def test_add_pre_and_postproc(self, topology, wbits, abits):
prev_chkpt_name = get_checkpoint_name(topology, wbits, abits, "import_and_tidy")
model = load_test_checkpoint_or_skip(prev_chkpt_name)
global_inp_name = model.graph.input[0].name
ishape = model.get_tensor_shape(global_inp_name)
# preprocessing: torchvision's ToTensor divides uint8 inputs by 255
totensor_pyt = ToTensor()
chkpt_preproc_name = get_checkpoint_name(topology, wbits, abits, "preproc")
bo.export_finn_onnx(totensor_pyt, ishape, chkpt_preproc_name)
assert os.path.isfile(chkpt_preproc_name)
# join preprocessing and core model
pre_model = ModelWrapper(chkpt_preproc_name)
model = model.transform(MergeONNXModels(pre_model))
# add input quantization annotation: UINT8 for all BNN-PYNQ models
global_inp_name = model.graph.input[0].name
model.set_tensor_datatype(global_inp_name, DataType.UINT8)
# postprocessing: insert Top-1 node at the end
model = model.transform(InsertTopK(k=1))
chkpt_name = get_checkpoint_name(topology, wbits, abits, "pre_post")
# tidy-up again
model = model.transform(InferShapes())
model = model.transform(FoldConstants())
model = model.transform(GiveUniqueNodeNames())
model = model.transform(GiveReadableTensorNames())
model = model.transform(InferDataTypes())
model = model.transform(RemoveStaticGraphInputs())
model.save(chkpt_name)
assert os.path.isfile(chkpt_name)
def test_brevitas_fc_onnx_export_and_exec(size, wbits, abits, pretrained):
if size == "LFC" and wbits == 2 and abits == 2:
pytest.skip(f"No LFC_{MAX_WBITS}W{MAX_ABITS}A present.")
if wbits > abits:
pytest.skip("No wbits > abits cases.")
nname = f"{size}_{wbits}W{abits}A"
finn_onnx = nname + ".onnx"
fc, _ = model_with_cfg(nname.lower(), pretrained=pretrained)
FINNManager.export_onnx(fc, FC_INPUT_SIZE, finn_onnx)
model = ModelWrapper(finn_onnx)
model = model.transform(GiveUniqueNodeNames())
model = model.transform(DoubleToSingleFloat())
model = model.transform(InferShapes())
model = model.transform(FoldConstants())
model = model.transform(RemoveStaticGraphInputs())
# load a random test vector
input_tensor = np.random.uniform(MIN_INP_VAL,
MAX_INP_VAL,
size=FC_INPUT_SIZE).astype(np.float32)
# run using FINN-based execution
input_dict = {"0": input_tensor}
output_dict = oxe.execute_onnx(model, input_dict)
produced = output_dict[list(output_dict.keys())[0]]
# run using PyTorch/Brevitas
input_tensor = torch.from_numpy(input_tensor).float()
# do forward pass in PyTorch/Brevitas
expected = fc.forward(input_tensor).detach().numpy()
assert np.isclose(produced, expected, atol=ATOL).all()
def test_brevitas_cnv_export_exec(wbits, abits):
if wbits > abits:
pytest.skip("No wbits > abits cases at the moment")
cnv = get_test_model_trained("CNV", wbits, abits)
bo.export_finn_onnx(cnv, (1, 3, 32, 32), export_onnx_path)
model = ModelWrapper(export_onnx_path)
model = model.transform(GiveUniqueNodeNames())
model = model.transform(InferShapes())
model = model.transform(FoldConstants())
model = model.transform(RemoveStaticGraphInputs())
assert len(model.graph.input) == 1
assert len(model.graph.output) == 1
fn = pk.resource_filename("finn", "data/cifar10/cifar10-test-data-class3.npz")
input_tensor = np.load(fn)["arr_0"].astype(np.float32)
input_tensor = input_tensor / 255
assert input_tensor.shape == (1, 3, 32, 32)
# run using FINN-based execution
input_dict = {model.graph.input[0].name: input_tensor}
output_dict = oxe.execute_onnx(model, input_dict, True)
produced = output_dict[model.graph.output[0].name]
# do forward pass in PyTorch/Brevitas
input_tensor = torch.from_numpy(input_tensor).float()
expected = cnv.forward(input_tensor).detach().numpy()
assert np.isclose(produced, expected, atol=1e-3).all()
assert np.argmax(produced) == 3
os.remove(export_onnx_path)
def test_brevitas_fc_onnx_export_and_exec(size, wbits, abits):
if size == "LFC" and wbits == 2 and abits == 2:
pytest.skip("No LFC-w2a2 present at the moment")
if wbits > abits:
pytest.skip("No wbits > abits cases at the moment")
nname = "%s_%dW%dA" % (size, wbits, abits)
finn_onnx = export_onnx_path + "/%s.onnx" % nname
fc = get_test_model_trained(size, wbits, abits)
bo.export_finn_onnx(fc, (1, 1, 28, 28), finn_onnx)
model = ModelWrapper(finn_onnx)
model = model.transform(InferShapes())
model = model.transform(FoldConstants())
model = model.transform(RemoveStaticGraphInputs())
assert len(model.graph.input) == 1
assert len(model.graph.output) == 1
# load one of the test vectors
raw_i = get_data("finn", "data/onnx/mnist-conv/test_data_set_0/input_0.pb")
input_tensor = onnx.load_tensor_from_string(raw_i)
# run using FINN-based execution
input_dict = {"0": nph.to_array(input_tensor)}
output_dict = oxe.execute_onnx(model, input_dict)
produced = output_dict[list(output_dict.keys())[0]]
# run using PyTorch/Brevitas
input_tensor = torch.from_numpy(nph.to_array(input_tensor)).float()
assert input_tensor.shape == (1, 1, 28, 28)
# do forward pass in PyTorch/Brevitas
expected = fc.forward(input_tensor).detach().numpy()
assert np.isclose(produced, expected, atol=1e-3).all()
def test_quartznet_asr_4b(pretrained):
finn_onnx = "quant_quartznet_perchannelscaling_4b.onnx"
quartznet = quant_quartznet_perchannelscaling_4b(pretrained,
export_mode=True)
quartznet.eval()
FINNManager.export(quartznet, QUARTZNET_POSTPROCESSED_INPUT_SIZE,
finn_onnx)
model = ModelWrapper(finn_onnx)
model = model.transform(GiveUniqueNodeNames())
model = model.transform(DoubleToSingleFloat())
model = model.transform(InferShapes())
model = model.transform(FoldConstants())
model = model.transform(RemoveStaticGraphInputs())
#load a random test vector
input_tensor = np.random.uniform(
MIN_INP_VAL, MAX_INP_VAL,
size=QUARTZNET_POSTPROCESSED_INPUT_SIZE).astype(np.float32)
# run using FINN-based execution
input_dict = {"0": input_tensor}
output_dict = oxe.execute_onnx(model, input_dict)
produced = output_dict[list(output_dict.keys())[0]]
# run using PyTorch/Brevitas
input_tensor = torch.from_numpy(input_tensor).float()
# do forward pass in PyTorch/Brevitas
expected = quartznet(input_tensor).detach().numpy()
assert np.isclose(produced, expected, atol=ATOL).all()
def test_brevitas_cnv_onnx_export_and_exec(wbits, abits, pretrained):
if wbits > abits:
pytest.skip("No wbits > abits cases.")
nname = f"CNV_{wbits}W{abits}A"
finn_onnx = nname + ".onnx"
cnv, _ = model_with_cfg(nname.lower(), pretrained=pretrained)
cnv.eval()
# load a random int test vector
input_a = np.random.randint(MIN_INP_VAL, MAX_INP_VAL, size=CNV_INPUT_SIZE).astype(np.float32)
scale = 1. / 255
input_t = torch.from_numpy(input_a * scale)
input_qt = QuantTensor(
input_t, scale=torch.tensor(scale), bit_width=torch.tensor(8.0), signed=False)
FINNManager.export(cnv, export_path=finn_onnx, input_t=input_qt)
model = ModelWrapper(finn_onnx)
model = model.transform(GiveUniqueNodeNames())
model = model.transform(DoubleToSingleFloat())
model = model.transform(InferShapes())
model = model.transform(FoldConstants())
model = model.transform(RemoveStaticGraphInputs())
# run using FINN-based execution
input_dict = {"0": input_a}
output_dict = oxe.execute_onnx(model, input_dict)
produced = output_dict[list(output_dict.keys())[0]]
# do forward pass in PyTorch/Brevitas
expected = cnv(input_t).detach().numpy()
assert np.isclose(produced, expected, atol=ATOL).all()
def test_streamline_fc(size, wbits, abits):
if size == "LFC" and wbits == 2 and abits == 2:
pytest.skip("No LFC-w2a2 present at the moment")
if wbits > abits:
pytest.skip("No wbits > abits cases at the moment")
nname = "%s_%dW%dA" % (size, wbits, abits)
finn_onnx = export_onnx_path + "/%s.onnx" % nname
fc = get_test_model_trained(size, wbits, abits)
bo.export_finn_onnx(fc, (1, 1, 28, 28), finn_onnx)
model = ModelWrapper(finn_onnx)
model = model.transform(InferShapes())
model = model.transform(FoldConstants())
model = model.transform(GiveUniqueNodeNames())
model = model.transform(GiveReadableTensorNames())
model = model.transform(RemoveStaticGraphInputs())
# load one of the test vectors
raw_i = get_data("finn", "data/onnx/mnist-conv/test_data_set_0/input_0.pb")
input_tensor = onnx.load_tensor_from_string(raw_i)
# run using FINN-based execution
input_dict = {"global_in": nph.to_array(input_tensor)}
expected_ctx = oxe.execute_onnx(model, input_dict, True)
expected = expected_ctx[model.graph.output[0].name]
model = model.transform(Streamline())
model = model.transform(RemoveUnusedTensors())
assert len(model.graph.initializer) == 11
assert len(model.graph.value_info) == 21
assert len(model.graph.quantization_annotation) == 20
produced_ctx = oxe.execute_onnx(model, input_dict, True)
produced = produced_ctx[model.graph.output[0].name]
assert np.isclose(expected, produced, atol=1e-3).all()
def test_brevitas_debug():
finn_onnx = "test_brevitas_debug.onnx"
fc = get_test_model_trained("TFC", 2, 2)
dbg_hook = bo.enable_debug(fc)
bo.export_finn_onnx(fc, (1, 1, 28, 28), finn_onnx)
model = ModelWrapper(finn_onnx)
model = model.transform(InferShapes())
model = model.transform(FoldConstants())
model = model.transform(RemoveStaticGraphInputs())
assert len(model.graph.input) == 1
assert len(model.graph.output) == 1
# load one of the test vectors
raw_i = get_data("finn", "data/onnx/mnist-conv/test_data_set_0/input_0.pb")
input_tensor = onnx.load_tensor_from_string(raw_i)
# run using FINN-based execution
input_dict = {"0": nph.to_array(input_tensor)}
output_dict = oxe.execute_onnx(model, input_dict, return_full_exec_context=True)
produced = output_dict[model.graph.output[0].name]
# run using PyTorch/Brevitas
input_tensor = torch.from_numpy(nph.to_array(input_tensor)).float()
assert input_tensor.shape == (1, 1, 28, 28)
# do forward pass in PyTorch/Brevitas
expected = fc.forward(input_tensor).detach().numpy()
assert np.isclose(produced, expected, atol=1e-3).all()
# check all tensors at debug markers
names_brevitas = set(dbg_hook.values.keys())
names_finn = set(output_dict.keys())
names_common = names_brevitas.intersection(names_finn)
assert len(names_common) == 16
for dbg_name in names_common:
tensor_pytorch = dbg_hook.values[dbg_name].detach().numpy()
tensor_finn = output_dict[dbg_name]
assert np.isclose(tensor_finn, tensor_pytorch, atol=1e-5).all()
os.remove(finn_onnx)
def test_streamline_cnv(size, wbits, abits):
if wbits > abits:
pytest.skip("No wbits > abits cases at the moment")
nname = "%s_%dW%dA" % (size, wbits, abits)
finn_onnx = export_onnx_path + "/%s.onnx" % nname
fc = get_test_model_trained(size, wbits, abits)
bo.export_finn_onnx(fc, (1, 3, 32, 32), finn_onnx)
model = ModelWrapper(finn_onnx)
model = model.transform(InferShapes())
model = model.transform(FoldConstants())
model = model.transform(GiveUniqueNodeNames())
model = model.transform(GiveReadableTensorNames())
model = model.transform(RemoveStaticGraphInputs())
# load one of the test vectors
fn = pk.resource_filename("finn",
"data/cifar10/cifar10-test-data-class3.npz")
input_tensor = np.load(fn)["arr_0"].astype(np.float32)
input_tensor = input_tensor / 255
assert input_tensor.shape == (1, 3, 32, 32)
# run using FINN-based execution
input_dict = {"global_in": input_tensor}
expected_ctx = oxe.execute_onnx(model, input_dict, True)
expected = expected_ctx[model.graph.output[0].name]
# model.save("orig_cnv.onnx")
model = model.transform(Streamline())
model = model.transform(RemoveUnusedTensors())
assert len(model.graph.initializer) == 21
assert len(model.graph.value_info) == 43
# model.save("streamlined_cnv.onnx")
assert len(model.graph.node) == 23
produced_ctx = oxe.execute_onnx(model, input_dict, True)
produced = produced_ctx[model.graph.output[0].name]
assert np.isclose(expected, produced, atol=1e-3).all()
assert model.graph.node[0].op_type == "MultiThreshold"
assert np.argmax(produced) == 3
def test_import_and_tidy(self, topology, wbits, abits):
prev_chkpt_name = get_checkpoint_name(topology, wbits, abits, "export")
model = load_test_checkpoint_or_skip(prev_chkpt_name)
model = model.transform(InferShapes())
model = model.transform(FoldConstants())
model = model.transform(GiveUniqueNodeNames())
model = model.transform(GiveReadableTensorNames())
model = model.transform(InferDataTypes())
model = model.transform(RemoveStaticGraphInputs())
chkpt = get_checkpoint_name(topology, wbits, abits, "import_and_tidy")
model.save(chkpt)
def tidy_up(model):
log("Basic transformations launched")
model = model.transform(InferShapes())
model = model.transform(FoldConstants())
model = model.transform(GiveUniqueNodeNames())
model = model.transform(GiveReadableTensorNames())
model = model.transform(InferDataTypes())
model = model.transform(RemoveStaticGraphInputs())
log("Basic transformations completed")
save(model, "0_tidy")
return model
def cleanup(self):
"Run cleanup transformations on the model."
transformed_model = self
cleanup_transforms = [
RemoveUnusedTensors(),
RemoveStaticGraphInputs(),
SortGraph(),
]
for trn in cleanup_transforms:
transformed_model = transformed_model.transform(
trn, cleanup=False, make_deepcopy=False)
return transformed_model
def step_resnet50_tidy(model: ModelWrapper, cfg: DataflowBuildConfig):
model = model.transform(GiveUniqueParameterTensors())
model = model.transform(InferShapes())
model = model.transform(FoldConstants())
model = model.transform(RemoveStaticGraphInputs())
model = model.transform(GiveUniqueNodeNames())
model = model.transform(GiveReadableTensorNames())
model = model.transform(InferDataTypes())
model = model.transform(InsertTopK())
model = model.transform(InferShapes())
model = model.transform(GiveUniqueNodeNames())
model = model.transform(GiveReadableTensorNames())
model = model.transform(InferDataTypes())
return model
def post_processing(model):
log("Starting Post Processing")
# Insert Top-1 node at the end
model = model.transform(InsertTopK(k=1))
# Tidy-up again
model = model.transform(InferShapes())
model = model.transform(FoldConstants())
model = model.transform(GiveUniqueNodeNames())
model = model.transform(GiveReadableTensorNames())
model = model.transform(InferDataTypes())
model = model.transform(RemoveStaticGraphInputs())
log("Finished Post Processing!")
save(model, "2_with_pre_post")
return model
def inference_cost(model_filename,
*,
output_json=None,
output_onnx=None,
preprocess=True,
discount_sparsity=True):
"""Print the inference cost estimate metric for given ONNX model.
Supports the Quant op for weight/activation quantization.
:param model_filename: Filename for ONNX model
:param output_json: Optional JSON filename to save the inference cost dict
:param output_onnx: Optional ONNX filename to save the final model after any
preprocessing
:param preprocess: If set, run preprocessing steps such as shape inference,
datatype inference and constant folding. Strongly recommended.
:param discount_sparsity: If set, will discount op cost of MAC ops with a
constant zero weight, and the mem cost of constant zero weights.
"""
print("Inference cost for " + model_filename)
model = ModelWrapper(model_filename)
if preprocess:
qnt_nodes = model.get_nodes_by_op_type("Quant")
for qnt_node in qnt_nodes:
qnt_node.domain = "finn.custom_op.general"
model = model.transform(InferShapes())
model = model.transform(GiveUniqueParameterTensors())
model = model.transform(InferDataTypes())
model = model.transform(FoldConstants())
model = model.transform(RemoveUnusedTensors())
model = model.transform(RemoveStaticGraphInputs())
model = model.transform(InferDataTypes())
model = model.transform(GiveUniqueNodeNames())
model = model.transform(GiveReadableTensorNames())
if output_onnx is not None:
model.save(output_onnx)
ret = model.analysis(lambda x: infca.inference_cost(x, discount_sparsity))
bops = compute_bops(ret)
mem_w_bits = compute_mem_bits(ret, "mem_w")
mem_o_bits = compute_mem_bits(ret, "mem_o")
ret["total_bops"] = bops
ret["total_mem_w_bits"] = mem_w_bits
ret["total_mem_o_bits"] = mem_o_bits
if "unsupported" in ret:
ret["unsupported"] = str(ret["unsupported"])
print(json.dumps(ret, sort_keys=True, indent=2))
if output_json is not None:
with open(output_json, "w") as f:
json.dump(ret, f, sort_keys=True, indent=2)
def step_tidy_up(model: ModelWrapper, cfg: DataflowBuildConfig):
"""Run the tidy-up step on given model. This includes shape and datatype
inference, constant folding, and giving nodes and tensors better names.
"""
model = model.transform(InferShapes())
model = model.transform(FoldConstants())
model = model.transform(GiveUniqueNodeNames())
model = model.transform(GiveReadableTensorNames())
model = model.transform(InferDataTypes())
model = model.transform(RemoveStaticGraphInputs())
if VerificationStepType.TIDY_UP_PYTHON in cfg._resolve_verification_steps():
verify_step(model, cfg, "initial_python", need_parent=False)
return model
def test_mobilenet_v1_4b(pretrained):
finn_onnx = "mobilenet_v1_4b.onnx"
mobilenet = quant_mobilenet_v1_4b(pretrained)
mobilenet.eval()
#load a random test vector
np.random.seed(SEED)
numpy_tensor = np.random.random(size=INPUT_SIZE).astype(np.float32)
# run using PyTorch/Brevitas
torch_tensor = torch.from_numpy(numpy_tensor).float()
# do forward pass in PyTorch/Brevitas
expected = mobilenet(torch_tensor).detach().numpy()
export_finn_onnx(mobilenet, INPUT_SIZE, finn_onnx)
model = ModelWrapper(finn_onnx)
model = model.transform(GiveUniqueNodeNames())
model = model.transform(DoubleToSingleFloat())
model = model.transform(InferShapes())
model = model.transform(FoldConstants())
model = model.transform(RemoveStaticGraphInputs())
# run using FINN-based execution
inp_name = model.graph.input[0].name
input_dict = {inp_name: numpy_tensor}
output_dict = oxe.execute_onnx(model, input_dict)
produced = output_dict[list(output_dict.keys())[0]]
assert np.isclose(produced, expected, atol=ATOL).all()
def test_brevitas_debug(QONNX_export, QONNX_FINN_conversion):
if (not QONNX_export) and QONNX_FINN_conversion:
pytest.skip(
"This test configuration is not valid and is thus skipped.")
finn_onnx = "test_brevitas_debug.onnx"
fc = get_test_model_trained("TFC", 2, 2)
ishape = (1, 1, 28, 28)
if QONNX_export:
dbg_hook = bo.enable_debug(fc, proxy_level=True)
BrevitasONNXManager.export(fc, ishape, finn_onnx)
# DebugMarkers have the brevitas.onnx domain, so that needs adjusting
model = ModelWrapper(finn_onnx)
dbg_nodes = model.get_nodes_by_op_type("DebugMarker")
for dbg_node in dbg_nodes:
dbg_node.domain = "finn.custom_op.general"
model.save(finn_onnx)
qonnx_cleanup(finn_onnx, out_file=finn_onnx)
if QONNX_FINN_conversion:
model = ModelWrapper(finn_onnx)
model = model.transform(ConvertQONNXtoFINN())
model.save(finn_onnx)
else:
dbg_hook = bo.enable_debug(fc)
bo.export_finn_onnx(fc, ishape, finn_onnx)
model = ModelWrapper(finn_onnx)
# DebugMarkers have the brevitas.onnx domain, so that needs adjusting
# ToDo: We should probably have transformation pass, which does this
# domain conversion for us?
dbg_nodes = model.get_nodes_by_op_type("DebugMarker")
for dbg_node in dbg_nodes:
dbg_node.domain = "finn.custom_op.general"
model = model.transform(InferShapes())
model = model.transform(FoldConstants())
model = model.transform(RemoveStaticGraphInputs())
model.save(finn_onnx)
model = ModelWrapper(finn_onnx)
assert len(model.graph.input) == 1
assert len(model.graph.output) == 1
# load one of the test vectors
raw_i = get_data("finn.data", "onnx/mnist-conv/test_data_set_0/input_0.pb")
input_tensor = onnx.load_tensor_from_string(raw_i)
# run using FINN-based execution
input_dict = {model.graph.input[0].name: nph.to_array(input_tensor)}
output_dict = oxe.execute_onnx(model,
input_dict,
return_full_exec_context=True)
produced = output_dict[model.graph.output[0].name]
# run using PyTorch/Brevitas
input_tensor = torch.from_numpy(nph.to_array(input_tensor)).float()
assert input_tensor.shape == (1, 1, 28, 28)
# do forward pass in PyTorch/Brevitas
expected = fc.forward(input_tensor).detach().numpy()
assert np.isclose(produced, expected, atol=1e-3).all()
# check all tensors at debug markers
names_brevitas = set(dbg_hook.values.keys())
names_finn = set(output_dict.keys())
names_common = names_brevitas.intersection(names_finn)
# The different exports return debug markers in different numbers and places
print(len(names_common))
if QONNX_export and not QONNX_FINN_conversion:
assert len(names_common) == 12
elif QONNX_export and QONNX_FINN_conversion:
assert len(names_common) == 8
else:
assert len(names_common) == 16
for dbg_name in names_common:
if QONNX_export:
tensor_pytorch = dbg_hook.values[dbg_name].value.detach().numpy()
else:
tensor_pytorch = dbg_hook.values[dbg_name].detach().numpy()
tensor_finn = output_dict[dbg_name]
assert np.isclose(tensor_finn, tensor_pytorch, atol=1e-5).all()
os.remove(finn_onnx)
def test_QONNX_to_FINN(model_name, wbits, abits):
if wbits > abits:
pytest.skip("No wbits > abits cases at the moment")
if model_name == "LFC" and wbits == 2 and abits == 2:
pytest.skip("No LFC-w2a2 present at the moment")
if model_name == "mobilenet" and (wbits != 2 or abits != 2):
pytest.skip("Mobilenet only runs at W2A2, though it's technically W4A4.")
# Get test config and model
ATOL = 1e-7
brev_model, in_shape, input_tensor = get_brev_model_and_sample_inputs(
model_name, wbits, abits
)
temp_dir = TemporaryDirectory()
qonnx_base_path = temp_dir.name + "/qonnx_{}.onnx"
finn_base_path = temp_dir.name + "/finn_{}.onnx"
# Get Brevitas output
torch_input_tensor = torch.from_numpy(input_tensor).float()
brev_output = brev_model.forward(torch_input_tensor).detach().numpy()
# Get "clean" FINN model and it's output
_ = bo.export_finn_onnx(brev_model, in_shape, finn_base_path.format("raw"))
model = ModelWrapper(finn_base_path.format("raw"))
model = model.transform(GiveUniqueNodeNames())
model = model.transform(InferShapes())
model = model.transform(FoldConstants())
model = model.transform(RemoveStaticGraphInputs())
model.save(finn_base_path.format("clean"))
model = ModelWrapper(finn_base_path.format("clean"))
input_dict = {model.graph.input[0].name: input_tensor}
output_dict = oxe.execute_onnx(model, input_dict, False)
finn_export_output = output_dict[model.graph.output[0].name]
# This test always fails on MobileNet for some reason
if model_name != "mobilenet":
assert np.isclose(
brev_output, finn_export_output, atol=ATOL
).all(), "The output of the Brevitas model and the FINN model should match."
# Get the equivalent QONNX model
b_onnx.function.DOMAIN_STRING = "finn.custom_op.general"
_ = b_onnx.manager.BrevitasONNXManager.export(
brev_model, in_shape, qonnx_base_path.format("raw")
)
cleanup(qonnx_base_path.format("raw"), out_file=qonnx_base_path.format("clean"))
# Compare output
model = ModelWrapper(qonnx_base_path.format("clean"))
input_dict = {model.graph.input[0].name: input_tensor}
output_dict = oxe.execute_onnx(model, input_dict, False)
qonnx_export_output = output_dict[model.graph.output[0].name]
assert np.isclose(
brev_output, qonnx_export_output, atol=ATOL
).all(), "The output of the Brevitas model and the QONNX model should match."
# This test always fails on MobileNet for some reason
if model_name != "mobilenet":
assert np.isclose(
qonnx_export_output, finn_export_output, atol=ATOL
).all(), "The output of the FINN model and the QONNX model should match."
# Run QONNX to FINN conversion
model = ModelWrapper(qonnx_base_path.format("clean"))
model = model.transform(ConvertQONNXtoFINN())
model.save(qonnx_base_path.format("whole_trafo"))
# Compare output
model = ModelWrapper(qonnx_base_path.format("whole_trafo"))
input_dict = {model.graph.input[0].name: input_tensor}
output_dict = oxe.execute_onnx(model, input_dict, False)
test_output = output_dict[model.graph.output[0].name]
assert np.isclose(test_output, finn_export_output, atol=ATOL).all(), (
"The output of the FINN model "
"and the QONNX -> FINN converted model should match."
)
# Run analysis passes on the converted model
model = ModelWrapper(qonnx_base_path.format("whole_trafo"))
_ = model.analysis(analysis_testing_for_no_quant_nodes)
temp_dir.cleanup()
def test_brevitas_compare_exported_mobilenet():
if "IMAGENET_VAL_PATH" not in os.environ.keys():
pytest.skip("Can't do validation without IMAGENET_VAL_PATH")
n_images = 10
debug_mode = False
export_onnx_path = make_build_dir("test_brevitas_mobilenet-v1_")
# export preprocessing
preproc_onnx = export_onnx_path + "/quant_mobilenet_v1_4b_preproc.onnx"
preproc = NormalizePreProc(mean, std, ch)
bo.export_finn_onnx(preproc, (1, 3, 224, 224), preproc_onnx)
preproc_model = ModelWrapper(preproc_onnx)
preproc_model = preproc_model.transform(InferShapes())
preproc_model = preproc_model.transform(GiveUniqueNodeNames())
preproc_model = preproc_model.transform(GiveUniqueParameterTensors())
preproc_model = preproc_model.transform(GiveReadableTensorNames())
# export the actual MobileNet-v1
finn_onnx = export_onnx_path + "/quant_mobilenet_v1_4b.onnx"
mobilenet = get_test_model_trained("mobilenet", 4, 4)
if debug_mode:
dbg_hook = bo.enable_debug(mobilenet)
bo.export_finn_onnx(mobilenet, (1, 3, 224, 224), finn_onnx)
model = ModelWrapper(finn_onnx)
model = model.transform(InferShapes())
model = model.transform(FoldConstants())
model = model.transform(RemoveStaticGraphInputs())
model = model.transform(InsertTopK())
# get initializer from Mul that will be absorbed into topk
a0 = model.get_initializer(model.get_nodes_by_op_type("Mul")[-1].input[1])
model = model.transform(absorb.AbsorbScalarMulAddIntoTopK())
model = model.transform(InferShapes())
model = model.transform(InferDataTypes())
model = model.transform(InferDataLayouts())
model = model.transform(GiveUniqueNodeNames())
model = model.transform(GiveUniqueParameterTensors())
model = model.transform(GiveReadableTensorNames())
model.save(export_onnx_path + "/quant_mobilenet_v1_4b_wo_preproc.onnx")
# create merged preprocessing + MobileNet-v1 model
model = model.transform(MergeONNXModels(preproc_model))
model.save(export_onnx_path + "/quant_mobilenet_v1_4b.onnx")
with open(
export_onnx_path + "/mobilenet_validation.csv", "w", newline=""
) as csvfile:
writer = csv.writer(csvfile)
writer.writerow(
[
"goldenID",
"brevitasTop5",
"brevitasTop5[%]",
"finnTop5",
"finnTop5[%]",
"top5equal",
"top5%equal",
]
)
csvfile.flush()
workload = imagenet_util.get_val_images(n_images, interleave_classes=True)
all_inds_ok = True
all_probs_ok = True
for (img_path, target_id) in workload:
img_np = imagenet_util.load_resize_crop(img_path)
img_torch = torch.from_numpy(img_np).float()
# do forward pass in PyTorch/Brevitas
input_tensor = preproc.forward(img_torch)
expected = mobilenet.forward(input_tensor).detach().numpy()
expected_topk = expected.flatten()
expected_top5 = np.argsort(expected_topk)[-5:]
expected_top5 = np.flip(expected_top5)
expected_top5_prob = []
for index in expected_top5:
expected_top5_prob.append(expected_topk[index])
idict = {model.graph.input[0].name: img_np}
odict = oxe.execute_onnx(model, idict, return_full_exec_context=True)
produced = odict[model.graph.output[0].name]
produced_prob = odict["TopK_0_out0"] * a0
inds_ok = (produced.flatten() == expected_top5).all()
probs_ok = np.isclose(produced_prob.flatten(), expected_top5_prob).all()
all_inds_ok = all_inds_ok and inds_ok
all_probs_ok = all_probs_ok and probs_ok
writer.writerow(
[
str(target_id),
str(expected_top5),
str(expected_top5_prob),
str(produced.flatten()),
str(produced_prob.flatten()),
str(inds_ok),
str(probs_ok),
]
)
csvfile.flush()
if ((not inds_ok) or (not probs_ok)) and debug_mode:
print("Results differ for %s" % img_path)
# check all tensors at debug markers
names_brevitas = set(dbg_hook.values.keys())
names_finn = set(odict.keys())
names_common = names_brevitas.intersection(names_finn)
for dbg_name in names_common:
if not np.isclose(
dbg_hook.values[dbg_name].detach().numpy(),
odict[dbg_name],
atol=1e-3,
).all():
print("Tensor %s differs between Brevitas and FINN" % dbg_name)
assert all_inds_ok and all_probs_ok
