def vcall(self,
module: 'values.Field',
graph: 'graphs.Graph',
inst: 'values.Object',
args: 'functions.FunctionArgInput',
context: 'functions.VEvalContext' = None,
line=-1):
func_args = self.args.merge_inputs(inst, args)
name = func_args.get_value().get_value(key='name')
obj = func_args.keywords['object']
attr = obj.get_field().get_attribute(name.internal_value,
graph.root_graph, False)
# property(getter)
if attr.has_obj() and isinstance(attr.get_obj().get_value(
), values.FuncValue) and attr.get_obj().get_value().func.is_property:
func_value = attr.get_obj().get_value()
ret = func_value.func.vcall(func_value.module, graph,
func_value.obj,
functions.FunctionArgInput(), context,
utils.LineProperty(line))
return ret
if attr.has_obj():
return attr
# if attr is not found
gotten_obj = obj.try_get_and_store_obj(name.internal_value,
graph.root_graph)
if gotten_obj is not None:
return obj.get_field().get_attribute(name.internal_value,
graph.root_graph, False)
return None
def veval_ast_attribute(astc: 'AstContext',
local_field: 'values.Field',
graph: 'Graph',
option: 'VEvalOption' = None) -> 'Attribute':
assert (isinstance(astc.nast, gast.gast.Attribute))
lineprop = utils.LineProperty(astc.lineno, astc.filename)
from_module = True
if option is not None and option.eval_as_written_target:
from_module = False
value = veval_ast(astc.c(astc.nast.value), local_field, graph)
value_ref = try_get_ref(value, 'attribute', lineprop)
if (value_ref is None):
utils.print_warning(
'Unknown or disabled attribute "{}" is accessed'.format(
get_ast_name_forcibly(astc.nast.value)), lineprop)
return None
attr = value_ref.get_field().get_attribute(astc.nast.attr,
graph.root_graph, from_module)
# property(getter)
if attr.has_obj() and isinstance(
attr.get_ref().get_value(),
values.FuncValue) and attr.get_ref().get_value().func.is_property:
func_value = attr.get_ref().get_value()
ret = func_value.func.vcall(func_value.module, graph, func_value.obj,
functions.FunctionArgInput(), lineprop)
return ret
if attr.has_obj():
return attr
# if attr is not found
gotten_obj = value_ref.try_get_and_store_obj(astc.nast.attr,
graph.root_graph)
if gotten_obj is not None:
return value_ref.get_field().get_attribute(astc.nast.attr,
graph.root_graph,
from_module)
if option is not None and option.eval_as_written_target:
return attr
# value is unknown
if value is None:
utils.print_warning(
'Assigning value {} is not found'.format(
get_ast_name_forcibly(astc.nast.value)), lineprop)
else:
utils.print_warning(
'Assigning value {} is not found'.format(
get_ast_name_forcibly(astc.nast.attr)), lineprop)
return None
def veval_ast_call(astc: 'AstContext', local_field: 'values.Field',
graph: 'Graph') -> 'Attribute':
assert (isinstance(astc.nast, gast.gast.Call))
lineprop = utils.LineProperty(astc.lineno, astc.filename)
func = veval_ast(astc.c(astc.nast.func), local_field, graph)
if func == None or not func.has_obj():
utils.print_warning(
'Unknown function "{}" is called'.format(
get_ast_name_forcibly(astc.nast.func)), lineprop)
return None
func_obj = try_get_ref(func, 'call', lineprop)
func_value = try_get_value(func, 'call', lineprop)
finput = functions.FunctionArgInput()
for arg in astc.nast.args:
arg_ = veval_ast(astc.c(arg), local_field, graph)
finput.inputs.append(try_get_ref(arg_, 'call', lineprop))
for keyword in astc.nast.keywords:
arg_ = veval_ast(astc.c(keyword.value), local_field, graph)
finput.keywords[keyword.arg] = try_get_ref(arg_, 'call', lineprop)
lineprop = utils.LineProperty(astc.lineno, astc.filename)
# check arguments
for o in finput.inputs:
if o is None:
utils.print_warning(
'Invalid arguments exists in "{}"'.format(
get_ast_name_forcibly(astc.nast.func)), lineprop)
return None
ret = None
if isinstance(func_value, values.FuncValue):
ret = func_value.func.vcall(func_value.module, graph, func_value.obj,
finput, lineprop)
return ret
elif isinstance(func_value, values.Instance):
# __call__
call_func_ref = func_obj.try_get_and_store_obj('__call__',
graph.root_graph)
if call_func_ref is not None:
func_value = call_func_ref.get_value()
ret = func_value.func.vcall(func_value.module, graph, func_obj,
finput, lineprop)
return ret
if config.show_warnings:
print('Unknown function is called in L.{}'.format(astc.lineno))
return None
def vcall(self,
module: 'values.Field',
graph: 'graphs.Graph',
inst: 'values.Object',
args: 'functions.FunctionArgInput',
option: 'vevaluator.VEvalContext' = None,
line=-1):
args = functions.FunctionArgInput()
args.inputs.append(inst)
args.keywords['self'] = inst
value = values.ListValue(self.owner.children)
return values.Object(value)
def vcall(self,
module: 'Field',
graph: 'Graph',
inst: 'values.ValueRef',
args: 'functions.FunctionArgInput',
line=-1):
args = functions.FunctionArgInput()
args.inputs.append(inst)
args.keywords['self'] = inst
node = nodes.NodeCall(self, args, line)
value = values.TupleValue()
value.name = '@F.{}.{}'.format(line, self.name)
node.set_outputs([value])
graph.add_node(node)
return values.ValueRef(value)
def vcall(self,
module: 'values.Field',
graph: 'graphs.Graph',
inst: 'values.Object',
args: 'functions.FunctionArgInput',
context: 'functions.VEvalContext' = None,
line=-1):
args = functions.FunctionArgInput()
args.inputs.append(inst)
args.keywords['self'] = inst
node = nodes.NodeCall(self, args, line)
value = values.NumberValue(None)
value.dtype = np.array(0).dtype
value.name = '@F.{}.{}'.format(line, self.name)
node.set_outputs([value])
graph.add_node(node)
return values.Object(value)
def veval_ast_listcomp(astc: 'AstContext', local_field: 'values.Field',
graph: 'Graph'):
'''
Ex. [x for x in xx]
[elt for target in iter]
'''
assert (isinstance(astc.nast, gast.gast.ListComp))
lineprop = utils.LineProperty(astc.lineno, astc.filename)
listcomp_guid = str(utils.get_guid())
listcomp_id = 'listcomp_' + listcomp_guid
body_id = 'listcomp_body_' + listcomp_guid
internal_counter_id = '@internal/listcomp_counter_' + listcomp_guid
internal_list_id = '@internal/listcomp_list_' + listcomp_guid
internal_cond_id = '@internal/listcomp_cond_' + listcomp_guid
generator = astc.nast.generators[0]
iter_value = try_get_value(
veval_ast(astc.c(generator.iter), local_field, graph), 'generator',
lineprop)
list_value = values.ListValue()
list_obj = values.ValueRef(list_value)
node_generate_list = nodes.NodeGenerate('List', [], lineprop)
node_generate_list.set_outputs([list_value])
graph.add_node(node_generate_list)
# body
target_name = ''
if isinstance(generator.target, gast.gast.Name):
target_name = generator.target.id
else:
if config.show_warnings:
print('This for is not supported. in L.{}'.format(astc.lineno))
return None
counter_value = values.NumberValue(None)
counter_value.dtype = np.array(0).dtype
counter_value.name = internal_counter_id
cond_value = values.BoolValue(None)
cond_value.name = internal_cond_id
# set values with internal name
local_field.get_attribute(internal_list_id).revise(list_obj)
values.push_history(listcomp_id)
body_graph = Graph()
body_graph.root_graph = graph.root_graph
body_graph.name = 'Body_' + listcomp_guid
node_forgen = nodes.NodeForGenerator(counter_value, iter_value)
target_ref = iter_value.get_iterator()
if target_ref is None:
target_ref = values.ValueRef(values.UnknownValue())
if config.show_warnings:
print('unknown iteratable type in L.{}'.format(lineprop))
target_value = target_ref.get_value()
node_forgen.set_outputs([target_ref.get_value()])
local_field.get_attribute(target_name).revise(target_ref)
body_graph.add_node(node_forgen)
elt = veval_ast(astc.c(astc.nast.elt), local_field, body_graph)
elt_obj = try_get_ref(elt, 'listcomp', lineprop)
finput = functions.FunctionArgInput()
finput.inputs.append(elt_obj)
append_value = local_field.get_attribute(internal_list_id).get_ref(
).get_field().get_attribute('append').get_ref().get_value()
append_value.func.vcall(
None, body_graph,
local_field.get_attribute(internal_list_id).get_ref(), finput,
lineprop)
value_inputs = values.get_inputs()
value_outputs = values.get_outputs()
values.pop_history()
inputs = []
outputs = []
# default input for subgraph's input
body_graph.add_input_value(counter_value)
body_graph.add_input_value(cond_value)
body_graph.add_input_value(iter_value)
# default output for subgraph's output
body_graph.add_output_value(cond_value)
body_graph.add_output_value(iter_value)
# default output
outputs.append(functions.generate_value_with_same_type(iter_value))
# generate pairs
value_pairs = {}
for v in value_inputs:
key = str(v.field.id) + '_' + v.name
if not (key in value_pairs.keys()):
value_pairs[key] = {}
value_pairs[key]['field'] = v.field
value_pairs[key]['name'] = v.name
value_pairs[key]['input_value'] = v.input_value
value_pairs[key]['input_body_value'] = v.value
for v in value_outputs:
key = str(v.field.id) + '_' + v.name
if not (key in value_pairs.keys()):
value_pairs[key] = {}
value_pairs[key]['field'] = v.field
value_pairs[key]['name'] = v.name
value_pairs[key]['output_body_value'] = v.value
value_pairs[key]['output_obj'] = v.obj
# remove iterator
removed_name = str(local_field.id) + '_' + target_value.name
del value_pairs[removed_name]
for k, v in value_pairs.items():
name = v['name']
field = v['field']
if 'input_body_value' in v:
inputs.append(v['input_value'])
body_graph.add_input_value(v['input_body_value'])
else:
temp_value1 = functions.generate_value_with_same_type(
v['output_body_value'])
temp_value2 = functions.generate_value_with_same_type(
v['output_body_value'])
inputs.append(temp_value1)
body_graph.add_input_value(temp_value2)
if 'output_body_value' in v:
body_graph.add_output_value(v['output_body_value'])
output_value = functions.generate_value_with_same_type(
v['output_body_value'])
outputs.append(output_value)
if 'output_obj' in v:
obj = v['output_obj']
obj.revise(output_value)
field.get_attribute(name).revise(obj)
elif field.get_attribute(name).has_obj():
field.get_attribute(name).get_ref().revise(output_value)
else:
field.get_attribute(name).revise(values.ValueRef(output_value))
else:
temp_value1 = v['input_body_value']
temp_value2 = functions.generate_value_with_same_type(
v['input_body_value'])
body_graph.add_output_value(temp_value1)
outputs.append(temp_value2)
node = nodes.NodeListcomp(iter_value, inputs, body_graph, astc.lineno)
node.set_outputs(outputs)
graph.add_node(node)
return local_field.get_attribute(internal_list_id).get_ref()
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)
