def test_convert_bipolar_matmul_to_xnorpopcountmatmul():
lfc = get_test_model_trained("LFC", 1, 1)
bo.export_finn_onnx(lfc, (1, 1, 28, 28), export_onnx_path)
model = ModelWrapper(export_onnx_path)
model = model.transform(InferShapes())
model = model.transform(FoldConstants())
model = model.transform(GiveUniqueNodeNames())
model = model.transform(GiveReadableTensorNames())
model = model.transform(ConvertSignToThres())
# 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(ConvertBipolarMatMulToXnorPopcount())
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()
os.remove(export_onnx_path)
def test_streamline_fc(size, wbits, abits):
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())
# 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())
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_fc_onnx_export_and_exec(size, wbits, abits):
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())
# 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_modelwrapper():
lfc = get_test_model_trained("LFC", 1, 1)
bo.export_finn_onnx(lfc, (1, 1, 28, 28), export_onnx_path)
model = ModelWrapper(export_onnx_path)
assert model.check_all_tensor_shapes_specified() is False
inp_name = model.graph.input[0].name
inp_shape = model.get_tensor_shape(inp_name)
assert inp_shape == [1, 1, 28, 28]
# find first matmul node
l0_mat_tensor_name = ""
l0_inp_tensor_name = ""
for node in model.graph.node:
if node.op_type == "MatMul":
l0_inp_tensor_name = node.input[0]
l0_mat_tensor_name = node.input[1]
break
assert l0_mat_tensor_name != ""
l0_weights = model.get_initializer(l0_mat_tensor_name)
assert l0_weights.shape == (784, 1024)
l0_weights_hist = Counter(l0_weights.flatten())
assert (l0_weights_hist[1.0] + l0_weights_hist[-1.0]) == 784 * 1024
l0_weights_rand = np.random.randn(784, 1024)
model.set_initializer(l0_mat_tensor_name, l0_weights_rand)
assert (model.get_initializer(l0_mat_tensor_name) == l0_weights_rand).all()
assert l0_inp_tensor_name != ""
inp_cons = model.find_consumer(l0_inp_tensor_name)
assert inp_cons.op_type == "MatMul"
out_prod = model.find_producer(l0_inp_tensor_name)
assert out_prod.op_type == "MultiThreshold"
inp_layout = model.get_tensor_layout(inp_name)
assert inp_layout is None
inp_layout = DataLayout.NCHW
model.set_tensor_layout(inp_name, inp_layout)
assert model.get_tensor_layout(inp_name) == inp_layout
inp_sparsity = model.get_tensor_sparsity(inp_name)
assert inp_sparsity is None
inp_sparsity = {"dw": {"kernel_shape": 3}}
model.set_tensor_sparsity(inp_name, inp_sparsity)
assert model.get_tensor_sparsity(inp_name) == inp_sparsity
os.remove(export_onnx_path)
def test_infer_datatypes_lfc():
lfc = get_test_model_trained("LFC", 1, 1)
bo.export_finn_onnx(lfc, (1, 1, 28, 28), export_onnx_path)
model = ModelWrapper(export_onnx_path)
model = model.transform(InferShapes())
model = model.transform(FoldConstants())
model = model.transform(GiveUniqueNodeNames())
model = model.transform(GiveReadableTensorNames())
model = model.transform(InferDataTypes())
assert model.get_tensor_datatype("MatMul_0_out0") == DataType.INT32
assert model.get_tensor_datatype("MatMul_1_out0") == DataType.INT32
assert model.get_tensor_datatype("MatMul_2_out0") == DataType.INT32
assert model.get_tensor_datatype("MatMul_3_out0") == DataType.INT32
assert model.get_tensor_datatype(
"MultiThreshold_0_out0") == DataType.BIPOLAR
assert model.get_tensor_datatype(
"MultiThreshold_1_out0") == DataType.BIPOLAR
assert model.get_tensor_datatype(
"MultiThreshold_2_out0") == DataType.BIPOLAR
assert model.get_tensor_datatype(
"MultiThreshold_3_out0") == DataType.BIPOLAR
os.remove(export_onnx_path)
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(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 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(InferShapes())
model = model.transform(FoldConstants())
fn = pk.resource_filename("finn.qnn-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_brevitas_fc_onnx_export_and_exec(size, wbits, abits, QONNX_export):
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_QONNX-%d" % (size, wbits, abits, QONNX_export)
finn_onnx = export_onnx_path + "/%s.onnx" % nname
fc = get_test_model_trained(size, wbits, abits)
ishape = (1, 1, 28, 28)
if QONNX_export:
BrevitasONNXManager.export(fc, ishape, finn_onnx)
qonnx_cleanup(finn_onnx, out_file=finn_onnx)
model = ModelWrapper(finn_onnx)
model = model.transform(ConvertQONNXtoFINN())
model.save(finn_onnx)
else:
bo.export_finn_onnx(fc, ishape, 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 = {model.graph.input[0].name: 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_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_topk_insert(k):
tfc = get_test_model_trained("TFC", 1, 1)
bo.export_finn_onnx(tfc, (1, 1, 28, 28), export_onnx_path)
model = ModelWrapper(export_onnx_path)
# do transformations (no topk)
model = model.transform(InferShapes())
model = model.transform(FoldConstants())
model = model.transform(GiveUniqueNodeNames())
model = model.transform(GiveReadableTensorNames())
model = model.transform(InferDataTypes())
# verification: generate random input, run through net, streamline,
# run again, check that output is top-k
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)
input_brevitas = torch.from_numpy(nph.to_array(input_tensor)).float()
output_golden = tfc.forward(input_brevitas).detach().numpy()
output_golden_topk = np.flip(output_golden.flatten().argsort())[:k]
output_golden_topk = output_golden_topk.flatten()
input_dict = {"global_in": nph.to_array(input_tensor)}
# insert top-k
model = model.transform(InsertTopK(k))
model = model.transform(GiveUniqueNodeNames())
model = model.transform(GiveReadableTensorNames())
model = model.transform(InferShapes())
# verify output of top-k
output_dict_topk = oxe.execute_onnx(model, input_dict)
output_pysim_topk = output_dict_topk[list(output_dict_topk.keys())[0]]
output_pysim_topk = output_pysim_topk.astype(np.int).flatten()
assert np.array_equal(output_golden_topk, output_pysim_topk)
os.remove(export_onnx_path)
def test_convert_to_hls_layers_tfc_w1a2():
tfc = get_test_model_trained("TFC", 1, 2)
bo.export_finn_onnx(tfc, (1, 1, 28, 28), export_onnx_path)
model = ModelWrapper(export_onnx_path)
model = model.transform(InferShapes())
model = model.transform(FoldConstants())
model = model.transform(GiveUniqueNodeNames())
model = model.transform(GiveReadableTensorNames())
model = model.transform(Streamline())
from finn.transformation.fpgadataflow.convert_to_hls_layers import (
InferQuantizedStreamingFCLayer,
)
model = model.transform(InferQuantizedStreamingFCLayer())
fc0 = model.graph.node[2]
assert fc0.op_type == "StreamingFCLayer_Batch"
assert model.get_tensor_shape(fc0.input[0]) == [1, 784]
assert model.get_tensor_shape(fc0.input[1]) == [784, 64]
assert model.get_tensor_shape(fc0.input[2]) == [64, 2]
fc1 = model.graph.node[3]
assert fc1.op_type == "StreamingFCLayer_Batch"
assert model.get_tensor_shape(fc1.input[0]) == [1, 64]
assert model.get_tensor_shape(fc1.input[1]) == [64, 64]
assert model.get_tensor_shape(fc1.input[2]) == [64, 2]
fc2 = model.graph.node[4]
assert fc2.op_type == "StreamingFCLayer_Batch"
assert model.get_tensor_shape(fc2.input[0]) == [1, 64]
assert model.get_tensor_shape(fc2.input[1]) == [64, 64]
assert model.get_tensor_shape(fc2.input[2]) == [64, 2]
fc3 = model.graph.node[5]
assert fc3.op_type == "StreamingFCLayer_Batch"
assert model.get_tensor_shape(fc3.input[0]) == [1, 64]
assert model.get_tensor_shape(fc3.input[1]) == [64, 10]
fc0w = getCustomOp(fc0)
fc0w.set_nodeattr("SIMD", 784)
fc0w.set_nodeattr("PE", 16)
fc1w = getCustomOp(fc1)
fc1w.set_nodeattr("SIMD", 16)
fc1w.set_nodeattr("PE", 16)
fc2w = getCustomOp(fc2)
fc2w.set_nodeattr("SIMD", 16)
fc2w.set_nodeattr("PE", 16)
fc3w = getCustomOp(fc3)
fc3w.set_nodeattr("SIMD", 16)
fc3w.set_nodeattr("PE", 10)
model = model.transform(PrepareCppSim())
model = model.transform(CompileCppSim())
model = model.transform(SetExecMode("cppsim"))
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)}
output_dict = oxe.execute_onnx(model, input_dict, True)
produced = output_dict[model.graph.output[0].name]
model = ModelWrapper(export_onnx_path)
model = model.transform(InferShapes())
model = model.transform(FoldConstants())
model = model.transform(GiveUniqueNodeNames())
model = model.transform(GiveReadableTensorNames())
model = model.transform(Streamline())
golden_output_dict = oxe.execute_onnx(model, input_dict, True)
expected = golden_output_dict[model.graph.output[0].name]
assert np.isclose(produced, expected, atol=1e-3).all()
os.remove(export_onnx_path)
def test_infer_data_layouts_cnv():
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(InferShapes())
model = model.transform(FoldConstants())
model = model.transform(GiveUniqueNodeNames())
model = model.transform(GiveReadableTensorNames())
model = model.transform(Streamline())
model = model.transform(InferDataLayouts())
assert model.get_tensor_layout("global_in") == DataLayout.NCHW
assert model.get_tensor_layout("Conv_0_out0") == DataLayout.NCHW
assert model.get_tensor_layout("MaxPool_0_out0") == DataLayout.NCHW
assert model.get_tensor_layout("MultiThreshold_6_out0") == DataLayout.NCHW
assert model.get_tensor_layout("Reshape_0_out0") == DataLayout.NC
assert model.get_tensor_layout("MatMul_0_out0") == DataLayout.NC
assert model.get_tensor_layout("global_out") == DataLayout.NC
model = model.transform(LowerConvsToMatMul())
model = model.transform(MakeMaxPoolNHWC())
model = model.transform(GiveUniqueNodeNames())
model = model.transform(GiveReadableTensorNames())
model = model.transform(InferDataLayouts())
assert model.get_tensor_layout("global_in") == DataLayout.NCHW
assert model.get_tensor_layout("Transpose_0_out0") == DataLayout.NHWC
assert model.get_tensor_layout("Im2Col_0_out0") == DataLayout.NHWC
# note: im2col output isn't really NHWC or any other common layout
# since the concept of channels changes with lowering... but it is
# conceptually close to NHWC since the innermost dim gets multiplied
assert model.get_tensor_layout("MatMul_0_out0") == DataLayout.NHWC
assert model.get_tensor_layout("Transpose_1_out0") == DataLayout.NCHW
assert model.get_tensor_layout("Transpose_2_out0") == DataLayout.NHWC
assert model.get_tensor_layout("MaxPoolNHWC_0_out0") == DataLayout.NHWC
assert model.get_tensor_layout("Reshape_0_out0") == DataLayout.NC
assert model.get_tensor_layout("global_out") == DataLayout.NC
model = model.transform(absorb.AbsorbTransposeIntoMultiThreshold())
model = model.transform(ConvertBipolarMatMulToXnorPopcount())
model = model.transform(Streamline())
model = model.transform(to_hls.InferBinaryStreamingFCLayer())
model = model.transform(to_hls.InferQuantizedStreamingFCLayer())
model = model.transform(to_hls.InferConvInpGen())
model = model.transform(to_hls.InferStreamingMaxPool())
model = model.transform(GiveUniqueNodeNames())
model = model.transform(GiveReadableTensorNames())
model = model.transform(InferDataLayouts())
assert model.get_tensor_layout("global_in") == DataLayout.NCHW
assert model.get_tensor_layout("Transpose_0_out0") == DataLayout.NHWC
# note: im2col output isn't really NHWC or any other common layout
# since the concept of channels changes with lowering... but it is
# conceptually close to NHWC since the innermost dim gets multiplied
assert (model.get_tensor_layout("ConvolutionInputGenerator_0_out0") ==
DataLayout.NHWC)
assert model.get_tensor_layout(
"StreamingFCLayer_Batch_3_out0") == DataLayout.NHWC
assert model.get_tensor_layout("Reshape_0_out0") == DataLayout.NC
assert model.get_tensor_layout(
"StreamingFCLayer_Batch_6_out0") == DataLayout.NC
assert model.get_tensor_layout("global_out") == DataLayout.NC
os.remove(export_onnx_path_cnv)
def test_end2end_tfc_w1a2_export():
import brevitas.onnx as bo
tfc = get_test_model_trained("TFC", 1, 2)
bo.export_finn_onnx(tfc, (1, 1, 28, 28),
build_dir + "/end2end_tfc_w1a2_export.onnx")
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
download=True,
transform=transform)
testloader = torch.utils.data.DataLoader(testset,
batch_size=128,
shuffle=False,
num_workers=1)
"""
model = cnv(2,2,8)
package = torch.load("model_best.pth.tar", map_location='cpu')
model_state_dict = package['state_dict']
model.load_state_dict(model_state_dict, strict=False)
"""
# LOAD MODEL
model = get_test_model_trained("CNV", 2, 2)
import torch.optim as optim
optimizer = optim.Adam(model.parameters(), lr=0.0001)
device = 'cuda:0'
criterion = nn.CrossEntropyLoss().to(device)
# TRAINING AND TESTING
def test():
model.to(device)
model.eval()
criterion.eval()
prec1_global = []
def test_nodes_topologically_sorted():
# test analysis pass (nodes_topologically_sorted) with different models
# test with data/onnx/finn-hls-model/tfc_w1_a1_after_conv_to_hls.onnx
raw_m = get_data(
"finn", "data/onnx/finn-hls-model/tfc_w1_a1_after_conv_to_hls.onnx")
model = ModelWrapper(raw_m)
ret = model.analysis(ta.nodes_topologically_sorted)
assert ret["nodes_topologically_sorted"] is True
# remove first node and add it at the end
graph = model.graph
first_node = graph.node[0]
graph.node.remove(first_node)
graph.node.append(first_node)
ret = model.analysis(ta.nodes_topologically_sorted)
assert ret["nodes_topologically_sorted"] is False
# test with data/onnx/mnist-conv/model.onnx
raw_m = get_data("finn", "data/onnx/mnist-conv/model.onnx")
model = ModelWrapper(raw_m)
ret = model.analysis(ta.nodes_topologically_sorted)
assert ret["nodes_topologically_sorted"] is True
# remove first node and add it at the end
graph = model.graph
first_node = graph.node[0]
graph.node.remove(first_node)
graph.node.append(first_node)
ret = model.analysis(ta.nodes_topologically_sorted)
assert ret["nodes_topologically_sorted"] is False
# test with manually created small network
Neg_node = oh.make_node("Neg", inputs=["in1"], outputs=["neg1"])
Round_node = oh.make_node("Round", inputs=["neg1"], outputs=["round1"])
Ceil_node = oh.make_node("Ceil", inputs=["neg1"], outputs=["ceil1"])
Add_node = oh.make_node("Add",
inputs=["round1", "ceil1"],
outputs=["out1"])
in1 = oh.make_tensor_value_info("in1", TensorProto.FLOAT, [4, 4])
out1 = oh.make_tensor_value_info("out1", TensorProto.FLOAT, [4, 4])
graph = oh.make_graph(
nodes=[Neg_node, Round_node, Ceil_node, Add_node],
name="simple_graph",
inputs=[in1],
outputs=[out1],
value_info=[
oh.make_tensor_value_info("neg1", TensorProto.FLOAT, [4, 4]),
oh.make_tensor_value_info("round1", TensorProto.FLOAT, [4, 4]),
oh.make_tensor_value_info("ceil1", TensorProto.FLOAT, [4, 4]),
],
)
onnx_model = oh.make_model(graph, producer_name="simple-model")
model = ModelWrapper(onnx_model)
ret = model.analysis(ta.nodes_topologically_sorted)
assert ret["nodes_topologically_sorted"] is True
# create same graph but with "wrong" node order
graph = oh.make_graph(
nodes=[Round_node, Ceil_node, Neg_node, Add_node],
name="simple_graph",
inputs=[in1],
outputs=[out1],
value_info=[
oh.make_tensor_value_info("neg1", TensorProto.FLOAT, [4, 4]),
oh.make_tensor_value_info("round1", TensorProto.FLOAT, [4, 4]),
oh.make_tensor_value_info("ceil1", TensorProto.FLOAT, [4, 4]),
],
)
onnx_model = oh.make_model(graph, producer_name="simple-model")
model = ModelWrapper(onnx_model)
ret = model.analysis(ta.nodes_topologically_sorted)
assert ret["nodes_topologically_sorted"] is False
# test with data/onnx/finn-hls-model/finn-hls-onnx-model.onnx
raw_m = get_data("finn",
"data/onnx/finn-hls-model/finn-hls-onnx-model.onnx")
model = ModelWrapper(raw_m)
ret = model.analysis(ta.nodes_topologically_sorted)
assert ret["nodes_topologically_sorted"] is True
# remove first node and add it at the end
graph = model.graph
first_node = graph.node[0]
graph.node.remove(first_node)
graph.node.append(first_node)
ret = model.analysis(ta.nodes_topologically_sorted)
assert ret["nodes_topologically_sorted"] is False
# test with cnv_w1a1
build_dir = "/tmp/" + os.environ["FINN_INST_NAME"]
cnv = get_test_model_trained("CNV", 1, 1)
bo.export_finn_onnx(cnv, (1, 3, 32, 32),
build_dir + "/end2end_cnv_w1a1_export.onnx")
model = ModelWrapper(build_dir + "/end2end_cnv_w1a1_export.onnx")
ret = model.analysis(ta.nodes_topologically_sorted)
assert ret["nodes_topologically_sorted"] is True
# remove first node and add it at the end
graph = model.graph
first_node = graph.node[0]
graph.node.remove(first_node)
graph.node.append(first_node)
ret = model.analysis(ta.nodes_topologically_sorted)
assert ret["nodes_topologically_sorted"] is False
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_brevitas_mobilenet():
# get single image as input and prepare image
img = Image.open("/workspace/finn/tests/brevitas/king_charles.jpg")
# resize smallest side of the image to 256 pixels and resize larger side
# with same ratio
img = resize_smaller_side(256, img)
# crop central 224*224 window
img = crop_center(224, img)
# save image as numpy array and as torch tensor to enable testing in
# brevitas/pytorch and finn and transpose from (H, W, C) to (C, H, W)
img_np = np.asarray(img).copy().astype(np.float32).transpose(2, 0, 1)
img_np = img_np.reshape(1, 3, 224, 224)
img_torch = torch.from_numpy(img_np).float()
# export preprocess
export_onnx_path = make_build_dir("test_brevitas_mobilenet-v1_")
preproc_onnx = export_onnx_path + "/quant_mobilenet_v1_4b_preproc.onnx"
mean = [0.485, 0.456, 0.406]
std = 0.226
ch = 3
preproc = NormalizePreProc(mean, std, ch)
bo.export_finn_onnx(preproc, (1, 3, 224, 224), preproc_onnx)
preproc_model = ModelWrapper(preproc_onnx)
# set input finn datatype to UINT8
preproc_model.set_tensor_datatype(preproc_model.graph.input[0].name, DataType.UINT8)
preproc_model = preproc_model.transform(InferShapes())
preproc_model = preproc_model.transform(GiveUniqueNodeNames())
preproc_model = preproc_model.transform(GiveUniqueParameterTensors())
preproc_model = preproc_model.transform(GiveReadableTensorNames())
finn_onnx = export_onnx_path + "/quant_mobilenet_v1_4b_exported.onnx"
mobilenet = get_test_model_trained("mobilenet", 4, 4)
bo.export_finn_onnx(mobilenet, (1, 3, 224, 224), finn_onnx)
# 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])
model = ModelWrapper(finn_onnx)
model = model.transform(InferShapes())
model = model.transform(FoldConstants())
model = model.transform(InsertTopK())
# get initializer from Mul that will be absorbed into topk
a0 = model.get_initializer(model.graph.node[-2].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")
model = model.transform(MergeONNXModels(preproc_model))
model.save(export_onnx_path + "/quant_mobilenet_v1_4b.onnx")
idict = {model.graph.input[0].name: img_np}
odict = oxe.execute_onnx(model, idict, True)
produced = odict[model.graph.output[0].name]
produced_prob = odict["TopK_0_out0"] * a0
assert (produced.flatten() == expected_top5).all()
assert np.isclose(produced_prob.flatten(), expected_top5_prob).all()
def test_convert_to_hls_layers_cnv_w1a1(fused_activation):
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(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 = model.transform(InferDataLayouts())
# model.save("golden.onnx")
# load one of the test vectors
fn = pk.resource_filename("finn.qnn-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]
# if we infer thresholding first, all MultiThresholds get converted to HLS
# subsequently, the FC inference will generate passthrough MVAUs
if not fused_activation:
model = model.transform(to_hls.InferThresholdingLayer())
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()
if fused_activation:
assert len(finn_nodes) == 18
else:
assert len(finn_nodes) == 26
thr_nodes = model.get_nodes_by_op_type("Thresholding_Batch")
assert len(thr_nodes) == 8
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)
def test_convert_to_hls_layers_tfc_w1a1():
tfc = get_test_model_trained("TFC", 1, 1)
bo.export_finn_onnx(tfc, (1, 1, 28, 28), export_onnx_path)
model = ModelWrapper(export_onnx_path)
model = model.transform(InferShapes())
model = model.transform(FoldConstants())
model = model.transform(GiveUniqueNodeNames())
model = model.transform(GiveReadableTensorNames())
model = model.transform(Streamline())
model = model.transform(ConvertBipolarMatMulToXnorPopcount())
model = model.transform(absorb.AbsorbAddIntoMultiThreshold())
model = model.transform(absorb.AbsorbMulIntoMultiThreshold())
model = model.transform(RoundAndClipThresholds())
model = model.transform(to_hls.InferBinaryStreamingFCLayer())
fc0 = model.graph.node[2]
assert fc0.op_type == "StreamingFCLayer_Batch"
assert model.get_tensor_shape(fc0.input[0]) == [1, 784]
assert model.get_tensor_shape(fc0.input[1]) == [784, 64]
assert model.get_tensor_shape(fc0.input[2]) == [64, 1]
fc1 = model.graph.node[3]
assert fc1.op_type == "StreamingFCLayer_Batch"
assert model.get_tensor_shape(fc1.input[0]) == [1, 64]
assert model.get_tensor_shape(fc1.input[1]) == [64, 64]
assert model.get_tensor_shape(fc1.input[2]) == [64, 1]
fc2 = model.graph.node[4]
assert fc2.op_type == "StreamingFCLayer_Batch"
assert model.get_tensor_shape(fc2.input[0]) == [1, 64]
assert model.get_tensor_shape(fc2.input[1]) == [64, 64]
assert model.get_tensor_shape(fc2.input[2]) == [64, 1]
fc3 = model.graph.node[5]
assert fc3.op_type == "StreamingFCLayer_Batch"
assert model.get_tensor_shape(fc3.input[0]) == [1, 64]
assert model.get_tensor_shape(fc3.input[1]) == [64, 10]
fc0w = getCustomOp(fc0)
fc0w.set_nodeattr("SIMD", 784)
fc0w.set_nodeattr("PE", 16)
fc1w = getCustomOp(fc1)
fc1w.set_nodeattr("SIMD", 16)
fc1w.set_nodeattr("PE", 16)
fc2w = getCustomOp(fc2)
fc2w.set_nodeattr("SIMD", 16)
fc2w.set_nodeattr("PE", 16)
fc3w = getCustomOp(fc3)
fc3w.set_nodeattr("SIMD", 16)
fc3w.set_nodeattr("PE", 10)
model = model.transform(PrepareCppSim())
model = model.transform(CompileCppSim())
model = model.transform(SetExecMode("cppsim"))
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)}
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 = tfc.forward(input_tensor).detach().numpy()
assert np.isclose(produced, expected, atol=1e-3).all()
os.remove(export_onnx_path)
