def ShouldRunTest(self, run_params):
if trt_convert.get_linked_tensorrt_version()[0] < 5:
return False
# Only test static engine mode, with or without calibration.
return (trt_test.IsQuantizationMode(run_params.precision_mode)
and not run_params.use_optimizer
and not run_params.dynamic_engine)
def _GetConfigProto(self, run_params, graph_state):
"""Get config proto based on specific settings."""
if graph_state != GraphState.ORIGINAL and run_params.use_optimizer:
conversion_params = self.GetConversionParams(run_params)
rewriter_cfg = trt_convert.tensorrt_rewriter_config(
conversion_params.rewriter_config,
conversion_params.max_batch_size,
conversion_params.max_workspace_size_bytes,
conversion_params.precision_mode,
conversion_params.minimum_segment_size,
conversion_params.is_dynamic_op,
conversion_params.maximum_cached_engines,
conversion_params.cached_engine_batch_sizes)
graph_options = config_pb2.GraphOptions(
rewrite_options=rewriter_cfg)
else:
graph_options = config_pb2.GraphOptions()
gpu_options = config_pb2.GPUOptions()
gpu_options.allow_growth = True
if trt_convert.get_linked_tensorrt_version()[0] == 3:
gpu_options.per_process_gpu_memory_fraction = 0.50
config = config_pb2.ConfigProto(gpu_options=gpu_options,
graph_options=graph_options)
return config
def testEval(self):
if not trt_convert.is_tensorrt_enabled():
return
model_dir = test.test_src_dir_path('contrib/tensorrt/test/testdata')
accuracy_tf_native = self._Run(
is_training=False,
use_trt=False,
batch_size=128,
num_epochs=None,
model_dir=model_dir)['accuracy']
logging.info('accuracy_tf_native: %f', accuracy_tf_native)
self.assertAllClose(accuracy_tf_native, 0.9662)
if trt_convert.get_linked_tensorrt_version()[0] < 5:
return
accuracy_tf_trt = self._Run(
is_training=False,
use_trt=True,
batch_size=128,
num_epochs=None,
model_dir=model_dir)['accuracy']
logging.info('accuracy_tf_trt: %f', accuracy_tf_trt)
self.assertAllClose(accuracy_tf_trt, 0.9677)
def _GetConfigProto(self, run_params, graph_state):
"""Get config proto based on specific settings."""
if graph_state != GraphState.ORIGINAL and run_params.use_optimizer:
conversion_params = self.GetConversionParams(run_params)
rewriter_cfg = trt_convert.tensorrt_rewriter_config(
conversion_params.rewriter_config, conversion_params.max_batch_size,
conversion_params.max_workspace_size_bytes,
conversion_params.precision_mode,
conversion_params.minimum_segment_size,
conversion_params.is_dynamic_op,
conversion_params.maximum_cached_engines,
conversion_params.cached_engine_batch_sizes)
graph_options = config_pb2.GraphOptions(rewrite_options=rewriter_cfg)
else:
graph_options = config_pb2.GraphOptions()
gpu_options = config_pb2.GPUOptions()
gpu_options.allow_growth = True
if trt_convert.get_linked_tensorrt_version()[0] == 3:
gpu_options.per_process_gpu_memory_fraction = 0.50
config = config_pb2.ConfigProto(
gpu_options=gpu_options, graph_options=graph_options)
return config
def _GetConfigProto(self,
params,
use_optimizer,
precision_mode=None,
is_dynamic_op=None):
"""Get config proto based on specific settings."""
if use_optimizer:
rewriter_cfg = rewriter_config_pb2.RewriterConfig()
rewriter_cfg.optimizers.extend(["constfold", "layout"])
custom_op = rewriter_cfg.custom_optimizers.add()
custom_op.name = "TensorRTOptimizer"
custom_op.parameter_map["minimum_segment_size"].i = 3
custom_op.parameter_map["max_batch_size"].i = max(
[dims[0] for dims in params.input_dims])
custom_op.parameter_map["is_dynamic_op"].b = is_dynamic_op
custom_op.parameter_map["max_workspace_size_bytes"].i = 1 << 25
custom_op.parameter_map["precision_mode"].s = self._ToBytes(
precision_mode)
graph_options = config_pb2.GraphOptions(rewrite_options=rewriter_cfg)
else:
graph_options = config_pb2.GraphOptions()
gpu_options = config_pb2.GPUOptions()
gpu_options.allow_growth = True
if trt_convert.get_linked_tensorrt_version()[0] == 3:
gpu_options.per_process_gpu_memory_fraction = 0.50
config = config_pb2.ConfigProto(
gpu_options=gpu_options, graph_options=graph_options)
return config
def _GetConfigProto(self, run_params, graph_state):
"""Get config proto based on specific settings."""
if graph_state != GraphState.ORIGINAL and run_params.use_optimizer:
rewriter_cfg = rewriter_config_pb2.RewriterConfig()
rewriter_cfg.optimizers.extend(["constfold", "layout"])
custom_op = rewriter_cfg.custom_optimizers.add()
custom_op.name = "TensorRTOptimizer"
trt_params = self.GetConversionParams(run_params)
custom_op.parameter_map["max_batch_size"].i = trt_params.max_batch_size
custom_op.parameter_map["max_workspace_size_bytes"].i = (
trt_params.max_workspace_size_bytes)
custom_op.parameter_map["precision_mode"].s = trt_params.precision_mode
custom_op.parameter_map["minimum_segment_size"].i = (
trt_params.minimum_segment_size)
custom_op.parameter_map["is_dynamic_op"].b = trt_params.is_dynamic_op
custom_op.parameter_map["maximum_cached_engines"].i = (
trt_params.maximum_cached_engines)
if trt_params.cached_engine_batches:
custom_op.parameter_map["cached_engine_batches"].list.i.extend(
trt_params.cached_engine_batches)
graph_options = config_pb2.GraphOptions(rewrite_options=rewriter_cfg)
else:
graph_options = config_pb2.GraphOptions()
gpu_options = config_pb2.GPUOptions()
gpu_options.allow_growth = True
if trt_convert.get_linked_tensorrt_version()[0] == 3:
gpu_options.per_process_gpu_memory_fraction = 0.50
config = config_pb2.ConfigProto(
gpu_options=gpu_options, graph_options=graph_options)
return config
def testEval(self):
if not trt_convert.is_tensorrt_enabled():
return
model_dir = test.test_src_dir_path('contrib/tensorrt/test/testdata')
accuracy_tf_native = self._Run(
is_training=False,
use_trt=False,
batch_size=128,
num_epochs=None,
model_dir=model_dir)['accuracy']
logging.info('accuracy_tf_native: %f', accuracy_tf_native)
self.assertAllClose(0.9662, accuracy_tf_native, rtol=1e-3, atol=1e-3)
if trt_convert.get_linked_tensorrt_version()[0] < 5:
return
accuracy_tf_trt = self._Run(
is_training=False,
use_trt=True,
batch_size=128,
num_epochs=None,
model_dir=model_dir)['accuracy']
logging.info('accuracy_tf_trt: %f', accuracy_tf_trt)
self.assertAllClose(0.9675, accuracy_tf_trt, rtol=1e-3, atol=1e-3)
def _GetConfigProto(self, run_params, graph_state):
"""Get config proto based on specific settings."""
if graph_state != GraphState.ORIGINAL and run_params.use_optimizer:
rewriter_cfg = rewriter_config_pb2.RewriterConfig()
rewriter_cfg.optimizers.extend(["constfold", "layout"])
custom_op = rewriter_cfg.custom_optimizers.add()
custom_op.name = "TensorRTOptimizer"
trt_params = self.GetConversionParams(run_params)
custom_op.parameter_map[
"max_batch_size"].i = trt_params.max_batch_size
custom_op.parameter_map["max_workspace_size_bytes"].i = (
trt_params.max_workspace_size_bytes)
custom_op.parameter_map[
"precision_mode"].s = trt_params.precision_mode
custom_op.parameter_map["minimum_segment_size"].i = (
trt_params.minimum_segment_size)
custom_op.parameter_map[
"is_dynamic_op"].b = trt_params.is_dynamic_op
custom_op.parameter_map["maximum_cached_engines"].i = (
trt_params.maximum_cached_engines)
if trt_params.cached_engine_batches:
custom_op.parameter_map["cached_engine_batches"].list.i.extend(
trt_params.cached_engine_batches)
graph_options = config_pb2.GraphOptions(
rewrite_options=rewriter_cfg)
else:
graph_options = config_pb2.GraphOptions()
gpu_options = config_pb2.GPUOptions()
gpu_options.allow_growth = True
if trt_convert.get_linked_tensorrt_version()[0] == 3:
gpu_options.per_process_gpu_memory_fraction = 0.50
config = config_pb2.ConfigProto(gpu_options=gpu_options,
graph_options=graph_options)
return config
def ShouldRunTest(self, run_params):
if trt_convert.get_linked_tensorrt_version()[0] < 5:
return False
# Test static/dynamic engine with/without calibration.
return (trt_test.IsQuantizationMode(run_params.precision_mode) and
not run_params.use_optimizer)
