
required during all the training steps only.
"""
batch_arg = BatchManagerArgument(
params=model.params,
grad_prev=grad_prev,
update_prev=update_prev,
g_states=StatesSlicingWrapper(g_states),
mse_loss=C(torch.zeros(1)),
sigma=C(torch.zeros(model.params.size().flat())),
#updates=OptimizerBatchManager.placeholder(model.params.flat.size()),
)
for iteration in iter_pbar:
data_ = data.sample()
# if grad_real_cur is not None: # if iteration % self.k == 0
# grad_cur_prev = grad_real_cur # back up previous grad
if (iteration +
1) % self.k == 0 or should_train or eval_test_grad_loss:
# get grad every step while training
# get grad one step earlier than teacher-forcing steps while testing
model_dt = C(model_cls(params=model.params.detach()))
model_loss_dt = model_dt(*data_)
# model.params.register_grad_cut_hook_()
model_loss_dt.backward()
# model.params.remove_grad_cut_hook_()
grad_real_cur = model_dt.params.flat.grad
assert grad_real_cur is not None
if should_train or eval_test_grad_loss:
# mse loss target
# grad_real_cur is not used as a mse target while testing
assert grad_real_cur is not None
batch_arg.update(
BatchManagerArgument(
grad_real_cur=grad_real_cur).immutable())
if iteration % self.k == 0 or (should_train
and not train_with_lazy_tf):
# teacher-forcing
assert grad_real_prev is not None
batch_arg.update(
BatchManagerArgument(
grad_real_prev=grad_real_prev).immutable().volatile())
grad_real_prev = None
########################################################################
batch_manager = OptimizerBatchManager(batch_arg=batch_arg)
for get, set in batch_manager.iterator():
if iteration % self.k == 0 or (should_train
and not train_with_lazy_tf):
grad_prev = get.grad_real_prev # teacher-forcing
else:
grad_prev = get.grad_prev # free-running
grad_cur, g_states = self(grad_prev.detach(),
get.update_prev.detach(),
get.params.detach(), get.g_states)
set.grad_prev(grad_cur)
set.g_states(g_states)
set.params(get.params - 1.0 * grad_cur.detach())
set.sigma(self.sigma)
if should_train or eval_test_grad_loss:
set.mse_loss(
self.mse_loss(grad_cur, get.grad_real_cur.detach()))
##########################################################################
batch_arg = batch_manager.batch_arg_to_set
grad_real = batch_manager.batch_arg_to_get.grad_real_cur.detach()
cosim_loss = nn.functional.cosine_similarity(batch_arg.grad_prev,
grad_real,
dim=0)
grad_loss = batch_arg.mse_loss * 100
cosim_loss = -torch.log((cosim_loss + 1) * 0.5) * 100
grad_losses += (grad_loss + cosim_loss)
#to_float = lambda x: float(x.data.cpu().numpy())
sigma = batch_arg.sigma.detach().mean()
iter_pbar.set_description(
'optim_iteration[loss(model):{:10.6f}/(mse):{:10.6f}/(cosim):{:10.6f}/(sigma):{:10.6f}]'
''.format(model_loss_dt.tolist(),
grad_loss.tolist()[0],
cosim_loss.tolist()[0], sigma.tolist()))
if should_train and iteration % unroll == 0:
w_decay_loss = 0
for p in self.parameters():
w_decay_loss += torch.sum(p * p)
w_decay_loss = (w_decay_loss * 0.5) * 1e-4
total_loss = grad_losses + w_decay_loss
meta_optimizer.zero_grad()
total_loss.backward()
meta_optimizer.step()
if not should_train or iteration % unroll == 0:
model_losses = 0
grad_losses = 0
batch_arg.detach_()
# batch_arg.params.detach_()
# batch_arg.g_states.detach_()
# batch_arg.u_states.detach_()
result_dict.append(
step_num=iteration,
loss=model_loss_dt,
grad_loss=grad_loss,
)
if not should_train:
result_dict.append(walltime=iter_watch.touch(),
**self.params_tracker(
grad_real=grad_real_cur,
grad_pred=batch_arg.grad_prev,
update=batch_arg.update_prev,
sigma=batch_arg.sigma,
))
if (iteration +
1) % self.k == 0 or should_train or eval_test_grad_loss:
grad_real_prev = grad_real_cur
grad_real_cur = None
return result_dict
def meta_optimize(self,
meta_optim,
data,
model_cls,
optim_it,
unroll,
out_mul,
k_obsrv=1,
no_mask=False,
writer=None,
mode='train'):
assert mode in ['train', 'valid', 'test']
if no_mask is True:
raise Exception(
"this module currently does NOT suport no_mask option")
self.set_mode(mode)
############################################################################
n_samples = k_obsrv
"""MSG: better postfix?"""
analyze_model = False
analyze_surface = False
############################################################################
if analyze_surface:
writer.new_subdirs('best', 'any', 'rand', 'inv', 'dense')
result_dict = ResultDict()
unroll_losses = 0
walltime = Walltime()
test_kld = torch.tensor(0.)
params = C(model_cls()).params
self.feature_gen.new()
self.step_gen.new()
iter_pbar = tqdm(range(1, optim_it + 1), 'Inner_loop')
set_size = {'layer_0': 500, 'layer_1': 10} # NOTE: make it smarter
for iter in iter_pbar:
debug_1 = sigint.is_active(iter == 1 or iter % 10 == 0)
debug_2 = sigstp.is_active()
best_loss = 9999999
best_params = None
with WalltimeChecker(walltime):
model_train = C(model_cls(params.detach()))
train_nll, train_acc = model_train(*data['in_train'].load())
train_nll.backward()
g = model_train.params.grad.flat.detach()
w = model_train.params.flat.detach()
feature, v_sqrt = self.feature_gen(g)
size = params.size().unflat()
kld = self.mask_gen(feature, size)
losses = []
lips = []
valid_mask_patience = 100
assert n_samples > 0
"""FIX LATER:
when n_samples == 0 it can behave like no_mask flag is on."""
for i in range(n_samples):
# step & mask genration
for j in range(valid_mask_patience):
mask = self.mask_gen.sample_mask()
mask = ParamsFlattener(mask)
if mask.is_valid_sparsity():
if j > 0:
print(
f'\n\n[!]Resampled {j + 1} times to get valid mask!'
)
break
if j == valid_mask_patience - 1:
raise Exception(
"[!]Could not sample valid mask for "
f"{j+1} trials.")
step_out = self.step_gen(feature, v_sqrt, debug=debug_1)
step = params.new_from_flat(step_out[0])
mask_layout = mask.expand_as(params)
step_sparse = step * mask_layout
params_sparse = params + step_sparse
params_pruned = params_sparse.prune(mask > 0.5)
if params_pruned.size().unflat()['mat_0'][1] == 0:
continue
# cand_loss = model(*outer_data_s)
sparse_model = C(model_cls(params_pruned.detach()))
loss, _ = sparse_model(*data['in_train'].load())
if (loss < best_loss) or i == 0:
best_loss = loss
best_params = params_sparse
best_pruned = params_pruned
best_mask = mask
if best_params is not None:
params = best_params
with WalltimeChecker(walltime if mode == 'train' else None):
model_test = C(model_cls(params))
test_nll, test_acc = utils.isnan(*model_test(
*data['in_test'].load()))
test_kld = kld / data['in_test'].full_size / unroll
## kl annealing function 'linear' / 'logistic' / None
test_kld2 = test_kld * kl_anneal_function(
anneal_function=None, step=iter, k=0.0025, x0=optim_it)
total_test = test_nll + test_kld2
if mode == 'train':
unroll_losses += total_test
if iter % unroll == 0:
meta_optim.zero_grad()
unroll_losses.backward()
nn.utils.clip_grad_value_(self.parameters(), 0.01)
meta_optim.step()
unroll_losses = 0
with WalltimeChecker(walltime):
if not mode == 'train' or iter % unroll == 0:
params = params.detach_()
##########################################################################
"""Analyzers"""
if analyze_model:
analyzers.model_analyzer(self,
mode,
model_train,
params,
model_cls,
set_size,
data,
iter,
optim_it,
analyze_mask=True,
sample_mask=True,
draw_loss=False)
if analyze_surface:
analyzers.surface_analyzer(params, best_mask, step, writer,
iter)
##########################################################################
result = dict(
train_nll=train_nll.tolist(),
test_nll=test_nll.tolist(),
train_acc=train_acc.tolist(),
test_acc=test_acc.tolist(),
test_kld=test_kld.tolist(),
walltime=walltime.time,
**best_mask.sparsity(overall=True),
**best_mask.sparsity(overall=False),
)
result_dict.append(result)
log_pbar(result, iter_pbar)
return result_dict, params
def loop(mode, data, outer_steps, inner_steps, log_steps, fig_epochs, inner_lr,
log_mask=True, unroll_steps=None, meta_batchsize=0, sampler=None,
epoch=1, outer_optim=None, save_path=None, is_RL=False):
#####################################################################
"""Args:
meta_batchsize(int): If meta_batchsize |m| > 0, gradients for multiple
unrollings from each episodes size of |m| will be accumulated in
sequence but updated all at once. (Which can be done in parallel when
VRAM is large enough, but will be simulated in this code.)
If meta_batchsize |m| = 0(default), then update will be performed after
each unrollings.
"""
assert mode in ['train', 'valid', 'test']
assert meta_batchsize >= 0
print(f'Start_of_{mode}.')
if mode == 'train' and unroll_steps is None:
raise Exception("unroll_steps has to be specied when mode='train'.")
if mode != 'train' and unroll_steps is not None:
raise Warning("unroll_steps has no effect when mode mode!='train'.")
train = True if mode == 'train' else False
force_base = True
# TODO: to gin configuration
fc_pulling = False # does not support for now
class_balanced = False
if class_balanced:
classes_per_episode = 10
samples_per_class = 3
else:
batch_size = 128
sample_split_ratio = 0.5
# sample_split_ratio = None
anneal_outer_steps = 50
concrete_resample = True
detach_param = False
split_method = {1: 'inclusive', 2: 'exclusive'}[1]
sampler_type = {1: 'pre_sampler', 2: 'post_sampler'}[2]
#####################################################################
mask_unit = {
0: MaskUnit.SAMPLE,
1: MaskUnit.CLASS,
}[0]
mask_dist = {
0: MaskDist.SOFT,
1: MaskDist.DISCRETE,
2: MaskDist.CONCRETE,
3: MaskDist.RL,
}[3 if is_RL else 2]
mask_mode = MaskMode(mask_unit, mask_dist)
#####################################################################
# scheduler
inner_step_scheduler = InnerStepScheduler(
outer_steps, inner_steps, anneal_outer_steps)
# 100 classes in total
if split_method == 'exclusive':
# meta_support, remainder = data.split_classes(0.1) # 10 classes
# meta_query = remainder.sample_classes(50) # 50 classes
meta_support, meta_query = data.split_classes(1 / 5) # 1(100) : 4(400)
# meta_support, meta_query = data.split_classes(0.3) # 30 : 70 classes
elif split_method == 'inclusive':
# subdata = data.sample_classes(10) # 50 classes
meta_support, meta_query = data.split_instances(0.5) # 5:5 instances
else:
raise Exception()
if train:
if meta_batchsize > 0:
# serial processing of meta-minibatch
update_epochs = meta_batchsize
update_steps = None
else:
# update at every unrollings
update_steps = unroll_steps
update_epochs = None
assert (update_epochs is None) != (update_steps is None)
# for result recordin
result_frame = ResultFrame()
if save_path:
writer = SummaryWriter(os.path.join(save_path, 'tfevent'))
##################################
if is_RL:
env = Environment()
policy = Policy(sampler)
neg_rw = None
##################################
for i in range(1, outer_steps + 1):
outer_loss = 0
result_dict = ResultDict()
# initialize sampler
sampler.initialize()
# initialize base learner
model_q = Model(len(meta_support), mode='metric')
if fc_pulling:
model_s = Model(len(meta_support), mode='fc')
params = C(model_s.get_init_params('ours'))
else:
model_s = model_q
params = C(model_s.get_init_params('ours'))
##################################
if is_RL:
policy.reset()
# env.reset()
##################################
if not train or force_base:
"""baseline 0: naive single task learning
baseline 1: single task learning with the same loss scale
"""
# baseline parameters
params_b0 = C(params.copy('b0'), 4)
params_b1 = C(params.copy('b1'), 4)
if class_balanced:
batch_sampler = BalancedBatchSampler.get(class_size, sample_size)
else:
batch_sampler = RandomBatchSampler.get(batch_size)
# episode iterator
episode_iterator = MetaEpisodeIterator(
meta_support=meta_support,
meta_query=meta_query,
batch_sampler=batch_sampler,
match_inner_classes=match_inner_classes,
inner_steps=inner_step_scheduler(i, verbose=True),
sample_split_ratio=sample_split_ratio,
num_workers=4,
pin_memory=True,
)
do_sample = True
for k, (meta_s, meta_q) in episode_iterator(do_sample):
outs = []
meta_s = meta_s.cuda()
#####################################################################
if do_sample:
do_sample = False
# task encoding (very first step and right after a meta-update)
with torch.set_grad_enabled(train):
def feature_fn():
# determein whether to untilize model features
if sampler_type == 'post_sampler':
return model_s(meta_s, params)
elif sampler_type == 'pre_sampler':
return None
# sampler do the work!
mask, lr_ = sampler(mask_mode, feature_fn)
# out_for_sampler = model_s(meta_s, params, debug=do_sample)
# mask_, lr = sampler(
# pairwise_dist=out_for_sampler.pairwise_dist,
# classwise_loss=out_for_sampler.loss,
# classwise_acc=out_for_sampler.acc,
# n_classes=out_for_sampler.n_classes,
# mask_mode=mask_mode,
# )
# sample from concrete distribution
# while keeping original distribution for simple resampling
if mask_mode.dist is MaskDist.CONCRETE:
mask = mask_.rsample()
else:
mask = mask_
if is_RL:
mask, neg_rw = policy.predict(mask)
if neg_rw:
# TODO fix iteratively [0 class]s are sampled.
do_sample = True
print('[!] select 0 class!')
policy.rewards.append(neg_rw)
import pdb
pdb.set_trace()
# policy.update()
continue
else:
# we can simply resample masks when using concrete distribution
# without seriously going through the sampler
if mask_mode.dist is MaskDist.CONCRETE and concrete_resample:
mask = mask_.rsample()
#####################################################################
# use learned learning rate if available
# lr = inner_lr if lr is None else lr # inner_lr: preset / lr: learned
# train on support set
params, mask = C([params, mask], 2)
#####################################################################
out_s = model_s(meta_s, params, mask=mask, mask_mode=mask_mode)
# out_s_loss_masked = out_s.loss_masked
outs.append(out_s)
# inner gradient step
out_s_loss_mean = out_s.loss.mean() if mask_mode == MaskMode.RL \
else out_s.loss_masked_mean
out_s_loss_mean, lr = C([out_s_loss_mean, lr], 2)
params = params.sgd_step(
out_s_loss_mean, lr,
second_order=False if is_RL else True,
detach_param=True if is_RL else detach_param
)
#####################################################################
# baseline
if not train or force_base:
out_s_b0 = model_s(meta_s, params_b0)
out_s_b1 = model_s(meta_s, params_b1)
outs.extend([out_s_b0, out_s_b1])
# attach mask to get loss_s
out_s_b1.attach_mask(mask)
# inner gradient step (baseline)
params_b0 = params_b0.sgd_step(out_s_b0.loss.mean(), inner_lr)
params_b1 = params_b1.sgd_step(out_s_b1.loss_scaled_mean, lr.detach())
meta_s = meta_s.cpu()
meta_q = meta_q.cuda()
# test on query set
with torch.set_grad_enabled(train):
params = C(params, 3)
out_q = model_q(meta_q, params)
outs.append(out_q)
# baseline
if not train or force_base:
with torch.no_grad():
# test on query set
out_q_b0 = model_q(meta_q, params_b0)
out_q_b1 = model_q(meta_q, params_b1)
outs.extend([out_q_b0, out_q_b1])
##################################
if is_RL:
reward = env.step(outs)
policy.rewards.append(reward)
outs.append(policy)
##################################
# record result
result_dict.append(
outer_step=epoch * i,
inner_step=k,
**ModelOutput.as_merged_dict(outs)
)
# append to the dataframe
result_frame = result_frame.append_dict(
result_dict.index_all(-1).mean_all(-1))
# logging
if k % log_steps == 0:
logger.step_info(
epoch, mode, i, outer_steps, k, inner_step_scheduler(i), lr)
logger.split_info(meta_support, meta_query, episode_iterator)
logger.colorized_mask(
mask, fmt="2d", vis_num=20, cond=not sig_1.is_active() and log_mask)
logger.outputs(outs, print_conf=sig_1.is_active())
logger.flush()
# to debug in the middle of running process.
if k == inner_steps and sig_2.is_active():
import pdb
pdb.set_trace()
# compute outer gradient
if train and (k % unroll_steps == 0 or k == inner_steps):
#####################################################################
if not is_RL:
outer_loss += out_q.loss.mean()
outer_loss.backward()
outer_loss = 0
params.detach_().requires_grad_()
mask = mask.detach()
lr = lr.detach()
do_sample = True
#####################################################################
if not train:
if not is_RL:
# when params is not leaf node created by user,
# requires_grad attribute is False by default.
params.requires_grad_()
# meta(outer) learning
if train and update_steps and k % update_steps == 0:
# when you have meta_batchsize == 0, update_steps == unroll_steps
if is_RL:
policy.update()
else:
outer_optim.step()
sampler.zero_grad()
### end of inner steps (k) ###
if train and update_epochs and i % update_epochs == 0:
# when you have meta_batchsize > 0, update_epochs == meta_batchsize
if is_RL:
policy.update()
else:
outer_optim.step()
sampler.zero_grad()
if update_epochs:
print(f'Meta-batchsize is {meta_batchsize}: Sampler updated.')
if train and update_steps:
print(f'Meta-batchsize is zero. Updating after every unrollings.')
# tensorboard
if save_path and train:
step = (epoch * (outer_steps - 1)) + i
res = ResultFrame(result_frame[result_frame['outer_step'] == i])
loss = res.get_best_loss().mean()
acc = res.get_best_acc().mean()
writer.add_scalars(
'Loss/train', {n: loss[n] for n in loss.index}, step)
writer.add_scalars('Acc/train', {n: acc[n] for n in acc.index}, step)
# dump figures
if save_path and i % fig_epochs == 0:
# meta_s.s.save_fig(f'imgs/meta_support', save_path, i)
# meta_q.s.save_fig(f'imgs/meta_query', save_path, i)
result_dict['ours_s_mask'].save_fig(f'imgs/masks', save_path, i)
result_dict.get_items(
['ours_s_mask', 'ours_s_loss', 'ours_s_loss_masked', 'b0_s_loss',
'b1_s_loss', 'ours_q_loss', 'b0_q_loss', 'b1_q_loss']
).save_csv(f'classwise/{mode}', save_path, i)
# distinguishable episodes
if not i == outer_steps:
print(f'Path for saving: {save_path}')
print(f'End_of_episode: {i}')
### end of episode (i) ###
print(f'End_of_{mode}.')
# del metadata
return sampler, result_frame
def meta_optimize(self,
meta_optimizer,
data_cls,
model_cls,
optim_it,
unroll,
out_mul,
should_train=True):
data = data_cls(training=should_train)
model = C(model_cls())
optimizer_states = C(
OptimizerStates.initial_zeros(
(self.n_layers, len(model.params), self.hidden_sz)))
if should_train:
self.train()
else:
self.eval()
result_dict = ResultDict()
unroll_losses = 0
update = C(torch.zeros(model.params.size().flat))
iter_pbar = tqdm(range(1, optim_it + 1), 'optim_iteration')
iter_watch = utils.StopWatch('optim_iteration')
batch_arg = BatchManagerArgument(
params=model.params,
states=StatesSlicingWrapper(optimizer_states),
updates=update,
)
for iteration in iter_pbar:
data_ = data.sample()
loss = model(*data_)
unroll_losses += loss
model_dt = C(model_cls(params=batch_arg.params.detach()))
model_dt_loss = model_dt(*data_)
assert loss == model_dt_loss
model_dt_loss.backward()
assert model_dt.params.flat.grad is not None
grad = model_dt.params.flat.grad
batch_arg.update(
BatchManagerArgument(grad=grad).immutable().volatile())
iter_pbar.set_description(f'optim_iteration[loss:{loss}]')
##########################################################################
batch_manager = OptimizerBatchManager(batch_arg=batch_arg)
for get, set in batch_manager.iterator():
updates, new_states = self(get.grad.detach(), get.states)
set.states(new_states)
set.params(get.params + updates * out_mul)
if not should_train:
set.updates(updates)
##########################################################################
batch_arg = batch_manager.batch_arg_to_set
if should_train and iteration % unroll == 0:
meta_optimizer.zero_grad()
unroll_losses.backward()
meta_optimizer.step()
if not should_train or iteration % unroll == 0:
unroll_losses = 0
batch_arg.detach_()
# model.params = model.params.detach()
# optimizer_states = optimizer_states.detach()
model = C(model_cls(params=batch_arg.params))
result_dict.append(loss=loss)
if not should_train:
result_dict.append(walltime=iter_watch.touch(),
**self.params_tracker(
grad=grad,
update=batch_arg.updates,
))
return result_dict
def test_normal(name, cfg):
"""function for test of static optimizers."""
cfg = Config.merge(
[cfg.problem, cfg.args, cfg.neural_optimizers[name].test_args])
data_cls = _get_attr_by_name(cfg.problem.data_cls)
model_cls = _get_attr_by_name(cfg.problem.model_cls)
optim_cls = _get_attr_by_name(cfg.problem.optim_cls)
writer = TFWriter(cfg.save_dir, name) if cfg.save_dir else None
tests_pbar = tqdm(range(cfg.n_test), 'outer_test')
result_all = ResultDict()
result_final = ResultDict() # test loss & acc for the whole testset
best_test_loss = 999999
# params_tracker = ParamsIndexTracker(n_tracks=10)
# Outer loop (n_test)
for j in tests_pbar:
data = data_cls().sample_meta_test()
model = C(model_cls())
optimizer = optim_cls(model.parameters(), **cfg.optim_args)
walltime = Walltime()
result = ResultDict()
iter_pbar = tqdm(range(cfg.iter_test), 'inner_train')
# Inner loop (iter_test)
for k in iter_pbar:
with WalltimeChecker(walltime):
# Inner-training loss
train_nll, train_acc = model(*data['in_train'].load())
optimizer.zero_grad()
train_nll.backward()
# before = model.params.detach('hard').flat
optimizer.step()
# Inner-test loss
test_nll, test_acc = model(*data['in_test'].load())
# update = model.params.detach('hard').flat - before
# grad = model.params.get_flat().grad
result_ = dict(
train_nll=train_nll.tolist(),
test_nll=test_nll.tolist(),
train_acc=train_acc.tolist(),
test_acc=test_acc.tolist(),
walltime=walltime.time,
)
log_pbar(result_, iter_pbar)
result.append(result_)
if cfg.save_dir:
save_figure(name, cfg.save_dir, writer, result, j, 'normal')
result_all.append(result)
result_final.append(
final_inner_test(model, data['in_test'], mode='test'))
result_final_mean = result_final.mean()
last_test_loss = result_final_mean['final_loss_mean']
last_test_acc = result_final_mean['final_acc_mean']
if last_test_loss < best_test_loss:
best_test_loss = last_test_loss
best_test_acc = last_test_acc
result_test = dict(
best_test_loss=best_test_loss,
best_test_acc=best_test_acc,
last_test_loss=last_test_loss,
last_test_acc=last_test_acc,
)
log_pbar(result_test, tests_pbar)
# TF-events for inner loop (train & test)
mean_j = result_all.mean(0)
if cfg.save_dir:
for i in range(cfg.iter_test):
mean = mean_j.getitem(i)
walltime = mean_j['walltime'][i]
log_tf_event(writer, 'meta_test_inner', mean, i, walltime)
result_all.save(name, cfg.save_dir)
result_all.save_as_csv(name, cfg.save_dir, 'meta_test_inner')
result_final.save_as_csv(name,
cfg.save_dir,
'meta_test_final',
trans_1d=True)
return result_all
def meta_optimize(self,
meta_optim,
data,
model_cls,
optim_it,
unroll,
out_mul,
k_obsrv=None,
no_mask=None,
writer=None,
mode='train'):
"""FIX LATER: do something about the dummy arguments."""
assert mode in ['train', 'valid', 'test']
self.set_mode(mode)
use_indexer = False
params = C(model_cls()).params
states = C(
OptimizerStates.initial_zeros(size=(self.n_layers, len(params),
self.hidden_sz),
rnn_cell=self.rnn_cell))
result_dict = ResultDict()
unroll_losses = 0
walltime = Walltime()
update = C(torch.zeros(params.size().flat()))
if use_indexer:
batch_arg = BatchManagerArgument(
params=params,
states=StatesSlicingWrapper(states),
updates=update,
)
iter_pbar = tqdm(range(1, optim_it + 1), 'inner_train')
for iter in iter_pbar:
with WalltimeChecker(walltime):
model_train = C(model_cls(params=params.detach()))
train_nll, train_acc = model_train(*data['in_train'].load())
train_nll.backward()
if model_train.params.flat.grad is not None:
g = model_train.params.flat.grad
else:
g = model_train._grad2params(model_train.params)
assert g is not None
if use_indexer:
# This indexer was originally for outer-level batch split
# in case that the whold parameters do not fit in the VRAM.
# Which is not good, unless unavoidable, since you cannot avoid
# additional time cost induced by the serial computation.
batch_arg.update(
BatchManagerArgument(grad=g).immutable().volatile())
indexer = DefaultIndexer(input=g, shuffle=False)
batch_manager = OptimizerBatchManager(batch_arg=batch_arg,
indexer=indexer)
######################################################################
for get, set in batch_manager.iterator():
updates, new_states = self(get.grad().detach(),
get.states())
set.states(new_states)
set.params(get.params() + updates * out_mul)
if not mode == 'train':
set.updates(updates)
######################################################################
batch_arg = batch_manager.batch_arg_to_set
params = batch_arg.params
else:
updates, states = self(g.detach(), states)
updates = params.new_from_flat(updates)
params = params + updates * out_mul
#params = params - params.new_from_flat(g) * 0.1
with WalltimeChecker(walltime if mode == 'train' else None):
model_test = C(model_cls(params=params))
test_nll, test_acc = model_test(*data['in_test'].load())
if mode == 'train':
unroll_losses += test_nll
if iter % unroll == 0:
meta_optim.zero_grad()
unroll_losses.backward()
nn.utils.clip_grad_value_(self.parameters(), 0.01)
meta_optim.step()
unroll_losses = 0
with WalltimeChecker(walltime):
if not mode == 'train' or iter % unroll == 0:
if use_indexer:
batch_arg.detach_()
else:
params.detach_()
states.detach_()
# model.params = model.params.detach()
# states = states.detach()
result = dict(
train_nll=train_nll.tolist(),
test_nll=test_nll.tolist(),
train_acc=train_acc.tolist(),
test_acc=test_acc.tolist(),
walltime=walltime.time,
)
result_dict.append(result)
log_pbar(result, iter_pbar)
return result_dict, params
def meta_optimize(self, meta_optimizer, data_cls, model_cls, optim_it, unroll,
out_mul, should_train=True):
data = data_cls(training=should_train)
model = C(model_cls())
optimizer_states = C(OptimizerStates.initial_zeros(
size=(self.n_layers, len(model.params), self.hidden_sz),
rnn_cell=self.rnn_cell))
if should_train:
self.train()
else:
self.eval()
result_dict = ResultDict()
walltime = 0
unroll_losses = 0
update = C(torch.zeros(model.params.size().flat()))
batch_arg = BatchManagerArgument(
params=model.params,
states=StatesSlicingWrapper(optimizer_states),
updates=update,
)
iter_pbar = tqdm(range(1, optim_it + 1), 'optim_iteration')
iter_watch = utils.StopWatch('optim_iteration')
for iteration in iter_pbar:
iter_watch.touch()
# entire_loop.touch()
# interval.touch()
data_ = data.sample()
if should_train:
model = C(model_cls(params=batch_arg.params))
loss = model(*data_)
unroll_losses += loss
# print("\n[1] : "+str(interval.touch('interval', False)))
# model_ff.touch()
model_detached = C(model_cls(params=batch_arg.params.detach()))
loss_detached = model_detached(*data_)
# model_ff.touch('interval', True)
if should_train:
assert loss == loss_detached
# model_bp.touch()
loss_detached.backward()
# model_bp.touch('interval', True)
# interval.touch()
assert model_detached.params.flat.grad is not None
grad = model_detached.params.flat.grad
batch_arg.update(BatchManagerArgument(grad=grad).immutable().volatile())
iter_pbar.set_description(f'optim_iteration[loss:{loss_detached}]')
# print("\n[2-1] : "+str(interval.touch('interval', False)))
##########################################################################
# indexer = DefaultIndexer(input=grad, shuffle=False)
indexer = TopKIndexer(input=grad, k_ratio=0.1, random_k_mode=False)# bound=model_dt.params.size().bound_layerwise())
#indexer = DefaultIndexer(input=grad)
batch_manager = OptimizerBatchManager(
batch_arg=batch_arg, indexer=indexer)
# print("\n[2-2] : "+str(interval.touch('interval', False)))
# optim_ff.touch()
# optim_ff_1.touch()
for get, set in batch_manager.iterator():
# optim_ff_1.touch('interval', True)
# optim_ff_2.touch()
# optim_ff_2.touch('interval', True)
# optim_ff_3.touch()
updates, new_states = self(get.grad().detach(), get.states())
# optim_ff_3.touch('interval', True)
# optim_ff_4.touch()
updates = updates * out_mul
set.states(new_states)
set.params(get.params() + updates)
if not should_train:
set.updates(updates)
# optim_ff_4.touch('interval', True)
# optim_ff.touch('interval', True)
##########################################################################
# interval.touch()
batch_arg = batch_manager.batch_arg_to_set
# print("\n[3] : "+str(interval.touch('interval', False)))
if should_train and iteration % unroll == 0:
# optim_bp.touch()
meta_optimizer.zero_grad()
unroll_losses.backward()
nn.utils.clip_grad_norm_(self.parameters(), 10)
meta_optimizer.step()
unroll_losses = 0
# optim_bp.touch('interval', True)
# interval.touch()
if not should_train or iteration % unroll == 0:
batch_arg.detach_()
# print("\n[4] : "+str(interval.touch('interval', False)))
# interval.touch()
# model.params = model.params.detach()
# optimizer_states = optimizer_states.detach()
walltime += iter_watch.touch('interval')
result_dict.append(loss=loss_detached)
# print("\n[5] : "+str(interval.touch('interval', False)))
if not should_train:
result_dict.append(
walltime=iter_watch.touch(),
**self.params_tracker(
grad=grad,
update=batch_arg.updates,
)
)
# entire_loop.touch('interval', True)
return result_dict
def meta_optimize(self, meta_optimizer, data, model_cls, optim_it, unroll,
out_mul, mode):
assert mode in ['train', 'valid', 'test']
if mode == 'train':
self.train()
# data.new_train_data()
inner_data = data.loaders['inner_train']
outer_data = data.loaders['inner_valid']
# inner_data = data.loaders['train']
# outer_data = inner_data
drop_mode = 'soft_drop'
elif mode == 'valid':
self.eval()
inner_data = data.loaders['valid']
outer_data = None
drop_mode = 'hard_drop'
elif mode == 'eval':
self.eval()
inner_data = data.loaders['test']
outer_data = None
drop_mode = 'hard_drop'
# data = dataset(mode=mode)
model = C(model_cls(sb_mode=self.sb_mode))
model(*data.pseudo_sample())
# layer_sz = sum([v for v in model.params.size().unflat(1, 'mat').values()])
mask_states = C(
OptimizerStates.initial_zeros(
(self.n_layers, len(model.activations.mean(0)),
self.hidden_sz)))
update_states = C(
OptimizerStates.initial_zeros(
(self.n_layers, len(model.params), self.hidden_sz)))
result_dict = ResultDict()
unroll_losses = 0
walltime = 0
update = C(torch.zeros(model.params.size().flat()))
batch_arg = BatchManagerArgument(
params=model.params,
states=StatesSlicingWrapper(update_states),
updates=model.params.new_zeros(model.params.size()),
)
iter_pbar = tqdm(range(1, optim_it + 1), 'optim_iteration')
iter_watch = utils.StopWatch('optim_iteration')
bp_watch = utils.StopWatch('bp')
loss_decay = 1.0
for iteration in iter_pbar:
iter_watch.touch()
model_detached = C(
model_cls(params=batch_arg.params.detach(),
sb_mode=self.sb_mode))
loss_detached = model_detached(*inner_data.load())
loss_detached.backward()
# import pdb; pdb.set_trace()
iter_pbar.set_description(f'optim_iteration[loss:{loss_detached}]')
# np.set_printoptions(suppress=True, precision=3, linewidth=200, edgeitems=20)
if debug_sigint.signal_on:
debug = iteration == 1 or iteration % 10 == 0
# debug = True
else:
debug = False
backprop_mask, mask_states, sparsity_loss = self.mask_generator(
activations=model_detached.activations.grad.detach(),
debug=debug,
states=mask_states)
# if debug_sigstp.signal_on and (iteration == 1 or iteration % 10 == 0):
# mask_list.append(backprop_mask)
# import pdb; pdb.set_trace()
# min = 0.001
# max = 0.01
unroll_losses += sparsity_loss * 0.01 * \
loss_decay # * ((max - min) * lambda_ + min)
# import pdb; pdb.set_trace()
# model_detached.apply_backprop_mask(backprop_mask.unflat, drop_mode)
# loss_detached.backward()
assert model_detached.params.grad is not None
if mode == 'train':
model = C(
model_cls(params=batch_arg.params, sb_mode=self.sb_mode))
loss = model(*outer_data.load())
# assert loss == loss_detached
if iteration == 1:
initial_loss = loss
loss_decay = 1.0
else:
loss_decay = (loss / initial_loss).detach()
unroll_losses += loss
# model_detached.apply_backprop_mask(backprop_mask.unflat, drop_mode)
#model_detached.params.apply_grad_mask(mask, 'soft_drop')
coord_mask = backprop_mask.expand_as_params(model_detached.params)
if drop_mode == 'soft_drop':
# grad = model_detached.params.grad.mul(coord_mask).flat
# model_detached.params = model_detached.params.mul(params_mask)
index = None
# updates, states = self.updater(grad, states, mask=coord_mask)
# updates = updates * out_mul * coord_mask
# import pdb; pdb.set_trace()
elif drop_mode == 'hard_drop':
index = (coord_mask.flat.squeeze() > 0.5).nonzero().squeeze()
grad = model_detached.params.grad.flat
batch_arg.update(BatchManagerArgument(grad=grad, mask=coord_mask))
##########################################################################
indexer = DefaultIndexer(input=grad, index=index, shuffle=False)
batch_manager = OptimizerBatchManager(batch_arg=batch_arg,
indexer=indexer)
for get, set in batch_manager.iterator():
inp = get.grad().detach()
#inp = [get.grad().detach(), get.tag().detach()]
if drop_mode == 'soft_drop':
updates, new_states = self.updater(
inp, get.states()) # , mask=get.mask())
# if iteration == 100:
# import pdb; pdb.set_trace()
# aaa = get.states() *2
updates = updates * out_mul * get.mask()
set.states(get.states() * (1 - get.mask()) +
new_states * get.mask())
# set.states(new_states)
elif drop_mode == 'hard_drop':
updates, new_states = self.updater(inp, get.states())
updates = updates * out_mul
set.states(new_states)
else:
raise RuntimeError('Unknown drop mode!')
set.params(get.params() + updates)
# if not mode != 'train':
# set.updates(updates)
##########################################################################
batch_arg = batch_manager.batch_arg_to_set
# updates = model.params.new_from_flat(updates)
if mode == 'train' and iteration % unroll == 0:
meta_optimizer.zero_grad()
unroll_losses.backward()
nn.utils.clip_grad_norm_(self.parameters(), 5)
meta_optimizer.step()
unroll_losses = 0
# if iteration == 1 or iteration % 10 == 0:
# import pdb; pdb.set_trace()
if not mode == 'train' or iteration % unroll == 0:
mask_states.detach_()
batch_arg.detach_()
# model.params = model.params.detach()
# update_states = update_states.detach()
walltime += iter_watch.touch('interval')
result_dict.append(loss=loss_detached)
if not mode == 'train':
result_dict.append(
walltime=walltime,
# **self.params_tracker(
# grad=grad,
# update=batch_arg.updates,
# )
)
return result_dict
def meta_optimize(self, meta_optimizer, data_cls, model_cls, optim_it, unroll,
out_mul, should_train=True):
target = data_cls(training=should_train)
optimizee = C(model_cls())
optimizer_states = OptimizerStates.zero_states(
self.n_layers, len(optimizee.params), self.hidden_sz)
if should_train:
self.train()
else:
self.eval()
coordinate_tracker = ParamsIndexTracker(
n_params=len(optimizee.params), n_tracks=30)
result_dict = ResultDict()
unroll_losses = 0
iter_pbar = tqdm(range(1, optim_it + 1), 'optim_iteration')
iter_watch = utils.StopWatch('optim_iteration')
for iteration in iter_pbar:
loss = optimizee(target)
unroll_losses += loss
loss.backward(retain_graph=should_train)
iter_pbar.set_description(f'optim_iteration[loss:{loss}]')
unit_size = torch.Size([len(optimizee.params), 1])
grid_size = torch.Size([len(optimizee.params), self.n_coord])
batch_manager = OptimizerBatchManager(
params=optimizee.params,
states=StatesSlicingWrapper(optimizer_states),
units=OptimizerBatchManager.placeholder(unit_size),
grid_abs=OptimizerBatchManager.placeholder(grid_size),
grid_rel=OptimizerBatchManager.placeholder(grid_size),
batch_size=30000,
)
##########################################################################
for get, set in batch_manager.iterator():
units, new_states = self.unit_maker(
C.detach(get.params.grad), get.states.clone())
units = units * out_mul / self.n_coord / 2
grid_abs, grid_rel = self.grid_maker(units, get.params)
set.states(new_states)
set.units(units)
set.grid_abs(grid_abs)
set.grid_rel(grid_rel)
##########################################################################
landscape = self.loss_observer(batch_manager.grid_abs, model_cls, target)
updates = self.step_maker(batch_manager.grid_rel, landscape)
optimizee.params.add_flat_(updates)
if should_train and iteration % unroll == 0:
meta_optimizer.zero_grad()
unroll_losses.backward()
meta_optimizer.step()
if not should_train or iteration % unroll == 0:
unroll_losses = 0
optimizee.params = optimizee.params.detach()
optimizer_states = optimizer_states.detach()
optimizee = C(model_cls(optimizee.params))
result_dict.append(loss=loss)
if not should_train:
result_dict.append(
grid=coordinate_tracker(batch_manager.units),
update=coordinate_tracker(updates),
walltime=iter_watch.touch(),
)
return result_dict
def meta_optimize(self, meta_optimizer, data, model_cls, optim_it, unroll,
out_mul, mode):
assert mode in ['train', 'valid', 'test']
if mode == 'train':
self.train()
# data.new_train_data()
inner_data = data.loaders['inner_train']
outer_data = data.loaders['inner_valid']
# inner_data = data.loaders['train']
# outer_data = inner_data
drop_mode = 'soft_drop'
elif mode == 'valid':
self.eval()
inner_data = data.loaders['valid']
outer_data = None
drop_mode = 'hard_drop'
elif mode == 'test':
self.eval()
inner_data = data.loaders['test']
outer_data = None
drop_mode = 'hard_drop'
# data = dataset(mode=mode)
model = C(model_cls(sb_mode=self.sb_mode))
model(*data.pseudo_sample())
params = model.params
result_dict = ResultDict()
unroll_losses = 0
walltime = 0
self.feature_gen.new()
################
mask_dict = ResultDict()
analyze_mask = False
sample_mask = False
draw_loss = False
################
iter_pbar = tqdm(range(1, optim_it + 1), 'optim_iteration')
iter_watch = utils.StopWatch('optim_iteration')
res1 = None
res2 = []
res3 = []
lamb = []
gamm_g = []
gamm_l1 = []
gamm_l2 = []
iters = []
for iteration in iter_pbar:
iter_watch.touch()
if debug_sigint.signal_on:
debug_1 = iteration == 1 or iteration % 10 == 0
else:
debug_1 = False
if debug_sigstp.signal_on:
debug_2 = True
else:
debug_2 = False
model_detached = C(model_cls(params.detach()))
inner_data_s = inner_data.load()
loss_detached = model_detached(*inner_data_s)
loss_detached.backward()
g = model_detached.params.grad.flat.detach()
w = model_detached.params.flat.detach()
cand_params = []
cand_losses = []
best_loss = 9999999
best_params = None
n_samples = 5
for m in range(1):
feature = self.feature_gen(g,
w,
n=n_samples,
m_update=(m == 0))
p_size = params.size().unflat()
mask_gen_out = self.mask_gen(feature, p_size, n=n_samples)
step_out = self.step_gen(feature, n=n_samples)
# step & mask genration
for i in range(n_samples):
mask, _, kld = mask_gen_out[i]
step = step_out[i]
mask = ParamsFlattener(mask)
mask_layout = mask.expand_as(params)
step = params.new_from_flat(step)
# prunning
step = step * mask_layout
params_ = params + step
# cand_params.append(params_.flat)
sparse_params = params_.prune(mask > 1e-6)
if sparse_params.size().unflat()['mat_0'][1] == 0:
continue
if debug_2:
import pdb
pdb.set_trace()
# cand_loss = model(*outer_data_s)
sparse_model = C(model_cls(sparse_params.detach()))
loss = sparse_model(*inner_data_s)
try:
if loss < best_loss:
best_loss = loss
best_params = params_
best_kld = kld
except:
best_params = params_
best_kld = kld
if best_params is not None:
params = best_params
best_kld = 0
if mode == 'train':
model = C(model_cls(params))
optim_loss = model(*outer_data.load())
if torch.isnan(optim_loss):
import pdb
pdb.set_trace()
unroll_losses += optim_loss + best_kld / outer_data.full_size
if mode == 'train' and iteration % unroll == 0:
meta_optimizer.zero_grad()
unroll_losses.backward()
# import pdb; pdb.set_trace()
nn.utils.clip_grad_value_(self.parameters(), 0.1)
meta_optimizer.step()
unroll_losses = 0
if not mode == 'train' or iteration % unroll == 0:
# self.mask_gen.detach_lambdas_()
params = params.detach_()
# import pdb; pdb.set_trace()
if params is None:
import pdb
pdb.set_trace()
iter_pbar.set_description(
f'optim_iteration'
f'[optim_loss:{loss_detached.tolist():5.5}')
# f' sparse_loss:{sparsity_loss.tolist():5.5}]')
# iter_pbar.set_description(f'optim_iteration[loss:{loss_dense.tolist()}/dist:{dist.tolist()}]')
# torch.optim.SGD(sparse_params, lr=0.1).step()
walltime += iter_watch.touch('interval')
##############################################################
text_dir = 'test/analyze_mask'
result_dir = 'test/drawloss'
sample_dir = 'test/mask_compare'
iter_interval = 10
sample_num = 10000
if mode == 'test' and analyze_mask:
analyzing_mask(self.mask_gen, layer_size, mode, iter,
iter_interval, text_dir)
if mode == 'test' and sample_mask:
mask_result = sampling_mask(self.mask_gen, layer_size,
model_train, params, sample_num,
mode, iter, iter_interval,
sample_dir)
if mask_result is not None:
mask_dict.append(mask_result)
if mode == 'test' and draw_loss:
plot_loss(model_cls=model_cls,
model=model_train,
params,
input_data=data['in_train'].load(),
dataset=data['in_train'],
feature_gen=self.feature_gen,
mask_gen=self.mask_gen,
step_gen=self.step_gen,
scale_way=None,
xmin=-2.0,
xmax=0.5,
num_x=20,
mode=mode,
iteration=iter,
iter_interval=iter_interval,
loss_dir=result_dir)
##############################################################
result_dict.append(loss=loss_detached)
if not mode == 'train':
result_dict.append(
walltime=walltime,
# **self.params_tracker(
# grad=grad,
# update=batch_arg.updates,
# )
)
return result_dict
def meta_optimize(self,
meta_optimizer,
data_cls,
model_cls,
optim_it,
unroll,
out_mul,
should_train=True):
data = data_cls(training=should_train)
model = C(model_cls())
states = C(
OptimizerStates.initial_zeros(
(self.n_layers, len(model.params), self.hidden_sz)))
if should_train:
self.train()
else:
self.eval()
result_dict = ResultDict()
model_dt = None
model_losses = 0
#grad_real_prev = None
iter_pbar = tqdm(range(1, optim_it + 1), 'optim_iteration')
iter_watch = utils.StopWatch('optim_iteration')
batch_arg = BatchManagerArgument(
params=model.params,
states=StatesSlicingWrapper(states),
update_prev=C(torch.zeros(model.params.size().flat())),
mu=C(torch.zeros(model.params.size().flat())),
sigma=C(torch.zeros(model.params.size().flat())),
)
for iteration in iter_pbar:
data_ = data.sample()
model_loss = model(*data_)
model_losses += model_loss
# if iteration == 21:
# import pdb; pdb.set_trace()
try:
make_dot(model_losses)
except:
import pdb
pdb.set_trace()
# if iteration == 1 or iteration == 21:
# import pdb; pdb.set_trace()
model_dt = C(model_cls(params=model.params.detach()))
model_loss_dt = model_dt(*data_)
# model.params.register_grad_cut_hook_()
model_loss_dt.backward()
assert model_dt.params.flat.grad is not None
grad_cur = model_dt.params.flat.grad
# model.params.remove_grad_cut_hook_()
batch_arg.update(BatchManagerArgument(grad_cur=grad_cur))
# NOTE: read_only, write_only
########################################################################
batch_manager = OptimizerBatchManager(batch_arg=batch_arg)
for get, set in batch_manager.iterator():
grad_cur = get.grad_cur.detach() # free-running
update_prev = get.update_prev.detach()
points_rel, states = self.point_sampler(
grad_cur, update_prev, get.states, self.n_points)
points_rel = points_rel * out_mul
points_abs = get.params.detach(
) + points_rel # NOTE: check detach
losses = self.loss_observer(points_abs, model_cls, data_)
update = self.point_selector(points_rel, losses)
set.params(get.params + update)
set.states(states)
set.update_prev(update)
set.mu(self.point_sampler.mu)
set.sigma(self.point_sampler.sigma)
##########################################################################
batch_arg = batch_manager.batch_arg_to_set
iter_pbar.set_description('optim_iteration[loss:{:10.6f}]'.format(
model_loss.tolist()))
if should_train and iteration % unroll == 0:
meta_optimizer.zero_grad()
try:
model_losses.backward()
except:
import pdb
pdb.set_trace()
meta_optimizer.step()
if not should_train or iteration % unroll == 0:
model_losses = 0
batch_arg.detach_()
# batch_arg.params.detach_()
# batch_arg.states.detach_()
# NOTE: just do batch_arg.detach_()
model = C(model_cls(params=batch_arg.params))
result_dict.append(
step_num=iteration,
loss=model_loss,
)
if not should_train:
result_dict.append(walltime=iter_watch.touch(),
**self.params_tracker(
grad=grad_cur,
update=batch_arg.update_prev,
mu=batch_arg.mu,
sigma=batch_arg.sigma,
))
return result_dict
def meta_optimize(self, meta_optimizer, data, model_cls, optim_it, unroll,
out_mul, mode):
assert mode in ['train', 'valid', 'test']
if mode == 'train':
self.train()
# data.new_train_data()
inner_data = data.loaders['inner_train']
outer_data = data.loaders['inner_valid']
# inner_data = data.loaders['train']
# outer_data = inner_data
drop_mode = 'soft_drop'
elif mode == 'valid':
self.eval()
inner_data = data.loaders['valid']
outer_data = None
drop_mode = 'hard_drop'
elif mode == 'eval':
self.eval()
inner_data = data.loaders['test']
outer_data = None
drop_mode = 'hard_drop'
# data = dataset(mode=mode)
model = C(model_cls(sb_mode=self.sb_mode))
model(*data.pseudo_sample())
# layer_sz = sum([v for v in model.params.size().unflat(1, 'mat').values()])
mask_states = C(
OptimizerStates.initial_zeros(
(self.n_layers, len(model.activations.mean(0)),
self.hidden_sz)))
update_states = C(
OptimizerStates.initial_zeros(
(self.n_layers, len(model.params), self.hidden_sz)))
result_dict = ResultDict()
unroll_losses = 0
walltime = 0
update = C(torch.zeros(model.params.size().flat()))
batch_arg = BatchManagerArgument(
params=model.params,
states=StatesSlicingWrapper(update_states),
updates=model.params.new_zeros(model.params.size()),
)
params = model.params
lr_sgd = 0.2
lr_adam = 0.001
sgd = torch.optim.SGD(model.parameters(), lr=lr_sgd)
adam = torch.optim.Adam(model.parameters())
adagrad = torch.optim.Adagrad(model.parameters())
# print('###', lr_adam, '####')
iter_pbar = tqdm(range(1, optim_it + 1), 'optim_iteration')
iter_watch = utils.StopWatch('optim_iteration')
bp_watch = utils.StopWatch('bp')
loss_decay = 1.0
dist_list = []
dense_rank = 0
sparse_rank = 0
dense_rank_list = []
sparse_rank_list = []
for iteration in iter_pbar:
iter_watch.touch()
# # conventional optimizers with API
# model.zero_grad()
# loss = model(*inner_data.load())
# loss.backward()
# adam.step()
# loss_detached = loss
model_detached = C(model_cls(params=params.detach()))
loss_detached = model_detached(*inner_data.load())
loss_detached.backward()
# # conventional optimizers with manual SGD
# params = model_detached.params + model_detached.params.grad * (-0.2)
# iter_pbar.set_description(f'optim_iteration[loss:{loss_detached}]')
masks = []
# sparse_old_loss = []
# sparse_new_loss = []
sparse_loss = []
dense_loss = []
candidates = []
mode = 2
n_sample = 10
if mode == 1:
lr = 1.0
elif mode == 2:
lr = 0.2
else:
raise Exception()
for i in range(n_sample):
if i == 9:
mask = MaskGenerator.topk(
model_detached.activations.grad.detach())
else:
mask = MaskGenerator.randk(
model_detached.activations.grad.detach())
if model == 1:
# recompute gradients again using prunned network
sparse_params = model_detached.params.prune(mask)
model_prunned = C(model_cls(params=sparse_params.detach()))
loss_prunned = model_prunned(*inner_data.load())
loss_prunned.backward()
grads = model_prunned.params.grad
elif mode == 2:
# use full gradients but sparsified
sparse_params = model_detached.params.prune(mask)
grads = model_detached.params.grad.prune(mask)
else:
raise Exception('unknown model.')
sparse_params = sparse_params + grads * (-lr)
model_prunned = C(model_cls(params=sparse_params.detach()))
loss_prunned = model_prunned(*inner_data.load())
# params = model_detached.params
params = model_detached.params.sparse_update(
mask, grads * (-lr))
candidates.append(params)
model_dense = C(model_cls(params=params))
loss_dense = model_dense(*inner_data.load())
sparse_loss.append(loss_prunned)
dense_loss.append(loss_dense)
sparse_loss = torch.stack(sparse_loss, 0)
dense_loss = torch.stack(dense_loss, 0)
min_id = (
sparse_loss.min() == sparse_loss).nonzero().squeeze().tolist()
params = candidates[min_id]
sparse_order = sparse_loss.sort()[1].float()
dense_order = dense_loss.sort()[1].float()
dist = (sparse_order - dense_order).abs().sum()
dist_list.append(dist)
dense_rank += (dense_order == 9).nonzero().squeeze().tolist()
sparse_rank += (sparse_order == 9).nonzero().squeeze().tolist()
dense_rank_list.append(
(dense_order == 9).nonzero().squeeze().tolist())
sparse_rank_list.append(
(sparse_order == 9).nonzero().squeeze().tolist())
iter_pbar.set_description(
f'optim_iteration[dense_loss:{loss_dense.tolist():5.5}/sparse_loss:{loss_prunned.tolist():5.5}]'
)
# iter_pbar.set_description(f'optim_iteration[loss:{loss_dense.tolist()}/dist:{dist.tolist()}]')
# torch.optim.SGD(sparse_params, lr=0.1).step()
result_dict.append(loss=loss_detached)
if not mode == 'train':
result_dict.append(
walltime=walltime,
# **self.params_tracker(
# grad=grad,
# update=batch_arg.updates,
# )
)
dist_mean = torch.stack(dist_list, 0).mean()
import pdb
pdb.set_trace()
return result_dict
def meta_optimize(self,
meta_optimizer,
data_cls,
model_cls,
optim_it,
unroll,
out_mul,
should_train=True):
data = data_cls(training=should_train)
model = C(model_cls())
g_states = C(
OptimizerStates.initial_zeros(
(self.n_layers, len(model.params), self.hidden_sz)))
u_states = C(
OptimizerStates.initial_zeros(
(self.n_layers, len(model.params), self.hidden_sz)))
if should_train:
self.train()
else:
self.eval()
grad_pred_tracker = ParamsIndexTracker(name='grad_pred',
n_tracks=10,
n_params=len(model.params))
grad_real_tracker = grad_pred_tracker.clone_new_name('grad_real')
update_tracker = grad_pred_tracker.clone_new_name('update')
result_dict = ResultDict()
model_dt = None
model_losses = 0
grad_losses = 0
train_with_lazy_tf = False
eval_test_grad_loss = True
grad_prev = C(torch.zeros(model.params.size().flat()))
update_prev = C(torch.zeros(model.params.size().flat()))
#grad_real_prev = None
grad_real_cur = None
grad_real_prev = C(torch.zeros(model.params.size().flat()))
iter_pbar = tqdm(range(1, optim_it + 1), 'optim_iteration')
iter_watch = utils.StopWatch('optim_iteration')
"""grad_real_prev:
teacher-forcing current input of RNN.
required during all of training steps + every 'k' test steps.
grad_real_cur:
MSE loss target for current output of RNN.
required during all the training steps only.
"""
batch_arg = BatchManagerArgument(
params=model.params,
grad_prev=grad_prev,
update_prev=update_prev,
g_states=StatesSlicingWrapper(g_states),
u_states=StatesSlicingWrapper(u_states),
mse_loss=C(torch.zeros(1)),
#updates=OptimizerBatchManager.placeholder(model.params.flat.size()),
)
for iteration in iter_pbar:
inp, out = data.sample()
model_loss = model(inp, out)
model_losses += model_loss
# if grad_real_cur is not None: # if iteration % self.k == 0
# grad_cur_prev = grad_real_cur # back up previous grad
if (iteration +
1) % self.k == 0 or should_train or eval_test_grad_loss:
# get grad every step while training
# get grad one step earlier than teacher-forcing steps while testing
model_dt = C(model_cls(params=model.params.detach()))
model_loss_dt = model_dt(inp, out)
# model.params.register_grad_cut_hook_()
model_loss_dt.backward()
# model.params.remove_grad_cut_hook_()
grad_real_cur = model_dt.params.flat.grad
assert grad_real_cur is not None
del model_dt, model_loss_dt
if should_train or eval_test_grad_loss:
# mse loss target
# grad_real_cur is not used as a mse target while testing
assert grad_real_cur is not None
batch_arg.update(
BatchManagerArgument(
grad_real_cur=grad_real_cur).immutable().volatile())
if iteration % self.k == 0 or (should_train
and not train_with_lazy_tf):
# teacher-forcing
assert grad_real_prev is not None
batch_arg.update(
BatchManagerArgument(
grad_real_prev=grad_real_prev).immutable().volatile())
grad_real_prev = None
########################################################################
batch_manager = OptimizerBatchManager(batch_arg=batch_arg)
for get, set in batch_manager.iterator():
if iteration % self.k == 0 or (should_train
and not train_with_lazy_tf):
grad_prev = get.grad_real_prev.detach() # teacher-forcing
else:
grad_prev = get.grad_prev.detach() # free-running
update_prev = get.update_prev.detach()
grad_cur, update_cur, g_states, u_states = self(
grad_prev, update_prev, get.g_states, get.u_states)
set.grad_prev(grad_cur)
set.update_prev(update_cur)
set.g_states(g_states)
set.u_states(u_states)
set.params(get.params + update_cur * out_mul)
if should_train or eval_test_grad_loss:
set.mse_loss(
self.mse_loss(grad_cur, get.grad_real_cur.detach()))
##########################################################################
batch_arg = batch_manager.batch_arg_to_set
grad_loss = batch_arg.mse_loss * 100
grad_losses += grad_loss
#to_float = lambda x: float(x.data.cpu().numpy())
iter_pbar.set_description(
'optim_iteration[loss(model):{:10.6f}/(grad):{:10.6f}]'.format(
model_loss.tolist(),
grad_loss.tolist()[0]))
if should_train and iteration % unroll == 0:
meta_optimizer.zero_grad()
losses = model_losses + grad_losses
losses.backward()
meta_optimizer.step()
if not should_train or iteration % unroll == 0:
model_losses = 0
grad_losses = 0
batch_arg.detach_()
# batch_arg.params.detach_()
# batch_arg.g_states.detach_()
# batch_arg.u_states.detach_()
model = C(model_cls(params=batch_arg.params))
result_dict.append(
step_num=iteration,
loss=model_loss,
grad_loss=grad_loss,
)
if not should_train:
result_dict.append(
walltime=iter_watch.touch(),
**grad_pred_tracker.track_num,
**grad_pred_tracker(batch_arg.grad_prev),
**grad_real_tracker(grad_real_cur),
**update_tracker(batch_arg.update_prev),
)
if (iteration +
1) % self.k == 0 or should_train or eval_test_grad_loss:
grad_real_prev = grad_real_cur
grad_real_cur = None
return result_dict
def test_neural(name, cfg, learned_params):
"""Function for meta-test of learned optimizers.
"""
cfg = Config.merge(
[cfg.problem, cfg.neural_optimizers[name].test_args, cfg.args])
data_cls = _get_attr_by_name(cfg.data_cls)
model_cls = _get_attr_by_name(cfg.model_cls)
optim_cls = _get_optim_by_name(cfg.optim_module)
writer = TFWriter(cfg.save_dir, name) if cfg.save_dir else None
optimizer = C(optim_cls())
if learned_params is not None:
optimizer.params = learned_params
meta_optim = None
unroll = 1
best_test_loss = 999999
result_all = ResultDict()
result_final = ResultDict() # test loss & acc for the whole testset
tests_pbar = tqdm(range(cfg.n_test), 'test')
# Meta-test
for j in tests_pbar:
data = data_cls().sample_meta_test()
result, params = optimizer.meta_optimize(cfg, meta_optim, data,
model_cls, writer, 'test')
result_all.append(result)
result_final.append(
final_inner_test(C(model_cls(params)),
data['in_test'],
mode='test'))
result_final_mean = result_final.mean()
last_test_loss = result_final_mean['final_loss_mean']
last_test_acc = result_final_mean['final_acc_mean']
if last_test_loss < best_test_loss:
best_test_loss = last_test_loss
best_test_acc = last_test_acc
result_test = dict(
best_test_loss=best_test_loss,
best_test_acc=best_test_acc,
last_test_loss=last_test_loss,
last_test_acc=last_test_acc,
)
log_pbar(result_test, tests_pbar)
if cfg.save_dir:
save_figure(name, cfg.save_dir, writer, result, j, 'test')
result_all_mean = result_all.mean(0)
# TF-events for inner loop (train & test)
if cfg.save_dir:
# test_result_mean = ResultDict()
for i in range(cfg.iter_test):
mean_i = result_all_mean.getitem(i)
walltime = result_all_mean['walltime'][i]
log_tf_event(writer, 'meta_test_inner', mean_i, i, walltime)
# test_result_mean.update(**mean_i, step=i, walltime=walltime)
result_all.save(name, cfg.save_dir)
result_all.save_as_csv(name, cfg.save_dir, 'meta_test_inner')
result_final.save_as_csv(name,
cfg.save_dir,
'meta_test_final',
trans_1d=True)
return result_all
def test_neural(name,
save_dir,
learned_params,
data_cls,
model_cls,
optim_module,
n_test=100,
iter_test=100,
preproc=False,
out_mul=1.0,
no_mask=False,
k_obsrv=10):
"""Function for meta-test of learned optimizers.
"""
data_cls = _get_attr_by_name(data_cls)
model_cls = _get_attr_by_name(model_cls)
optim_cls = _get_optim_by_name(optim_module)
writer = TFWriter(save_dir, name) if save_dir else None
optimizer = C(optim_cls())
if learned_params is not None:
optimizer.params = learned_params
meta_optim = None
unroll = 1
best_test_loss = 999999
result_all = ResultDict()
result_final = ResultDict() # test loss & acc for the whole testset
tests_pbar = tqdm(range(n_test), 'test')
# Meta-test
for j in tests_pbar:
data = data_cls().sample_meta_test()
result, params = optimizer.meta_optimize(meta_optim, data, model_cls,
iter_test, unroll, out_mul,
k_obsrv, no_mask, writer,
'test')
result_all.append(result)
result_final.append(
final_inner_test(C(model_cls(params)),
data['in_test'],
mode='test'))
result_final_mean = result_final.mean()
last_test_loss = result_final_mean['loss_mean']
last_test_acc = result_final_mean['acc_mean']
if last_test_loss < best_test_loss:
best_test_loss = last_test_loss
best_test_acc = last_test_acc
result_test = dict(
best_test_loss=best_test_loss,
best_test_acc=best_test_acc,
last_test_loss=last_test_loss,
last_test_acc=last_test_acc,
)
log_pbar(result_test, tests_pbar)
if save_dir:
save_figure(name, save_dir, writer, result, j, 'test')
result_all_mean = result_all.mean(0)
# TF-events for inner loop (train & test)
if save_dir:
# test_result_mean = ResultDict()
for i in range(iter_test):
mean_i = result_all_mean.getitem(i)
walltime = result_all_mean['walltime'][i]
log_tf_event(writer, 'meta_test_inner', mean_i, i, walltime)
# test_result_mean.update(**mean_i, step=i, walltime=walltime)
result_all.save(name, save_dir)
result_all.save_as_csv(name, save_dir, 'meta_test_inner')
result_final.save_as_csv(name,
save_dir,
'meta_test_final',
trans_1d=True)
return result_all
def train_neural(name, cfg):
"""Function for meta-training and meta-validation of learned optimizers.
"""
cfg = Config.merge(
[cfg.problem, cfg.neural_optimizers[name].train_args, cfg.args])
# NOTE: lr will be overwritten here.
# Options
lr_scheduling = False # learning rate scheduling
tr_scheduling = False # Truncation scheduling
seperate_lr = False
print(f'data_cls: {cfg.data_cls}')
print(f'model_cls: {cfg.model_cls}')
data_cls = _get_attr_by_name(cfg.data_cls)
model_cls = _get_attr_by_name(cfg.model_cls)
optim_cls = _get_optim_by_name(cfg.optim_module)
meta_optim = cfg.meta_optim
optimizer = C(optim_cls())
if len([p for p in optimizer.parameters()]) == 0:
return None # no need to be trained if there's no parameter
writer = TFWriter(cfg.save_dir, name) if cfg.save_dir else None
# TODO: handle variable arguments according to different neural optimziers
"""meta-optimizer"""
meta_optim = {'sgd': 'SGD', 'adam': 'Adam'}[meta_optim.lower()]
if cfg.mask_lr == 0.0:
print(f'meta optimizer: {meta_optim} / lr: {cfg.lr} / wd: {cfg.wd}\n')
meta_optim = getattr(torch.optim, meta_optim)(optimizer.parameters(),
lr=cfg.lr,
weight_decay=cfg.wd)
# scheduler = torch.optim.lr_scheduler.MultiStepLR(
# meta_optim, milestones=[1, 2, 3, 4], gamma=0.1)
else:
print(f'meta optimizer: {meta_optim} / gen_lr: {cfg.lr} '
f'/ mask_lr:{cfg.mask_lr} / wd: {cfg.wd}\n')
p_feat = optimizer.feature_gen.parameters()
p_step = optimizer.step_gen.parameters()
p_mask = optimizer.mask_gen.parameters()
meta_optim = getattr(torch.optim, meta_optim)([
{
'params': p_feat,
'lr': cfg.lr,
'weight_decay': cfg.wd
},
{
'params': p_step,
'lr': cfg.lr,
'weight_decay': cfg.wd
},
{
'params': p_mask,
'lr': cfg.mask_lr
},
])
if lr_scheduling:
lr_scheduler = ReduceLROnPlateau(meta_optim,
mode='min',
factor=0.5,
patience=10,
verbose=True)
# tr_scheduler = Truncation(
# cfg.unroll, mode='min', factor=1.5, patience=1, max_len=50, verbose=True)
data = data_cls()
best_params = None
best_valid_loss = 999999
# best_converge = 999999
train_result_mean = ResultDict()
valid_result_mean = ResultDict()
epoch_pbar = tqdm(range(cfg.n_epoch), 'epoch')
for i in epoch_pbar:
train_pbar = tqdm(range(cfg.n_train), 'outer_train')
# Meta-training
for j in train_pbar:
train_data = data.sample_meta_train()
result, _ = optimizer.meta_optimize(cfg, meta_optim, train_data,
model_cls, writer, 'train')
result_mean = result.mean()
log_pbar(result_mean, train_pbar)
if cfg.save_dir:
step = (cfg.n_train * i) + j
train_result_mean.append(result_mean, step=step)
log_tf_event(writer, 'meta_train_outer', result_mean, step)
if cfg.save_dir:
save_figure(name, cfg.save_dir, writer, result, i, 'train')
valid_pbar = tqdm(range(cfg.n_valid), 'outer_valid')
result_all = ResultDict()
result_final = ResultDict()
# Meta-validation
for j in valid_pbar:
valid_data = data.sample_meta_valid()
result, params = optimizer.meta_optimize(cfg, meta_optim,
valid_data, model_cls,
writer, 'valid')
result_all.append(**result)
result_final.append(
final_inner_test(C(model_cls(params)),
valid_data['in_test'],
mode='valid'))
log_pbar(result.mean(), valid_pbar)
result_final_mean = result_final.mean()
result_all_mean = result_all.mean().w_postfix('mean')
last_valid_loss = result_final_mean['final_loss_mean']
last_valid_acc = result_final_mean['final_acc_mean']
# Learning rate scheduling
if lr_scheduling: # and last_valid < 0.5:
lr_scheduler.step(last_valid_loss, i)
# Truncation scheduling
if tr_scheduling:
tr_scheduler.step(last_valid_loss, i)
# Save TF events and figures
if cfg.save_dir:
step = cfg.n_train * (i + 1)
result_mean = ResultDict(
**result_all_mean,
**result_final_mean,
**get_lr(meta_optim),
step=step) #trunc_len=tr_scheduler.len, step=step)
valid_result_mean.append(result_mean)
log_tf_event(writer, 'meta_valid_outer', result_mean, step)
save_figure(name, cfg.save_dir, writer, result_all.mean(0), i,
'valid')
# Save the best snapshot
if last_valid_loss < best_valid_loss:
best_valid_loss = last_valid_loss
best_valid_acc = last_valid_acc
best_params = copy.deepcopy(optimizer.params)
if cfg.save_dir:
optimizer.params.save(name, cfg.save_dir)
# Update epoch progress bar
result_epoch = dict(
best_valid_loss=best_valid_loss,
best_valid_acc=best_valid_acc,
last_valid_loss=last_valid_loss,
last_valid_acc=last_valid_acc,
)
log_pbar(result_epoch, epoch_pbar)
if cfg.save_dir:
train_result_mean.save_as_csv(name, cfg.save_dir, 'meta_train_outer',
True)
valid_result_mean.save_as_csv(name, cfg.save_dir, 'meta_valid_outer',
True)
return best_params
def loop(mode,
outer_steps,
inner_steps,
log_steps,
fig_epochs,
inner_lr,
log_mask=True,
unroll_steps=None,
meta_batchsize=0,
sampler=None,
epoch=1,
outer_optim=None,
save_path=None):
"""Args:
meta_batchsize(int): If meta_batchsize |m| > 0, gradients for multiple
unrollings from each episodes size of |m| will be accumulated in
sequence but updated all at once. (Which can be done in parallel when
VRAM is large enough, but will be simulated in this code.)
If meta_batchsize |m| = 0(default), then update will be performed after
each unrollings.
"""
assert mode in ['train', 'valid', 'test']
assert meta_batchsize >= 0
print(f'Start_of_{mode}.')
if mode == 'train' and unroll_steps is None:
raise Exception("unroll_steps has to be specied when mode='train'.")
if mode != 'train' and unroll_steps is not None:
raise Warning("unroll_steps has no effect when mode mode!='train'.")
train = True if mode == 'train' else False
force_base = True
metadata = MetaMultiDataset(split=mode)
mask_based = 'query'
mask_type = 5
mask_sample = False
mask_scale = True
easy_ratio = 17 / 50 # 0.3
scale_manual = 0.8 # 1.0 when lr=0.001 and 0.8 when lr=0.00125
inner_lr *= scale_manual
if train:
if meta_batchsize > 0:
# serial processing of meta-minibatch
update_epochs = meta_batchsize
update_steps = None
else:
# update at every unrollings
update_steps = unroll_steps
update_epochs = None
assert (update_epochs is None) != (update_steps is None)
# for result recordin
result_frame = ResultFrame()
if save_path:
writer = SummaryWriter(os.path.join(save_path, 'tfevent'))
for i in range(1, outer_steps + 1):
outer_loss = 0
for j, epi in enumerate(metadata.loader(n_batches=1), 1):
# initialize base learner
model = Model(epi.n_classes)
params = model.get_init_params('ours')
epi.s = C(epi.s)
epi.q = C(epi.q)
# baseline parameters
params_b0 = C(params.copy('b0'))
params_b1 = C(params.copy('b1'))
params_b2 = C(params.copy('b2'))
result_dict = ResultDict()
for k in range(1, inner_steps + 1):
# feed support set (baseline)
out_s_b0 = model(epi.s, params_b0, None)
out_s_b1 = model(epi.s, params_b1, None)
out_s_b2 = model(epi.s, params_b2, None)
if mask_based == 'support':
out = out_s_b1
elif mask_based == 'query':
with torch.no_grad():
# test on query set
out_q_b1 = model(epi.q, params_b1, mask=None)
out = out_q_b1
else:
print('WARNING')
# attach mask to get loss_s
if mask_type == 1:
mask = (out.loss.exp().mean().log() - out.loss).exp()
elif mask_type == 2:
mask = (out.loss.exp().mean().log() / out.loss)
elif mask_type == 3:
mask = out.loss.mean() / out.loss
elif mask_type == 4:
mask = out.loss.min() / out.loss
elif mask_type == 5 or mask_type == 6:
mask_scale = False
# weight by magnitude
if mask_type == 5:
mask = [scale_manual
] * 5 + [(1 - easy_ratio) * scale_manual] * 5
# weight by ordering
elif mask_type == 6:
if k < inner_steps * easy_ratio:
mask = [scale_manual] * 5 + [0.0] * 5
else:
mask = [scale_manual] * 5 + [scale_manual] * 5
# sampling from 0 < p < 1
if mask_sample:
mask = [np.random.binomial(1, m) for m in mask]
mask = C(torch.tensor(mask).float())
else:
print('WARNING')
if mask_scale:
mask = (mask / (mask.max() + 0.05))
# to debug in the middle of running process.class
if sig_2.is_active():
import pdb
pdb.set_trace()
mask = mask.unsqueeze(1)
out_s_b1.attach_mask(mask)
out_s_b2.attach_mask(mask)
# lll = out_s_b0.loss * 0.5
params_b0 = params_b0.sgd_step(out_s_b0.loss.mean(), inner_lr,
'no_grad')
params_b1 = params_b1.sgd_step(out_s_b1.loss_masked_mean,
inner_lr, 'no_grad')
params_b2 = params_b2.sgd_step(out_s_b2.loss_scaled_mean,
inner_lr, 'no_grad')
with torch.no_grad():
# test on query set
out_q_b0 = model(epi.q, params_b0, mask=None)
out_q_b1 = model(epi.q, params_b1, mask=None)
out_q_b2 = model(epi.q, params_b2, mask=None)
# record result
result_dict.append(outer_step=epoch * i,
inner_step=k,
**out_s_b0.as_dict(),
**out_s_b1.as_dict(),
**out_s_b2.as_dict(),
**out_q_b0.as_dict(),
**out_q_b1.as_dict(),
**out_q_b2.as_dict())
### end of inner steps (k) ###
# append to the dataframe
result_frame = result_frame.append_dict(
result_dict.index_all(-1).mean_all(-1))
# logging
if k % log_steps == 0:
# print info
msg = Printer.step_info(epoch, mode, i, outer_steps, k,
inner_steps, inner_lr)
msg += Printer.way_shot_query(epi)
# # print mask
if not sig_1.is_active() and log_mask:
msg += Printer.colorized_mask(mask,
fmt="3d",
vis_num=20)
# print outputs (loss, acc, etc.)
msg += Printer.outputs([out_s_b0, out_s_b1, out_s_b2],
sig_1.is_active())
msg += Printer.outputs([out_q_b0, out_q_b1, out_q_b2],
sig_1.is_active())
print(msg)
### end of meta minibatch (j) ###
# tensorboard
if save_path and train:
step = (epoch * (outer_steps - 1)) + i
res = ResultFrame(
result_frame[result_frame['outer_step'] == i])
loss = res.get_best_loss().mean()
acc = res.get_best_acc().mean()
writer.add_scalars('Loss/train',
{n: loss[n]
for n in loss.index}, step)
writer.add_scalars('Acc/train', {n: acc[n]
for n in acc.index}, step)
# dump figures
if save_path and i % fig_epochs == 0:
epi.s.save_fig(f'imgs/support', save_path, i)
epi.q.save_fig(f'imgs/query', save_path, i)
# result_dict['ours_s_mask'].save_fig(f'imgs/masks', save_path, i)
result_dict.get_items([
'b0_s_loss', 'b1_s_loss', 'b2_s_loss'
'b0_q_loss', 'b1_q_loss', 'b2_q_loss'
]).save_csv(f'classwise/{mode}', save_path, i)
# distinguishable episodes
if not i == outer_steps:
print(f'Path for saving: {save_path}')
print(f'End_of_episode: {i}')
import pdb
pdb.set_trace()
### end of episode (i) ###
print(f'End_of_{mode}.')
# del metadata
return sampler, result_frame
def meta_optimize(self,
meta_optimizer,
data,
model_cls,
optim_it,
unroll,
out_mul,
tf_writer=None,
mode='train'):
assert mode in ['train', 'valid', 'test']
if mode == 'train':
self.train()
else:
self.eval()
result_dict = ResultDict()
unroll_losses = 0
walltime = 0
params = C(model_cls()).params
# lr_sgd = 0.2
# lr_adam = 0.001
# sgd = torch.optim.SGD(model.parameters(), lr=lr_sgd)
# adam = torch.optim.Adam(model.parameters())
# adagrad = torch.optim.Adagrad(model.parameters())
iter_pbar = tqdm(range(1, optim_it + 1), 'inner_train')
iter_watch = utils.StopWatch('inner_train')
lr = 0.1
topk = False
n_sample = 10
topk_decay = True
recompute_grad = False
self.topk_ratio = 0.5
# self.topk_ratio = 1.0
set_size = {'layer_0': 500, 'layer_1': 10} # NOTE: do it smarter
for iteration in iter_pbar:
# # conventional optimizers with API
# model.zero_grad()
# loss = model(*inner_data.load())
# loss.backward()
# adam.step()
# loss_detached = loss
iter_watch.touch()
model_train = C(model_cls(params=params.detach()))
train_nll = model_train(*data['in_train'].load())
train_nll.backward()
params = model_train.params.detach()
grad = model_train.params.grad.detach()
act = model_train.activations.detach()
# conventional optimizers with manual SGD
params_good = params + grad * (-lr)
grad_good = grad
best_loss_sparse = 999999
for i in range(n_sample):
mask_gen = MaskGenerator.topk if topk else MaskGenerator.randk
mask = mask_gen(grad=grad,
set_size=set_size,
topk=self.topk_ratio)
if recompute_grad:
# recompute gradients again using prunned network
params_sparse = params.prune(mask)
model_sparse = C(model_cls(params=params_sparse.detach()))
loss_sparse = model_sparse(*data['in_train'].load())
loss_sparse.backward()
grad_sparse = model_sparse.params.grad
else:
# use full gradients but sparsified
params_sparse = params.prune(mask)
grad_sparse = grad.prune(mask)
params_sparse_ = params_sparse + grad_sparse * (-lr) # SGD
model_sparse = C(model_cls(params=params_sparse_.detach()))
loss_sparse = model_sparse(*data['in_train'].load())
if loss_sparse < best_loss_sparse:
best_loss_sparse = loss_sparse
best_params_sparse = params_sparse_
best_grads = grad_sparse
best_mask = mask
# below lines are just for evaluation purpose (excluded from time cost)
walltime += iter_watch.touch('interval')
# decayu topk_ratio
if topk_decay:
self.topk_ratio *= 0.999
# update!
params = params.sparse_update(best_mask, best_grads * (-lr))
# generalization performance
model_test = C(model_cls(params=params.detach()))
test_nll = model_test(*data['in_test'].load())
# result dict
result = dict(
train_nll=train_nll.tolist(),
test_nll=test_nll.tolist(),
sparse_0=self.topk_ratio,
sparse_1=self.topk_ratio,
)
result_dict.append(result)
log_pbar(result, iter_pbar)
# desc = [f'{k}: {v:5.5}' for k, v in result.items()]
# iter_pbar.set_description(f"inner_train [ {' / '.join(desc)} ]")
return result_dict
def loop(mode,
data,
outer_steps,
inner_steps,
log_steps,
fig_epochs,
inner_lr,
log_mask=True,
unroll_steps=None,
meta_batchsize=0,
sampler=None,
epoch=1,
outer_optim=None,
save_path=None):
"""Args:
meta_batchsize(int): If meta_batchsize |m| > 0, gradients for multiple
unrollings from each episodes size of |m| will be accumulated in
sequence but updated all at once. (Which can be done in parallel when
VRAM is large enough, but will be simulated in this code.)
If meta_batchsize |m| = 0(default), then update will be performed after
each unrollings.
"""
assert mode in ['train', 'valid', 'test']
assert meta_batchsize >= 0
print(f'Start_of_{mode}.')
if mode == 'train' and unroll_steps is None:
raise Exception("unroll_steps has to be specied when mode='train'.")
if mode != 'train' and unroll_steps is not None:
raise Warning("unroll_steps has no effect when mode mode!='train'.")
samples_per_class = 3
train = True if mode == 'train' else False
force_base = True
model_mode = 'metric'
# split_method = 'inclusive'
split_method = 'exclusive'
# 100 classes in total
if split_method == 'exclusive':
meta_support, remainder = data.split_classes(0.1) # 10 classes
meta_query = remainder.sample_classes(50) # 50 classes
elif split_method == 'inclusive':
subdata = data.sample_classes(10) # 50 classes
meta_support, meta_query = subdata.split_instances(
0.5) # 5:5 instances
else:
raise Exception()
if train:
if meta_batchsize > 0:
# serial processing of meta-minibatch
update_epochs = meta_batchsize
update_steps = None
else:
# update at every unrollings
update_steps = unroll_steps
update_epochs = None
assert (update_epochs is None) != (update_steps is None)
# for result recordin
result_frame = ResultFrame()
if save_path:
writer = SummaryWriter(os.path.join(save_path, 'tfevent'))
for i in range(1, outer_steps + 1):
fc_lr = 0.01
metric_lr = 0.01
import pdb
pdb.set_trace()
model_fc = Model(len(meta_support), mode='fc')
model_metric = Model(len(meta_support), mode='metric')
# baseline parameters
params_fc = C(model_fc.get_init_params('fc'))
params_metric = C(model_metric.get_init_params('metric'))
result_dict = ResultDict()
episode_iterator = EpisodeIterator(
support=meta_support,
query=meta_query,
split_ratio=0.5,
resample_every_iteration=True,
inner_steps=inner_steps,
samples_per_class=samples_per_class,
num_workers=2,
pin_memory=True,
)
episode_iterator.sample_episode()
for k, (meta_s, meta_q) in enumerate(episode_iterator, 1):
# import pdb; pdb.set_trace()
# feed support set
# [standard network]
out_s_fc = model_fc(meta_s, params_fc, None)
with torch.no_grad():
# embedding space metric
params_fc_m = C(params_fc.copy('fc_m'))
out_s_fc_m = model_metric(meta_s, params_fc_m, mask=None)
# [prototypical network]
out_s_metric = model_metric(meta_s, params_metric, None)
# inner gradient step (baseline)
params_fc = params_fc.sgd_step(out_s_fc.loss.mean(), fc_lr,
'no_grad')
params_metric = params_metric.sgd_step(out_s_metric.loss.mean(),
metric_lr, 'no_grad')
# test on query set
with torch.no_grad():
if split_method == 'inclusive':
out_q_fc = model_fc(meta_q, params_fc, mask=None)
params_fc_m = C(params_fc.copy('fc_m'))
out_q_fc_m = model_metric(meta_q, params_fc_m, mask=None)
out_q_metric = model_metric(meta_q, params_metric, mask=None)
if split_method == 'inclusive':
outs = [
out_s_fc, out_s_fc_m, out_s_metric, out_q_fc, out_q_fc_m,
out_q_metric
]
elif split_method == 'exclusive':
outs = [
out_s_fc, out_s_fc_m, out_s_metric, out_q_fc_m,
out_q_metric
]
# record result
result_dict.append(outer_step=epoch * i,
inner_step=k,
**ModelOutput.as_merged_dict(outs))
### end of inner steps (k) ###
# append to the dataframe
result_frame = result_frame.append_dict(
result_dict.index_all(-1).mean_all(-1))
# logging
if k % log_steps == 0:
# print info
msg = Printer.step_info(epoch, mode, i, outer_steps, k,
inner_steps, inner_lr)
# msg += Printer.way_shot_query(epi)
# print mask
if not sig_1.is_active() and log_mask:
# print outputs (loss, acc, etc.)
msg += Printer.outputs(outs, True)
print(msg)
# to debug in the middle of running process.
if k == inner_steps and sig_2.is_active():
import pdb
pdb.set_trace()
# # tensorboard
# if save_path and train:
# step = (epoch * (outer_steps - 1)) + i
# res = ResultFrame(result_frame[result_frame['outer_step'] == i])
# loss = res.get_best_loss().mean()
# acc = res.get_best_acc().mean()
# writer.add_scalars(
# 'Loss/train', {n: loss[n] for n in loss.index}, step)
# writer.add_scalars('Acc/train', {n: acc[n] for n in acc.index}, step)
#
# # dump figures
# if save_path and i % fig_epochs == 0:
# meta_s.s.save_fig(f'imgs/meta_support', save_path, i)
# meta_q.s.save_fig(f'imgs/meta_query', save_path, i)
# result_dict['ours_s_mask'].save_fig(f'imgs/masks', save_path, i)
# result_dict.get_items(['ours_s_mask', 'ours_s_loss',
# 'ours_s_loss_masked', 'b0_s_loss', 'b1_s_loss',
# 'ours_q_loss', 'b0_q_loss', 'b1_q_loss']).save_csv(
# f'classwise/{mode}', save_path, i)
# distinguishable episodes
if not i == outer_steps:
print(f'Path for saving: {save_path}')
print(f'End_of_episode: {i}')
### end of episode (i) ###
print(f'End_of_{mode}.')
# del metadata
return sampler, result_frame