def update(self, preds: torch.Tensor, target: torch.Tensor):
do_print = self.print_every is not None and not self.counter % self.print_every
if do_print:
utils.hprint(f'BetterAccuracy is set to print every {self.print_every} and you at {self.counter}:')
print(f"BetterAccuracy: preds: \n{preds}")
print(f"BetterAccuracy: target: \n{target}")
print()
self.counter += 1
# preds, target = _input_format_classification(preds, target, self.threshold)
assert preds.shape == target.shape, f'preds.shape = {preds.shape} != target.shape = {target.shape}'
preds = preds.argmax(dim=1)
target = target.argmax(dim=1)
if do_print:
print(f"BetterAccuracy: preds post argmax: \n{preds}")
print(f"BetterAccuracy: target post argmax: \n{target}")
print()
assert target.dim() == 1, f'got target of shape {target.shape}'
eqs = preds.eq(target)
if do_print:
print(
f"BetterAccuracy: new_correct: {eqs.sum()}, "
f" numel: {target.numel()}, "
f" shape[0]: {target.shape[0]}, "
f" ignore: {target.eq(self.Y_VALUE_TO_IGNORE).sum()}"
)
self.correct = self.correct + torch.sum(eqs)
self.total = self.total + target.shape[0]
def on_train_start(self, trainer: pl.Trainer, pl_module: pl.LightningModule):
field_to_class_counts = self.get_field_to_class_counts(dataloader=pl_module.train_dataloader())
if self.verbose:
utils.hprint('ClassCounterCallback Class Counts:')
utils.print_dict(field_to_class_counts)
print()
if self.hp is None:
if self.verbose:
print(f' Not setting head_params.pos_class_weights because you did not pass hp to my init')
else:
if self.hp.type != 'weighted':
raise NotImplementedError(
f'hp == {self.hp} but this is only implemented for WeightedHeadParams'
)
for field_name, class_counts_df in field_to_class_counts.items():
if field_name not in self.field_name_to_head_name:
# we might not be using all fields in heads
continue
head_name = self.field_name_to_head_name[field_name]
head = pl_module.head.heads[head_name]
if head.did_set_pos_class_weights:
pos_class_weights = head.pos_class_weights
if self.verbose:
weights_str = ', '.join([f'{e:.2f}' for e in pos_class_weights])
print(f' head_params["{field_name}"].pos_class_weights was already set to [{weights_str}]')
print()
else:
pos_class_weights = class_counts_df.loc[self.INV_PORTIONS].values
max_inds = np.where(pos_class_weights > self.max_pos_class_weight)[0]
pos_class_weights[max_inds] = self.max_pos_class_weight
head.set_pos_class_weights(
torch.tensor(pos_class_weights, dtype=torch.float, device=pl_module.device)
)
if self.verbose:
weights_str = ', '.join([f'{e:.2f}' for e in pos_class_weights])
print(f' Setting head_params["{field_name}"].pos_class_weights = [{weights_str}]')
print()
pl_module.log_lossmetrics_dict(
phase=utils.Phase.train,
d={self.CLASS_COUNTS: field_to_class_counts},
do_log_to_progbar=False,
)
def run(self, fast_dev_run=False, use_gpus=False, log_to_file=False):
search_dict = self.search_params.to_ray_tune_search_dict()
# see tune.utils.UtilMonitor
search_dict['log_sys_usage'] = True
# noinspection PyTypeChecker
analysis = tune.run(
run_or_experiment=self._get_train_fn(fast_dev_run=fast_dev_run,
include_gpus=use_gpus),
name=self.search_params.exp.get_project_exp_name(),
stop=self.get_tune_stopper(self.search_params.opt.num_epochs),
config=search_dict,
resources_per_trial=self.get_resources_per_trial(
self.search_params, include_gpu=use_gpus),
num_samples=self.search_params.tune.num_hp_samples,
sync_config=tune.SyncConfig(
upload_dir=self.search_params.metrics.output_dir),
loggers=self.get_tune_loggers(),
log_to_file=log_to_file,
keep_checkpoints_num=2,
checkpoint_score_attr=
f'{self.search_params.opt.search_mode}-{self.search_params.opt.search_metric}',
fail_fast=False,
scheduler=self.get_tune_scheduler(self.search_params),
verbose=2,
progress_reporter=self.get_cli_reporter(),
reuse_actors=False,
)
utils.hprint("done with tune.run")
param_hash = self.search_params.get_short_hash(num_chars=8)
analysis_file = self.search_params.metrics.output_dir / f'tune_analysis_{param_hash}.pkl'
print(f"Saving {analysis_file}")
utils.save_pickle(analysis_file, analysis)
best_trial = analysis.get_best_trial(
self.search_params.opt.search_metric,
self.search_params.opt.search_mode, "last-5-avg")
print(f'best_trial.last_result: {best_trial.last_result}')
print("Best trial config: {}".format(best_trial.config))
print("Best trial final search_metric: {}".format(
best_trial.last_result[self.search_params.opt.search_metric]))
def _train_fn(self,
config: Dict,
checkpoint_dir=None,
fast_dev_run=False,
include_gpus=False):
utils.hprint('Starting train function with config:')
utils.print_dict(config)
del config['tune']
hp = self._param_class.from_dict(config)
assert isinstance(hp, self._param_class)
if checkpoint_dir:
# see https://docs.ray.io/en/master/tune/user-guide.html#checkpointing
raise NotImplementedError(
f"Got checkpoint_dir in trian_fn: {checkpoint_dir}")
net = self._factored_lightning_module_class.from_hp(hp=hp)
utils.set_seeds(hp.data.seed)
# noinspection PyTypeChecker
trainer = pl.Trainer(
logger=torch_mod.get_pl_logger(hp=hp.exp,
tune=tune,
offline_mode=fast_dev_run),
default_root_dir=tune.get_trial_dir(),
callbacks=self.extra_pl_callbacks + self.get_tune_callbacks(),
max_epochs=hp.opt.num_epochs,
gpus=hp.data.num_gpus if include_gpus else None,
weights_summary='full',
fast_dev_run=fast_dev_run,
accumulate_grad_batches=1,
profiler='simple',
deterministic=True,
log_every_n_steps=hp.metrics.num_steps_per_metric_log,
)
fit_out = trainer.fit(net, datamodule=net.dm)
utils.print_dict(config)
utils.hprint('Done with tune_runner._train_fn')
return fit_out
def run(
net: pl.LightningModule,
dm: pl.LightningDataModule,
hp: ModelBertConvTransTClass.Params,
fast_dev_run=False,
do_find_lr=False,
):
print("model run about to create trainer")
trainer = pl.Trainer(
logger=True if fast_dev_run else torch_mod.get_pl_logger(hp.exp),
default_root_dir=hp.metrics.output_dir,
callbacks=[metrics_mod.CounterTimerCallback()],
max_epochs=hp.opt.num_epochs,
gpus=hp.data.num_gpus,
weights_summary='full',
fast_dev_run=fast_dev_run,
accumulate_grad_batches=1,
profiler=True,
deterministic=True,
auto_lr_find=do_find_lr,
log_every_n_steps=hp.metrics.num_steps_per_metric_log,
)
print("model run done creating trainer")
if do_find_lr:
utils.hprint("Starting trainer.tune:")
lr_tune_out = trainer.tune(net, datamodule=dm)
print(f' Tune out: {lr_tune_out}')
else:
utils.hprint("Starting trainer.fit:")
print(f' Dataset file: {hp.data.get_dataset_file()}')
trainer.fit(net, datamodule=dm)
utils.hprint('Done with model run fn')
def _train_fn(self,
config: Dict,
checkpoint_dir=None,
fast_dev_run=False,
include_gpus=False):
utils.hprint('Starting train function with config:')
utils.print_dict(config)
print()
utils.set_pandas_disp(width=200)
hp = self._model_param_class.from_dict(config)
assert isinstance(hp, self._model_param_class)
print(' hp:', hp)
if checkpoint_dir:
# see https://docs.ray.io/en/master/tune/user-guide.html#checkpointing
raise NotImplementedError(
f"Got checkpoint_dir in trian_fn: {checkpoint_dir}")
utils.hprint("About to create net in TuneRunner")
net = hp.build()
# import torch.autograd.profiler as profiler
# with profiler.profile(record_shapes=True, use_cuda=True, profile_memory=True) as prof:
# net = self._factored_lightning_module_class.from_hp(hp=hp)
# print(prof.key_averages().table(sort_by="cpu_time_total", row_limit=1000))
utils.set_seeds(hp.data.seed)
# noinspection PyTypeChecker
trainer = pl.Trainer(
logger=logs_mod.get_pl_logger(hp=hp.exp, tune=tune),
default_root_dir=tune.get_trial_dir(),
callbacks=self.extra_pl_callbacks +
self.get_pl_callbacks_for_tune(),
max_epochs=hp.opt.num_epochs,
gpus=hp.data.num_gpus if include_gpus else None,
weights_summary='full',
fast_dev_run=fast_dev_run,
accumulate_grad_batches=1,
profiler='simple',
deterministic=True,
log_every_n_steps=hp.logs.num_steps_per_metric_log,
log_gpu_memory=hp.logs.log_gpu_memory,
)
utils.hprint('About to start tune_runner\'s trainer.fit...')
fit_out = trainer.fit(net, datamodule=net.dm)
utils.hprint('Done with tune_runner._train_fn')
return fit_out
def run(self, fast_dev_run=False, use_gpus=False):
utils.set_seeds(self.search_params.data.seed)
search_dict = self.search_params.to_ray_tune_search_dict()
# see tune.utils.UtilMonitor
search_dict['log_sys_usage'] = True
output_str = str(self.search_params.logs.output_dir)
if output_str.startswith('s3://') or output_str.startswith(
'gs://') or output_str.startswith('hdfs://'):
sync_config = tune.SyncConfig(
upload_dir=self.search_params.logs.output_dir)
else:
sync_config = None
analysis = tune.run(
run_or_experiment=self._get_train_fn(fast_dev_run=fast_dev_run,
include_gpus=use_gpus),
name=self.search_params.exp.get_project_exp_name(),
stop=self.get_tune_stopper(self.search_params.opt.num_epochs),
config=search_dict,
resources_per_trial=self.get_resources_per_trial(
self.search_params, include_gpu=use_gpus),
num_samples=self.tune_hp.num_hp_samples,
sync_config=sync_config,
loggers=self.get_tune_loggers(),
log_to_file=self.tune_hp.log_to_file
and not self.tune_hp.ray_local_mode,
keep_checkpoints_num=2,
checkpoint_score_attr=
f'{self.search_params.opt.search_mode}-{self.search_params.opt.search_metric}',
fail_fast=False,
scheduler=self.get_tune_scheduler(self.search_params,
self.tune_hp),
verbose=2,
progress_reporter=self.get_cli_reporter(),
reuse_actors=False,
)
utils.hprint("done with tune.run")
param_hash = self.search_params.get_short_hash(num_chars=8)
analysis_file = self.search_params.logs.output_dir / f'tune_analysis_{param_hash}.cloudpickle'
print(f"Saving {analysis_file}")
utils.save_cloudpickle(analysis_file, analysis)
best_trial = analysis.get_best_trial(
self.search_params.opt.search_metric,
self.search_params.opt.search_mode, "last-5-avg")
utils.hprint('best_trial.last_result', do_include_pre_break_line=True)
utils.print_dict(best_trial.last_result)
utils.hprint('best_trial.config', do_include_pre_break_line=True)
utils.print_dict(best_trial.config)
def run(
net: pl.LightningModule,
hp: TotalParams,
fast_dev_run=False,
do_find_lr=False,
callbacks=None,
):
utils.set_seeds(hp.data.seed)
utils.set_pandas_disp()
if callbacks is None:
callbacks = [
logs_mod.CounterTimerLrCallback(),
logs_mod.VocabLengthCallback(),
]
print("model run about to create trainer")
trainer = pl.Trainer(
logger=True if fast_dev_run else logs_mod.get_pl_logger(hp.exp),
default_root_dir=hp.logs.output_dir,
callbacks=callbacks,
max_epochs=hp.opt.num_epochs,
gpus=hp.data.num_gpus,
weights_summary='full',
fast_dev_run=fast_dev_run,
accumulate_grad_batches=1,
profiler='simple',
deterministic=True,
auto_lr_find=do_find_lr,
log_every_n_steps=hp.logs.num_steps_per_metric_log,
)
print("model run done creating trainer")
if do_find_lr:
utils.hprint("Starting trainer.tune:")
lr_tune_out = trainer.tune(net, datamodule=net.dm)
print(f' Tune out: {lr_tune_out}')
else:
utils.hprint("Starting trainer.fit:")
print(f' Dataset file: {hp.data.dataset_file}')
trainer.fit(net, datamodule=net.dm)
utils.hprint('Done with model run fn')
hp.heads.num_features = params.Discrete([32, 64, 128, 256])
hp.heads.num_layers = params.Integer(2, 5)
hp.heads.num_groups = params.Discrete([8, 16, 32, 64])
hp.heads.num_blocks_per_residual = params.Integer(1, 5)
hp.heads.num_blocks_per_dropout = params.Integer(1, 5)
hp.heads.requires_grad = True
hp.tune = TuneRunner.TuneParams()
hp.tune.asha_grace_period = 16
hp.tune.asha_reduction_factor = 2
hp.tune.num_hp_samples = 100
hostname = socket.gethostname()
is_local_run = hostname.endswith('.local')
# noinspection PyTypeChecker
tune_runner = TuneRunner(
search_params=hp,
factored_lightning_module_class=model_bert_trans_conv_tclass.
ModelBertConvTransTClass,
extra_pl_callbacks=None,
ray_local_mode=False,
)
tune_runner.run(
fast_dev_run=False,
use_gpus=not is_local_run,
log_to_file=True,
)
utils.hprint("done with tune_runner.run")
hp.trans.num_layers = 6
hp.trans.num_query_features = None
hp.trans.fc_dim_mult = 2
hp.fc.num_features = 128
hp.fc.num_layers = 2
hp.fc.num_groups = 16
hp.fc.num_blocks_per_residual = 2
hp.fc.num_blocks_per_dropout = 2
hp.fc.requires_grad = True
hp.heads.num_features = 128
hp.heads.num_layers = 4
hp.heads.num_groups = 16
hp.heads.num_blocks_per_residual = 2
hp.heads.num_blocks_per_dropout = 2
hp.heads.requires_grad = True
dm = data.TablestakesDataModule(hp.data)
net = ModelBertConvTransTClass(
hp=hp,
data_module=dm,
metrics_tracker=metrics_mod.ClassificationMetricsTracker(hp.metrics),
opt=factored.OptimizersMaker(hp.opt),
)
utils.hprint('About to start model run:')
utils.print_dict(hp.to_dict())
run(net, dm, hp, fast_dev_run, do_find_lr=False)
},
)
hp.verbose = False
tune_hp = TuneParams()
tune_hp.asha_grace_period = 4
tune_hp.asha_reduction_factor = 2
tune_hp.num_hp_samples = 2
tune_hp.log_to_file = False
tune_hp.ray_local_mode = False
hostname = socket.gethostname()
is_local_run = hostname.endswith('.local')
utils.hprint('About to start model run:')
utils.print_dict(hp.to_dict())
tune_runner = TuneRunner(
model_hp=hp,
tune_hp=tune_hp,
# factored_lightning_module_class=ts_model.TablestakesBertConvTransTClassModel,
extra_pl_callbacks=None,
)
tune_runner.run(
fast_dev_run=False,
use_gpus=not is_local_run,
)
utils.hprint("done with tune_runner.run")
