def run(force_rerun: bool, params_path: Optional[str],
changed_params_path: Optional[str], device_id: Optional[int]) -> None:
module = get_params_module(params_path, 'cl_default_config')
config = module.classifier_config
if changed_params_path:
with open(changed_params_path, 'r') as f:
patch = dict(jsons.loads(f.read()))
config = patch_config(config, patch)
if gpu.gpu_available():
gpu_id_to_use = device_id if device_id is not None else get_current_device(
)
logger.debug(f'Using gpu with id: {gpu_id_to_use}')
with device(gpu_id_to_use):
run_on_device(config, force_rerun)
else:
run_on_device(config, force_rerun)
def create_nn_architecture(fs: FS,
text_field: Field,
level_label: Field,
data: Data,
arch: Arch,
threshold: float,
path: str = None):
splits = ContextsDataset.splits(text_field,
level_label,
fs.path_to_model_dataset,
context_len=arch.bptt,
threshold=threshold,
data=data)
text_data = TextData.from_splits(fs.path_to_model if not path else path,
splits, arch.bs)
opt_fn = partial(torch.optim.Adam, betas=(0.7, 0.99))
if arch.qrnn and not gpu.gpu_available():
logger.warning(
"Cuda not available, not using qrnn. Using lstm instead")
arch.qrnn = False
dropout_multiplier = arch.drop.multiplier
rnn_learner = text_data.get_model(
opt_fn,
arch.bptt + 1,
arch.bptt,
arch.em_sz,
arch.nh,
arch.nl,
dropout=arch.drop.out * dropout_multiplier,
dropouti=arch.drop.outi * dropout_multiplier,
wdrop=arch.drop.w * dropout_multiplier,
dropoute=arch.drop.oute * dropout_multiplier,
dropouth=arch.drop.outh * dropout_multiplier,
bidir=arch.bidir,
qrnn=arch.qrnn)
# reguarizing LSTM paper -- penalizing large activations -- reduce overfitting
rnn_learner.reg_fn = partial(seq2seq_reg,
alpha=arch.reg_fn.alpha,
beta=arch.reg_fn.beta)
rnn_learner.clip = arch.clip
logger.info(f'Dictionary size is: {len(text_field.vocab.itos)}')
return rnn_learner
def run(find_lr: bool, force_rerun: bool, params_path: Optional[str],
changed_params_path: Optional[str], device_id: Optional[int]) -> None:
if find_lr:
module = get_params_module(params_path, 'lm_lr_default_config')
config = module.lm_lr_config
else:
module = get_params_module(params_path, 'lm_default_config')
config = module.lm_config
if changed_params_path:
with open(changed_params_path, 'r') as f:
patch = dict(jsons.loads(f.read()))
config = patch_config(config, patch)
logger.info(f'Using config: {jsons.dumps(config)}')
if gpu.gpu_available():
gpu_id_to_use = device_id if device_id is not None else get_current_device(
)
with device(gpu_id_to_use):
run_on_device(config, find_lr, force_rerun)
else:
run_on_device(config, find_lr, force_rerun)
#!/usr/bin/env python
from logrec.util import gpu
if __name__ == '__main__':
if gpu.gpu_available():
print(gpu.get_n_gpus())
else:
print(-1)