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_end2end_mobilenet_streamline():
model = load_test_checkpoint_or_skip(build_dir +
"/end2end_mobilenet_tidy.onnx")
model = model.transform(Streamline())
additional_streamline_transformations = [
DoubleToSingleFloat(),
reorder.MoveMulPastDWConv(),
absorb.AbsorbMulIntoMultiThreshold(),
ChangeDataLayoutQuantAvgPool2d(),
InferDataLayouts(),
reorder.MoveTransposePastScalarMul(),
absorb.AbsorbTransposeIntoFlatten(),
reorder.MoveFlattenPastAffine(),
reorder.MoveFlattenPastTopK(),
reorder.MoveScalarMulPastMatMul(),
CollapseRepeatedMul(),
RemoveIdentityOps(),
RoundAndClipThresholds(),
]
for trn in additional_streamline_transformations:
model = model.transform(trn)
model = model.transform(GiveUniqueNodeNames())
model = model.transform(GiveReadableTensorNames())
model = model.transform(InferDataTypes())
model.save(build_dir + "/end2end_mobilenet_streamlined.onnx")
assert (len(model.get_nodes_by_op_type("Add")) == 1
) # only final quantized bias Add op remains
assert len(model.get_nodes_by_op_type("Mul")) == 0 # no Mul ops remain
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 transform(
self, transformation, make_deepcopy=True, cleanup=True, fix_float64=True
):
"""Applies given Transformation repeatedly until no more changes can be made
and returns a transformed ModelWrapper instance.
- make_deepcopy : operates on a new (deep)copy of model.
- fix_float64 : DoubleToSingleFloat correction before starting
- cleanup : execute cleanup transformations before returning
"""
transformed_model = self
if make_deepcopy:
transformed_model = copy.deepcopy(self)
if fix_float64:
(transformed_model, model_was_changed) = DoubleToSingleFloat().apply(
transformed_model
)
model_was_changed = True
while model_was_changed:
(transformed_model, model_was_changed) = transformation.apply(
transformed_model
)
if cleanup:
transformed_model.cleanup()
return transformed_model
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(DoubleToSingleFloat())
model = model.transform(InferShapes())
model = model.transform(FoldConstants())
model = model.transform(GiveUniqueNodeNames())
model = model.transform(GiveReadableTensorNames())
# 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.save("streamlined_cnv.onnx")
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_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(DoubleToSingleFloat())
model = model.transform(InferShapes())
model = model.transform(FoldConstants())
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_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_end2end_cnv_w1a1_import_and_tidy():
model = ModelWrapper(build_dir + "/end2end_cnv_w1a1_export.onnx")
model = model.transform(DoubleToSingleFloat())
model = model.transform(InferShapes())
model = model.transform(FoldConstants())
model = model.transform(GiveUniqueNodeNames())
model = model.transform(GiveReadableTensorNames())
model.save(build_dir + "/end2end_cnv_w1a1_tidy.onnx")
def step_resnet50_streamline(model: ModelWrapper, cfg: DataflowBuildConfig):
for iter_id in range(4):
model = step_resnet50_streamline_linear(model, cfg)
model = step_resnet50_streamline_nonlinear(model, cfg)
# big loop tidy up
model = model.transform(RemoveUnusedTensors())
model = model.transform(GiveReadableTensorNames())
model = model.transform(InferDataTypes())
model = model.transform(SortGraph())
model = model.transform(DoubleToSingleFloat())
return model
def test_conv_lowering_cnv_w1a1():
cnv = get_test_model_trained("CNV", 1, 1)
bo.export_finn_onnx(cnv, (1, 3, 32, 32), export_onnx_path)
model = ModelWrapper(export_onnx_path)
model = model.transform(DoubleToSingleFloat())
model = model.transform(InferShapes())
model = model.transform(FoldConstants())
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)
# execute imported model to get expected answer
input_dict = {"0": input_tensor}
output_dict_e = oxe.execute_onnx(model, input_dict)
expected = output_dict_e[list(output_dict_e.keys())[0]]
# execute transformed model and compare
model = model.transform(LowerConvsToMatMul())
output_dict_p = oxe.execute_onnx(model, input_dict)
produced = output_dict_p[list(output_dict_p.keys())[0]]
assert np.isclose(produced, expected).all()
assert np.argmax(produced) == 3
os.remove(export_onnx_path)
def step_mobilenet_streamline(model: ModelWrapper, cfg: DataflowBuildConfig):
model = model.transform(Streamline())
additional_streamline_transformations = [
DoubleToSingleFloat(),
reorder.MoveMulPastDWConv(),
absorb.AbsorbMulIntoMultiThreshold(),
ChangeDataLayoutQuantAvgPool2d(),
InferDataLayouts(),
reorder.MoveTransposePastScalarMul(),
absorb.AbsorbTransposeIntoFlatten(),
reorder.MoveFlattenPastAffine(),
reorder.MoveFlattenPastTopK(),
reorder.MoveScalarMulPastMatMul(),
CollapseRepeatedMul(),
RemoveIdentityOps(),
RoundAndClipThresholds(),
]
for trn in additional_streamline_transformations:
model = model.transform(trn)
model = model.transform(GiveUniqueNodeNames())
model = model.transform(GiveReadableTensorNames())
model = model.transform(InferDataTypes())
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 step_resnet50_convert_to_hls(model: ModelWrapper,
cfg: DataflowBuildConfig):
model.set_tensor_datatype(model.graph.input[0].name, DataType["UINT8"])
model = model.transform(InferDataLayouts())
try:
from finn.transformation.fpgadataflow.infer_doublepacked_dsp import InferDoublePackedConv
model = model.transform(InferDoublePackedConv([1]))
except:
print(
" FINN Experimental not available. Using non-packed convolution ")
model = model.transform(DoubleToSingleFloat())
model = model.transform(InferDataTypes())
model = model.transform(SortGraph())
to_hls_transformations = [
to_hls.InferAddStreamsLayer, LowerConvsToMatMul,
to_hls.InferChannelwiseLinearLayer, to_hls.InferPool_Batch,
AbsorbTransposeIntoMultiThreshold, RoundAndClipThresholds,
to_hls.InferQuantizedStreamingFCLayer, to_hls.InferThresholdingLayer,
AbsorbConsecutiveTransposes, to_hls.InferConvInpGen,
to_hls.InferDuplicateStreamsLayer, to_hls.InferLabelSelectLayer
]
for trn in to_hls_transformations:
model = model.transform(trn())
model = model.transform(InferDataLayouts())
model = model.transform(GiveUniqueNodeNames())
model = model.transform(InferDataTypes())
model = model.transform(RemoveCNVtoFCFlatten())
model = model.transform(GiveReadableTensorNames())
model = model.transform(RemoveUnusedTensors())
model = model.transform(SortGraph())
return model
def test_batchnorm_to_affine_cnv_w1a1():
lfc = get_test_model_trained("CNV", 1, 1)
bo.export_finn_onnx(lfc, (1, 3, 32, 32), export_onnx_path)
model = ModelWrapper(export_onnx_path)
model = model.transform(DoubleToSingleFloat())
model = model.transform(InferShapes())
model = model.transform(FoldConstants())
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)
input_dict = {"0": input_tensor}
output_dict = oxe.execute_onnx(model, input_dict)
expected = output_dict[list(output_dict.keys())[0]]
new_model = model.transform(BatchNormToAffine())
# check that there are no BN nodes left
op_types = list(map(lambda x: x.op_type, new_model.graph.node))
assert "BatchNormalization" not in op_types
output_dict_p = oxe.execute_onnx(new_model, input_dict)
produced = output_dict_p[list(output_dict_p.keys())[0]]
assert np.isclose(expected, produced).all()
assert np.argmax(produced) == 3
os.remove(export_onnx_path)
def test_convert_to_hls_layers_cnv_w1a1():
cnv = get_test_model_trained("CNV", 1, 1)
bo.export_finn_onnx(cnv, (1, 3, 32, 32), export_onnx_path_cnv)
model = ModelWrapper(export_onnx_path_cnv)
model = model.transform(DoubleToSingleFloat())
model = model.transform(InferShapes())
model = model.transform(FoldConstants())
model = model.transform(GiveUniqueNodeNames())
model = model.transform(GiveReadableTensorNames())
model = model.transform(Streamline())
model = model.transform(LowerConvsToMatMul())
model = model.transform(MakeMaxPoolNHWC())
model = model.transform(absorb.AbsorbTransposeIntoMultiThreshold())
model = model.transform(ConvertBipolarMatMulToXnorPopcount())
model = model.transform(Streamline())
# model.save("golden.onnx")
# 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)
# generate expected value from streamlined net
input_dict = {"global_in": input_tensor}
expected_ctx = oxe.execute_onnx(model, input_dict, True)
expected = expected_ctx[model.graph.output[0].name]
model = model.transform(to_hls.InferBinaryStreamingFCLayer())
model = model.transform(to_hls.InferQuantizedStreamingFCLayer())
for node in model.graph.node:
if node.op_type == "StreamingFCLayer_Batch":
inst = getCustomOp(node)
inst.set_nodeattr("mem_mode", "decoupled")
mw = inst.get_nodeattr("MW")
mh = inst.get_nodeattr("MH")
if mh % 4 == 0:
pe = mh // 4
else:
pe = mh
inst.set_nodeattr("PE", pe)
if mw % 16 == 0:
simd = mw // 16
else:
simd = mw
inst.set_nodeattr("SIMD", simd)
model = model.transform(to_hls.InferConvInpGen())
model = model.transform(to_hls.InferStreamingMaxPool())
# check topology status
finn_nodes = model.get_finn_nodes()
assert len(finn_nodes) == 18
non_finn_nodes = model.get_non_finn_nodes()
assert len(non_finn_nodes) == 4
exp_non_finn_nodes = ["Transpose", "Reshape", "Mul", "Add"]
assert [x.op_type for x in non_finn_nodes] == exp_non_finn_nodes
fc_nodes = model.get_nodes_by_op_type("StreamingFCLayer_Batch")
assert len(fc_nodes) == 9
swg_nodes = model.get_nodes_by_op_type("ConvolutionInputGenerator")
assert len(swg_nodes) == 6
mp_nodes = model.get_nodes_by_op_type("StreamingMaxPool_Batch")
assert len(mp_nodes) == 2
# model.save("cnv-pre-compile.onnx")
model = model.transform(PrepareCppSim())
model = model.transform(CompileCppSim())
model = model.transform(SetExecMode("cppsim"))
# model.save("cnv-post-compile.onnx")
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 np.argmax(produced) == 3
os.remove(export_onnx_path_cnv)
