def analyze_args(self, func):
sig = inspect.signature(func)
# TODO: replace with https://docs.python.org/3.4/library/inspect.html#inspect.getfullargspec
argspec = inspect.getargspec(
func) # TODO: Doesn't support keyword only args.
parameter_count = 0
for k, v in sig.parameters.items():
# TODO improve it
if k == 'kwargs':
continue
parameter_count += 1
isSelfRemoved = parameter_count != len(argspec[0])
if isSelfRemoved:
self.add_arg(argspec[0][0], None)
for k, v in sig.parameters.items():
# TODO improve it
if k == 'kwargs':
continue
self.add_arg(v.name, values.parse_instance(None, v.name,
v.default))
def __init__(self, module: 'Field', inst):
super().__init__(module, inst, None)
self.is_chainer_link = True
self.children = []
for child in inst.children():
child_ = values.parse_instance(module, '', child, inst)
self.children.append(child_)
def analyze_args(self, func):
sig = inspect.signature(func)
argspec = inspect.getfullargspec(func)
parameter_count = 0
for k, v in sig.parameters.items():
# TODO improve it
if k == 'kwargs':
continue
parameter_count += 1
isSelfRemoved = parameter_count != len(argspec.args) + len(argspec.kwonlyargs)
if isSelfRemoved:
self.add_arg(argspec.args[0], None)
for k, v in sig.parameters.items():
# TODO improve it
if k == 'kwargs':
continue
self.add_arg(v.name, values.parse_instance(None, v.name, v.default))
def convert_model(model: 'chainer.Chain', args=[]):
# reset values
values.reset_field_and_attributes()
utils.reset_guid()
values.function_converters.clear()
values.builtin_function_converters.clear()
values.instance_converters.clear()
def instance_converter(m, i):
if links_builtin.is_builtin_chainer_link(i):
return links_builtin.ChainerLinkInstance(m, i)
if isinstance(i, chainer.ChainList):
module = values.Object(values.ModuleValue(sys.modules[i.__module__]))
return links_builtin.ChainerChainListInstance(module, i)
if isinstance(i, chainer.Link):
module = values.Object(values.ModuleValue(sys.modules[i.__module__]))
return links_builtin.ChainerChainInstance(module, i)
return None
values.instance_converters.append(instance_converter)
custom_functions_module = values.Object(values.ModuleValue(custom_functions))
# onnx
functions_onnx_module = values.Object(values.ModuleValue(functions_onnx))
def ret_same(funcArgs):
return functions.generate_value_with_same_type(funcArgs.keywords['x'].get_value())
values.function_converters[functions_onnx.onnx_abs] = values.FuncValue(functions_builtin.ChainerFunction(functions_onnx.onnx_abs, ret_value_func=ret_same), None, module=functions_onnx_module)
# chainer
c_variable = values.FuncValue(functions_ndarray.NDArrayFunction(), None)
values.function_converters[chainer.Variable] = c_variable
# chainer.functions
def add_chainer_function(func, ret_value_func = None):
if ret_value_func is None:
f = values.FuncValue(
functions_builtin.ChainerFunction(func), None)
else:
f = values.FuncValue(
functions_builtin.ChainerFunction(func, ret_value_func=ret_value_func), None)
values.function_converters[func] = f
def ret_tuple(funcArgs = None):
ret = values.TupleValue()
ret.vtype = values.TensorValue
return ret
# register unsupported functions to show error when unsupported functions are called
for f in F.__dict__.items():
if inspect.isfunction(f[1]):
values.function_converters[f[1]] = values.FuncValue(functions.UnimplementedFunction(f[1]), None)
# activation
add_chainer_function(F.elu)
add_chainer_function(F.leaky_relu)
add_chainer_function(F.log_softmax)
add_chainer_function(F.relu)
add_chainer_function(F.selu)
add_chainer_function(F.sigmoid)
add_chainer_function(F.softmax)
add_chainer_function(F.tanh)
add_chainer_function(F.softmax_cross_entropy)
add_chainer_function(F.pad_sequence)
add_chainer_function(F.average_pooling_2d)
add_chainer_function(F.unpooling_2d)
add_chainer_function(F.reshape)
add_chainer_function(F.transpose)
add_chainer_function(F.split_axis, ret_value_func=ret_tuple)
add_chainer_function(F.hstack)
add_chainer_function(F.vstack)
add_chainer_function(F.stack)
add_chainer_function(F.separate, ret_value_func=ret_tuple)
add_chainer_function(F.squeeze)
add_chainer_function(F.swapaxes)
add_chainer_function(F.dropout)
add_chainer_function(F.concat)
add_chainer_function(F.matmul)
add_chainer_function(F.max_pooling_2d)
add_chainer_function(F.resize_images)
add_chainer_function(F.broadcast_to)
add_chainer_function(F.expand_dims)
add_chainer_function(F.local_response_normalization)
add_chainer_function(F.mean)
add_chainer_function(F.average)
add_chainer_function(F.sum)
add_chainer_function(F.maximum)
add_chainer_function(F.minimum)
add_chainer_function(F.max)
add_chainer_function(F.min)
values.function_converters[F.absolute] = values.FuncValue(functions.UserDefinedFunction(custom_functions.chainer_absolute), None, module=custom_functions_module)
add_chainer_function(F.sin)
add_chainer_function(F.sinh)
add_chainer_function(F.sign)
add_chainer_function(F.cos)
add_chainer_function(F.cosh)
add_chainer_function(F.tan)
add_chainer_function(F.tanh)
add_chainer_function(F.arcsin)
add_chainer_function(F.arccos)
add_chainer_function(F.arctan)
add_chainer_function(F.exp)
add_chainer_function(F.log)
add_chainer_function(F.sqrt)
add_chainer_function(F.clip)
values.function_converters[F.argmax] = values.FuncValue(functions_builtin.ChainerArgminmaxFunction(F.argmax), None)
values.function_converters[F.argmin] = values.FuncValue(functions_builtin.ChainerArgminmaxFunction(F.argmin), None)
values.function_converters[F.clipped_relu] = values.FuncValue(functions.UserDefinedFunction(custom_functions.chainer_clipped_relu), None, module=custom_functions_module)
if int(chainer.__version__[0]) >= 6:
add_chainer_function(F.roi_max_pooling_2d)
add_chainer_function(F.roi_average_pooling_2d)
add_chainer_function(F.roi_max_align_2d)
add_chainer_function(F.roi_average_align_2d)
# numpy
f_array = values.FuncValue(functions_ndarray.NDArrayFunction(), None)
f_zeros = values.FuncValue(functions_ndarray.NDArrayZerosFunction(), None)
f_full = values.FuncValue(functions_ndarray.NDArrayFullFunction(), None)
f_ceil = values.FuncValue(functions_ndarray.NDArrayCeilFunction(), None)
f_cumsum = values.FuncValue(functions_ndarray.NDArrayCumsumFunction(), None)
f_maximum = values.FuncValue(functions_ndarray.NDArrayChainerFunction(functions_ndarray.dummy_maximum), None)
f_minimum = values.FuncValue(functions_ndarray.NDArrayChainerFunction(functions_ndarray.dummy_minimum), None)
f_argmax = values.FuncValue(functions_ndarray.NDarrayArgminmaxFunction(functions_ndarray.dummy_argmax), None)
f_argmin = values.FuncValue(functions_ndarray.NDarrayArgminmaxFunction(functions_ndarray.dummy_argmin), None)
f_round = values.FuncValue(functions_ndarray.NDarrayRoundFunction(functions_ndarray.dummy_round), None)
f_sqrt = values.FuncValue(functions_ndarray.NDarraySqrtFunction(functions_ndarray.dummy_sqrt), None)
f_stack = values.FuncValue(functions_ndarray.NDarrayStackFunction(functions_ndarray.dummy_stack), None)
f_reshape = values.FuncValue(functions_ndarray.NDarrayReshapeFunction(functions_ndarray.dummy_reshape), None)
f_transpose = values.FuncValue(functions_ndarray.NDarrayTransposeFunction(functions_ndarray.dummy_transpose), None)
f_int32 = values.FuncValue(functions_ndarray.NDArrayInt32(), None)
f_float32 = values.FuncValue(functions_ndarray.NDArrayFloat32(), None)
values.function_converters[np.array] = f_array
values.function_converters[np.zeros] = f_zeros
values.function_converters[np.full] = f_full
values.function_converters[np.ceil] = f_ceil
values.function_converters[np.cumsum] = f_cumsum
values.function_converters[np.int32] = f_int32
values.function_converters[np.float32] = f_float32
values.function_converters[np.maximum] = f_maximum
values.function_converters[np.minimum] = f_minimum
values.function_converters[np.argmax] = f_argmax
values.function_converters[np.argmin] = f_argmin
values.function_converters[np.round] = f_round
values.function_converters[np.sqrt] = f_sqrt
values.function_converters[np.stack] = f_stack
values.function_converters[np.reshape] = f_reshape
values.function_converters[np.transpose] = f_transpose
values.function_converters[np.clip] = values.FuncValue(functions.UserDefinedFunction(custom_functions.numpy_clip), None, module=custom_functions_module)
values.function_converters[np.absolute] = values.FuncValue(functions.UserDefinedFunction(custom_functions.numpy_absolute), None, module=custom_functions_module)
values.function_converters[custom_functions.check_attribute_value] = values.FuncValue(functions.CheckAttributeValueFunction(), None, module=custom_functions_module)
values.function_converters[custom_functions.check_attribute_scalar] = values.FuncValue(functions.CheckAttributeScalarFunction(), None, module=custom_functions_module)
values.builtin_function_converters['abs'] = values.FuncValue(functions.UserDefinedFunction(custom_functions.builtin_absolute), None, module=custom_functions_module)
m_range = values.FuncValue(functions_builtin.RangeFunction(), None)
values.builtin_function_converters['range'] = m_range
m_len = values.FuncValue(functions_builtin.LenFunction(), None)
values.builtin_function_converters['len'] = m_len
values.function_converters[six.moves.range] = m_range
m_list = values.FuncValue(functions_builtin.ListFunction(), None)
values.builtin_function_converters['list'] = m_list
m_print = values.FuncValue(functions_builtin.PrintFunction(), None)
values.builtin_function_converters['print'] = m_print
m_getattr = values.FuncValue(functions_builtin.GetAttrFunction(), None)
values.builtin_function_converters['getattr'] = m_getattr
m_hasattr = values.FuncValue(functions_builtin.HasAttrFunction(), None)
values.builtin_function_converters['hasattr'] = m_hasattr
m_to_gpu = values.FuncValue(functions_builtin.CopyFunction(cuda.to_gpu), None)
values.function_converters[cuda.to_gpu] = m_to_gpu
m_to_cpu = values.FuncValue(functions_builtin.CopyFunction(cuda.to_cpu), None)
values.function_converters[cuda.to_cpu] = m_to_cpu
# generate VEvalFlag functions
def add_veval_flag_function(name:'str', func):
f = values.FuncValue(functions_builtin.VEvalContextFunction(func), None)
values.builtin_function_converters[name] = f
add_veval_flag_function('eval_as_written_target', flags.eval_as_written_target)
add_veval_flag_function('ignore_branch', flags.ignore_branch)
add_veval_flag_function('for_unroll', flags.for_unroll)
# generate default module
default_module = values.Object(values.ModuleValue(sys.modules[model.__module__]))
model_inst = values.parse_instance(default_module, '', model)
forward_func = model_inst.try_get_and_store_obj('forward', None)
# convert args
finput = functions.FunctionArgInput()
value_args = []
ind = 0
node_input = nodes.NodeInput('input')
for arg in args:
varg = values.parse_instance(default_module, '', arg, None)
varg.name = 'in_' + str(ind)
varg.get_value().name = 'in_' + str(ind)
# make value unknown
varg.get_value().internal_value = None
finput.inputs.append(varg)
value_args.append(varg.get_value())
ind += 1
node_input.set_outputs(value_args)
graph = Graph()
graph.root_graph = graph
graph.add_node(node_input)
forward_func_value = forward_func.get_value()
ret = forward_func_value.func.vcall(
default_module, graph, forward_func_value.obj, finput).get_obj()
assert(ret is None or isinstance(ret, values.Object))
def try_get_value(value) -> 'values.Value':
if isinstance(value, values.Value):
return value
if isinstance(value, values.Object):
return value.get_value()
if isinstance(value, values.Attribute):
return value.get_obj().get_value()
if ret is None or isinstance(ret, values.NoneValue):
if config.show_warnings:
print('Failed to compile. output is None.')
return (value_args, None, graph)
ret_ = []
if isinstance(ret.get_value(), values.TupleValue):
if ret.get_value().internal_value is not None:
for v in ret.get_value().internal_value:
assert(v is not None)
ret_.append(try_get_value(v))
else:
ret_ = [ret.get_value()]
elif isinstance(ret, list):
ret_ = [r.get_value() for r in ret]
else:
ret_ = [ret.get_value()]
for v in value_args:
graph.add_input_value(v)
for v in ret_:
graph.add_output_value(v)
return (value_args, ret_, graph)
def convert_model(model: 'chainer.Chain', args=[]):
# reset values
values.reset_field_and_attributes()
utils.reset_guid()
values.function_converters.clear()
values.builtin_function_converters.clear()
values.instance_converters.clear()
def instance_converter(m, i):
if links_builtin.is_builtin_chainer_link(i):
return links_builtin.ChainerLinkInstance(m, i)
if isinstance(i, chainer.ChainList):
module = values.ValueRef(
values.ModuleValue(sys.modules[i.__module__]))
return links_builtin.ChainerChainListInstance(module, i)
if isinstance(i, chainer.Link):
module = values.ValueRef(
values.ModuleValue(sys.modules[i.__module__]))
return links_builtin.ChainerChainInstance(module, i)
return None
values.instance_converters.append(instance_converter)
# chainer
c_variable = values.FuncValue(functions_ndarray.NDArrayFunction(), None)
values.function_converters[chainer.Variable] = c_variable
# chainer.functions
def add_chainer_funtion(name: 'str', func, ret_value_func=None):
if ret_value_func is None:
f = values.FuncValue(functions_builtin.ChainerFunction(func), None)
else:
f = values.FuncValue(
functions_builtin.ChainerFunction(
func, ret_value_func=ret_value_func), None)
values.function_converters[func] = f
def ret_tuple():
ret = values.TupleValue()
ret.vtype = values.TensorValue
return ret
add_chainer_funtion('relu', F.relu)
add_chainer_funtion('softmax', F.softmax)
add_chainer_funtion('softmax_cross_entropy', F.softmax_cross_entropy)
add_chainer_funtion('pad_sequence', F.pad_sequence)
add_chainer_funtion('average_pooling_2d', F.average_pooling_2d)
add_chainer_funtion('unpooling_2d', F.unpooling_2d)
add_chainer_funtion('reshape', F.reshape)
add_chainer_funtion('split_axis', F.split_axis, ret_value_func=ret_tuple)
add_chainer_funtion('hstack', F.hstack)
add_chainer_funtion('vstack', F.vstack)
add_chainer_funtion('stack', F.stack)
add_chainer_funtion('separate', F.separate, ret_value_func=ret_tuple)
add_chainer_funtion('squeeze', F.squeeze)
add_chainer_funtion('swapaxes', F.swapaxes)
add_chainer_funtion('dropout', F.dropout)
add_chainer_funtion('concat', F.concat)
add_chainer_funtion('matmul', F.matmul)
add_chainer_funtion('max_pooling_2d', F.max_pooling_2d)
add_chainer_funtion('resize_images', F.resize_images)
add_chainer_funtion('tanh', F.tanh)
add_chainer_funtion('sigmoid', F.sigmoid)
add_chainer_funtion('broadcast_to', F.broadcast_to)
add_chainer_funtion('expand_dims', F.expand_dims)
add_chainer_funtion('local_response_normalization',
F.local_response_normalization)
add_chainer_funtion('mean', F.mean)
add_chainer_funtion('average', F.average)
add_chainer_funtion('sum', F.sum)
if int(chainer.__version__[0]) >= 6:
add_chainer_funtion('roi_max_pooling_2d', F.roi_max_pooling_2d)
add_chainer_funtion('roi_average_pooling_2d', F.roi_average_pooling_2d)
add_chainer_funtion('roi_max_align_2d', F.roi_max_align_2d)
add_chainer_funtion('roi_average_align_2d', F.roi_average_align_2d)
# numpy
f_array = values.FuncValue(functions_ndarray.NDArrayFunction(), None)
f_zeros = values.FuncValue(functions_ndarray.NDArrayZerosFunction(), None)
f_full = values.FuncValue(functions_ndarray.NDArrayFullFunction(), None)
f_ceil = values.FuncValue(functions_ndarray.NDArrayCeilFunction(), None)
f_cumsum = values.FuncValue(functions_ndarray.NDArrayCumsumFunction(),
None)
f_int32 = values.FuncValue(functions_ndarray.NDArrayInt32(), None)
f_float32 = values.FuncValue(functions_ndarray.NDArrayFloat32(), None)
values.function_converters[np.array] = f_array
values.function_converters[np.zeros] = f_zeros
values.function_converters[np.full] = f_full
values.function_converters[np.ceil] = f_ceil
values.function_converters[np.cumsum] = f_cumsum
values.function_converters[np.int32] = f_int32
values.function_converters[np.float32] = f_float32
m_range = values.FuncValue(functions_builtin.RangeFunction(), None)
values.builtin_function_converters['range'] = m_range
m_len = values.FuncValue(functions_builtin.LenFunction(), None)
values.builtin_function_converters['len'] = m_len
values.function_converters[six.moves.range] = m_range
m_list = values.FuncValue(functions_builtin.ListFunction(), None)
values.builtin_function_converters['list'] = m_list
m_to_gpu = values.FuncValue(functions_builtin.CopyFunction(cuda.to_gpu),
None)
values.function_converters[cuda.to_gpu] = m_to_gpu
m_to_cpu = values.FuncValue(functions_builtin.CopyFunction(cuda.to_cpu),
None)
values.function_converters[cuda.to_cpu] = m_to_cpu
# generate default module
default_module = values.ValueRef(
values.ModuleValue(sys.modules[model.__module__]))
model_inst = values.parse_instance(default_module, '', model)
forward_func = model_inst.try_get_and_store_obj('forward', None)
# convert args
finput = functions.FunctionArgInput()
value_args = []
ind = 0
node_input = nodes.NodeInput('input')
for arg in args:
varg = values.parse_instance(default_module, '', arg, None)
varg.name = 'in_' + str(ind)
varg.get_value().name = 'in_' + str(ind)
# make value unknown
varg.get_value().internal_value = None
finput.inputs.append(varg)
value_args.append(varg.get_value())
ind += 1
node_input.set_outputs(value_args)
graph = Graph()
graph.root_graph = graph
graph.add_node(node_input)
forward_func_value = forward_func.get_value()
ret = forward_func_value.func.vcall(default_module, graph,
forward_func_value.obj, finput)
assert (ret is None or isinstance(ret, values.ValueRef))
def try_get_value(value) -> 'values.Value':
if isinstance(value, values.Value):
return value
if isinstance(value, values.ValueRef):
return value.get_value()
if isinstance(value, values.Attribute):
return value.get_ref().get_value()
if ret is None or isinstance(ret, values.NoneValue):
if config.show_warnings:
print('Failed to compile. output is None.')
return (value_args, None, graph)
ret_ = []
if isinstance(ret.get_value(), values.TupleValue):
if ret.get_value().internal_value is not None:
for v in ret.get_value().internal_value:
assert (v is not None)
ret_.append(try_get_value(v))
else:
ret_ = [ret.get_value()]
elif isinstance(ret, list):
ret_ = [r.get_value() for r in ret]
else:
ret_ = [ret.get_value()]
for v in value_args:
graph.add_input_value(v)
for v in ret_:
graph.add_output_value(v)
return (value_args, ret_, graph)