def conv2d_leaky_relu_nhwc_oihw_test(
in_shape,
w_shape,
conv_padding,
conv_strides,
conv_dilation,
kernel_layout="OIHW",
targets=["DPUCZDX8G-zcu104"],
) -> None:
for target in targets:
xgraph = _create_conv2d_leaky_relu_nhwc_oihw(
in_shape,
w_shape,
conv_padding,
conv_strides,
conv_dilation,
kernel_layout,
target,
)
def inputs_func(iter):
inputs = np.ones(in_shape, dtype=np.float32)
return {"in1": inputs}
work_dir = os.path.join(FILE_PATH, "work")
build_dir = os.path.join(FILE_PATH, "build")
quantize_func = TARGET_REGISTRY.get_target_quantizer(target)
q_xgraph = quantize_func(xgraph, inputs_func, work_dir=work_dir)
opt_xgraph = px.optimize(q_xgraph, target)
c_xgraph = px.compile(
opt_xgraph, target, work_dir=work_dir, build_dir=build_dir
)
c_output = c_xgraph.get_compiler_output()
assert list(c_output.keys()) == ["xp0"]
assert c_output.get_in_map("xp0") == {"xinput0": "xinput0:0"}
assert c_output.get_out_map("xp0") == {"lr1": "lr1:0"}
assert len(c_output.get_code_files("xp0")) == 1
shutil.rmtree(work_dir)
shutil.rmtree(build_dir)
def xcompiler_conv2d_leaky_relu_nhwc_oihw_test(
in_shape,
w_shape,
conv_padding,
conv_strides,
conv_dilation,
kernel_layout="OIHW",
targets=["DPUCZDX8G-zcu104"],
expected_nb_subgraphs=3,
) -> None:
for target in targets:
xgraph = _create_conv2d_leaky_relu_nhwc_oihw(
in_shape,
w_shape,
conv_padding,
conv_strides,
conv_dilation,
kernel_layout,
target,
)
def inputs_func(iter):
inputs = np.ones(in_shape, dtype=np.float32)
return {"in1": inputs}
work_dir = os.path.join(FILE_PATH, "work")
build_dir = os.path.join(FILE_PATH, "build")
quantize_func = TARGET_REGISTRY.get_target_quantizer(target)
q_xgraph = quantize_func(xgraph, inputs_func, work_dir=work_dir)
opt_xgraph = px.optimize(q_xgraph, target)
c_xgraph = px.compile(opt_xgraph,
target,
work_dir=work_dir,
build_dir=build_dir)
g = xir.Graph.deserialize(os.path.join(build_dir, "xp0.xmodel"))
# TODO subgraphs[1].get_attr("device") -> *** RuntimeError: bad any_cast
subgraphs = get_child_subgraphs(g)
assert (len(subgraphs) == expected_nb_subgraphs
), "Expected {0} subgraphs but got: {1}".format(
expected_nb_subgraphs, len(subgraphs))
bn_handling=BatchNormHandling.MERGE_AND_QUANTIZE)
# Finetune the model
# . . .
# Export to ONNX
onnx_filename = 'dpuv2_resnet18.onnx'
export_dpuv2_onnx(model,
input_shape=IN_SIZE,
input_t=inp,
export_path=onnx_filename)
# Load ONNX into PyXIR
onnx_model = onnx.load(onnx_filename)
xgraph = from_onnx(onnx_model)
xgraph = pyxir.partition(xgraph, [target])
xgraph = pyxir.optimize(xgraph, target)
work_dir = os.path.join(file_dir, f'{target}_quant_trained_resnet18_workdir')
inputs = np.random.randn(*IN_SIZE)
def inputs_func(iter):
return {'inp.1': inputs}
xgraph = pyxir.quantize(xgraph, target, inputs_func, work_dir=work_dir)
pyxir.build(xgraph,
target,
work_dir=work_dir,
build_dir=work_dir,
runtime='cpu-np')
def xcompiler_resnetv1_block_test(
in_shape,
pool_size,
pool_strides,
w1_shape,
w2_shape,
w3_shape,
w4_shape,
c1_padding=[0, 0, 0, 0],
c2_padding=[0, 0, 0, 0],
c3_padding=[0, 0, 0, 0],
c4_padding=[0, 0, 0, 0],
c1_strides=[1, 1],
c2_strides=[1, 1],
c3_strides=[1, 1],
c4_strides=[1, 1],
c1_dilation=[1, 1],
c2_dilation=[1, 1],
c3_dilation=[1, 1],
c4_dilation=[1, 1],
kernel_layout="OIHW",
target="DPUCAHX8H-u50",
expected_nb_subgraphs=3,
):
xgraph = _create_resnetv1_block(
in_shape,
pool_size,
pool_strides,
w1_shape,
w2_shape,
w3_shape,
w4_shape,
c1_padding,
c2_padding,
c3_padding,
c4_padding,
c1_strides,
c2_strides,
c3_strides,
c4_strides,
c1_dilation,
c2_dilation,
c3_dilation,
c4_dilation,
kernel_layout,
target,
)
def inputs_func(iter):
inputs = np.ones(in_shape, dtype=np.float32)
return {"in1": inputs}
work_dir = os.path.join(FILE_PATH, "work")
build_dir = os.path.join(FILE_PATH, "build")
quantize_func = TARGET_REGISTRY.get_target_quantizer(target)
q_xgraph = quantize_func(xgraph, inputs_func, work_dir=work_dir)
opt_xgraph = px.optimize(q_xgraph, target)
c_xgraph = px.compile(opt_xgraph, target, work_dir=work_dir, build_dir=build_dir)
c_output = c_xgraph.get_compiler_output()
g = xir.Graph.deserialize(os.path.join(build_dir, "xp0.xmodel"))
# TODO subgraphs[1].get_attr("device") -> *** RuntimeError: bad any_cast
subgraphs = get_child_subgraphs(g)
assert (
len(subgraphs) == expected_nb_subgraphs
), "Expected {0} subgraphs but got: {1}".format(
expected_nb_subgraphs, len(subgraphs)
)
shutil.rmtree(work_dir)
shutil.rmtree(build_dir)
def xcompiler_conv2d_pool2d_nhwc_oihw_test(
in_shape,
w_shape,
conv_padding,
conv_strides,
conv_dilation,
pool_type,
pool_size,
pool_padding=[0, 0],
pool_strides=[1, 1],
conv_groups=1,
conv_invalid=False,
kernel_layout="OIHW",
targets=["DPUCAHX8H-u50"],
expected_nb_subgraphs=3,
):
for target in targets:
xgraph = _create_conv2d_pool2d_nhwc_oihw(
in_shape,
w_shape,
conv_padding,
conv_strides,
conv_dilation,
pool_type,
pool_size,
pool_padding,
pool_strides,
conv_groups,
conv_invalid,
kernel_layout,
target,
)
def inputs_func(iter):
inputs = np.ones(in_shape, dtype=np.float32)
return {"in1": inputs}
work_dir = os.path.join(FILE_PATH, "work")
build_dir = os.path.join(FILE_PATH, "build")
quantize_func = TARGET_REGISTRY.get_target_quantizer(target)
q_xgraph = quantize_func(xgraph, inputs_func, work_dir=work_dir)
opt_xgraph = px.optimize(q_xgraph, target)
c_xgraph = px.compile(
opt_xgraph, target, work_dir=work_dir, build_dir=build_dir
)
c_output = c_xgraph.get_compiler_output()
assert list(c_output.keys()) == ["xp0"]
assert c_output.get_in_map("xp0") == {"xinput0": "xinput0"}
assert c_output.get_out_map("xp0") == {"pool1": "pool1"}
assert len(c_output.get_code_files("xp0")) == 1
g = xir.Graph.deserialize(os.path.join(build_dir, "xp0.xmodel"))
# TODO subgraphs[1].get_attr("device") -> *** RuntimeError: bad any_cast
subgraphs = get_child_subgraphs(g)
assert len(subgraphs) == expected_nb_subgraphs
dpu_subgraph = subgraphs[1]
# import pdb; pdb.set_trace()
# assert len(dpu_subgraph.get_children()) == 3
shutil.rmtree(work_dir)
shutil.rmtree(build_dir)
def prequantize_onnx_model(onnx_model, target, inputs_func, out_file,
**kwargs):
xgraph = _from_onnx(onnx_model)
xgraph = px.partition(xgraph, [target])
xgraph = px.optimize(xgraph, target)
q_xgraph = px.quantize(xgraph, target, inputs_func, **kwargs)
# Move quant_info information from XGraph to ONNX model
tensor_quant_info = {}
for X in q_xgraph.get_layers():
if "vai_quant" in X.attrs and X.attrs["vai_quant"] != []:
tensor_name = X.attrs['onnx_id']
if tensor_name in tensor_quant_info:
raise NotImplementedError("Quantization for ONNX tensor: {}"
" already provided. Merging of"
" multiple tensor quantization info"
" parameters not supported yet")
tensor_quant_info[tensor_name] = {
"vai_quant": X.attrs["vai_quant"]
}
for vai_quant_elem in X.attrs['vai_quant']:
tensor_quant_info[tensor_name][vai_quant_elem] = \
X.attrs[vai_quant_elem]
for node in onnx_model.graph.node:
node_w = NodeWrapper(node)
tensor_name = node_w.get_outputs()[0]
if tensor_name in tensor_quant_info:
node_w.add_attribute(
'vai_quant',
list(tensor_quant_info[tensor_name]['vai_quant']),
'STRINGS'
)
for vai_quant_elem in tensor_quant_info[tensor_name]['vai_quant']:
node_w.add_attribute(
vai_quant_elem,
tensor_quant_info[tensor_name][vai_quant_elem],
'INTS'
)
# q_output = q_xgraph.get_quantizer_output()
# for qkey in q_output.keys():
# quant_file = q_output.get_q_file(qkey)
# quant_info_file = q_output.get_q_info(qkey)
# quant_orig_pb = q_output.get_orig_pb(qkey)
# if not os.path.isfile(quant_info_file):
# raise ValueError("quant file: {} for qkey: {} does not exist"
# .format(quant_info_file, qkey))
# meta = onnx_model.metadata_props.add()
# meta.key = "vitis_ai_quant--q_file--" + qkey
# meta.value = str(quant_file)
# meta = onnx_model.metadata_props.add()
# meta.key = "vitis_ai_quant--q_info--" + qkey
# meta.value = str(quant_info_file)
# meta = onnx_model.metadata_props.add()
# meta.key = "vitis_ai_quant--orig_pb--" + qkey
# meta.value = str(quant_orig_pb)
onnx.save(onnx_model, out_file)