def allocate_buffers(engine):
input_buffers = OrderedDict()
output_buffers = OrderedDict()
bindings = []
stream = cuda.Stream()
G_LOGGER.verbose("Using batch size: " +
str(engine.max_batch_size) +
" during buffer allocation")
for binding in engine:
shape = (engine.max_batch_size, ) + tuple(
engine.get_binding_shape(binding))
dtype = engine.get_binding_dtype(binding)
device_mem = cuda.DeviceBuffer(shape=shape,
dtype=trt.nptype(dtype))
G_LOGGER.extra_verbose("Tensor: "
"{:40} | Allocated: {:}".format(
binding, device_mem))
if engine.binding_is_input(binding):
input_buffers[binding] = TrtLegacyRunner.HostDeviceMem(
None, device_mem)
else:
host_mem = np.empty(shape=shape, dtype=trt.nptype(dtype))
output_buffers[binding] = TrtLegacyRunner.HostDeviceMem(
host_mem, device_mem)
return input_buffers, output_buffers, stream
def is_output_node(node):
# Make sure that we're not using hanging nodes as outputs - must have at least one input.
if len(node_output_map[node.name]) != 0 or len(node.input) == 0:
return False
# Tensors with no shape cannot be outputs and TensorFlow doesn't like certain ops as outputs.
EXCLUDE_OPS = [
"Switch",
"FusedBatchNorm",
"Assert",
"NextIteration",
"Enter",
"LoopCond",
"Exit",
"Print",
"Assign",
"NoOp",
"ReadVariableOp",
"VarIsInitializedOp",
"Const"
]
# Additionally, we sometimes need to exclude entire namespaces e.g. while loops.
EXCLUDE_NAMESPACES = ["while", "Assert"]
if any([ex_op in node.op for ex_op in EXCLUDE_OPS]) or any([ns in node.name for ns in EXCLUDE_NAMESPACES]):
G_LOGGER.extra_verbose("Excluding {:}, op {:} is not a valid output op or is part of an excluded namespace "
"(Note: excluded namespaces: {:})".format(node.name, node.op, EXCLUDE_NAMESPACES))
return False
return True
def infer_impl(self, feed_dict):
G_LOGGER.extra_verbose("Received feed_dict: {:}".format(feed_dict))
start = time.time()
inference_outputs = self.sess.run(self.output_names,
feed_dict=feed_dict,
options=self.run_options,
run_metadata=self.run_metadata)
end = time.time()
out_dict = OrderedDict()
for name, out in zip(self.output_names, inference_outputs):
out_dict[name] = out
self.inference_time = end - start
def generate_timeline():
from tensorflow.python.client import timeline
t1 = timeline.Timeline(self.run_metadata.step_stats)
return t1.generate_chrome_trace_format()
if self.timeline_dir is not None:
misc.lazy_write(contents=generate_timeline,
path=os.path.join(
self.timeline_dir,
"run-{:}".format(self.num_inferences)),
mode="w")
self.num_inferences += 1
return out_dict
def build_profile(builder, network, profile):
trt_profile = builder.create_optimization_profile()
unused_keys = set(profile.keys())
for idx in range(network.num_inputs):
inp = network.get_input(idx)
if inp.name in unused_keys:
unused_keys.remove(inp.name)
with G_LOGGER.verbosity(): # WAR for spam from TRT
is_shape_tensor = inp.is_shape_tensor
if is_shape_tensor:
if inp.name in profile:
shapes = profile[inp.name]
trt_profile.set_shape_input(inp.name, shapes.min, shapes.opt, shapes.max)
G_LOGGER.extra_verbose("Input shape-tensor: {:24} | Setting values to min: {:}, opt: {:}, max: {:}".format(inp.name, shapes.min, shapes.opt, shapes.max))
else:
G_LOGGER.warning("input shape-tensor: {:24} | No values provided. Assuming this is not a dynamic shape-tensor.".format(inp.name), mode=LogMode.ONCE)
elif misc.is_shape_dynamic(inp.shape):
shapes = profile[inp.name]
trt_profile.set_shape(inp.name, shapes.min, shapes.opt, shapes.max)
G_LOGGER.extra_verbose("Input tensor: {:24} | Setting shape to min: {:}, opt: {:}, max: {:}".format(inp.name, shapes.min, shapes.opt, shapes.max))
if unused_keys:
G_LOGGER.warning("Some inputs provided in the profile were unused: {:}".format(list(unused_keys)))
return check_profile(trt_profile)
def log_mismatches(mismatches):
try:
with G_LOGGER.indent():
G_LOGGER.super_verbose("Mismatched indices:\n{:}".format(np.argwhere(mismatches)))
G_LOGGER.extra_verbose("Runner: {:40} | Mismatched values:\n{:}".format(iter_result0.runner_name, out0[mismatches]))
G_LOGGER.extra_verbose("Runner: {:40} | Mismatched values:\n{:}".format(iter_result1.runner_name, out1[mismatches]))
except:
G_LOGGER.warning("Failing to log mismatches - this may be because the outputs are of different shapes")
def is_not_nan(output):
nans = np.isnan(output)
if np.any(nans):
G_LOGGER.error("Encountered one or more NaNs")
G_LOGGER.error(
"Note: Use -vv or set logging verbosity to EXTRA_VERBOSE to display locations of NaNs",
mode=LogMode.ONCE)
G_LOGGER.extra_verbose("Note: NaNs at:\n{:}".format(nans))
return False
return True
def from_engine(engine):
buffers = Buffers()
bindings_per_profile = trt_util.get_bindings_per_profile(engine)
for idx in range(bindings_per_profile):
binding = engine[idx]
dtype = trt.nptype(engine.get_binding_dtype(binding))
buffers.device_buffers[binding] = cuda.DeviceBuffer(dtype=dtype)
if not engine.binding_is_input(binding):
buffers.outputs[binding] = np.empty(shape=tuple(), dtype=dtype)
G_LOGGER.extra_verbose("Created device buffers: {:}".format(
buffers.device_buffers))
return buffers
def is_finite(output):
non_finite = np.logical_not(np.isfinite(output))
if np.any(non_finite):
G_LOGGER.error("Encountered one or more non-finite values")
G_LOGGER.error(
"Note: Use -vv or set logging verbosity to EXTRA_VERBOSE to display non-finite values",
mode=LogMode.ONCE)
G_LOGGER.extra_verbose(
"Note: non-finite values at:\n{:}".format(non_finite))
G_LOGGER.extra_verbose("Note: non-finite values:\n{:}".format(
output[non_finite]))
return False
return True
def __call__(self):
"""
Creates a TensorFlow session.
Returns:
tf.Session: The TensorFlow session.
"""
config, _ = misc.try_call(self.config)
(graph, output_names), _ = misc.try_call(self.graph)
with graph.as_default() as graph, tf.compat.v1.Session(
graph=graph, config=config).as_default() as sess:
G_LOGGER.verbose(
"Using TensorFlow outputs: {:}".format(output_names))
G_LOGGER.extra_verbose(
"Initializing variables in TensorFlow Graph")
sess.run(tf.compat.v1.initializers.global_variables())
return sess, output_names
def make_buffers(engine):
"""
Creates empty host and device buffers for the specified engine.
Always uses binding names from Profile 0.
"""
device_buffers = OrderedDict()
host_output_buffers = OrderedDict()
for idx in range(trt_util.get_bindings_per_profile(engine)):
binding = engine[idx]
dtype = trt.nptype(engine.get_binding_dtype(binding))
device_buffers[binding] = cuda.DeviceBuffer(dtype=dtype)
if not engine.binding_is_input(binding):
host_output_buffers[binding] = np.empty(shape=tuple(),
dtype=dtype)
G_LOGGER.extra_verbose(
"Created device buffers: {:}".format(device_buffers))
return device_buffers, host_output_buffers
def execute_runner(runner, loader_cache):
with runner as active_runner:
input_metadata = active_runner.get_input_metadata()
G_LOGGER.info("Runner: {:40} | Input Metadata: {:}".format(
active_runner.name, input_metadata),
mode=LogMode.ONCE)
# DataLoaderCache will ensure that the feed_dict does not contain any extra entries
# based on the provided input_metadata.
loader_cache.set_input_metadata(input_metadata)
if warm_up:
G_LOGGER.start(
"Runner: {:40} | Running {:} warm-up runs".format(
active_runner.name, warm_up))
try:
feed_dict = loader_cache[0]
except IndexError:
G_LOGGER.warning(
"{:} warm-up runs were requested, but data loader did not supply any data. "
"Skipping warm-up runs".format(warm_up))
else:
G_LOGGER.ultra_verbose(
"Warm-up Input Buffers:\n{:}".format(
misc.indent_block(feed_dict)))
# First do a few warm-up runs, and don't time them.
for i in range(warm_up):
active_runner.infer(feed_dict=feed_dict)
# Then, actual iterations.
index = 0
iteration_results = []
output_metadata = TensorMetadata()
for index, feed_dict in enumerate(loader_cache):
G_LOGGER.extra_verbose(
lambda: "Runner: {:40} | Feeding inputs:\n{:}".format(
active_runner.name, misc.indent_block(feed_dict)))
outputs = active_runner.infer(feed_dict=feed_dict)
runtime = active_runner.last_inference_time()
# Without a deep copy here, outputs will always reference the output of the last run
iteration_results.append(
IterationResult(outputs=copy.deepcopy(outputs),
runtime=runtime,
runner_name=active_runner.name))
if index == 0:
for name, out in outputs.items():
output_metadata.add(name, out.dtype, out.shape)
G_LOGGER.info(
"Runner: {:40} | Output Metadata: {:}".format(
active_runner.name, output_metadata),
mode=LogMode.ONCE)
G_LOGGER.extra_verbose(
lambda:
"Runner: {:40} | Inference Time: {:.3f} ms | Received outputs:\n{:}"
.format(active_runner.name, runtime * 1000.0,
misc.indent_block(outputs)))
G_LOGGER.finish(
"Runner: {:40} | Completed {:} iterations.".format(
active_runner.name, index + 1))
return iteration_results
def run(runners,
data_loader=None,
warm_up=None,
use_subprocess=None,
subprocess_timeout=None,
subprocess_polling_interval=None,
save_inputs_path=None):
"""
Runs the supplied runners sequentially.
Args:
runners (List[BaseRunner]):
A list of runners to run.
data_loader (Generator -> OrderedDict[str, numpy.ndarray]):
A generator or iterable that yields a dictionary that maps input names to input numpy buffers.
In the simplest case, this can be a `List[Dict[str, numpy.ndarray]]` .
In case you don't know details about the inputs ahead of time, you can access the
`input_metadata` property in your data loader, which will be set to an `TensorMetadata`
instance by this function.
Note that this does not work for generators or lists.
The number of iterations run by this function is controlled by the number of items supplied
by the data loader.
Defaults to an instance of `DataLoader`.
warm_up (int):
The number of warm up runs to perform for each runner before timing.
Defaults to 0.
use_subprocess (bool):
Whether each runner should be run in a subprocess. This allows each runner to have exclusive
access to the GPU. When using a subprocess, runners and loaders will never be modified.
subprocess_timeout (int):
The timeout before a subprocess is killed automatically. This is useful for handling processes
that never terminate. A value of None disables the timeout. Defaults to None.
subprocess_polling_interval (int):
The polling interval, in seconds, for checking whether a subprocess has completed or crashed.
In rare cases, omitting this parameter when subprocesses are enabled may cause this function
to hang indefinitely if the subprocess crashes.
A value of 0 disables polling. Defaults to 30 seconds.
save_inputs_path (str):
[EXPERIMENTAL] Path at which to save inputs used during inference. This will include all inputs generated by
the provided data_loader, and will be saved as a pickled List[Dict[str, numpy.ndarray]].
Returns:
RunResults:
A mapping of runner names to the results of their inference.
The ordering of `runners` is preserved in this mapping.
"""
warm_up = misc.default_value(warm_up, 0)
data_loader = misc.default_value(data_loader, DataLoader())
use_subprocess = misc.default_value(use_subprocess, False)
subprocess_polling_interval = misc.default_value(
subprocess_polling_interval, 30)
loader_cache = DataLoaderCache(data_loader,
save_inputs_path=save_inputs_path)
def execute_runner(runner, loader_cache):
with runner as active_runner:
input_metadata = active_runner.get_input_metadata()
G_LOGGER.info("Runner: {:40} | Input Metadata: {:}".format(
active_runner.name, input_metadata),
mode=LogMode.ONCE)
# DataLoaderCache will ensure that the feed_dict does not contain any extra entries
# based on the provided input_metadata.
loader_cache.set_input_metadata(input_metadata)
if warm_up:
G_LOGGER.start(
"Runner: {:40} | Running {:} warm-up runs".format(
active_runner.name, warm_up))
try:
feed_dict = loader_cache[0]
except IndexError:
G_LOGGER.warning(
"{:} warm-up runs were requested, but data loader did not supply any data. "
"Skipping warm-up runs".format(warm_up))
else:
G_LOGGER.ultra_verbose(
"Warm-up Input Buffers:\n{:}".format(
misc.indent_block(feed_dict)))
# First do a few warm-up runs, and don't time them.
for i in range(warm_up):
active_runner.infer(feed_dict=feed_dict)
# Then, actual iterations.
index = 0
iteration_results = []
output_metadata = TensorMetadata()
for index, feed_dict in enumerate(loader_cache):
G_LOGGER.extra_verbose(
lambda: "Runner: {:40} | Feeding inputs:\n{:}".format(
active_runner.name, misc.indent_block(feed_dict)))
outputs = active_runner.infer(feed_dict=feed_dict)
runtime = active_runner.last_inference_time()
# Without a deep copy here, outputs will always reference the output of the last run
iteration_results.append(
IterationResult(outputs=copy.deepcopy(outputs),
runtime=runtime,
runner_name=active_runner.name))
if index == 0:
for name, out in outputs.items():
output_metadata.add(name, out.dtype, out.shape)
G_LOGGER.info(
"Runner: {:40} | Output Metadata: {:}".format(
active_runner.name, output_metadata),
mode=LogMode.ONCE)
G_LOGGER.extra_verbose(
lambda:
"Runner: {:40} | Inference Time: {:.3f} ms | Received outputs:\n{:}"
.format(active_runner.name, runtime * 1000.0,
misc.indent_block(outputs)))
G_LOGGER.finish(
"Runner: {:40} | Completed {:} iterations.".format(
active_runner.name, index + 1))
return iteration_results
# Wraps execute_runner to use a queue.
def execute_runner_with_queue(runner_queue, runner, loader_cache):
iteration_results = None
try:
iteration_results = execute_runner(runner, loader_cache)
except:
# Cannot send the exception back, as it is not necessarily pickleable
import traceback
G_LOGGER.error(traceback.format_exc())
misc.try_send_on_queue(runner_queue, iteration_results)
# After finishing, send the updated loader_cache back.
misc.try_send_on_queue(runner_queue, loader_cache)
# Do all inferences in one loop, then comparisons at a later stage.
# We run each runner in a separate process so that we can provide exclusive GPU access for each runner.
run_results = RunResults()
for runner in runners:
G_LOGGER.start(
"Runner: {:40} | Activating and starting inference".format(
runner.name))
if use_subprocess:
runner_queue = Queue()
process = Process(target=execute_runner_with_queue,
args=(runner_queue, runner, loader_cache))
process.start()
# If a subprocess hangs in a certain way, then process.join could block forever. Hence,
# we need to keep polling the process to make sure it really is alive.
iteration_results = None
while process.is_alive() and iteration_results is None:
try:
iteration_results = misc.try_receive_on_queue(
runner_queue,
timeout=subprocess_polling_interval / 2)
# Receive updated loader cache, or fall back if it could not be sent.
loader_cache = misc.try_receive_on_queue(
runner_queue,
timeout=subprocess_polling_interval / 2)
except queue.Empty:
G_LOGGER.extra_verbose(
"Polled subprocess - still running")
try:
assert iteration_results is not None
run_results.append((runner.name, iteration_results))
process.join(subprocess_timeout)
except:
G_LOGGER.critical(
"Runner: {:40} | Terminated prematurely. Check the exception logged above. "
"If there is no exception logged above, make sure not to use the --use-subprocess "
"flag or set use_subprocess=False in Comparator.run()."
.format(runner.name))
finally:
process.terminate()
if loader_cache is None:
G_LOGGER.critical(
"Could not send data loader cache to runner subprocess. Please try disabling subprocesses "
"by removing the --use-subprocess flag, or setting use_subprocess=False in Comparator.run()"
)
else:
run_results.append(
(runner.name, execute_runner(runner, loader_cache)))
G_LOGGER.verbose("Successfully ran: {:}".format(
[r.name for r in runners]))
return run_results
def compare_accuracy(run_results,
fail_fast=False,
comparisons=None,
compare_func=None):
"""
Args:
run_results (RunResults): The result of Comparator.run()
fail_fast (bool): Whether to exit after the first failure
comparisons (List[Tuple[int, int]]):
Comparisons to perform, specified by runner indexes. For example, [(0, 1), (1, 2)]
would compare the first runner with the second, and the second with the third.
By default, this compares each result to the subsequent one.
compare_func (Callable(IterationResult, IterationResult) -> OrderedDict[str, bool]):
A function that takes in two IterationResults, and returns a dictionary that maps output
names to a boolean (or anything convertible to a boolean) indicating whether outputs matched.
The order of arguments to this function is guaranteed to be the same as the ordering of the
tuples contained in `comparisons`.
Returns:
AccuracyResult:
A summary of the results of the comparisons. The order of the keys (i.e. runner pairs) is
guaranteed to be the same as the order of `comparisons`. For more details, see the AccuracyResult
docstring (e.g. help(AccuracyResult)).
"""
def find_mismatched(match_dict):
return [
name for name, matched in match_dict.items()
if not bool(matched)
]
compare_func = misc.default_value(compare_func,
CompareFunc.basic_compare_func())
comparisons = misc.default_value(
comparisons, Comparator.default_comparisons(run_results))
accuracy_result = AccuracyResult()
for runner0_index, runner1_index in comparisons:
(runner0_name, results0), (
runner1_name, results1
) = run_results[runner0_index], run_results[runner1_index]
G_LOGGER.start("Accuracy Comparison | {:} vs. {:}".format(
runner0_name, runner1_name))
with G_LOGGER.indent():
runner_pair = (runner0_name, runner1_name)
accuracy_result[runner_pair] = []
num_iters = min(len(results0), len(results1))
for iteration, (result0,
result1) in enumerate(zip(results0, results1)):
if num_iters > 1:
G_LOGGER.info("Iteration: {:}".format(iteration))
with contextlib.ExitStack() as stack:
if num_iters > 1:
stack.enter_context(G_LOGGER.indent())
iteration_match_dict = compare_func(result0, result1)
accuracy_result[runner_pair].append(
iteration_match_dict)
mismatched_outputs = find_mismatched(iteration_match_dict)
if fail_fast and mismatched_outputs:
return accuracy_result
G_LOGGER.extra_verbose(
"Finished comparing {:} with {:}".format(
runner0_name,
runner1_name,
))
passed, failed, total = accuracy_result.stats(runner_pair)
pass_rate = accuracy_result.percentage(runner_pair) * 100.0
if num_iters > 1 or len(comparisons) > 1:
msg = "Accuracy Summary | {:} vs. {:} | Passed: {:}/{:} iterations | Pass Rate: {:}%".format(
runner0_name, runner1_name, passed, total, pass_rate)
if passed == total:
G_LOGGER.finish(msg)
else:
G_LOGGER.error(msg)
return accuracy_result
def get_output_metadata(graph, layerwise=False):
graphdef = graph.as_graph_def()
node_output_map = map_node_outputs(graphdef)
def is_output_node(node):
# Make sure that we're not using hanging nodes as outputs - must have at least one input.
if len(node_output_map[node.name]) != 0 or len(node.input) == 0:
return False
# Tensors with no shape cannot be outputs and TensorFlow doesn't like certain ops as outputs.
EXCLUDE_OPS = [
"Switch",
"FusedBatchNorm",
"Assert",
"NextIteration",
"Enter",
"LoopCond",
"Exit",
"Print",
"Assign",
"NoOp",
"ReadVariableOp",
"VarIsInitializedOp",
"Const"
]
# Additionally, we sometimes need to exclude entire namespaces e.g. while loops.
EXCLUDE_NAMESPACES = ["while", "Assert"]
if any([ex_op in node.op for ex_op in EXCLUDE_OPS]) or any([ns in node.name for ns in EXCLUDE_NAMESPACES]):
G_LOGGER.extra_verbose("Excluding {:}, op {:} is not a valid output op or is part of an excluded namespace "
"(Note: excluded namespaces: {:})".format(node.name, node.op, EXCLUDE_NAMESPACES))
return False
return True
# For layerwise mode, every layer becomes an output.
if layerwise:
output_nodes = list(graphdef.node)
G_LOGGER.verbose("Running in layerwise mode. Marking {:} layers as potential outputs".format(len(output_nodes)))
else:
output_nodes = [node for node in graphdef.node if is_output_node(node)]
G_LOGGER.extra_verbose("Found likely output nodes: {:}".format(output_nodes))
output_tensors = []
for node in output_nodes:
tensor_name = node.name + ":0"
try:
tensor = graph.get_tensor_by_name(tensor_name)
output_tensors.append(tensor)
except KeyError:
G_LOGGER.warning("Could not import: {:}. Skipping.".format(tensor_name))
if len(output_tensors) != len(output_nodes):
G_LOGGER.warning("Excluded {:} ops that don't seem like outputs. Use -vv/--super-verbose, or set "
"logging verbosity to EXTRA_VERBOSE to view them.".format(len(output_nodes) - len(output_tensors)))
G_LOGGER.extra_verbose("Found output op types in graph: {:}".format(set([tensor.op.type for tensor in output_tensors])))
G_LOGGER.verbose("Retrieved TensorFlow output_tensors: {:}".format(output_tensors))
return get_tensor_metadata(output_tensors)
def check_outputs_match(out0, out0_name, out1, out1_name, per_out_rtol, per_out_atol):
def compute_max(buffer):
if misc.is_empty_shape(buffer.shape):
return 0
return np.amax(buffer)
# Returns index of max value
def compute_argmax(buffer):
if misc.is_empty_shape(buffer.shape):
return 0
return np.unravel_index(np.argmax(buffer), buffer.shape)
def compute_min(buffer):
if misc.is_empty_shape(buffer.shape):
return 0
return np.amin(buffer)
# Returns index of min value
def compute_argmin(buffer):
if misc.is_empty_shape(buffer.shape):
return 0
return np.unravel_index(np.argmin(buffer), buffer.shape)
def compute_mean(buffer):
if misc.is_empty_shape(buffer.shape):
return 0
return np.mean(buffer)
def compute_required():
# The purpose of this function is to determine the minimum tolerances such that
# the outputs would be considered a match.
# The NumPy formula for np.isclose is absolute(out0 - out1) <= (per_out_atol + per_out_rtol * absolute(out1))
# So, for both absolute/relative tolerance, given either one,
# we can compute the required value for the other:
# per_out_atol = absolute(out0 - out1)
# atol_if_rtol = absolute(out0 - out1) - per_out_rtol * absolute(out1)
# per_out_rtol = (absolute(out0 - out1) - per_out_atol) / absolute(out1)
if np.issubdtype(out0.dtype, np.bool_) and np.issubdtype(out1.dtype, np.bool_):
absdiff = np.logical_xor(out0, out1)
else:
absdiff = np.abs(out0 - out1)
absout1 = np.abs(out1)
max_absdiff = max(compute_max(absdiff), 0.0)
required_atol_if_rtol = max(compute_max(absdiff - per_out_rtol * absout1), 0.0)
# Suppress divide by 0 warnings
with np.testing.suppress_warnings() as sup:
sup.filter(RuntimeWarning)
reldiff = np.maximum(absdiff - per_out_atol, 0.0) / absout1
max_reldiff = max(compute_max(reldiff), 0.0)
return max_absdiff, required_atol_if_rtol, max_reldiff, compute_mean(absdiff), compute_mean(reldiff)
def log_mismatches(mismatches):
try:
with G_LOGGER.indent():
G_LOGGER.super_verbose("Mismatched indices:\n{:}".format(np.argwhere(mismatches)))
G_LOGGER.extra_verbose("Runner: {:40} | Mismatched values:\n{:}".format(iter_result0.runner_name, out0[mismatches]))
G_LOGGER.extra_verbose("Runner: {:40} | Mismatched values:\n{:}".format(iter_result1.runner_name, out1[mismatches]))
except:
G_LOGGER.warning("Failing to log mismatches - this may be because the outputs are of different shapes")
try:
mismatches = np.logical_not(np.isclose(output0, output1, rtol=per_out_rtol, atol=per_out_atol))
except Exception as err:
G_LOGGER.warning("Failed to compare outputs with:\n{:}\nSkipping".format(err))
return False
G_LOGGER.super_verbose("Runner: {:40} | Output: {:} (dtype={:}, shape={:}):\n{:}".format(
iter_result0.runner_name, out0_name, out0.dtype, out0.shape, misc.indent_block(out0)))
G_LOGGER.super_verbose("Runner: {:40} | Output: {:} (dtype={:}, shape={:}):\n{:}".format(
iter_result1.runner_name, out1_name, out1.dtype, out1.shape, misc.indent_block(out1)))
failed = np.any(mismatches)
try:
max_absdiff, required_atol_if_rtol, max_reldiff, mean_absdiff, mean_reldiff = compute_required()
except Exception as err:
max_absdiff, required_atol_if_rtol, max_reldiff, mean_absdiff, mean_reldiff = None, None, None, None, None
G_LOGGER.warning("Could not determine required tolerances due to an error:\n{:}".format(err))
log_msg = ""
else:
log_msg = "Required tolerances: [atol={:.5g}] OR [rtol={:.5g}, atol={:.5g}] OR [rtol={:.5g}, atol={:.5g}] | Mean Error: Absolute={:.5g}, Relative={:.5g}\n".format(
max_absdiff, per_out_rtol, required_atol_if_rtol, max_reldiff, per_out_atol, mean_absdiff, mean_reldiff)
log_msg += "Runner: {:40} | Stats: mean={:.5g}, min={:.5g} at {:}, max={:.5g} at {:}\n".format(
iter_result0.runner_name, compute_mean(out0), compute_min(out0), compute_argmin(out0), compute_max(out0), compute_argmax(out0))
log_msg += "Runner: {:40} | Stats: mean={:.5g}, min={:.5g} at {:}, max={:.5g} at {:}\n".format(
iter_result1.runner_name, compute_mean(out1), compute_min(out1), compute_argmin(out1), compute_max(out1), compute_argmax(out1))
G_LOGGER.info(log_msg)
if failed:
log_mismatches(mismatches)
G_LOGGER.error("FAILED | Difference exceeds tolerance (rtol={:}, atol={:})".format(per_out_rtol, per_out_atol))
else:
G_LOGGER.finish("PASSED | Difference is within tolerance (rtol={:}, atol={:})".format(per_out_rtol, per_out_atol))
G_LOGGER.extra_verbose("Finished comparing: '{:}' (dtype={:}, shape={:}) [{:}] and '{:}' (dtype={:}, shape={:}) [{:}]"
.format(out0_name, out0.dtype, out0.shape, iter_result0.runner_name, out1_name, out1.dtype, out1.shape, iter_result1.runner_name))
return OutputCompareResult(not failed, max_absdiff, max_reldiff)
def compare_output(iter_result0, iter_result1):
"""
Compare the outputs of two runners from a single iteration.
This function will always iterate over the output names of the first IterationResult,
and attempt to find corresponding output names in the second.
If no corresponding output name is found, the output is skipped.
If all output names are skipped, then this function raises an error.
Args:
iter_result0 (IterationResult): The result of the first runner.
iter_result1 (IterationResult): The result of the second runner.
Returns:
OrderedDict[str, OutputCompareResult]:
The name of the outputs compared, derived from the first IterationResult,
and whether they matched. If an output name is not found, it is omitted from this dictionary.
Raises:
PolygraphyException: If all output names are skipped, and thus no outputs are compared.
"""
# Returns whether the outputs match
def check_outputs_match(out0, out0_name, out1, out1_name, per_out_rtol, per_out_atol):
def compute_max(buffer):
if misc.is_empty_shape(buffer.shape):
return 0
return np.amax(buffer)
# Returns index of max value
def compute_argmax(buffer):
if misc.is_empty_shape(buffer.shape):
return 0
return np.unravel_index(np.argmax(buffer), buffer.shape)
def compute_min(buffer):
if misc.is_empty_shape(buffer.shape):
return 0
return np.amin(buffer)
# Returns index of min value
def compute_argmin(buffer):
if misc.is_empty_shape(buffer.shape):
return 0
return np.unravel_index(np.argmin(buffer), buffer.shape)
def compute_mean(buffer):
if misc.is_empty_shape(buffer.shape):
return 0
return np.mean(buffer)
def compute_required():
# The purpose of this function is to determine the minimum tolerances such that
# the outputs would be considered a match.
# The NumPy formula for np.isclose is absolute(out0 - out1) <= (per_out_atol + per_out_rtol * absolute(out1))
# So, for both absolute/relative tolerance, given either one,
# we can compute the required value for the other:
# per_out_atol = absolute(out0 - out1)
# atol_if_rtol = absolute(out0 - out1) - per_out_rtol * absolute(out1)
# per_out_rtol = (absolute(out0 - out1) - per_out_atol) / absolute(out1)
if np.issubdtype(out0.dtype, np.bool_) and np.issubdtype(out1.dtype, np.bool_):
absdiff = np.logical_xor(out0, out1)
else:
absdiff = np.abs(out0 - out1)
absout1 = np.abs(out1)
max_absdiff = max(compute_max(absdiff), 0.0)
required_atol_if_rtol = max(compute_max(absdiff - per_out_rtol * absout1), 0.0)
# Suppress divide by 0 warnings
with np.testing.suppress_warnings() as sup:
sup.filter(RuntimeWarning)
reldiff = np.maximum(absdiff - per_out_atol, 0.0) / absout1
max_reldiff = max(compute_max(reldiff), 0.0)
return max_absdiff, required_atol_if_rtol, max_reldiff, compute_mean(absdiff), compute_mean(reldiff)
def log_mismatches(mismatches):
try:
with G_LOGGER.indent():
G_LOGGER.super_verbose("Mismatched indices:\n{:}".format(np.argwhere(mismatches)))
G_LOGGER.extra_verbose("Runner: {:40} | Mismatched values:\n{:}".format(iter_result0.runner_name, out0[mismatches]))
G_LOGGER.extra_verbose("Runner: {:40} | Mismatched values:\n{:}".format(iter_result1.runner_name, out1[mismatches]))
except:
G_LOGGER.warning("Failing to log mismatches - this may be because the outputs are of different shapes")
try:
mismatches = np.logical_not(np.isclose(output0, output1, rtol=per_out_rtol, atol=per_out_atol))
except Exception as err:
G_LOGGER.warning("Failed to compare outputs with:\n{:}\nSkipping".format(err))
return False
G_LOGGER.super_verbose("Runner: {:40} | Output: {:} (dtype={:}, shape={:}):\n{:}".format(
iter_result0.runner_name, out0_name, out0.dtype, out0.shape, misc.indent_block(out0)))
G_LOGGER.super_verbose("Runner: {:40} | Output: {:} (dtype={:}, shape={:}):\n{:}".format(
iter_result1.runner_name, out1_name, out1.dtype, out1.shape, misc.indent_block(out1)))
failed = np.any(mismatches)
try:
max_absdiff, required_atol_if_rtol, max_reldiff, mean_absdiff, mean_reldiff = compute_required()
except Exception as err:
max_absdiff, required_atol_if_rtol, max_reldiff, mean_absdiff, mean_reldiff = None, None, None, None, None
G_LOGGER.warning("Could not determine required tolerances due to an error:\n{:}".format(err))
log_msg = ""
else:
log_msg = "Required tolerances: [atol={:.5g}] OR [rtol={:.5g}, atol={:.5g}] OR [rtol={:.5g}, atol={:.5g}] | Mean Error: Absolute={:.5g}, Relative={:.5g}\n".format(
max_absdiff, per_out_rtol, required_atol_if_rtol, max_reldiff, per_out_atol, mean_absdiff, mean_reldiff)
log_msg += "Runner: {:40} | Stats: mean={:.5g}, min={:.5g} at {:}, max={:.5g} at {:}\n".format(
iter_result0.runner_name, compute_mean(out0), compute_min(out0), compute_argmin(out0), compute_max(out0), compute_argmax(out0))
log_msg += "Runner: {:40} | Stats: mean={:.5g}, min={:.5g} at {:}, max={:.5g} at {:}\n".format(
iter_result1.runner_name, compute_mean(out1), compute_min(out1), compute_argmin(out1), compute_max(out1), compute_argmax(out1))
G_LOGGER.info(log_msg)
if failed:
log_mismatches(mismatches)
G_LOGGER.error("FAILED | Difference exceeds tolerance (rtol={:}, atol={:})".format(per_out_rtol, per_out_atol))
else:
G_LOGGER.finish("PASSED | Difference is within tolerance (rtol={:}, atol={:})".format(per_out_rtol, per_out_atol))
G_LOGGER.extra_verbose("Finished comparing: '{:}' (dtype={:}, shape={:}) [{:}] and '{:}' (dtype={:}, shape={:}) [{:}]"
.format(out0_name, out0.dtype, out0.shape, iter_result0.runner_name, out1_name, out1.dtype, out1.shape, iter_result1.runner_name))
return OutputCompareResult(not failed, max_absdiff, max_reldiff)
output_status = OrderedDict() # OrderedDict[str, bool] Maps output names to whether they matched.
if not check_shapes:
G_LOGGER.info("Strict shape checking disabled. Will attempt to match output shapes before comparisons")
def default_find_output_func(output_name, index, iter_result):
found_name = misc.find_in_dict(output_name, iter_result, index)
if found_name is None:
return None
elif found_name != output_name:
exact_match = misc.find_in_dict(found_name, iter_result0)
if exact_match == found_name:
G_LOGGER.verbose("Will not compare {:} with {:}, since the former already has an exact match: {:}".format(
found_name, output_name, exact_match))
return None # If the found output is being compared against another output already, skip this non-exact match
G_LOGGER.warning("Output names did not match exactly. Assuming {:} output: {:} "
"corresponds to output: {:}".format(
iter_result.runner_name, found_name, output_name))
return [found_name]
nonlocal find_output_func
find_output_func = misc.default_value(find_output_func, default_find_output_func)
for index, (out0_name, output0) in enumerate(iter_result0.items()):
out1_names = misc.default_value(find_output_func(out0_name, index, iter_result1), [])
if len(out1_names) > 1:
G_LOGGER.info("Will attempt to compare output: '{:}' [{:}] with multiple outputs: '{:}' [{:}]".format(
out0_name, iter_result0.runner_name, list(out1_names), iter_result1.runner_name))
for out1_name in out1_names:
if out1_name is None or out1_name not in iter_result1:
G_LOGGER.warning("For output: '{:}' [{:}], skipping corresponding output: '{:}' [{:}], "
"since the output was not found".format(out0_name, iter_result0.runner_name,
out1_name, iter_result1.runner_name))
continue
output1 = iter_result1[out1_name]
G_LOGGER.start("Comparing Output: '{:}' (dtype={:}, shape={:}) with '{:}' (dtype={:}, shape={:})".format(
out0_name, output0.dtype, output0.shape, out1_name, output1.dtype, output1.shape))
G_LOGGER.extra_verbose("Note: Comparing {:} vs. {:}".format(iter_result0.runner_name, iter_result1.runner_name))
def get_tol(tol_dict):
if isinstance(tol_dict, numbers.Number):
return tol_dict
if out0_name in tol_dict:
return tol_dict[out0_name]
elif "" in tol_dict:
return tol_dict[""]
G_LOGGER.critical("Could not find a tolerance for output: '{:}' in the provided tolerance map: {:}.\n"
"Note: Use a key of `""` in the map to specify a default tolerance.".format(out0_name, tol_dict))
with G_LOGGER.indent():
if check_shapes and output0.shape != output1.shape:
G_LOGGER.error("Will not compare outputs of different shapes. Note: Output shapes are "
"{:} and {:}.".format(output0.shape, output1.shape))
G_LOGGER.error("Note: Use --no-strict-shape-checking or set check_shapes=False to "
"attempt to compare values anyway.", mode=LogMode.ONCE)
outputs_match = False
else:
output1 = misc.try_match_shape(output1, output0.shape)
output0 = output0.reshape(output1.shape)
outputs_match = check_outputs_match(output0, out0_name, output1, out1_name,
per_out_rtol=get_tol(rtol), per_out_atol=get_tol(atol))
output_status[out0_name] = outputs_match
if fail_fast and not outputs_match:
return output_status
mismatched_output_names = [name for name, matched in output_status.items() if not matched]
if mismatched_output_names:
G_LOGGER.error("FAILED | Mismatched outputs: {:}".format(mismatched_output_names))
# This is useful for catching cases were Polygraphy does something wrong with the runner output buffers
if not output_status and (bool(iter_result0.keys()) or bool(iter_result1.keys())):
r0_name = iter_result0.runner_name
r0_outs = list(iter_result0.keys())
r1_name = iter_result1.runner_name
r1_outs = list(iter_result1.keys())
G_LOGGER.critical("All outputs were skipped, no common outputs found! Note:\n{:} outputs: "
"{:}\n{:} outputs: {:}".format(r0_name, r0_outs, r1_name, r1_outs))
return output_status
