def _construct_job(self):
def job():
batch = self._cfg.data()
return self._model.validation_step(batch)
job.__name__ = self._model.__class__.__name__ + "_Model_eval_job"
deco = api_oneflow_function(type="predict", function_config=self._cfg.exe_cfg)
return deco(job)
def _construct_numpy_job(self, batch: Tuple[np.ndarray, ...] = None):
def job(*input_batch):
return self._model.validation_step(batch=input_batch)
_infer_job_signature(self._cfg.data, batch, 0, job)
job.__name__ = self._model.__class__.__name__ + "_Model_eval_numpy_job"
deco = api_oneflow_function(type="predict", function_config=self._cfg.exe_cfg)
return deco(job)
def _construct_job(self, optimizer_idx: int = 0):
def job():
batch = self._cfg.data(0, optimizer_idx)
outputs = self._model.training_step(batch=batch,
optimizer_idx=optimizer_idx)
loss = None
if isinstance(outputs, tuple) and len(outputs) > 0:
loss = outputs[0]
else:
loss = outputs
self._opts[optimizer_idx].minimize(loss)
return outputs
job.__name__ = (self._model.__class__.__name__ + "_Model_train_job_" +
str(optimizer_idx))
deco = api_oneflow_function(type="train",
function_config=self._cfg.exe_cfg)
return deco(job)
def _construct_numpy_job(self, batch, optimizer_idx):
def job(*input_batch):
outputs = self._model.training_step(batch=input_batch,
optimizer_idx=optimizer_idx)
loss = None
if isinstance(outputs, tuple) and len(outputs) > 0:
loss = outputs[0]
else:
loss = outputs
self._opts[optimizer_idx].minimize(loss)
return outputs
_infer_job_signature(self._cfg.data, batch, optimizer_idx, job)
job.__name__ = (self._model.__class__.__name__ +
"_Model_train_numpy_job_" + str(optimizer_idx))
deco = api_oneflow_function(type="train",
function_config=self._cfg.exe_cfg)
return deco(job)