def build_train_evaluate_model(self,
limit_train_batches: int,
limit_val_batches: int,
epoch: int,
batch_size: int,
model_config: BiDirectionalConfig,
precision: int,
gpus: int,
folders: bool):
cuda = gpus > 0
train_path, val_path, test_path = self.download_data(
DatasetConfig(
target_dir=self.target_dir,
manifest_dir=self.manifest_dir
),
folders=folders
)
train_cfg = self.create_training_config(
limit_train_batches=limit_train_batches,
limit_val_batches=limit_val_batches,
max_epochs=epoch,
batch_size=batch_size,
train_path=train_path,
val_path=val_path,
model_config=model_config,
precision=precision,
gpus=gpus
)
print("Running Training DeepSpeech Model Smoke Test")
train(train_cfg)
# Expected final model path after training
print(os.listdir(self.model_dir))
model_path = self.model_dir + '/last.ckpt'
assert os.path.exists(model_path)
lm_configs = [
LMConfig(), # Test Greedy
LMConfig(
decoder_type=DecoderType.beam
) # Test Beam Decoder
]
print("Running Inference Smoke Tests")
for lm_config in lm_configs:
self.eval_model(
model_path=model_path,
test_path=test_path,
cuda=cuda,
precision=precision,
lm_config=lm_config
)
self.inference(test_path=test_path,
model_path=model_path,
cuda=cuda,
precision=precision,
lm_config=lm_config)
def build_train_evaluate_model(self, epoch: int, batch_size: int,
model_config: BiDirectionalConfig,
use_half: bool, cuda: bool):
train_manifest, val_manifest, test_manifest = self.download_data(
DatasetConfig(target_dir=self.target_dir,
manifest_dir=self.manifest_dir))
train_cfg = self.create_training_config(epoch=epoch,
batch_size=batch_size,
train_manifest=train_manifest,
val_manifest=val_manifest,
model_config=model_config,
cuda=cuda)
print("Running Training DeepSpeech Model Smoke Test")
train(train_cfg)
# Expected final model path after training
model_path = self.model_dir + '/deepspeech_final.pth'
assert os.path.exists(model_path)
lm_configs = [
LMConfig(), # Test Greedy
LMConfig(decoder_type=DecoderType.beam) # Test Beam Decoder
]
print("Running Inference Smoke Tests")
for lm_config in lm_configs:
self.eval_model(model_path=model_path,
test_manifest=test_manifest,
cuda=cuda,
use_half=use_half,
lm_config=lm_config)
self.inference(test_manifest=test_manifest,
model_path=model_path,
cuda=cuda,
use_half=use_half,
lm_config=lm_config)
def hydra_main(cfg: DeepSpeechConfig):
train(cfg=cfg)
def hydra_main(cfg):
train(cfg=cfg)
def hydra_main(cfg: DictConfig):
train(cfg=cfg)
def main(args):
training.train(args)