def compute_reward(self, total_macs, total_nnz, log_stats=True):
"""Compute the reward.
We use the validation dataset (the size of the validation dataset is
configured when the data-loader is instantiated)"""
distiller.log_weights_sparsity(self.model, -1, loggers=[self.pylogger])
compression = distiller.model_numel(self.model, param_dims=[4]) / self.original_model_size
# Fine-tune (this is a nop if self.amc_cfg.num_ft_epochs==0)
accuracies = self.net_wrapper.train(self.amc_cfg.num_ft_epochs, self.episode)
self.ft_stats_logger.add_record([self.episode, accuracies])
top1, top5, vloss = self.net_wrapper.validate()
reward = self.amc_cfg.reward_fn(self, top1, top5, vloss, total_macs)
if log_stats:
macs_normalized = total_macs/self.original_model_macs
msglogger.info("Total parameters left: %.2f%%" % (compression*100))
msglogger.info("Total compute left: %.2f%%" % (total_macs/self.original_model_macs*100))
stats = ('Performance/EpisodeEnd/',
OrderedDict([('Loss', vloss),
('Top1', top1),
('Top5', top5),
('reward', reward),
('total_macs', int(total_macs)),
('macs_normalized', macs_normalized*100),
('log(total_macs)', math.log(total_macs)),
('total_nnz', int(total_nnz))]))
distiller.log_training_progress(stats, None, self.episode, steps_completed=0, total_steps=1,
log_freq=1, loggers=[self.tflogger, self.pylogger])
return reward, top1
def compute_reward(self):
"""The ADC paper defines reward = -Error"""
distiller.log_weights_sparsity(self.model, -1, loggers=[self.pylogger])
compression = distiller.model_numel(
self.model, param_dims=[4]) / self.dense_model_size
_, total_macs, total_nnz = collect_conv_details(
self.model, self.dataset)
msglogger.info("Total parameters left: %.2f%%" % (compression * 100))
msglogger.info("Total compute left: %.2f%%" %
(total_macs / self.dense_model_macs * 100))
top1, top5, vloss = self.validate_fn(model=self.model,
epoch=self.debug_stats['episode'])
reward = self.reward_fn(top1, top5, vloss, total_macs)
stats = (
'Peformance/Validation/',
OrderedDict([
('Loss', vloss),
('Top1', top1),
('Top5', top5),
('reward', reward),
('total_macs', int(total_macs)),
('log(total_macs)', math.log(total_macs)),
#('log(total_macs/self.dense_model_macs)', math.log(total_macs/self.dense_model_macs)),
('total_nnz', int(total_nnz))
]))
distiller.log_training_progress(stats,
None,
self.debug_stats['episode'],
steps_completed=0,
total_steps=1,
log_freq=1,
loggers=[self.tflogger, self.pylogger])
return reward, top1, total_macs, total_nnz
def compute_reward(self):
"""The ADC paper defines reward = -Error"""
distiller.log_weights_sparsity(self.model, -1, loggers=[self.pylogger])
compression = distiller.model_numel(
self.model, param_dims=[4]) / self.dense_model_size
_, total_macs, total_nnz = collect_conv_details(
self.model, self.dataset)
msglogger.info("Total parameters left: %.2f%%" % (compression * 100))
msglogger.info("Total compute left: %.2f%%" %
(total_macs / self.dense_model_macs * 100))
top1, top5, vloss = self.validate_fn(model=self.model,
epoch=self.debug_stats['episode'])
#reward = -1 * (1 - top1/100)
if self.desired_reduction is not None:
reward = top1 / 100
else:
reward = -1 * (1 - top1 / 100) * math.log(total_macs)
#reward = -1 * (1-top1/100) * math.log(total_macs/self.dense_model_macs)
#
#reward = -1 * (1-top1/100) + math.log(total_macs/self.dense_model_macs)
#reward = 4*top1/100 - math.log(total_macs)
#reward = reward * total_macs/213201664
#reward = reward - 5 * total_macs/213201664
#reward = -1 * vloss * math.sqrt(math.log(total_macs))
#reward = top1 / math.log(total_macs)
#alpha = 0.9
#reward = -1 * ( (1-alpha)*(top1/100) + 10*alpha*(total_macs/self.dense_model_macs) )
#alpha = 0.99
#reward = -1 * ( (1-alpha)*(top1/100) + alpha*(total_macs/self.dense_model_macs) )
#reward = vloss * math.log(total_macs)
#reward = -1 * vloss * (total_macs / self.dense_model_macs)
#reward = top1 * (self.dense_model_macs / total_macs)
#reward = -1 * math.log(total_macs)
#reward = -1 * vloss
stats = (
'Peformance/Validation/',
OrderedDict([
('Loss', vloss),
('Top1', top1),
('Top5', top5),
('reward', reward),
('total_macs', int(total_macs)),
('log(total_macs)', math.log(total_macs)),
#('log(total_macs/self.dense_model_macs)', math.log(total_macs/self.dense_model_macs)),
('total_nnz', int(total_nnz))
]))
distiller.log_training_progress(stats,
None,
self.debug_stats['episode'],
steps_completed=0,
total_steps=1,
log_freq=1,
loggers=[self.tflogger, self.pylogger])
return reward, top1
def get_experiment_performance_summary(chkpt_fname, dataset, arch, validate_fn):
model = create_model(False, dataset, arch)
model, compression_scheduler, start_epoch = apputils.load_checkpoint(model, chkpt_fname)
dummy_input = get_dummy_input(dataset)
perf_df = distiller.model_performance_summary(model, dummy_input, 1)
total_macs = perf_df['MACs'].sum()
top1, top5, vloss = validate_fn(model=model, epoch=-1)
return total_macs, distiller.model_numel(model), top1
def compute_reward(self):
"""The ADC paper defines reward = -Error"""
distiller.log_weights_sparsity(self.model, -1, loggers=[self.pylogger])
if PERFORM_THINNING:
_, total_macs, total_nnz = collect_conv_details(
self.model, self.app_args.dataset)
compression = distiller.model_numel(
self.model, param_dims=[4]) / self.dense_model_size
else:
_, total_macs, total_nnz = collect_conv_details(
self.model, self.app_args.dataset)
compression = 1 - distiller.model_sparsity(self.model) / 100
# What a hack!
total_nnz *= compression
msglogger.info("Total parameters left: %.2f%%" % (compression * 100))
msglogger.info("Total compute left: %.2f%%" %
(total_macs / self.dense_model_macs * 100))
# Train for zero or more epochs
optimizer = torch.optim.SGD(
self.model.parameters(),
lr=self.app_args.optimizer_data['lr'],
momentum=self.app_args.optimizer_data['momentum'],
weight_decay=self.app_args.optimizer_data['weight_decay'])
for _ in range(NUM_TRAINING_EPOCHS):
self.services.train_fn(
model=self.model,
compression_scheduler=self.create_scheduler(),
optimizer=optimizer,
epoch=self.debug_stats['episode'])
# Validate
top1, top5, vloss = self.services.validate_fn(
model=self.model, epoch=self.debug_stats['episode'])
reward = self.amc_cfg.reward_fn(top1, top5, vloss, total_macs)
stats = ('Peformance/Validation/',
OrderedDict([('Loss', vloss), ('Top1', top1), ('Top5', top5),
('reward', reward),
('total_macs', int(total_macs)),
('macs_normalized',
total_macs / self.dense_model_macs * 100),
('log(total_macs)', math.log(total_macs)),
('total_nnz', int(total_nnz))]))
distiller.log_training_progress(stats,
None,
self.debug_stats['episode'],
steps_completed=0,
total_steps=1,
log_freq=1,
loggers=[self.tflogger, self.pylogger])
return reward, top1, total_macs, total_nnz
def compute_reward(self, log_stats=True):
"""Compute the reward"""
distiller.log_weights_sparsity(self.model, -1, loggers=[self.pylogger])
total_macs, total_nnz = self.net_wrapper.get_model_resources_requirements(
self.model)
if self.amc_cfg.perform_thinning:
compression = distiller.model_numel(
self.model, param_dims=[4]) / self.dense_model_size
else:
compression = 1 - distiller.model_sparsity(self.model) / 100
# What a hack!
total_nnz *= compression
accuracies = self.net_wrapper.train(self.amc_cfg.num_ft_epochs,
self.episode)
self.ft_stats_file.add_record([self.episode, accuracies])
top1, top5, vloss = self.net_wrapper.validate()
reward = self.amc_cfg.reward_fn(self, top1, top5, vloss, total_macs)
if log_stats:
macs_normalized = total_macs / self.dense_model_macs
msglogger.info("Total parameters left: %.2f%%" %
(compression * 100))
msglogger.info("Total compute left: %.2f%%" %
(total_macs / self.dense_model_macs * 100))
stats = ('Performance/EpisodeEnd/',
OrderedDict([('Loss', vloss), ('Top1', top1),
('Top5', top5), ('reward', reward),
('total_macs', int(total_macs)),
('macs_normalized', macs_normalized * 100),
('log(total_macs)', math.log(total_macs)),
('total_nnz', int(total_nnz))]))
distiller.log_training_progress(
stats,
None,
self.episode,
steps_completed=0,
total_steps=1,
log_freq=1,
loggers=[self.tflogger, self.pylogger])
return reward, top1, total_macs, total_nnz