import sys
sys.path.insert(0, os.path.abspath('../machine-tasks'))
from tasks import get_task
from loss import L1Loss
comet_args = {
'project_name': 'attentive-guidance',
'workspace': 'andresespinosapc',
}
if os.environ.get('COMET_DISABLE'):
comet_args['disabled'] = True
comet_args['api_key'] = ''
if os.environ.get('COMET_OFFLINE'):
comet_args['api_key'] = ''
comet_args['offline_directory'] = 'comet_offline'
experiment = OfflineExperiment(**comet_args)
else:
experiment = Experiment(**comet_args)
def log_comet_parameters(opt):
opt_dict = vars(opt)
for key in opt_dict.keys():
experiment.log_parameter(key, opt_dict[key])
TASK_DEFAULT_PARAMS = {
'task_defaults': {
'batch_size': 128,
'k': 3,
'max_len': 60,
def run_experiment_iter(i, experiment, train_iter, nExp, agent_list, env,
video, user_seed, experiment_name, log_params, debug,
project_name, sps, sps_es, **kwargs):
"""
Function used to paralelize the run_experiment calculations.
Parameters
----------
i : int
Index of the agent being trained.
Raises
------
NotImplementedError
In case Comet is used, raises this error to signal where user intervention
is required (namely to set the api_key and the workspace).
Returns
-------
rewards : array
An array with the cumulative rewards, where each column corresponds to
an agent (random seed), and each row to a training iteration.
arms : array
An array with the number of agent arms, where each column corresponds
to an agent (random seed), and each row to a training iteration.
agent : Agent
The trained agent.
"""
if debug:
start = time.time()
print("Experiment {0} out of {1}...".format(i + 1, nExp))
if not user_seed:
seed = int.from_bytes(os.urandom(4), 'big')
else:
seed = user_seed
if experiment_name:
raise NotImplementedError(
"Before using Comet, you need to come here and set your API key")
experiment = Experiment(api_key=None,
project_name=project_name,
workspace=None,
display_summary=False,
offline_directory="offline")
experiment.add_tag(experiment_name)
experiment.set_name("{0}_{1}".format(experiment_name, i))
# Sometimes adding the tag fails
log_params["experiment_tag"] = experiment_name
experiment.log_parameters(log_params)
agent = agent_list[i]
if sps_es: # This one overrides sps
rewards, arms, agent = run_sps_es_experiment(agent,
env,
train_iter,
seed=seed,
video=video,
experiment=experiment,
**kwargs)
elif sps:
rewards, arms, agent = run_sps_experiment(agent,
env,
train_iter,
seed=seed,
video=video,
experiment=experiment,
**kwargs)
else:
rewards, arms, agent = run_aql_experiment(agent,
env,
train_iter,
seed=seed,
video=video,
experiment=experiment,
**kwargs)
agent_list[i] = agent
if experiment:
experiment.end()
if debug:
end = time.time()
elapsed = end - start
units = "secs"
if elapsed > 3600:
elapsed /= 3600
units = "hours"
elif elapsed > 60:
elapsed /= 60
units = "mins"
print("Time elapsed: {0:.02f} {1}".format(elapsed, units))
return rewards, arms, agent
def BSN_Train_PEM(opt):
model = PEM(opt)
model = torch.nn.DataParallel(model).cuda()
optimizer = optim.Adam(model.parameters(),
lr=opt["pem_training_lr"],
weight_decay=opt["pem_weight_decay"])
print('Total params: %.2fM' %
(sum(p.numel() for p in model.parameters()) / 1000000.0))
def collate_fn(batch):
batch_data = torch.cat([x[0] for x in batch])
batch_iou = torch.cat([x[1] for x in batch])
return batch_data, batch_iou
train_dataset = ProposalDataSet(opt, subset="train")
train_sampler = ProposalSampler(train_dataset.proposals,
train_dataset.indices,
max_zero_weight=opt['pem_max_zero_weight'])
global_step = 0
train_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=model.module.batch_size,
shuffle=False,
sampler=train_sampler,
num_workers=opt['data_workers'],
pin_memory=True,
drop_last=False,
collate_fn=collate_fn if not opt['pem_do_index'] else None)
subset = "validation" if opt['dataset'] == 'activitynet' else "test"
test_loader = torch.utils.data.DataLoader(
ProposalDataSet(opt, subset=subset),
batch_size=model.module.batch_size,
shuffle=True,
num_workers=opt['data_workers'],
pin_memory=True,
drop_last=False,
collate_fn=collate_fn if not opt['pem_do_index'] else None)
milestones = [int(k) for k in opt['pem_lr_milestones'].split(',')]
scheduler = torch.optim.lr_scheduler.MultiStepLR(
optimizer, milestones=milestones, gamma=opt['pem_step_gamma'])
if opt['log_to_comet']:
comet_exp = CometExperiment(api_key="hIXq6lDzWzz24zgKv7RYz6blo",
project_name="bsnpem",
workspace="cinjon",
auto_metric_logging=True,
auto_output_logging=None,
auto_param_logging=False)
elif opt['local_comet_dir']:
comet_exp = OfflineExperiment(api_key="hIXq6lDzWzz24zgKv7RYz6blo",
project_name="bsnpem",
workspace="cinjon",
auto_metric_logging=True,
auto_output_logging=None,
auto_param_logging=False,
offline_directory=opt['local_comet_dir'])
else:
comet_exp = None
if comet_exp:
comet_exp.log_parameters(opt)
comet_exp.set_name(opt['name'])
test_PEM(test_loader, model, -1, -1, comet_exp, opt)
for epoch in range(opt["pem_epoch"]):
global_step = train_PEM(train_loader, model, optimizer, epoch,
global_step, comet_exp, opt)
test_PEM(test_loader, model, epoch, global_step, comet_exp, opt)
scheduler.step()
from a2c_ppo_acktr.model import Policy
from a2c_ppo_acktr.storage import RolloutStorage
from a2c_ppo_acktr.utils import get_vec_normalize, update_linear_schedule
from a2c_ppo_acktr.visualize import visdom_plot
args = get_args()
assert args.algo in ['a2c', 'ppo', 'acktr']
if args.recurrent_policy:
assert args.algo in ['a2c', 'ppo'], \
'Recurrent policy is not implemented for ACKTR'
if args.comet == "offline":
experiment = OfflineExperiment(project_name="recurrent-value", workspace="nishanthvanand",
disabled=args.disable_log, offline_directory="../comet_offline",
parse_args=False)
elif args.comet == "online":
experiment = Experiment(api_key="tSACzCGFcetSBTapGBKETFARf",
project_name="recurrent-value", workspace="nishanthvanand",
disabled=args.disable_log,
parse_args=False)
else:
raise ValueError
experiment.log_parameters(vars(args))
num_updates = int(args.num_env_steps) // args.num_steps // args.num_processes
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
)
checkpoint_callback = skopt.callbacks.CheckpointSaver(
f'D:\\FINKI\\8_dps\\Project\\MODELS\\skopt_checkpoints\\{EXPERIMENT_ID}.pkl'
)
hyperparameters_optimizer.fit(X_train,
y_train,
callback=[checkpoint_callback])
skopt.dump(hyperparameters_optimizer, f'saved_models\\{EXPERIMENT_ID}.pkl')
y_pred = hyperparameters_optimizer.best_estimator_.predict(X_test)
for i in range(len(hyperparameters_optimizer.cv_results_['params'])):
exp = OfflineExperiment(
api_key='A8Lg71j9LtIrsv0deBA0DVGcR',
project_name=ALGORITHM,
workspace="8_dps",
auto_output_logging='native',
offline_directory=
f'D:\\FINKI\\8_dps\\Project\\MODELS\\comet_ml_offline_experiments\\{EXPERIMENT_ID}'
)
exp.set_name(f'{EXPERIMENT_ID}_{i + 1}')
exp.add_tags([
DS,
SEGMENTS_LENGTH,
])
for k, v in hyperparameters_optimizer.cv_results_.items():
if k == "params": exp.log_parameters(dict(v[i]))
else: exp.log_metric(k, v[i])
exp.end()
def BSN_Train_TEM(opt):
global_step = 0
epoch = 0
if opt['do_representation']:
model = TEM(opt)
optimizer = optim.Adam(model.parameters(),
lr=opt["tem_training_lr"],
weight_decay=opt["tem_weight_decay"])
global_step, epoch = _maybe_load_checkpoint(
model, optimizer, global_step, epoch,
os.path.join(opt["checkpoint_path"], opt['name']))
if opt['representation_checkpoint']:
# print(model.representation_model.backbone.inception_5b_3x3.weight[0][0])
if opt['do_random_model']:
print('DOING RANDOM MDOEL!!!')
else:
print('DOING Pretrianed modelll!!!')
partial_load(opt['representation_checkpoint'], model)
# print(model.representation_model.backbone.inception_5b_3x3.weight[0][0])
if not opt['no_freeze']:
for param in model.representation_model.parameters():
param.requires_grad = False
print(len([p for p in model.representation_model.parameters()]))
else:
model = TEM(opt)
optimizer = optim.Adam(model.parameters(),
lr=opt["tem_training_lr"],
weight_decay=opt["tem_weight_decay"])
global_step, epoch = _maybe_load_checkpoint(
model, optimizer, global_step, epoch,
os.path.join(opt["checkpoint_path"], opt['name']))
model = torch.nn.DataParallel(model).cuda()
# summary(model, (2, 3, 224, 224))
print(' Total params: %.2fM' %
(sum(p.numel() for p in model.parameters()) / 1000000.0))
if opt['dataset'] == 'gymnastics':
# default image_dir is '/checkpoint/cinjon/spaceofmotion/sep052019/rawframes.426x240.12'
img_loading_func = get_img_loader(opt)
train_data_set = GymnasticsImages(opt,
subset='Train',
img_loading_func=img_loading_func,
image_dir=opt['gym_image_dir'],
video_info_path=os.path.join(
opt['video_info'],
'Train_Annotation.csv'))
train_sampler = GymnasticsSampler(train_data_set, opt['sampler_mode'])
test_data_set = GymnasticsImages(opt,
subset="Val",
img_loading_func=img_loading_func,
image_dir=opt['gym_image_dir'],
video_info_path=os.path.join(
opt['video_info'],
'Val_Annotation.csv'))
elif opt['dataset'] == 'gymnasticsfeatures':
# feature_dirs should roughly look like:
# /checkpoint/cinjon/spaceofmotion/sep052019/tsn.1024.426x240.12.no-oversample/csv/rgb,/checkpoint/cinjon/spaceofmotion/sep052019/tsn.1024.426x240.12.no-oversample/csv/flow
feature_dirs = opt['feature_dirs'].split(',')
train_data_set = GymnasticsFeatures(opt,
subset='Train',
feature_dirs=feature_dirs,
video_info_path=os.path.join(
opt['video_info'],
'Train_Annotation.csv'))
test_data_set = GymnasticsFeatures(opt,
subset='Val',
feature_dirs=feature_dirs,
video_info_path=os.path.join(
opt['video_info'],
'Val_Annotation.csv'))
train_sampler = None
elif opt['dataset'] == 'thumosfeatures':
feature_dirs = opt['feature_dirs'].split(',')
train_data_set = ThumosFeatures(opt,
subset='Val',
feature_dirs=feature_dirs)
test_data_set = ThumosFeatures(opt,
subset="Test",
feature_dirs=feature_dirs)
train_sampler = None
elif opt['dataset'] == 'thumosimages':
img_loading_func = get_img_loader(opt)
train_data_set = ThumosImages(
opt,
subset='Val',
img_loading_func=img_loading_func,
image_dir=
'/checkpoint/cinjon/thumos/rawframes.TH14_validation_tal.30',
video_info_path=os.path.join(opt['video_info'],
'Val_Annotation.csv'))
test_data_set = ThumosImages(
opt,
subset='Test',
img_loading_func=img_loading_func,
image_dir='/checkpoint/cinjon/thumos/rawframes.TH14_test_tal.30',
video_info_path=os.path.join(opt['video_info'],
'Test_Annotation.csv'))
train_sampler = None
elif opt['dataset'] == 'activitynet':
train_sampler = None
representation_module = opt['representation_module']
train_transforms = get_video_transforms(representation_module,
opt['do_augment'])
test_transforms = get_video_transforms(representation_module, False)
train_data_set = VideoDataset(opt,
train_transforms,
subset='train',
fraction=0.3)
# We use val because we don't have annotations for test.
test_data_set = VideoDataset(opt,
test_transforms,
subset='val',
fraction=0.3)
print('train_loader / val_loader sizes: ', len(train_data_set),
len(test_data_set))
train_loader = torch.utils.data.DataLoader(
train_data_set,
batch_size=model.module.batch_size,
shuffle=False if train_sampler else True,
sampler=train_sampler,
num_workers=opt['data_workers'],
pin_memory=True,
drop_last=False)
test_loader = torch.utils.data.DataLoader(
test_data_set,
batch_size=model.module.batch_size,
shuffle=False,
num_workers=opt['data_workers'],
pin_memory=True,
drop_last=False)
# test_loader = None
milestones = [int(k) for k in opt['tem_lr_milestones'].split(',')]
scheduler = torch.optim.lr_scheduler.MultiStepLR(
optimizer, milestones=milestones, gamma=opt['tem_step_gamma'])
if opt['log_to_comet']:
comet_exp = CometExperiment(api_key="hIXq6lDzWzz24zgKv7RYz6blo",
project_name="bsn",
workspace="cinjon",
auto_metric_logging=True,
auto_output_logging=None,
auto_param_logging=False)
elif opt['local_comet_dir']:
comet_exp = OfflineExperiment(api_key="hIXq6lDzWzz24zgKv7RYz6blo",
project_name="bsn",
workspace="cinjon",
auto_metric_logging=True,
auto_output_logging=None,
auto_param_logging=False,
offline_directory=opt['local_comet_dir'])
else:
comet_exp = None
if comet_exp:
comet_exp.log_parameters(opt)
comet_exp.set_name(opt['name'])
# test_TEM(test_loader, model, optimizer, 0, 0, comet_exp, opt)
for epoch in range(epoch + 1, opt["tem_epoch"] + 1):
global_step = train_TEM(train_loader, model, optimizer, epoch,
global_step, comet_exp, opt)
test_TEM(test_loader, model, optimizer, epoch, global_step, comet_exp,
opt)
if opt['dataset'] == 'activitynet':
test_loader.dataset._subset_dataset(.3)
train_loader.dataset._subset_dataset(.3)
scheduler.step()
def _create_experiment(self, parameters, pid, trial, count,
experiment_kwargs):
# type: (Any, Any, Any, Any, Dict[str, Any]) -> BaseExperiment
"""
Instantiates an Experiment, OfflineExperiment, or
callable.
"""
from comet_ml import Experiment, OfflineExperiment
LOGGER.debug("Creating a %r with %r parameters", self.experiment_class,
parameters)
if not experiment_kwargs:
# Fallback on deprecated experiment kwargs given at Optimizer creation
experiment_kwargs = self.experiment_kwargs
# Inject the API Key if not set
if "api_key" not in experiment_kwargs and self.api_key is not None:
experiment_kwargs["api_key"] = self.api_key
if self.experiment_class == "Experiment":
exp = Experiment(**experiment_kwargs) # type: BaseExperiment
elif self.experiment_class == "OfflineExperiment":
exp = OfflineExperiment(**
experiment_kwargs) # type: BaseExperiment
elif callable(self.experiment_class):
exp = self.experiment_class(**experiment_kwargs)
else:
raise OptimizerException("Invalid experiment_class: %s" %
self.experiment_class)
exp._set_optimizer(self, pid, trial, count)
if self.predictor_spec:
LOGGER.debug("Setting predictor in experiment: %s",
self.predictor_spec)
if isinstance(exp, OfflineExperiment):
mode = "local"
else:
# This should be at least global as it is set in Optimizer._fill_defaults_predictor
mode = self.predictor_spec["mode"]
# Old version of the Python SDK used to store predictor_spec with an
# empty optimizer_id, which would leads to TypeError: type object
# got multiple values for keyword argument 'optimizer_id'
# Also pop mode as it is passed explicitely
predictor_spec = self.predictor_spec.copy()
predictor_spec.pop("optimizer_id", None)
predictor_spec.pop("mode", None)
try:
exp.set_predictor(
Predictor(exp,
optimizer_id=self.id,
mode=mode,
**predictor_spec))
except Exception:
LOGGER.warning(
"Failure to create Predictor, experiment will run without Predictor",
exc_info=True,
)
exp.log_parameters(parameters)
# Log optimizer static information:
exp.log_other("optimizer_id", self.id)
exp.log_other("optimizer_pid", pid)
exp.log_other("optimizer_trial", trial)
return exp
class CometLogger(LightningLoggerBase):
r"""
Log using `comet.ml <https://www.comet.ml>`_.
"""
def __init__(self,
api_key: Optional[str] = None,
save_dir: Optional[str] = None,
workspace: Optional[str] = None,
project_name: Optional[str] = None,
rest_api_key: Optional[str] = None,
experiment_name: Optional[str] = None,
experiment_key: Optional[str] = None,
**kwargs):
r"""
Requires either an API Key (online mode) or a local directory path (offline mode)
.. code-block:: python
# ONLINE MODE
from pytorch_lightning.loggers import CometLogger
# arguments made to CometLogger are passed on to the comet_ml.Experiment class
comet_logger = CometLogger(
api_key=os.environ["COMET_API_KEY"],
workspace=os.environ["COMET_WORKSPACE"], # Optional
project_name="default_project", # Optional
rest_api_key=os.environ["COMET_REST_API_KEY"], # Optional
experiment_name="default" # Optional
)
trainer = Trainer(logger=comet_logger)
.. code-block:: python
# OFFLINE MODE
from pytorch_lightning.loggers import CometLogger
# arguments made to CometLogger are passed on to the comet_ml.Experiment class
comet_logger = CometLogger(
save_dir=".",
workspace=os.environ["COMET_WORKSPACE"], # Optional
project_name="default_project", # Optional
rest_api_key=os.environ["COMET_REST_API_KEY"], # Optional
experiment_name="default" # Optional
)
trainer = Trainer(logger=comet_logger)
Args:
api_key (str): Required in online mode. API key, found on Comet.ml
save_dir (str): Required in offline mode. The path for the directory to save local comet logs
workspace (str): Optional. Name of workspace for this user
project_name (str): Optional. Send your experiment to a specific project.
Otherwise will be sent to Uncategorized Experiments.
If project name does not already exists Comet.ml will create a new project.
rest_api_key (str): Optional. Rest API key found in Comet.ml settings.
This is used to determine version number
experiment_name (str): Optional. String representing the name for this particular experiment on Comet.ml.
experiment_key (str): Optional. If set, restores from existing experiment.
"""
super().__init__()
self._experiment = None
# Determine online or offline mode based on which arguments were passed to CometLogger
if api_key is not None:
self.mode = "online"
self.api_key = api_key
elif save_dir is not None:
self.mode = "offline"
self.save_dir = save_dir
else:
# If neither api_key nor save_dir are passed as arguments, raise an exception
raise MisconfigurationException(
"CometLogger requires either api_key or save_dir during initialization."
)
log.info(f"CometLogger will be initialized in {self.mode} mode")
self.workspace = workspace
self.project_name = project_name
self.experiment_key = experiment_key
self._kwargs = kwargs
if rest_api_key is not None:
# Comet.ml rest API, used to determine version number
self.rest_api_key = rest_api_key
self.comet_api = API(self.rest_api_key)
else:
self.rest_api_key = None
self.comet_api = None
if experiment_name:
try:
self.name = experiment_name
except TypeError as e:
log.exception(
"Failed to set experiment name for comet.ml logger")
@property
def experiment(self) -> CometBaseExperiment:
r"""
Actual comet object. To use comet features do the following.
Example::
self.logger.experiment.some_comet_function()
"""
if self._experiment is not None:
return self._experiment
if self.mode == "online":
if self.experiment_key is None:
self._experiment = CometExperiment(
api_key=self.api_key,
workspace=self.workspace,
project_name=self.project_name,
**self._kwargs)
self.experiment_key = self._experiment.get_key()
else:
self._experiment = CometExistingExperiment(
api_key=self.api_key,
workspace=self.workspace,
project_name=self.project_name,
previous_experiment=self.experiment_key,
**self._kwargs)
else:
self._experiment = CometOfflineExperiment(
offline_directory=self.save_dir,
workspace=self.workspace,
project_name=self.project_name,
**self._kwargs)
return self._experiment
@rank_zero_only
def log_hyperparams(self, params: Union[Dict[str, Any],
Namespace]) -> None:
params = self._convert_params(params)
params = self._flatten_dict(params)
self.experiment.log_parameters(params)
@rank_zero_only
def log_metrics(self,
metrics: Dict[str, Union[torch.Tensor, float]],
step: Optional[int] = None) -> None:
# Comet.ml expects metrics to be a dictionary of detached tensors on CPU
for key, val in metrics.items():
if is_tensor(val):
metrics[key] = val.cpu().detach()
self.experiment.log_metrics(metrics, step=step)
def reset_experiment(self):
self._experiment = None
@rank_zero_only
def finalize(self, status: str) -> None:
r"""
When calling self.experiment.end(), that experiment won't log any more data to Comet. That's why, if you need
to log any more data you need to create an ExistingCometExperiment. For example, to log data when testing your
model after training, because when training is finalized CometLogger.finalize is called.
This happens automatically in the CometLogger.experiment property, when self._experiment is set to None
i.e. self.reset_experiment().
"""
self.experiment.end()
self.reset_experiment()
@property
def name(self) -> str:
return self.experiment.project_name
@name.setter
def name(self, value: str) -> None:
self.experiment.set_name(value)
@property
def version(self) -> str:
return self.experiment.id
def _get_comet_experiment():
experiment = OfflineExperiment(project_name='general',
workspace='benjaminbenoit',
offline_directory="../transformer_net_comet_experiences")
experiment.set_name("TransformerNet")
return experiment
def get_params():
parser = argparse.ArgumentParser(description='Perm')
# Hparams
padd = parser.add_argument
padd('--batch-size',
type=int,
default=64,
metavar='N',
help='input batch size for training (default: 64)')
padd('--latent_dim',
type=int,
default=20,
metavar='N',
help='Latent dim for VAE')
padd('--lr',
type=float,
default=0.01,
metavar='LR',
help='learning rate (default: 0.01)')
padd('--momentum',
type=float,
default=0.5,
metavar='M',
help='SGD momentum (default: 0.5)')
padd('--latent_size',
type=int,
default=50,
metavar='N',
help='Size of latent distribution (default: 50)')
padd('--estimator',
default='reinforce',
const='reinforce',
nargs='?',
choices=['reinforce', 'lax'],
help='Grad estimator for noise (default: %(default)s)')
padd('--reward',
default='soft',
const='soft',
nargs='?',
choices=['soft', 'hard'],
help='Reward for grad estimator (default: %(default)s)')
# Training
padd('--epochs',
type=int,
default=10,
metavar='N',
help='number of epochs to train (default: 10)')
padd('--PGD_steps',
type=int,
default=40,
metavar='N',
help='max gradient steps (default: 30)')
padd('--max_iter',
type=int,
default=20,
metavar='N',
help='max gradient steps (default: 30)')
padd('--max_batches',
type=int,
default=None,
metavar='N',
help=
'max number of batches per epoch, used for debugging (default: None)')
padd('--epsilon',
type=float,
default=0.5,
metavar='M',
help='Epsilon for Delta (default: 0.1)')
padd('--LAMBDA',
type=float,
default=100,
metavar='M',
help='Lambda for L2 lagrange penalty (default: 0.1)')
padd('--nn_temp',
type=float,
default=1.0,
metavar='M',
help='Starting diff. nearest neighbour temp (default: 1.0)')
padd('--temp_decay_rate',
type=float,
default=0.9,
metavar='M',
help='Nearest neighbour temp decay rate (default: 0.9)')
padd('--temp_decay_schedule',
type=float,
default=100,
metavar='M',
help='How many batches before decay (default: 100)')
padd('--bb_steps',
type=int,
default=2000,
metavar='N',
help='Max black box steps per sample(default: 1000)')
padd('--attack_epochs',
type=int,
default=10,
metavar='N',
help='Max numbe of epochs to train G')
padd('--seed',
type=int,
default=1,
metavar='S',
help='random seed (default: 1)')
padd('--batch_size', type=int, default=256, metavar='S', help='Batch size')
padd('--embedding_dim', type=int, default=300, help='embedding_dim')
padd('--embedding_type',
type=str,
default="non-static",
help='embedding_type')
padd('--test_batch_size',
type=int,
default=128,
metavar='N',
help='Test Batch size. 256 requires 12GB GPU memory')
padd('--test',
default=False,
action='store_true',
help='just test model and print accuracy')
padd('--deterministic_G',
default=False,
action='store_true',
help='Auto-encoder, no VAE')
padd('--resample_test',
default=False,
action='store_true',
help='Load model and test resampling capability')
padd('--resample_iterations',
type=int,
default=100,
metavar='N',
help='How many times to resample (default: 100)')
padd('--clip_grad',
default=True,
action='store_true',
help='Clip grad norm')
padd('--train_vae', default=False, action='store_true', help='Train VAE')
padd('--train_ae', default=False, action='store_true', help='Train AE')
padd('--use_flow',
default=False,
action='store_true',
help='Add A NF to Generator')
padd('--carlini_loss',
default=False,
action='store_true',
help='Use CW loss function')
padd('--vanilla_G',
default=False,
action='store_true',
help='Vanilla G White Box')
padd('--prepared_data',
default='dataloader/prepared_data.pickle',
help='Test on a single data')
# Imported Model Params
padd('--emsize', type=int, default=300, help='size of word embeddings')
padd('--nhidden',
type=int,
default=300,
help='number of hidden units per layer in LSTM')
padd('--nlayers', type=int, default=2, help='number of layers')
padd('--noise_radius',
type=float,
default=0.2,
help='stdev of noise for autoencoder (regularizer)')
padd('--noise_anneal',
type=float,
default=0.995,
help='anneal noise_radius exponentially by this every 100 iterations')
padd('--hidden_init',
action='store_true',
help="initialize decoder hidden state with encoder's")
padd('--arch_i',
type=str,
default='300-300',
help='inverter architecture (MLP)')
padd('--arch_g',
type=str,
default='300-300',
help='generator architecture (MLP)')
padd('--arch_d',
type=str,
default='300-300',
help='critic/discriminator architecture (MLP)')
padd('--arch_conv_filters',
type=str,
default='500-700-1000',
help='encoder filter sizes for different convolutional layers')
padd('--arch_conv_strides',
type=str,
default='1-2-2',
help='encoder strides for different convolutional layers')
padd('--arch_conv_windows',
type=str,
default='3-3-3',
help='encoder window sizes for different convolutional layers')
padd('--z_size',
type=int,
default=100,
help='dimension of random noise z to feed into generator')
padd('--temp',
type=float,
default=1,
help='softmax temperature (lower --> more discrete)')
padd('--enc_grad_norm',
type=bool,
default=True,
help='norm code gradient from critic->encoder')
padd('--train_emb', type=bool, default=True, help='Train Glove Embeddings')
padd('--gan_toenc',
type=float,
default=-0.01,
help='weight factor passing gradient from gan to encoder')
padd('--dropout',
type=float,
default=0.0,
help='dropout applied to layers (0 = no dropout)')
padd('--useJS',
type=bool,
default=True,
help='use Jenson Shannon distance')
padd('--perturb_z',
type=bool,
default=True,
help='perturb noise space z instead of hidden c')
padd('--max_seq_len', type=int, default=200, help='max_seq_len')
padd('--gamma', type=float, default=0.95, help='Discount Factor')
padd('--model',
type=str,
default="lstm_arch",
help='classification model name')
padd('--distance_func',
type=str,
default="cosine",
help='NN distance function')
padd('--hidden_dim', type=int, default=128, help='hidden_dim')
padd('--burn_in', type=int, default=500, help='Train VAE burnin')
padd('--beta', type=float, default=0., help='Entropy reg')
padd('--embedding_training',
type=bool,
default=False,
help='embedding_training')
padd('--seqgan_reward',
action='store_true',
default=False,
help='use seq gan reward')
padd('--train_classifier',
action='store_true',
default=False,
help='Train Classifier from scratch')
padd('--diff_nn',
action='store_true',
default=False,
help='Backprop through Nearest Neighbors')
# Bells
padd('--no-cuda',
action='store_true',
default=False,
help='disables CUDA training')
padd('--data_parallel',
action='store_true',
default=False,
help="Use multiple GPUs")
padd('--save_adv_samples',
action='store_true',
default=False,
help='Write adversarial samples to disk')
padd('--nearest_neigh_all',
action='store_true',
default=False,
help='Evaluate near. neig. for whole evaluation set')
padd("--comet",
action="store_true",
default=False,
help='Use comet for logging')
padd(
"--offline_comet",
action="store_true",
default=False,
help=
'Use comet offline. To upload, after training run: comet-upload file.zip'
)
padd("--comet_username",
type=str,
default="joeybose",
help='Username for comet logging')
padd("--comet_apikey", type=str,\
default="Ht9lkWvTm58fRo9ccgpabq5zV",help='Api for comet logging')
padd('--debug', default=False, action='store_true', help='Debug')
padd('--debug_neighbour',
default=False,
action='store_true',
help='Debug nearest neighbour training')
padd('--load_model',
default=False,
action='store_true',
help='Whether to load a checkpointed model')
padd('--save_model',
default=False,
action='store_true',
help='Whether to checkpointed model')
padd('--model_path', type=str, default="saved_models/lstm_torchtext2.pt",\
help='where to save/load target model')
padd('--adv_model_path', type=str, default="saved_models/adv_model.pt",\
help='where to save/load adversarial')
padd('--no_load_embedding',
action='store_false',
default=True,
help='load Glove embeddings')
padd('--namestr', type=str, default='BMD Text', \
help='additional info in output filename to describe experiments')
padd('--dataset', type=str, default="imdb", help='dataset')
padd('--clip', type=float, default=1, help='gradient clipping, max norm')
padd('--use_glove', type=str, default="true", help='gpu number')
args = parser.parse_args()
args.classes = 2
args.sample_file = "temp/adv_samples.txt"
use_cuda = not args.no_cuda and torch.cuda.is_available()
torch.manual_seed(args.seed)
# Check if settings file
if os.path.isfile("settings.json"):
with open('settings.json') as f:
data = json.load(f)
args.comet_apikey = data["apikey"]
args.comet_username = data["username"]
# Prep file to save adversarial samples
if args.save_adv_samples:
now = datetime.datetime.now()
if os.path.exists(args.sample_file):
os.remove(args.sample_file)
with open(args.sample_file, 'w') as f:
f.write("Adversarial samples starting:\n{}\n".format(now))
# Comet logging
args.device = torch.device("cuda" if use_cuda else "cpu")
if args.comet and not args.offline_comet:
experiment = Experiment(api_key=args.comet_apikey,
project_name="black-magic-design",
workspace=args.comet_username)
elif args.offline_comet:
offline_path = "temp/offline_comet"
if not os.path.exists(offline_path):
os.makedirs(offline_path)
from comet_ml import OfflineExperiment
experiment = OfflineExperiment(project_name="black-magic-design",
workspace=args.comet_username,
offline_directory=offline_path)
# To upload offline comet, run: comet-upload file.zip
if args.comet or args.offline_comet:
experiment.set_name(args.namestr)
def log_text(self, msg):
# Change line breaks for html breaks
msg = msg.replace('\n', '<br>')
self.log_html("<p>{}</p>".format(msg))
experiment.log_text = MethodType(log_text, experiment)
args.experiment = experiment
return args
hidden_size = configuration['hidden_size']
n_layers = configuration['n_layers']
flattened = False
device = 'cuda' if torch.cuda.is_available() else 'cpu'
# Set all seeds for full reproducibility
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
# Set up Comet Experiment tracking # Replace this with appropriate comet
# workspaces
experiment = OfflineExperiment("z15Um8oxWZwiXQXZxZKGh48cl",
workspace='swechhachoudhary',
offline_directory="../swechhas_experiments")
experiment.set_name(
name=args.config +
"_dim={}_split={}".format(latent_dim, train_unlabeled_split))
experiment.log_parameters(configuration)
if encoding_model == 'pca':
encoding_model = PCAEncoder(seed)
flattened = True
elif encoding_model == 'vae':
encoding_model = VAE(latent_dim=latent_dim).to(device)
flattened = True
elif encoding_model == "ae":
encoding_model = AE(latent_dim=latent_dim).to(device)
if args.debug is False:
ipdb.set_trace = lambda: None
# Comet logging
args.device = torch.device("cuda" if use_cuda else "cpu")
if args.comet and not args.offline_comet:
experiment = Experiment(api_key=args.comet_apikey,
project_name="black-magic-design",
workspace=args.comet_username)
elif args.offline_comet:
offline_path = "temp/offline_comet"
if not os.path.exists(offline_path):
os.makedirs(offline_path)
from comet_ml import OfflineExperiment
experiment = OfflineExperiment(project_name="black-magic-design",
workspace=args.comet_username,
offline_directory=offline_path)
# To upload offline comet, run: comet-upload file.zip
if args.comet or args.offline_comet:
experiment.set_name(args.namestr)
def log_text(self, msg):
# Change line breaks for html breaks
msg = msg.replace('\n', '<br>')
self.log_html("<p>{}</p>".format(msg))
experiment.log_text = MethodType(log_text, experiment)
args.experiment = experiment
main(args)
class CometLogger(LightningLoggerBase):
r"""
Log using `Comet.ml <https://www.comet.ml>`_. Install it with pip:
.. code-block:: bash
pip install comet-ml
Comet requires either an API Key (online mode) or a local directory path (offline mode).
**ONLINE MODE**
Example:
>>> import os
>>> from pytorch_lightning import Trainer
>>> from pytorch_lightning.loggers import CometLogger
>>> # arguments made to CometLogger are passed on to the comet_ml.Experiment class
>>> comet_logger = CometLogger(
... api_key=os.environ.get('COMET_API_KEY'),
... workspace=os.environ.get('COMET_WORKSPACE'), # Optional
... save_dir='.', # Optional
... project_name='default_project', # Optional
... rest_api_key=os.environ.get('COMET_REST_API_KEY'), # Optional
... experiment_name='default' # Optional
... )
>>> trainer = Trainer(logger=comet_logger)
**OFFLINE MODE**
Example:
>>> from pytorch_lightning.loggers import CometLogger
>>> # arguments made to CometLogger are passed on to the comet_ml.Experiment class
>>> comet_logger = CometLogger(
... save_dir='.',
... workspace=os.environ.get('COMET_WORKSPACE'), # Optional
... project_name='default_project', # Optional
... rest_api_key=os.environ.get('COMET_REST_API_KEY'), # Optional
... experiment_name='default' # Optional
... )
>>> trainer = Trainer(logger=comet_logger)
Args:
api_key: Required in online mode. API key, found on Comet.ml. If not given, this
will be loaded from the environment variable COMET_API_KEY or ~/.comet.config
if either exists.
save_dir: Required in offline mode. The path for the directory to save local
comet logs. If given, this also sets the directory for saving checkpoints.
workspace: Optional. Name of workspace for this user
project_name: Optional. Send your experiment to a specific project.
Otherwise will be sent to Uncategorized Experiments.
If the project name does not already exist, Comet.ml will create a new project.
rest_api_key: Optional. Rest API key found in Comet.ml settings.
This is used to determine version number
experiment_name: Optional. String representing the name for this particular experiment on Comet.ml.
experiment_key: Optional. If set, restores from existing experiment.
offline: If api_key and save_dir are both given, this determines whether
the experiment will be in online or offline mode. This is useful if you use
save_dir to control the checkpoints directory and have a ~/.comet.config
file but still want to run offline experiments.
"""
def __init__(self,
api_key: Optional[str] = None,
save_dir: Optional[str] = None,
workspace: Optional[str] = None,
project_name: Optional[str] = None,
rest_api_key: Optional[str] = None,
experiment_name: Optional[str] = None,
experiment_key: Optional[str] = None,
offline: bool = False,
**kwargs):
if not _COMET_AVAILABLE:
raise ImportError(
'You want to use `comet_ml` logger which is not installed yet,'
' install it with `pip install comet-ml`.')
super().__init__()
self._experiment = None
# Determine online or offline mode based on which arguments were passed to CometLogger
api_key = api_key or get_api_key(None, get_config())
if api_key is not None and save_dir is not None:
self.mode = "offline" if offline else "online"
self.api_key = api_key
self._save_dir = save_dir
elif api_key is not None:
self.mode = "online"
self.api_key = api_key
elif save_dir is not None:
self.mode = "offline"
self._save_dir = save_dir
else:
# If neither api_key nor save_dir are passed as arguments, raise an exception
raise MisconfigurationException(
"CometLogger requires either api_key or save_dir during initialization."
)
log.info(f"CometLogger will be initialized in {self.mode} mode")
self.workspace = workspace
self.project_name = project_name
self.experiment_key = experiment_key
self._kwargs = kwargs
if rest_api_key is not None:
# Comet.ml rest API, used to determine version number
self.rest_api_key = rest_api_key
self.comet_api = API(self.rest_api_key)
else:
self.rest_api_key = None
self.comet_api = None
if experiment_name:
self.experiment.set_name(experiment_name)
self._kwargs = kwargs
@property
@rank_zero_experiment
def experiment(self) -> CometBaseExperiment:
r"""
Actual Comet object. To use Comet features in your
:class:`~pytorch_lightning.core.lightning.LightningModule` do the following.
Example::
self.logger.experiment.some_comet_function()
"""
if self._experiment is not None:
return self._experiment
if self.mode == "online":
if self.experiment_key is None:
self._experiment = CometExperiment(
api_key=self.api_key,
workspace=self.workspace,
project_name=self.project_name,
**self._kwargs)
self.experiment_key = self._experiment.get_key()
else:
self._experiment = CometExistingExperiment(
api_key=self.api_key,
workspace=self.workspace,
project_name=self.project_name,
previous_experiment=self.experiment_key,
**self._kwargs)
else:
self._experiment = CometOfflineExperiment(
offline_directory=self.save_dir,
workspace=self.workspace,
project_name=self.project_name,
**self._kwargs)
return self._experiment
@rank_zero_only
def log_hyperparams(self, params: Union[Dict[str, Any],
Namespace]) -> None:
params = self._convert_params(params)
params = self._flatten_dict(params)
self.experiment.log_parameters(params)
@rank_zero_only
def log_metrics(self,
metrics: Dict[str, Union[torch.Tensor, float]],
step: Optional[int] = None) -> None:
assert rank_zero_only.rank == 0, 'experiment tried to log from global_rank != 0'
# Comet.ml expects metrics to be a dictionary of detached tensors on CPU
for key, val in metrics.items():
if is_tensor(val):
metrics[key] = val.cpu().detach()
self.experiment.log_metrics(metrics, step=step)
def reset_experiment(self):
self._experiment = None
@rank_zero_only
def finalize(self, status: str) -> None:
r"""
When calling ``self.experiment.end()``, that experiment won't log any more data to Comet.
That's why, if you need to log any more data, you need to create an ExistingCometExperiment.
For example, to log data when testing your model after training, because when training is
finalized :meth:`CometLogger.finalize` is called.
This happens automatically in the :meth:`~CometLogger.experiment` property, when
``self._experiment`` is set to ``None``, i.e. ``self.reset_experiment()``.
"""
self.experiment.end()
self.reset_experiment()
@property
def save_dir(self) -> Optional[str]:
return self._save_dir
@property
def name(self) -> str:
return str(self.experiment.project_name)
@property
def version(self) -> str:
return self.experiment.id
def __getstate__(self):
state = self.__dict__.copy()
state["_experiment"] = None
return state
class CometLogger(Logger):
def __init__(
self,
batch_size: int,
snapshot_dir: Optional[str] = None,
snapshot_mode: str = "last",
snapshot_gap: int = 1,
exp_set: Optional[str] = None,
use_print_exp: bool = False,
saved_exp: Optional[str] = None,
**kwargs,
):
"""
:param kwargs: passed to comet's Experiment at init.
"""
if use_print_exp:
self.experiment = PrintExperiment()
else:
from comet_ml import Experiment, ExistingExperiment, OfflineExperiment
if saved_exp:
self.experiment = ExistingExperiment(
previous_experiment=saved_exp, **kwargs
)
else:
try:
self.experiment = Experiment(**kwargs)
except ValueError: # no API key
log_dir = Path.home() / "logs"
log_dir.mkdir(exist_ok=True)
self.experiment = OfflineExperiment(offline_directory=str(log_dir))
self.experiment.log_parameter("complete", False)
if exp_set:
self.experiment.log_parameter("exp_set", exp_set)
if snapshot_dir:
snapshot_dir = Path(snapshot_dir) / self.experiment.get_key()
# log_traj_window (int): How many trajectories to hold in deque for computing performance statistics.
self.log_traj_window = 100
self._cum_metrics = {
"n_unsafe_actions": 0,
"constraint_used": 0,
"cum_completed_trajs": 0,
"logging_time": 0,
}
self._new_completed_trajs = 0
self._last_step = 0
self._start_time = self._last_time = time()
self._last_snapshot_upload = 0
self._snaphot_upload_time = 30 * 60
super().__init__(batch_size, snapshot_dir, snapshot_mode, snapshot_gap)
def log_fast(
self,
step: int,
traj_infos: Sequence[Dict[str, float]],
opt_info: Optional[Tuple[Sequence[float], ...]] = None,
test: bool = False,
) -> None:
if not traj_infos:
return
start = time()
self._new_completed_trajs += len(traj_infos)
self._cum_metrics["cum_completed_trajs"] += len(traj_infos)
# TODO: do we need to support sum(t[k]) if key in k?
# without that, this doesn't include anything from extra eval samplers
for key in self._cum_metrics:
if key == "cum_completed_trajs":
continue
self._cum_metrics[key] += sum(t.get(key, 0) for t in traj_infos)
self._cum_metrics["logging_time"] += time() - start
def log(
self,
step: int,
traj_infos: Sequence[Dict[str, float]],
opt_info: Optional[Tuple[Sequence[float], ...]] = None,
test: bool = False,
):
self.log_fast(step, traj_infos, opt_info, test)
start = time()
with (self.experiment.test() if test else nullcontext()):
step *= self.batch_size
if opt_info is not None:
# grad norm is left on the GPU for some reason
# https://github.com/astooke/rlpyt/issues/163
self.experiment.log_metrics(
{
k: np.mean(v)
for k, v in opt_info._asdict().items()
if k != "gradNorm"
},
step=step,
)
if traj_infos:
agg_vals = {}
for key in traj_infos[0].keys():
if key in self._cum_metrics:
continue
agg_vals[key] = sum(t[key] for t in traj_infos) / len(traj_infos)
self.experiment.log_metrics(agg_vals, step=step)
if not test:
now = time()
self.experiment.log_metrics(
{
"new_completed_trajs": self._new_completed_trajs,
"steps_per_second": (step - self._last_step)
/ (now - self._last_time),
},
step=step,
)
self._last_time = now
self._last_step = step
self._new_completed_trajs = 0
self.experiment.log_metrics(self._cum_metrics, step=step)
self._cum_metrics["logging_time"] += time() - start
def log_metric(self, name, val):
self.experiment.log_metric(name, val)
def log_parameters(self, parameters):
self.experiment.log_parameters(parameters)
def log_config(self, config):
self.experiment.log_parameter("config", json.dumps(convert_dict(config)))
def upload_snapshot(self):
if self.snapshot_dir:
self.experiment.log_asset(self._previous_snapshot_fname)
def save_itr_params(
self, step: int, params: Dict[str, Any], metric: Optional[float] = None
) -> None:
super().save_itr_params(step, params, metric)
now = time()
if now - self._last_snapshot_upload > self._snaphot_upload_time:
self._last_snapshot_upload = now
self.upload_snapshot()
def shutdown(self, error: bool = False) -> None:
if not error:
self.upload_snapshot()
self.experiment.log_parameter("complete", True)
self.experiment.end()
def main(args):
print('Pretrain? ', not args.not_pretrain)
print(args.model)
start_time = time.time()
if opt['local_comet_dir']:
comet_exp = OfflineExperiment(api_key="hIXq6lDzWzz24zgKv7RYz6blo",
project_name="selfcifar",
workspace="cinjon",
auto_metric_logging=True,
auto_output_logging=None,
auto_param_logging=False,
offline_directory=opt['local_comet_dir'])
else:
comet_exp = CometExperiment(api_key="hIXq6lDzWzz24zgKv7RYz6blo",
project_name="selfcifar",
workspace="cinjon",
auto_metric_logging=True,
auto_output_logging=None,
auto_param_logging=False)
comet_exp.log_parameters(vars(args))
comet_exp.set_name(args.name)
# Build model
# path = "/misc/kcgscratch1/ChoGroup/resnick/spaceofmotion/zeping/bsn"
linear_cls = NonLinearModel if args.do_nonlinear else LinearModel
if args.model == "amdim":
hparams = load_hparams_from_tags_csv(
'/checkpoint/cinjon/amdim/meta_tags.csv')
# hparams = load_hparams_from_tags_csv(os.path.join(path, "meta_tags.csv"))
model = AMDIMModel(hparams)
if not args.not_pretrain:
# _path = os.path.join(path, "_ckpt_epoch_434.ckpt")
_path = '/checkpoint/cinjon/amdim/_ckpt_epoch_434.ckpt'
model.load_state_dict(torch.load(_path)["state_dict"])
else:
print("AMDIM not loading checkpoint") # Debug
linear_model = linear_cls(AMDIM_OUTPUT_DIM, args.num_classes)
elif args.model == "ccc":
model = CCCModel(None)
if not args.not_pretrain:
# _path = os.path.join(path, "TimeCycleCkpt14.pth")
_path = '/checkpoint/cinjon/spaceofmotion/bsn/TimeCycleCkpt14.pth'
checkpoint = torch.load(_path)
base_dict = {
'.'.join(k.split('.')[1:]): v
for k, v in list(checkpoint['state_dict'].items())
}
model.load_state_dict(base_dict)
else:
print("CCC not loading checkpoint") # Debug
linear_model = linaer_cls(CCC_OUTPUT_DIM,
args.num_classes) #.to(device)
elif args.model == "corrflow":
model = CORRFLOWModel(None)
if not args.not_pretrain:
_path = '/checkpoint/cinjon/spaceofmotion/supercons/corrflow.kineticsmodel.pth'
# _path = os.path.join(path, "corrflow.kineticsmodel.pth")
checkpoint = torch.load(_path)
base_dict = {
'.'.join(k.split('.')[1:]): v
for k, v in list(checkpoint['state_dict'].items())
}
model.load_state_dict(base_dict)
else:
print("CorrFlow not loading checkpoing") # Debug
linear_model = linear_cls(CORRFLOW_OUTPUT_DIM, args.num_classes)
elif args.model == "resnet":
if not args.not_pretrain:
resnet = torchvision.models.resnet50(pretrained=True)
else:
resnet = torchvision.models.resnet50(pretrained=False)
print("ResNet not loading checkpoint") # Debug
modules = list(resnet.children())[:-1]
model = nn.Sequential(*modules)
linear_model = linear_cls(RESNET_OUTPUT_DIM, args.num_classes)
else:
raise Exception("model type has to be amdim, ccc, corrflow or resnet")
if torch.cuda.device_count() > 1:
model = nn.DataParallel(model).to(device)
linear_model = nn.DataParallel(linear_model).to(device)
else:
model = model.to(device)
linear_model = linear_model.to(device)
# model = model.to(device)
# linear_model = linear_model.to(device)
# Freeze model
for p in model.parameters():
p.requires_grad = False
model.eval()
if args.optimizer == "Adam":
optimizer = optim.Adam(linear_model.parameters(),
lr=args.lr,
weight_decay=args.weight_decay)
print("Optimizer: Adam with weight decay: {}".format(
args.weight_decay))
elif args.optimizer == "SGD":
optimizer = optim.SGD(linear_model.parameters(),
lr=args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
print("Optimizer: SGD with weight decay: {} momentum: {}".format(
args.weight_decay, args.momentum))
else:
raise Exception("optimizer should be Adam or SGD")
optimizer.zero_grad()
# Set up log dir
now = datetime.datetime.now()
log_dir = '/checkpoint/cinjon/spaceofmotion/bsn/cifar-%d-weights/%s/%s' % (
args.num_classes, args.model, args.name)
# log_dir = "{}{:%Y%m%dT%H%M}".format(args.model, now)
# log_dir = os.path.join("weights", log_dir)
if not os.path.exists(log_dir):
os.makedirs(log_dir)
print("Saving to {}".format(log_dir))
batch_size = args.batch_size * torch.cuda.device_count()
# CIFAR-10
if args.num_classes == 10:
data_path = ("/private/home/cinjon/cifar-data/cifar-10-batches-py")
_train_dataset = CIFAR_dataset(glob(os.path.join(data_path, "data*")),
args.num_classes, args.model, True)
# _train_acc_dataset = CIFAR_dataset(
# glob(os.path.join(data_path, "data*")),
# args.num_classes,
# args.model,
# False)
train_dataloader = data.DataLoader(_train_dataset,
shuffle=True,
batch_size=batch_size,
num_workers=args.num_workers)
# train_split = int(len(_train_dataset) * 0.8)
# train_dev_split = int(len(_train_dataset) - train_split)
# train_dataset, train_dev_dataset = data.random_split(
# _train_dataset, [train_split, train_dev_split])
# train_acc_dataloader = data.DataLoader(
# train_dataset, shuffle=False, batch_size=batch_size, num_workers=args.num_workers)
# train_dev_acc_dataloader = data.DataLoader(
# train_dev_dataset, shuffle=False, batch_size=batch_size, num_workers=args.num_workers)
# train_dataset = data.Subset(_train_dataset, list(range(train_split)))
# train_dataloader = data.DataLoader(
# train_dataset, shuffle=True, batch_size=batch_size, num_workers=args.num_workers)
# train_acc_dataset = data.Subset(
# _train_acc_dataset, list(range(train_split)))
# train_acc_dataloader = data.DataLoader(
# train_acc_dataset, shuffle=False, batch_size=batch_size, num_workers=args.num_workers)
# train_dev_acc_dataset = data.Subset(
# _train_acc_dataset, list(range(train_split, len(_train_acc_dataset))))
# train_dev_acc_dataloader = data.DataLoader(
# train_dev_acc_dataset, shuffle=False, batch_size=batch_size, num_workers=args.num_workers)
_val_dataset = CIFAR_dataset([os.path.join(data_path, "test_batch")],
args.num_classes, args.model, False)
val_dataloader = data.DataLoader(_val_dataset,
shuffle=False,
batch_size=batch_size,
num_workers=args.num_workers)
# val_split = int(len(_val_dataset) * 0.8)
# val_dev_split = int(len(_val_dataset) - val_split)
# val_dataset, val_dev_dataset = data.random_split(
# _val_dataset, [val_split, val_dev_split])
# val_dataloader = data.DataLoader(
# val_dataset, shuffle=False, batch_size=batch_size, num_workers=args.num_workers)
# val_dev_dataloader = data.DataLoader(
# val_dev_dataset, shuffle=False, batch_size=batch_size, num_workers=args.num_workers)
# CIFAR-100
elif args.num_classes == 100:
data_path = ("/private/home/cinjon/cifar-data/cifar-100-python")
_train_dataset = CIFAR_dataset([os.path.join(data_path, "train")],
args.num_classes, args.model, True)
train_dataloader = data.DataLoader(_train_dataset,
shuffle=True,
batch_size=batch_size)
_val_dataset = CIFAR_dataset([os.path.join(data_path, "test")],
args.num_classes, args.model, False)
val_dataloader = data.DataLoader(_val_dataset,
shuffle=False,
batch_size=batch_size)
else:
raise Exception("num_classes should be 10 or 100")
best_acc = 0.0
best_epoch = 0
# Training
for epoch in range(1, args.epochs + 1):
current_lr = max(3e-4, args.lr *\
math.pow(0.5, math.floor(epoch / args.lr_interval)))
linear_model.train()
if args.optimizer == "Adam":
optimizer = optim.Adam(linear_model.parameters(),
lr=current_lr,
weight_decay=args.weight_decay)
elif args.optimizer == "SGD":
optimizer = optim.SGD(
linear_model.parameters(),
lr=current_lr,
momentum=args.momentum,
weight_decay=args.weight_decay,
)
####################################################
# Train
t = time.time()
train_acc = 0
train_loss_sum = 0.0
for iter, input in enumerate(train_dataloader):
if time.time(
) - start_time > args.time * 3600 - 300 and comet_exp is not None:
comet_exp.end()
sys.exit(-1)
imgs = input[0].to(device)
if args.model != "resnet":
imgs = imgs.unsqueeze(1)
lbls = input[1].flatten().to(device)
# output = model(imgs)
# output = linear_model(output)
output = linear_model(model(imgs))
loss = F.cross_entropy(output, lbls)
train_loss_sum += float(loss.data)
train_acc += int(sum(torch.argmax(output, dim=1) == lbls))
optimizer.zero_grad()
loss.backward()
optimizer.step()
# log_text = "train epoch {}/{}\titer {}/{} loss:{} {:.3f}s/iter"
if iter % 1500 == 0:
log_text = "train epoch {}/{}\titer {}/{} loss:{}"
print(log_text.format(epoch, args.epochs, iter + 1,
len(train_dataloader), loss.data,
time.time() - t),
flush=False)
t = time.time()
train_acc /= len(_train_dataset)
train_loss_sum /= len(train_dataloader)
with comet_exp.train():
comet_exp.log_metrics({
'acc': train_acc,
'loss': train_loss_sum
},
step=(epoch + 1) * len(train_dataloader),
epoch=epoch + 1)
print("train acc epoch {}/{} loss:{} train_acc:{}".format(
epoch, args.epochs, train_loss_sum, train_acc),
flush=True)
#######################################################################
# Train acc
# linear_model.eval()
# train_acc = 0
# train_loss_sum = 0.0
# for iter, input in enumerate(train_acc_dataloader):
# imgs = input[0].to(device)
# if args.model != "resnet":
# imgs = imgs.unsqueeze(1)
# lbls = input[1].flatten().to(device)
#
# # output = model(imgs)
# # output = linear_model(output)
# output = linear_model(model(imgs))
# loss = F.cross_entropy(output, lbls)
# train_loss_sum += float(loss.data)
# train_acc += int(sum(torch.argmax(output, dim=1) == lbls))
#
# print("train acc epoch {}/{}\titer {}/{} loss:{} {:.3f}s/iter".format(
# epoch,
# args.epochs,
# iter+1,
# len(train_acc_dataloader),
# loss.data,
# time.time() - t),
# flush=True)
# t = time.time()
#
#
# train_acc /= len(train_acc_dataset)
# train_loss_sum /= len(train_acc_dataloader)
# print("train acc epoch {}/{} loss:{} train_acc:{}".format(
# epoch, args.epochs, train_loss_sum, train_acc), flush=True)
#######################################################################
# Train dev acc
# # linear_model.eval()
# train_dev_acc = 0
# train_dev_loss_sum = 0.0
# for iter, input in enumerate(train_dev_acc_dataloader):
# imgs = input[0].to(device)
# if args.model != "resnet":
# imgs = imgs.unsqueeze(1)
# lbls = input[1].flatten().to(device)
#
# output = model(imgs)
# output = linear_model(output)
# # output = linear_model(model(imgs))
# loss = F.cross_entropy(output, lbls)
# train_dev_loss_sum += float(loss.data)
# train_dev_acc += int(sum(torch.argmax(output, dim=1) == lbls))
#
# print("train dev acc epoch {}/{}\titer {}/{} loss:{} {:.3f}s/iter".format(
# epoch,
# args.epochs,
# iter+1,
# len(train_dev_acc_dataloader),
# loss.data,
# time.time() - t),
# flush=True)
# t = time.time()
#
# train_dev_acc /= len(train_dev_acc_dataset)
# train_dev_loss_sum /= len(train_dev_acc_dataloader)
# print("train dev epoch {}/{} loss:{} train_dev_acc:{}".format(
# epoch, args.epochs, train_dev_loss_sum, train_dev_acc), flush=True)
#######################################################################
# Val dev
# # linear_model.eval()
# val_dev_acc = 0
# val_dev_loss_sum = 0.0
# for iter, input in enumerate(val_dev_dataloader):
# imgs = input[0].to(device)
# if args.model != "resnet":
# imgs = imgs.unsqueeze(1)
# lbls = input[1].flatten().to(device)
#
# output = model(imgs)
# output = linear_model(output)
# loss = F.cross_entropy(output, lbls)
# val_dev_loss_sum += float(loss.data)
# val_dev_acc += int(sum(torch.argmax(output, dim=1) == lbls))
#
# print("val dev epoch {}/{} iter {}/{} loss:{} {:.3f}s/iter".format(
# epoch,
# args.epochs,
# iter+1,
# len(val_dev_dataloader),
# loss.data,
# time.time() - t),
# flush=True)
# t = time.time()
#
# val_dev_acc /= len(val_dev_dataset)
# val_dev_loss_sum /= len(val_dev_dataloader)
# print("val dev epoch {}/{} loss:{} val_dev_acc:{}".format(
# epoch, args.epochs, val_dev_loss_sum, val_dev_acc), flush=True)
#######################################################################
# Val
linear_model.eval()
val_acc = 0
val_loss_sum = 0.0
for iter, input in enumerate(val_dataloader):
if time.time(
) - start_time > args.time * 3600 - 300 and comet_exp is not None:
comet_exp.end()
sys.exit(-1)
imgs = input[0].to(device)
if args.model != "resnet":
imgs = imgs.unsqueeze(1)
lbls = input[1].flatten().to(device)
output = model(imgs)
output = linear_model(output)
loss = F.cross_entropy(output, lbls)
val_loss_sum += float(loss.data)
val_acc += int(sum(torch.argmax(output, dim=1) == lbls))
# log_text = "val epoch {}/{} iter {}/{} loss:{} {:.3f}s/iter"
if iter % 1500 == 0:
log_text = "val epoch {}/{} iter {}/{} loss:{}"
print(log_text.format(epoch, args.epochs, iter + 1,
len(val_dataloader), loss.data,
time.time() - t),
flush=False)
t = time.time()
val_acc /= len(_val_dataset)
val_loss_sum /= len(val_dataloader)
print("val epoch {}/{} loss:{} val_acc:{}".format(
epoch, args.epochs, val_loss_sum, val_acc))
with comet_exp.test():
comet_exp.log_metrics({
'acc': val_acc,
'loss': val_loss_sum
},
step=(epoch + 1) * len(train_dataloader),
epoch=epoch + 1)
if val_acc > best_acc:
best_acc = val_acc
best_epoch = epoch
linear_save_path = os.path.join(log_dir,
"{}.linear.pth".format(epoch))
model_save_path = os.path.join(log_dir,
"{}.model.pth".format(epoch))
torch.save(linear_model.state_dict(), linear_save_path)
torch.save(model.state_dict(), model_save_path)
# Check bias and variance
print(
"Epoch {} lr {} total: train_loss:{} train_acc:{} val_loss:{} val_acc:{}"
.format(epoch, current_lr, train_loss_sum, train_acc, val_loss_sum,
val_acc),
flush=True)
# print("Epoch {} lr {} total: train_acc:{} train_dev_acc:{} val_dev_acc:{} val_acc:{}".format(
# epoch, current_lr, train_acc, train_dev_acc, val_dev_acc, val_acc), flush=True)
print("The best epoch: {} acc: {}".format(best_epoch, best_acc))
if __name__ == '__main__':
SEED = 1234
random.seed(SEED)
torch.manual_seed(SEED)
torch.cuda.manual_seed(SEED)
torch.backends.cudnn.deterministic = True
config = get_config(sys.argv[1])
# experiment = Experiment("wXwnV8LZOtVfxqnRxr65Lv7C2")
comet_dir_path = os.path.join(config["result_directory"], config["model"])
makedirs(comet_dir_path)
experiment = OfflineExperiment(
project_name="DeepGenomics",
offline_directory=comet_dir_path)
experiment.log_parameters(config)
if torch.cuda.is_available():
# torch.cuda.set_device(str(os.environ["CUDA_VISIBLE_DEVICES"]))
device = torch.device('cuda:{}'.format(os.environ["CUDA_VISIBLE_DEVICES"]))
else:
device = torch.device('cpu')
print(device)
number_of_examples = len(get_filenames(os.path.join(config["data"], "x")))
list_ids = [str(i) for i in range(number_of_examples)]
random.shuffle(list_ids)
t_ind, v_ind = round(number_of_examples * 0.7), round(number_of_examples * 0.9)
train_indices, validation_indices, test_indices = list_ids[:t_ind], list_ids[t_ind:v_ind], list_ids[v_ind:]
params = {'batch_size': config["training"]["batch_size"],
hyper_params = {
"sequence_length": 28,
"input_size": 28,
"hidden_size": 128,
"num_layers": 2,
"num_classes": 10,
"batch_size": 100,
"num_epochs": 3,
"learning_rate": 0.01
}
optimizer = Optimizer("pA3Hqc1pEswNvXOPtSoRobt7C")
experiment = OfflineExperiment(project_name="horoma",
offline_directory="./experiments",
disabled=False)
experiment.log_parameters(hyper_params)
# MNIST Dataset
train_dataset = dsets.MNIST(root='./data/',
train=True,
transform=transforms.ToTensor(),
download=True)
test_dataset = dsets.MNIST(root='./data/',
train=False,
transform=transforms.ToTensor())
# Data Loader (Input Pipeline)
train_loader = torch.utils.data.DataLoader(
class CometLogger(LightningLoggerBase):
r"""
Log using `Comet.ml <https://www.comet.ml>`_.
Install it with pip:
.. code-block:: bash
pip install comet-ml
Comet requires either an API Key (online mode) or a local directory path (offline mode).
**ONLINE MODE**
.. code-block:: python
import os
from pytorch_lightning import Trainer
from pytorch_lightning.loggers import CometLogger
# arguments made to CometLogger are passed on to the comet_ml.Experiment class
comet_logger = CometLogger(
api_key=os.environ.get('COMET_API_KEY'),
workspace=os.environ.get('COMET_WORKSPACE'), # Optional
save_dir='.', # Optional
project_name='default_project', # Optional
rest_api_key=os.environ.get('COMET_REST_API_KEY'), # Optional
experiment_name='default' # Optional
)
trainer = Trainer(logger=comet_logger)
**OFFLINE MODE**
.. code-block:: python
from pytorch_lightning.loggers import CometLogger
# arguments made to CometLogger are passed on to the comet_ml.Experiment class
comet_logger = CometLogger(
save_dir='.',
workspace=os.environ.get('COMET_WORKSPACE'), # Optional
project_name='default_project', # Optional
rest_api_key=os.environ.get('COMET_REST_API_KEY'), # Optional
experiment_name='default' # Optional
)
trainer = Trainer(logger=comet_logger)
Args:
api_key: Required in online mode. API key, found on Comet.ml. If not given, this
will be loaded from the environment variable COMET_API_KEY or ~/.comet.config
if either exists.
save_dir: Required in offline mode. The path for the directory to save local
comet logs. If given, this also sets the directory for saving checkpoints.
project_name: Optional. Send your experiment to a specific project.
Otherwise will be sent to Uncategorized Experiments.
If the project name does not already exist, Comet.ml will create a new project.
rest_api_key: Optional. Rest API key found in Comet.ml settings.
This is used to determine version number
experiment_name: Optional. String representing the name for this particular experiment on Comet.ml.
experiment_key: Optional. If set, restores from existing experiment.
offline: If api_key and save_dir are both given, this determines whether
the experiment will be in online or offline mode. This is useful if you use
save_dir to control the checkpoints directory and have a ~/.comet.config
file but still want to run offline experiments.
prefix: A string to put at the beginning of metric keys.
\**kwargs: Additional arguments like `workspace`, `log_code`, etc. used by
:class:`CometExperiment` can be passed as keyword arguments in this logger.
"""
LOGGER_JOIN_CHAR = '-'
def __init__(self,
api_key: Optional[str] = None,
save_dir: Optional[str] = None,
project_name: Optional[str] = None,
rest_api_key: Optional[str] = None,
experiment_name: Optional[str] = None,
experiment_key: Optional[str] = None,
offline: bool = False,
prefix: str = '',
**kwargs):
if comet_ml is None:
raise ImportError(
"You want to use `comet_ml` logger which is not installed yet,"
" install it with `pip install comet-ml`.")
super().__init__()
self._experiment = None
# Determine online or offline mode based on which arguments were passed to CometLogger
api_key = api_key or comet_ml.config.get_api_key(
None, comet_ml.config.get_config())
if api_key is not None and save_dir is not None:
self.mode = "offline" if offline else "online"
self.api_key = api_key
self._save_dir = save_dir
elif api_key is not None:
self.mode = "online"
self.api_key = api_key
self._save_dir = None
elif save_dir is not None:
self.mode = "offline"
self._save_dir = save_dir
else:
# If neither api_key nor save_dir are passed as arguments, raise an exception
raise MisconfigurationException(
"CometLogger requires either api_key or save_dir during initialization."
)
log.info(f"CometLogger will be initialized in {self.mode} mode")
self._project_name = project_name
self._experiment_key = experiment_key
self._experiment_name = experiment_name
self._prefix = prefix
self._kwargs = kwargs
self._future_experiment_key = None
if rest_api_key is not None:
# Comet.ml rest API, used to determine version number
self.rest_api_key = rest_api_key
self.comet_api = API(self.rest_api_key)
else:
self.rest_api_key = None
self.comet_api = None
self._kwargs = kwargs
@property
@rank_zero_experiment
def experiment(self):
r"""
Actual Comet object. To use Comet features in your
:class:`~pytorch_lightning.core.lightning.LightningModule` do the following.
Example::
self.logger.experiment.some_comet_function()
"""
if self._experiment is not None:
return self._experiment
if self._future_experiment_key is not None:
os.environ["COMET_EXPERIMENT_KEY"] = self._future_experiment_key
try:
if self.mode == "online":
if self._experiment_key is None:
self._experiment = CometExperiment(
api_key=self.api_key,
project_name=self._project_name,
**self._kwargs,
)
self._experiment_key = self._experiment.get_key()
else:
self._experiment = CometExistingExperiment(
api_key=self.api_key,
project_name=self._project_name,
previous_experiment=self._experiment_key,
**self._kwargs,
)
else:
self._experiment = CometOfflineExperiment(
offline_directory=self.save_dir,
project_name=self._project_name,
**self._kwargs,
)
finally:
if self._future_experiment_key is not None:
os.environ.pop("COMET_EXPERIMENT_KEY")
self._future_experiment_key = None
if self._experiment_name:
self._experiment.set_name(self._experiment_name)
return self._experiment
@rank_zero_only
def log_hyperparams(self, params: Union[Dict[str, Any],
Namespace]) -> None:
params = self._convert_params(params)
params = self._flatten_dict(params)
self.experiment.log_parameters(params)
@rank_zero_only
def log_metrics(self,
metrics: Dict[str, Union[torch.Tensor, float]],
step: Optional[int] = None) -> None:
assert rank_zero_only.rank == 0, "experiment tried to log from global_rank != 0"
# Comet.ml expects metrics to be a dictionary of detached tensors on CPU
for key, val in metrics.items():
if is_tensor(val):
metrics[key] = val.cpu().detach()
metrics_without_epoch = metrics.copy()
epoch = metrics_without_epoch.pop('epoch', None)
metrics_without_epoch = self._add_prefix(metrics_without_epoch)
self.experiment.log_metrics(metrics_without_epoch,
step=step,
epoch=epoch)
def reset_experiment(self):
self._experiment = None
@rank_zero_only
def finalize(self, status: str) -> None:
r"""
When calling ``self.experiment.end()``, that experiment won't log any more data to Comet.
That's why, if you need to log any more data, you need to create an ExistingCometExperiment.
For example, to log data when testing your model after training, because when training is
finalized :meth:`CometLogger.finalize` is called.
This happens automatically in the :meth:`~CometLogger.experiment` property, when
``self._experiment`` is set to ``None``, i.e. ``self.reset_experiment()``.
"""
self.experiment.end()
self.reset_experiment()
@property
def save_dir(self) -> Optional[str]:
return self._save_dir
@property
def name(self) -> str:
# Don't create an experiment if we don't have one
if self._experiment is not None and self._experiment.project_name is not None:
return self._experiment.project_name
if self._project_name is not None:
return self._project_name
return "comet-default"
@property
def version(self) -> str:
# Don't create an experiment if we don't have one
if self._experiment is not None:
return self._experiment.id
if self._experiment_key is not None:
return self._experiment_key
if "COMET_EXPERIMENT_KEY" in os.environ:
return os.environ["COMET_EXPERIMENT_KEY"]
if self._future_experiment_key is not None:
return self._future_experiment_key
# Pre-generate an experiment key
self._future_experiment_key = comet_ml.generate_guid()
return self._future_experiment_key
def __getstate__(self):
state = self.__dict__.copy()
# Save the experiment id in case an experiment object already exists,
# this way we could create an ExistingExperiment pointing to the same
# experiment
state[
"_experiment_key"] = self._experiment.id if self._experiment is not None else None
# Remove the experiment object as it contains hard to pickle objects
# (like network connections), the experiment object will be recreated if
# needed later
state["_experiment"] = None
return state
encode = configuration['encode']
cluster = configuration['cluster']
latent_dim = configuration['latent_dim']
flattened = False # Default
device = 'cuda' if torch.cuda.is_available() else 'cpu'
# Set all seeds for full reproducibility
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
# Set up Comet Experiment tracking
experiment = OfflineExperiment(project_name='general',
workspace='timothynest', # Replace this with appropriate comet workspace
offline_directory="experiments")
experiment.set_name(
name=args.config + "_dim={}_overlapped={}".format(latent_dim, train_split))
experiment.log_parameters(configuration)
# Initialize necessary objects
if clustering_model == 'kmeans':
clustering_model = KMeansClustering(n_clusters, seed)
elif clustering_model == 'gmm':
clustering_model = GMMClustering(n_clusters, seed)
elif clustering_model == 'svm':
clustering_model = SVMClustering(seed)
else:
print('No clustering model specified. Using Kmeans.')
clustering_model = KMeansClustering(n_clusters, seed)