def _fit_on_prepared_data(self, backend, train_rows, val_rows, metadata, avg_row_size, dataset_idx=None): self._check_params(metadata) keras_utils = self._get_keras_utils() run_id = self.getRunId() if run_id is None: run_id = 'keras_' + str(int(time.time())) if self._has_checkpoint(run_id): serialized_model = self._load_model_from_checkpoint(run_id) else: serialized_model = self._compile_model(keras_utils) # Workaround: # https://stackoverflow.com/questions/50583056/is-there-any-way-to-set-java-opts-for-tensorflow-process/50615570 env = {'LIBHDFS_OPTS': '-Xms2048m -Xmx2048m'} trainer = remote.RemoteTrainer(self, metadata, keras_utils, run_id, dataset_idx) handle = backend.run(trainer, args=(serialized_model, train_rows, val_rows, avg_row_size), env=env) return self._create_model(handle, run_id, metadata)
def _fit_on_prepared_data(self, backend, train_rows, val_rows, metadata, avg_row_size, dataset_idx=None): self._check_params(metadata) keras_utils = self._get_keras_utils() run_id = self.getRunId() if run_id is None: run_id = 'keras_' + str(int(time.time())) if self._has_checkpoint(run_id): serialized_model = self._load_model_from_checkpoint(run_id) else: serialized_model = self._compile_model(keras_utils) trainer = remote.RemoteTrainer(self, metadata, keras_utils, run_id, dataset_idx) handle = backend.run(trainer, args=(serialized_model, train_rows, val_rows, avg_row_size), env=self.getBackendEnv()) return self._create_model(handle, run_id, metadata)