def test_device_buffer_order_matches_bindings(self):
model = ONNX_MODELS["reducable"]
engine = engine_from_network(NetworkFromOnnxBytes(model.loader))
with engine, TrtRunner(engine) as runner:
dev_buf_order = list(runner.device_buffers.keys())
for binding, dev_buf_name in zip(engine, dev_buf_order):
assert binding == dev_buf_name
def onnx_to_trt(self, output_fpath: str, input_fpath: str,
network_metadata: NetworkMetadata):
"""
Converts ONNX file to TRT engine.
Since TensorRT already supplies converter functions and scripts,
a default implementation is already provided.
Arg:
output_fpath (str): File location of the generated ONNX file.
input_fpath (str): Input file location of the generated ONNX file.
network_metadata (NetworkMetadata): Network metadata of the network being converted.
Returns:
TRTEngineFile: Newly generated engine.
"""
result = self.trt_engine_class(output_fpath, network_metadata)
self.trt_inference_config = CreateConfig(
fp16=network_metadata.precision.fp16,
max_workspace_size=result.DEFAULT_TRT_WORKSPACE_MB * 1024 * 1024,
profiles=result.get_dynamic_shape_profiles(),
strict_types=result.use_strict_types())
g_logger_verbosity = (PG_LOGGER.EXTRA_VERBOSE if G_LOGGER.level
== G_LOGGER.DEBUG else PG_LOGGER.WARNING)
with PG_LOGGER.verbosity(g_logger_verbosity):
network_definition = result.get_network_definition(
network_from_onnx_path(input_fpath))
trt_engine = engine_from_network(network_definition,
config=self.trt_inference_config)
save_engine(trt_engine, output_fpath)
return result
def test_context(self):
model = ONNX_MODELS["identity"]
engine = engine_from_network(NetworkFromOnnxBytes(model.loader))
with engine, TrtRunner(engine.create_execution_context) as runner:
model.check_runner(runner)
assert not runner.owns_engine
assert runner.owns_context
def test_calibrator_device_buffers_multiinput(self,
multi_input_builder_network,
mode):
def generate_dev_data(num_batches):
with cuda.DeviceArray(shape=(1, ), dtype=np.float32) as x:
for _ in range(num_batches):
x.copy_from(np.ones((1, ), dtype=np.float32))
xdata = {
"array": x,
"view": cuda.DeviceView(x.ptr, x.shape, x.dtype),
"pointer": x.ptr
}[mode]
yield {
"X0": xdata,
"Y0": np.zeros((1, ), dtype=np.float32)
}
builder, network = multi_input_builder_network
NUM_BATCHES = 2
calibrator = Calibrator(generate_dev_data(NUM_BATCHES))
create_config = CreateConfig(int8=True, calibrator=calibrator)
with engine_from_network((builder, network), create_config):
assert calibrator.num_batches == NUM_BATCHES
self.check_calibrator_cleanup(calibrator)
def main():
# In Polygraphy, loaders and runners take ownership of objects if they are provided
# via the return values of callables. For example, we don't need to worry about object
# lifetimes when we use lazy loaders.
#
# Since we are immediately evaluating, we take ownership of objects, and are responsible for freeing them.
builder, network, parser = network_from_onnx_path("identity.onnx")
# Extend the network with an identity layer.
prev_output = network.get_output(0)
network.unmark_output(prev_output)
output = network.add_identity(prev_output).get_output(0)
output.name = "output"
network.mark_output(output)
# Create a TensorRT IBuilderConfig so that we can build the engine with FP16 enabled.
config = create_config(builder, network, fp16=True)
# We can free everything we constructed above once we're done building the engine.
# NOTE: In TensorRT 8.0, we do *not* need to use a context manager here.
with builder, network, parser, config:
engine = engine_from_network((builder, network), config)
# NOTE: In TensorRT 8.0, we do *not* need to use a context manager to free `engine`.
with engine, TrtRunner(engine) as runner:
inp_data = np.ones((1, 1, 2, 2), dtype=np.float32)
# NOTE: The runner owns the output buffers and is free to reuse them between `infer()` calls.
# Thus, if you want to store results from multiple inferences, you should use `copy.deepcopy()`.
outputs = runner.infer(feed_dict={"x": inp_data})
assert np.array_equal(outputs["output"], inp_data) # It's an identity model!
print("Inference succeeded!")
def test_calibrator_with_path_name_cache(self, identity_builder_network):
builder, network = identity_builder_network
data = [{"x": np.ones((1, 1, 2, 2), dtype=np.float32)}]
with tempfile.NamedTemporaryFile() as cache:
calibrator = Calibrator(data, cache=cache.name)
create_config = CreateConfig(int8=True, calibrator=calibrator)
with engine_from_network((builder, network), create_config):
check_file_non_empty(cache.name)
self.check_calibrator_cleanup(calibrator)
def test_calibrator_generator_data(self, identity_builder_network):
builder, network = identity_builder_network
NUM_BATCHES = 2
calibrator = Calibrator(generate_data(NUM_BATCHES))
create_config = CreateConfig(int8=True, calibrator=calibrator)
with engine_from_network((builder, network), create_config):
assert calibrator.num_batches == NUM_BATCHES
self.check_calibrator_cleanup(calibrator)
def test_multithreaded_runners_from_engine(self):
model = ONNX_MODELS["identity"]
engine = engine_from_network(NetworkFromOnnxBytes(model.loader))
with engine, TrtRunner(engine) as runner0, TrtRunner(engine) as runner1:
t1 = threading.Thread(target=model.check_runner, args=(runner0,))
t2 = threading.Thread(target=model.check_runner, args=(runner1,))
t1.start()
t2.start()
t1.join()
t2.join()
def test_calibrator_caches_without_explicit_cache(
self, identity_builder_network):
builder, network = identity_builder_network
data = [{"x": np.ones((1, 1, 2, 2), dtype=np.float32)}]
calibrator = Calibrator(data)
# First, populate the cache
create_config = CreateConfig(int8=True, calibrator=calibrator)
with engine_from_network((builder, network), create_config):
pass
# Check that the internal cache is populated
assert calibrator.read_calibration_cache()
self.check_calibrator_cleanup(calibrator)
def test_calibrator_basic(self, identity_builder_network, BaseClass):
if mod.version(trt.__version__) < mod.version(
"7.0") and BaseClass == trt.IInt8LegacyCalibrator:
pytest.skip("Bug in TRT 6 causes NaNs with legacy calibrator")
builder, network = identity_builder_network
NUM_BATCHES = 2
data = [{"x": np.ones((1, 1, 2, 2), dtype=np.float32)}] * NUM_BATCHES
calibrator = Calibrator(data, BaseClass=BaseClass)
create_config = CreateConfig(int8=True, calibrator=calibrator)
with engine_from_network((builder, network), create_config):
assert calibrator.num_batches == NUM_BATCHES
self.check_calibrator_cleanup(calibrator)
def test_calibrator_invalid_input_fails(self, identity_builder_network,
names):
builder, network = identity_builder_network
data = [{
name: np.ones((1, 1, 2, 2), dtype=np.float32)
for name in names
}]
calibrator = Calibrator(data)
create_config = CreateConfig(int8=True, calibrator=calibrator)
with pytest.raises(PolygraphyException):
with engine_from_network((builder, network), create_config):
pass
def test_calibrator_rechecks_cache_on_reset(self,
identity_builder_network):
builder, network = identity_builder_network
data = [{"x": np.ones((1, 1, 2, 2), dtype=np.float32)}]
with tempfile.NamedTemporaryFile(mode="wb+") as cache:
calibrator = Calibrator(data, cache=cache.name)
# First, populate the cache
create_config = CreateConfig(int8=True, calibrator=calibrator)
with engine_from_network((builder, network), create_config):
pass
# Ensure that now the calibrator will read from the cache when reset
calibrator.reset()
assert not calibrator.has_cached_scales
assert len(calibrator.read_calibration_cache()) == get_file_size(
cache.name)
self.check_calibrator_cleanup(calibrator)
def test_infer_overhead(self, copy_inputs, copy_outputs):
inp = np.ones(shape=(1, 2, 1024, 1024), dtype=np.float32)
dev_inp = cuda.DeviceArray(shape=inp.shape, dtype=inp.dtype).copy_from(inp)
out = np.zeros(shape=(1, 2, 1024, 1024), dtype=np.float32) # Using identity model!
dev_out = cuda.DeviceArray(shape=out.shape, dtype=out.dtype)
stream = cuda.Stream()
model = ONNX_MODELS["dynamic_identity"]
profiles = [
Profile().add("X", (1, 2, 1024, 1024), (1, 2, 1024, 1024), (1, 2, 1024, 1024)),
]
inp_name = list(model.input_metadata.keys())[0]
with engine_from_network(
network_from_onnx_bytes(model.loader), CreateConfig(profiles=profiles)
) as engine, engine.create_execution_context() as context, TrtRunner(context) as runner, dev_inp, dev_out:
# Inference outside the TrtRunner
def infer():
if copy_inputs:
dev_inp.copy_from(inp, stream=stream)
context.execute_async_v2(bindings=[dev_inp.ptr, dev_out.ptr], stream_handle=stream.ptr)
if copy_outputs:
dev_out.copy_to(out, stream=stream)
stream.synchronize()
native_time = time_func(infer)
feed_dict = {inp_name: (inp if copy_inputs else dev_inp)}
runner_time = time_func(
lambda: runner.infer(feed_dict, check_inputs=False, copy_outputs_to_host=copy_outputs)
)
# The overhead should be less than 0.5ms, or the runtime should be within 5%
print("Absolute difference: {:.5g}".format(runner_time - native_time))
print("Relative difference: {:.5g}".format(runner_time / native_time))
assert (runner_time - native_time) < 0.5e-3 or runner_time <= (native_time * 1.05)
def test_serialize_engine(self, identity_network):
with engine_from_network(identity_network) as engine:
serialized_engine = bytes_from_engine(engine)
assert isinstance(serialized_engine, bytes)
def test_shape_output(self):
model = ONNX_MODELS["reshape"]
engine = engine_from_network(NetworkFromOnnxBytes(model.loader))
with engine, TrtRunner(engine.create_execution_context) as runner:
model.check_runner(runner)
def main():
# A Profile maps each input tensor to a range of shapes.
#
# TIP: To save lines, calls to `add` can be chained:
# profile.add("input0", ...).add("input1", ...)
#
# Of course, you may alternatively write this as:
# profile.add("input0", ...)
# profile.add("input1", ...)
#
profiles = [
# The low-latency case. For best performance, min == opt == max.
Profile().add("X",
min=(1, 3, 28, 28),
opt=(1, 3, 28, 28),
max=(1, 3, 28, 28)),
# The dynamic batching case. We use `4` for the opt batch size since that's our most common case.
Profile().add("X",
min=(1, 3, 28, 28),
opt=(4, 3, 28, 28),
max=(32, 3, 28, 28)),
# The offline case. For best performance, min == opt == max.
Profile().add("X",
min=(128, 3, 28, 28),
opt=(128, 3, 28, 28),
max=(128, 3, 28, 28)),
]
# See examples/api/06_immediate_eval_api for details on immediately evaluated functional loaders like `engine_from_network`.
engine = engine_from_network(NetworkFromOnnxPath("dynamic_identity.onnx"),
config=CreateConfig(profiles=profiles))
# We'll save the engine so that we can inspect it with `inspect model`.
# This should make it easy to see how the engine bindings are laid out.
save_engine(engine, "dynamic_identity.engine")
# We'll create, but not activate, three separate runners, each with a separate context.
#
# TIP: By providing a context directly, as opposed to via a lazy loader,
# we can ensure that the runner will *not* take ownership of it.
#
low_latency = TrtRunner(engine.create_execution_context())
# NOTE: The following two lines will cause TensorRT to display errors since profile 0
# is already in use by the first execution context. We'll suppress them using G_LOGGER.verbosity().
#
with G_LOGGER.verbosity(G_LOGGER.CRITICAL):
dynamic_batching = TrtRunner(engine.create_execution_context())
offline = TrtRunner(engine.create_execution_context())
# NOTE: We could update the profile index here (e.g. `context.active_optimization_profile = 2`),
# but instead, we'll use TrtRunner's `set_profile()` API when we later activate the runner.
# Finally, we can activate the runners as we need them.
#
# NOTE: Since the context and engine are already created, the runner will only need to
# allocate input and output buffers during activation.
input_img = np.ones((1, 3, 28, 28), dtype=np.float32) # An input "image"
with low_latency:
outputs = low_latency.infer({"X": input_img})
assert np.array_equal(outputs["Y"],
input_img) # It's an identity model!
print("Low latency runner succeeded!")
# While we're serving requests online, we might decide that we need dynamic batching
# for a moment.
#
# NOTE: We're assuming that activating runners will be cheap here, so we can bring up
# the dynamic batching runner just-in-time.
#
# TIP: If activating the runner is not cheap (e.g. input/output buffers are large),
# it might be better to keep the runner active the whole time.
#
with dynamic_batching:
# NOTE: The very first time we activate this runner, we need to set
# the profile index (it's 0 by default). We need to do this *only once*.
# Alternatively, we could have set the profile index in the context directly (see above).
#
dynamic_batching.set_profile(
1
) # Use the second profile, which is intended for dynamic batching.
# We'll create fake batches by repeating our fake input image.
small_input_batch = np.repeat(input_img, 4,
axis=0) # Shape: (4, 3, 28, 28)
outputs = dynamic_batching.infer({"X": small_input_batch})
assert np.array_equal(outputs["Y"], small_input_batch)
# If we need dynamic batching again later, we can activate the runner once more.
#
# NOTE: This time, we do *not* need to set the profile.
#
with dynamic_batching:
# NOTE: We can use any shape that's in the range of the profile without
# additional setup - Polygraphy handles the details behind the scenes!
#
large_input_batch = np.repeat(input_img, 16,
axis=0) # Shape: (16, 3, 28, 28)
outputs = dynamic_batching.infer({"X": large_input_batch})
assert np.array_equal(outputs["Y"], large_input_batch)
print("Dynamic batching runner succeeded!")
with offline:
# NOTE: We must set the profile to something other than 0 or 1 since both of those
# are now in use by the `low_latency` and `dynamic_batching` runners respectively.
#
offline.set_profile(
2
) # Use the third profile, which is intended for the offline case.
large_offline_batch = np.repeat(input_img, 128,
axis=0) # Shape: (128, 3, 28, 28)
outputs = offline.infer({"X": large_offline_batch})
assert np.array_equal(outputs["Y"], large_offline_batch)
print("Offline runner succeeded!")
