def get_npu_classifier(self):
if self.config.chip == 'npu':
from npu_bridge.estimator.npu.npu_config import NPURunConfig
from npu_bridge.estimator.npu.npu_estimator import NPUEstimator
## refer to links
# https://support.huaweicloud.com/tensorflowdevg-cann330alphaXtraining/atlasmprtg_13_0032.html
# https://support.huaweicloud.com/tensorflowdevg-cann330alphaXtraining/atlastfadapi_07_0004.html
if self.config.npu_profiling:
from npu_bridge.estimator.npu.npu_config import ProfilingConfig
work_dir = os.getcwd()
profiling_dir = os.path.join(work_dir, "npu_profiling")
if not os.path.exists(profiling_dir):
os.makedirs(profiling_dir)
profiling_options = '{"output":"%s","task_trace": "on","aicpu": "on"}' % (
profiling_dir)
profiling_config = ProfilingConfig(
enable_profiling=True, profiling_options=profiling_options)
run_config = NPURunConfig(
profiling_config=profiling_config, ## 配置profiling参数
save_checkpoints_steps=112590,
session_config=self.sess.estimator_config,
model_dir=self.config.log_dir,
keep_checkpoint_max=self.config.max_checkpoint_to_save)
else:
run_config = NPURunConfig(
save_checkpoints_steps=112590,
session_config=self.sess.estimator_config,
model_dir=self.config.log_dir,
keep_checkpoint_max=self.config.max_checkpoint_to_save)
classifier = NPUEstimator(
model_fn=self.model.get_estimator_model_func,
config=run_config)
elif self.config.chip in ['gpu', 'cpu']:
run_config = tf.estimator.RunConfig(
save_checkpoints_steps=112590,
session_config=self.sess.estimator_config,
model_dir=self.config.log_dir,
keep_checkpoint_max=self.config.max_checkpoint_to_save)
classifier = tf.estimator.Estimator(
model_fn=self.model.get_estimator_model_func,
config=run_config)
training_hooks = []
training_hooks.append(self.logger)
return classifier, training_hooks
def _init_npu_estimator(self, sess_config):
model_dir = self.get_local_worker_path()
config = NPURunConfig(model_dir=model_dir,
save_checkpoints_steps=self.config.save_steps,
log_step_count_steps=self.config.report_freq,
session_config=sess_config,
enable_data_pre_proc=True,
iterations_per_loop=1)
self.estimator = NPUEstimator(model_fn=self.model_fn, config=config)
def _refresh_model_dir_and_session_config(config, model_dir):
"""Overwrite estimator config by `model_dir` and `session_config` if needed.
Args:
config: Original estimator config.
model_dir: Estimator model checkpoint directory.
Returns:
Overwritten estimator config.
Raises:
ValueError: Model directory inconsistent between `model_dir` and `config`.
"""
if config is None or not isinstance(config, NPURunConfig):
raise ValueError(
'config must be an instance of `NPURunConfig`, but provided %s.' %
config)
if config.session_config is None:
session_config = run_config.get_default_session_config()
config = NPURunConfig.replace(config, session_config=session_config)
model_dir = compat_internal.path_to_str(model_dir)
if model_dir is not None:
if (getattr(config, 'model_dir', None) is not None
and config.model_dir != model_dir):
raise ValueError(
"`model_dir` are set both in constructor and `NPURunConfig`, but with "
"different values. In constructor: '{}', in `NPURunConfig`: "
"'{}' ".format(model_dir, config.model_dir))
if model_dir:
config = NPURunConfig.replace(config, model_dir=model_dir)
elif getattr(config, 'model_dir', None) is None:
model_dir = tempfile.mkdtemp()
logging.warning('Using temporary folder as model directory: %s',
model_dir)
config = NPURunConfig.replace(config, model_dir=model_dir)
return config
def main(unused_argv):
# set the log
set_log()
# Load training and eval data
mnist = input_data.read_data_sets(FLAGS.data_dir, False)
train_data = mnist.train.images # Returns np.array
train_labels = np.asarray(mnist.train.labels, dtype=np.int32)
eval_data = mnist.test.images # Returns np.array
eval_labels = np.asarray(mnist.test.labels, dtype=np.int32)
# profiling_config = ProfilingConfig(enable_profiling=True,
# enable_options=["training_trace","task_trace"])
npu_config = NPURunConfig(iterations_per_loop=100,
save_checkpoints_steps=10,
model_dir=FLAGS.model_dir,
session_config=tf.ConfigProto(
allow_soft_placement=True,
log_device_placement=False)
#,profiling_config=profiling_config
)
mnist_classifier = NPUEstimator(model_fn=cnn_model_fn,
config=npu_config,
params={},
job_start_file=FLAGS.job_start_file)
# Set up logging for predictions
# Log the values in the "Softmax" tensor with label "probabilities"
tensors_to_log = {"probabilities": "softmax_tensor"}
logging_hook = tf.train.LoggingTensorHook(
tensors=tensors_to_log, every_n_iter=50) # Train the model
# print("Train the model...")
mnist_classifier.train(input_fn=train_input_fn(train_data, train_labels),
steps=FLAGS.train_steps,
hooks=[logging_hook])
# Evaluate the model and print results
print("Evaluate the model...")
eval_results = mnist_classifier.evaluate(input_fn=eval_input_fn(
eval_data, eval_labels),
steps=FLAGS.eval_steps)
print("eval_results: ", eval_results)
def get_npu_classifier(self):
from npu_bridge.estimator.npu.npu_config import NPURunConfig
from npu_bridge.estimator.npu.npu_estimator import NPUEstimator
run_config = NPURunConfig(
hcom_parallel=True,
precision_mode="allow_mix_precision",
enable_data_pre_proc=True,
save_checkpoints_steps=self.args.nsteps_per_epoch,
session_config=self.sess.estimator_config,
model_dir=self.args.log_dir,
iterations_per_loop=self.args.iterations_per_loop,
keep_checkpoint_max=5)
classifier = NPUEstimator(model_fn=self.model.get_estimator_model_func,
config=run_config)
training_hooks = []
training_hooks.append(self.logger)
return classifier, training_hooks