def from_float(cls, mod):
r"""Create a qat module from a float module or qparams_dict
Args: `mod` a float module, either produced by torch.ao.quantization utilities
or directly from user
"""
assert type_before_parametrizations(mod) == cls._FLOAT_MODULE, (
" qat." + cls.__name__ + ".from_float only works for " +
cls._FLOAT_MODULE.__name__)
assert hasattr(
mod, "qconfig"), "Input float module must have qconfig defined"
assert mod.qconfig, "Input float module must have a valid qconfig"
if type_before_parametrizations(mod) == LinearReLU:
mod = mod[0]
qconfig = mod.qconfig
qat_linear = cls(mod.in_features,
mod.out_features,
bias=mod.bias is not None,
qconfig=qconfig)
if is_parametrized(mod, "weight"):
transfer_parametrizations_and_params(mod, qat_linear, "weight")
else:
qat_linear.weight = mod.weight
if is_parametrized(mod, "bias"):
transfer_parametrizations_and_params(mod, qat_linear, "bias")
else:
qat_linear.bias = mod.bias
return qat_linear
def swap_module(mod, mapping, custom_module_class_mapping):
r"""Swaps the module if it has a quantized counterpart and it has an
`observer` attached.
Args:
mod: input module
mapping: a dictionary that maps from nn module to nnq module
Return:
The corresponding quantized module of `mod`
"""
new_mod = mod
if hasattr(mod, 'qconfig') and mod.qconfig is not None:
swapped = False
if type_before_parametrizations(mod) in custom_module_class_mapping:
new_mod = custom_module_class_mapping[type_before_parametrizations(
mod)].from_observed(mod)
swapped = True
elif type_before_parametrizations(mod) in mapping:
qmod = mapping[type_before_parametrizations(mod)]
if hasattr(qmod, '_IS_REFERENCE') and qmod._IS_REFERENCE:
assert mod.qconfig is not None
weight_post_process = mod.qconfig.weight()
weight_post_process(mod.weight)
weight_qparams = get_qparam_dict(weight_post_process)
new_mod = qmod.from_float(mod, weight_qparams)
else:
new_mod = qmod.from_float(mod)
swapped = True
if swapped:
# Preserve module's pre forward hooks. They'll be called on quantized input
for pre_hook_fn in mod._forward_pre_hooks.values():
new_mod.register_forward_pre_hook(pre_hook_fn)
# Preserve module's post forward hooks except _observer_forward_hook
# After convert they'll work with quantized output
for hook_fn in mod._forward_hooks.values():
if hook_fn is not _observer_forward_hook:
new_mod.register_forward_hook(hook_fn)
# respect device affinity when swapping modules
devices = get_unique_devices_(mod)
assert len(devices) <= 1, (
"swap_module only works with cpu or single-device CUDA modules, "
"but got devices {}".format(devices))
device = next(iter(devices)) if len(devices) > 0 else None
if device:
new_mod.to(device)
return new_mod
def _get_special_act_post_process(module: torch.nn.Module) -> Optional[Callable]:
r""" Get the special activation post process for `module`, this has
higher priority than the activation post process in `qconfig`
e.g.
input: torch.nn.Sigmoid
output: default_affine_fixed_qparam_fake_quant
"""
return DEFAULT_MODULE_TO_ACT_POST_PROCESS.get(type_before_parametrizations(module), None)
def from_float(cls, mod):
r"""Create a quantized module from an observed float module
Args:
mod (Module): a float module, either produced by torch.ao.quantization
utilities or provided by the user
"""
if hasattr(mod, 'weight_fake_quant'):
if type_before_parametrizations(mod) == nniqat.LinearBn1d:
mod.weight, mod.bias = fuse_linear_bn_weights(
mod.weight, mod.bias, mod.bn.running_mean,
mod.bn.running_var, mod.bn.eps, mod.bn.weight, mod.bn.bias)
weight_post_process = mod.weight_fake_quant
activation_post_process = mod.activation_post_process
else:
# This function does not participate in JIT, so it is OK to ignore
# the type mismatch in assignment. Also, mypy has an issue with
# iterables not being implemented, so we are ignoring those too.
if not isinstance(cls._FLOAT_MODULE, Iterable):
cls._FLOAT_MODULE = [cls._FLOAT_MODULE
] # type: ignore[assignment]
supported_modules = ', '.join([
float_mod.__name__ for float_mod in cls._FLOAT_MODULE
]) # type: ignore[attr-defined]
error_msg = 'nnq.{}.from_float only works for {}, but got: {}'.format(
cls.__name__, supported_modules, type(mod))
assert type_before_parametrizations(
mod) in cls._FLOAT_MODULE, error_msg.format(
) # type: ignore[attr-defined]
assert hasattr(
mod, 'qconfig'), 'Input float module must have qconfig defined'
activation_post_process = mod.activation_post_process
if type_before_parametrizations(mod) == nni.LinearReLU:
mod = mod[0]
weight_post_process = mod.qconfig.weight()
weight_post_process(mod.weight)
dtype = weight_post_process.dtype
act_scale, act_zp = activation_post_process.calculate_qparams()
assert dtype == torch.qint8, 'Weight observer must have dtype torch.qint8'
qweight = _quantize_weight(mod.weight.float(), weight_post_process)
qlinear = cls(mod.in_features, mod.out_features, dtype=dtype)
qlinear.set_weight_bias(qweight, mod.bias)
qlinear.scale = float(act_scale)
qlinear.zero_point = int(act_zp)
return qlinear
def is_match(modules, node, pattern, max_uses=sys.maxsize):
""" Matches a node in fx against a pattern
"""
if isinstance(pattern, tuple):
self_match, *arg_matches = pattern
if self_match is getattr:
assert len(
pattern) == 2, 'Expecting getattr pattern to have two elements'
arg_matches = []
else:
self_match = pattern
arg_matches = []
if isinstance(self_match, type) and issubclass(self_match, MatchAllNode):
return True
if not isinstance(node, Node) or len(node.users) > max_uses:
return False
if isinstance(self_match, type) and issubclass(self_match,
torch.nn.Module):
if node.op != 'call_module':
return False
if not type_before_parametrizations(
modules[node.target]) == self_match:
return False
elif callable(self_match):
if node.op != 'call_function' or node.target is not self_match:
return False
elif node.target is getattr:
if node.args[1] != pattern[1]:
return False
elif isinstance(self_match, str):
if node.op != 'call_method' or node.target != self_match:
return False
elif node.target != self_match:
return False
if not arg_matches:
return True
if len(arg_matches) != len(node.args):
return False
return all(
is_match(modules, node, arg_match, max_uses=1)
for node, arg_match in zip(node.args, arg_matches))
def _convert(module,
mapping=None,
inplace=False,
is_reference=False,
convert_custom_config_dict=None):
r"""Converts submodules in input module to a different module according to `mapping`
by calling `from_float` method on the target module class
Args:
module: input module
mapping: a dictionary that maps from source module type to target
module type, can be overwritten to allow swapping user defined
Modules
inplace: carry out model transformations in-place, the original module
is mutated
is_reference: a flag to enable quantized reference module
"""
if mapping is None:
mapping = get_default_static_quant_reference_module_mappings() if is_reference \
else get_default_static_quant_module_mappings()
if convert_custom_config_dict is None:
convert_custom_config_dict = {}
custom_module_class_mapping = convert_custom_config_dict.get(
"observed_to_quantized_custom_module_class", {})
if not inplace:
module = copy.deepcopy(module)
reassign = {}
for name, mod in module.named_children():
# both fused modules and observed custom modules are
# swapped as one unit
if not isinstance(mod, _FusedModule) and \
type_before_parametrizations(mod) not in custom_module_class_mapping:
_convert(
mod,
mapping,
True, # inplace
is_reference,
convert_custom_config_dict)
reassign[name] = swap_module(mod, mapping, custom_module_class_mapping)
for key, value in reassign.items():
module._modules[key] = value
return module
def _propagate_qconfig_helper(module,
qconfig_dict,
qconfig_parent=None,
prefix='',
prepare_custom_config_dict=None):
r"""This is a helper function for `propagate_qconfig_`
Args:
module: input module
qconfig_dict: dictionary that maps from name of submodule to quantization
configuration
qconfig_parent: quantization config of parent module, we will fallback to
this config when there is no specified config for current
module
prefix: corresponding prefix of the current module, used as key in
qconfig_dict
prepare_custom_config_dict: dictionary for custom handling of modules
see docs for :func:`~torch.ao.quantization.prepare_fx`
Return:
None, module is modified inplace with qconfig attached
"""
module_qconfig = qconfig_dict.get(type_before_parametrizations(module),
qconfig_parent)
module_qconfig = qconfig_dict.get(prefix, module_qconfig)
module_qconfig = getattr(module, 'qconfig', module_qconfig)
torch.ao.quantization.qconfig.assert_valid_qconfig(module_qconfig, module)
qconfig_with_device_check = add_module_to_qconfig_obs_ctr(
module_qconfig, module)
module.qconfig = qconfig_with_device_check
for name, child in module.named_children():
module_prefix = prefix + '.' + name if prefix else name
# do no not propagate qconfig to child if child is non traceable
if prepare_custom_config_dict is None or not (
name in prepare_custom_config_dict.get(
"non_traceable_module_name", [])
or type(child) in prepare_custom_config_dict.get(
"non_traceable_module_class", [])):
_propagate_qconfig_helper(child, qconfig_dict,
qconfig_with_device_check, module_prefix)
def fuse_known_modules(mod_list, is_qat, additional_fuser_method_mapping=None):
r"""Returns a list of modules that fuses the operations specified
in the input module list.
Fuses only the following sequence of modules:
conv, bn
conv, bn, relu
conv, relu
linear, bn
linear, relu
For these sequences, the first element in the output module list performs
the fused operation. The rest of the elements are set to nn.Identity()
"""
types = tuple(type_before_parametrizations(m) for m in mod_list)
fuser_method = get_fuser_method(types, additional_fuser_method_mapping)
if fuser_method is None:
raise NotImplementedError("Cannot fuse modules: {}".format(types))
new_mod: List[Optional[nn.Module]] = [None] * len(mod_list)
fused = fuser_method(is_qat, *mod_list)
# NOTE: forward hooks not processed in the two following for loops will be lost after the fusion
# Move pre forward hooks of the base module to resulting fused module
for handle_id, pre_hook_fn in mod_list[0]._forward_pre_hooks.items():
fused.register_forward_pre_hook(pre_hook_fn)
del mod_list[0]._forward_pre_hooks[handle_id]
# Move post forward hooks of the last module to resulting fused module
for handle_id, hook_fn in mod_list[-1]._forward_hooks.items():
fused.register_forward_hook(hook_fn)
del mod_list[-1]._forward_hooks[handle_id]
new_mod[0] = fused
for i in range(1, len(mod_list)):
identity = nn.Identity()
identity.training = mod_list[0].training
new_mod[i] = identity
return new_mod
def __init__(self, conv, bn):
assert type_before_parametrizations(conv) == Conv3d and type_before_parametrizations(bn) == BatchNorm3d, \
'Incorrect types for input modules{}{}'.format(
type_before_parametrizations(conv), type_before_parametrizations(bn))
super().__init__(conv, bn)
def __init__(self, conv, bn, relu):
assert type_before_parametrizations(conv) == Conv2d and type_before_parametrizations(bn) == BatchNorm2d and \
type_before_parametrizations(relu) == ReLU, 'Incorrect types for input modules{}{}{}' \
.format(type_before_parametrizations(conv), type_before_parametrizations(bn), type_before_parametrizations(relu))
super().__init__(conv, bn, relu)
def __init__(self, linear, relu):
assert type_before_parametrizations(linear) == Linear and type_before_parametrizations(relu) == ReLU, \
'Incorrect types for input modules{}{}'.format(
type_before_parametrizations(linear), type_before_parametrizations(relu))
super().__init__(linear, relu)
def __init__(self, conv, relu):
assert type_before_parametrizations(conv) == Conv3d and type_before_parametrizations(relu) == ReLU, \
'Incorrect types for input modules{}{}'.format(
type_before_parametrizations(conv), type_before_parametrizations(relu))
super().__init__(conv, relu)
def __init__(self, linear, bn):
assert type_before_parametrizations(linear) == Linear and type_before_parametrizations(bn) == BatchNorm1d, \
'Incorrect types for input modules{}{}'.format(type_before_parametrizations(linear), type_before_parametrizations(bn))
super().__init__(linear, bn)
def convert_weighted_module(
node: Node,
modules: Dict[str, torch.nn.Module],
observed_node_names: Set[str],
qconfig_map: Dict[str, QConfigAny],
backend_config_dict: Dict[str, Any]):
""" Convert a weighted module to reference quantized module in the model
If the QConfig of a QAT module is not set, the module will still be converted to
a float module.
Args:
- node: The call_module node of the observed standalone module
- modules: named_module of original model
- observed_node_names: names for the set of observed fx node, we can skip
this conversion if the node is not observed
"""
original_module = modules[str(node.target)]
qconfig: QConfigAny = original_module.qconfig # type: ignore[assignment]
weight_post_process = None
qat_module_classes = get_qat_module_classes(backend_config_dict)
if isinstance(
original_module,
qat_module_classes):
# Converting qat module to a float module, we need to attch
# weight fake_quant to the module, weight fake_quant is assumed to be run during
# QAT so we don't need to run it again here
weight_post_process = original_module.weight_fake_quant
original_module = original_module.to_float() # type: ignore[operator]
# change qat module to float module
parent_name, name = _parent_name(node.target)
setattr(modules[parent_name], name, original_module)
is_observed = node.name in observed_node_names
# If a qconfig is not defined for this node, then skip converting to a reference module
if qconfig is None or has_none_qconfig(node, qconfig_map) or not is_observed:
return
# skip converting to reference quantized module if the qconfig is not supported
pattern_to_dtype_configs = get_pattern_to_dtype_configs(backend_config_dict)
dtype_configs = pattern_to_dtype_configs.get(type(original_module), [])
if not is_qconfig_supported_by_dtype_configs(qconfig, dtype_configs):
return
# TODO: rename weight_is_statically_quantized to weight_is_int8_quantized
is_weight_quantized = weight_is_quantized(qconfig)
# the condition for swapping the module to reference quantized module is:
# weights need to be quantized
if not is_weight_quantized:
return
fused_module = None
float_module = original_module
# extract the inidividual float_module and fused module
if isinstance(original_module, torch.nn.intrinsic._FusedModule):
fused_module = float_module
float_module = fused_module[0] # type: ignore[index]
# TODO: move this to the reference quantized module
# weight_qparams or weight_qparams dict
wq_or_wq_dict = {}
if isinstance(float_module, torch.nn.RNNCellBase):
weight_post_process_ih = qconfig.weight() # type: ignore[union-attr, operator]
weight_post_process_hh = qconfig.weight() # type: ignore[union-attr, operator]
weight_post_process_ih(float_module.weight_ih)
weight_post_process_hh(float_module.weight_hh)
weight_qparams_ih = get_qparam_dict(weight_post_process_ih)
weight_qparams_hh = get_qparam_dict(weight_post_process_hh)
wq_or_wq_dict = {
"weight_ih": weight_qparams_ih,
"weight_hh": weight_qparams_hh,
}
elif isinstance(float_module, torch.nn.LSTM):
# format for wq_or_wq_dict (flattened attributes):
# {"weight_ih_l0_scale": ..., "weight_ih_l0_qscheme": ..., ...}
for wn in float_module._flat_weights_names:
if hasattr(float_module, wn) and wn.startswith("weight"):
weight = getattr(float_module, wn)
weight_post_process = qconfig.weight() # type: ignore[union-attr, operator]
if weight_post_process.dtype == torch.qint8: # type: ignore[union-attr]
weight_post_process(weight) # type: ignore[operator, misc]
wq_or_wq_dict[wn] = get_qparam_dict(weight_post_process)
else:
# weight_post_process is None means the original module is not a QAT module
# we need to get weight_post_process from qconfig in this case
if weight_post_process is None:
weight_post_process = qconfig.weight() # type: ignore[union-attr, operator]
# run weight observer
# TODO: This is currently a hack for QAT to get the right shapes for scale and zero point.
# In the future, we should require the user to calibrate the model after calling prepare
# Issue: https://github.com/pytorch/pytorch/issues/73941
weight_post_process(float_module.weight) # type: ignore[operator]
wq_or_wq_dict = get_qparam_dict(weight_post_process)
# We use the same reference module for all modes of quantization: static, dynamic, weight_only
# root_module_to_quantized_reference_module: module mapping from root (floating point) module class
# to quantized reference module class, e.g. nn.Conv2d to nn.quantized._reference.Conv2d
root_module_to_quantized_reference_module = get_root_module_to_quantized_reference_module(backend_config_dict)
ref_qmodule_cls = root_module_to_quantized_reference_module.get(type_before_parametrizations(float_module), None)
assert (
ref_qmodule_cls is not None
), f"No reference quantized module class configured for {type_before_parametrizations(float_module)}"
ref_qmodule = ref_qmodule_cls.from_float(float_module, wq_or_wq_dict) # type: ignore[attr-defined]
if fused_module is not None:
fused_module[0] = ref_qmodule # type: ignore[operator]
else:
parent_name, name = _parent_name(node.target)
setattr(modules[parent_name], name, ref_qmodule)
def add_observer_(module,
qconfig_propagation_list=None,
non_leaf_module_list=None,
device=None,
custom_module_class_mapping=None):
r"""Add observer for the leaf child of the module.
This function insert observer module to all leaf child module that
has a valid qconfig attribute.
Args:
module: input module with qconfig attributes for all the leaf modules that we want to quantize
qconfig_propagation_list: a list of quantizable modules that will have observers added to them
if they are leaf nodes
device: parent device, if any
non_leaf_module_list: list of non-leaf modules we want to add observer
Return:
None, module is modified inplace with added observer modules and forward_hooks
"""
if qconfig_propagation_list is None:
qconfig_propagation_list = get_default_qconfig_propagation_list()
if custom_module_class_mapping is None:
custom_module_class_mapping = {}
# respect device affinity when adding observers
if device is None:
devices = get_unique_devices_(module)
assert len(devices) <= 1, (
"add_observer_ only works with cpu or single-device CUDA modules, "
"but got devices {}".format(devices))
device = next(iter(devices)) if len(devices) > 0 else None
def get_activation_post_process(qconfig,
device,
special_act_post_process=None):
activation = qconfig.activation(
) if special_act_post_process is None else special_act_post_process()
if device is not None:
activation.to(device)
return activation
def needs_observation(m):
return hasattr(m, 'qconfig') and m.qconfig is not None
def insert_activation_post_process(m, special_act_post_process=None):
""" Adds an activation post process module and register
a pre or post hook that calls the module
"""
# We don't insert observer/fake_quantize for DeQuantStub
if needs_observation(m) and not isinstance(m, DeQuantStub):
# observer and hook will be gone after we swap the module
m.add_module(
'activation_post_process',
get_activation_post_process(m.qconfig, device,
special_act_post_process))
# Register observer as the first entry in the hook list
# All post forward hooks are preserved and will be executed after the observer before convert
register_activation_post_process_hook(
m, pre_hook=activation_is_memoryless(m.qconfig))
for name, child in module.named_children():
# TODO remove Dropout special after codebase stable
if type_before_parametrizations(child) in [nn.Dropout]:
continue
elif type_before_parametrizations(child) in [
nnq.FloatFunctional, nnq.QFunctional
]:
if needs_observation(child):
child.activation_post_process = get_activation_post_process(
child.qconfig, device)
elif isinstance(child, _FusedModule):
# activation_post_process are now added directly to nn.Sequentail/_FusedModule
if needs_observation(child):
insert_activation_post_process(child)
elif _has_special_act_post_process(child):
special_act_post_process = _get_special_act_post_process(child)
insert_activation_post_process(child, special_act_post_process)
elif non_leaf_module_list is not None and type_before_parametrizations(
child) in non_leaf_module_list:
if needs_observation(child):
insert_activation_post_process(child)
elif needs_observation(child) and type_before_parametrizations(
child) in custom_module_class_mapping:
observed_child = custom_module_class_mapping[
type_before_parametrizations(child)].from_float(child)
setattr(module, name, observed_child)
# TODO: These are the modules that cannot be observed
# Once there are more, we should move them to a separate list
if custom_module_class_mapping[type_before_parametrizations(
child)] not in no_observer_set():
insert_activation_post_process(observed_child)
else:
add_observer_(child, qconfig_propagation_list,
non_leaf_module_list, device,
custom_module_class_mapping)
# Insert observers only for leaf nodes, note that this observer is for
# the output of the module, for input QuantStub will observe them
if has_no_children_ignoring_parametrizations(module) and not isinstance(module, torch.nn.Sequential) \
and type_before_parametrizations(module) in qconfig_propagation_list:
insert_activation_post_process(module)
def convert(
model: GraphModule, is_reference: bool = False,
convert_custom_config: Union[ConvertCustomConfig, Dict[str, Any], None] = None,
is_standalone_module: bool = False,
_remove_qconfig_flag: bool = True,
qconfig_mapping: Union[QConfigMapping, Dict[str, Any], None] = None,
backend_config_dict: Optional[Dict[str, Any]] = None) -> torch.nn.Module:
"""
We will convert an observed model (a module with observer calls) to a reference
quantized model, the rule is simple:
1. for each observer module call in the graph, we'll convert it to calls to
quantize and dequantize functions based on the observer instance
2. for weighted operations like linear/conv, we need to convert them to reference
quantized module, this requires us to know whether the dtype configured for the
weight is supported in the backend, this is done in prepare step and the result
is stored in observed_node_names, we can decide whether we need to swap the
module based on this set
standalone_module means it a submodule that is not inlined in
parent module, and will be quantized separately as one unit.
Returns a quantized standalone module, whether input/output is quantized is
specified by prepare_custom_config, with
input_quantized_idxs, output_quantized_idxs, please
see docs for prepare_fx for details
"""
if convert_custom_config is None:
convert_custom_config = ConvertCustomConfig()
if isinstance(convert_custom_config, Dict):
warnings.warn(
"Passing a convert_custom_config_dict to convert is deprecated and will not be supported "
"in a future version. Please pass in a ConvertCustomConfig instead.")
convert_custom_config = ConvertCustomConfig.from_dict(convert_custom_config)
if isinstance(qconfig_mapping, Dict):
warnings.warn(
"Passing a QConfig dictionary to convert is deprecated and will not be supported "
"in a future version. Please pass in a QConfigMapping instead.")
qconfig_mapping = QConfigMapping.from_dict(qconfig_mapping) if qconfig_mapping else None
qconfig_mapping = copy.deepcopy(qconfig_mapping)
assert(qconfig_mapping is None or isinstance(qconfig_mapping, QConfigMapping))
node_name_to_scope, prepare_custom_config, observed_node_names = restore_state(model)
qconfig_map: Dict[str, QConfigAny] = model._qconfig_map # type: ignore[assignment]
# mapping from fully qualified module name to module instance
# for example,
# {
# '': Model(...),
# 'linear': Linear(...),
# 'linear.weight_fake_quant': PerChannelMinMaxObserver(...),
# }
# We use remove_duplicate=False here because torch.cat uses
# the same activation_post_process module instance but different names
modules = dict(model.named_modules(remove_duplicate=False))
# TODO refactor this code once we update the prepare logic to have additional information on
# which graph nodes have been observed and share that with convert to decide which observers to ignore.
if qconfig_mapping:
prepare_qconfig_mapping: QConfigMapping = model._qconfig_mapping # type: ignore[assignment]
modules_copy = copy.deepcopy(modules)
if model._is_qat:
update_qconfig_for_qat(qconfig_mapping, {})
update_qconfig_for_fusion(model, qconfig_mapping)
compare_prepare_convert_qconfig_mappings(prepare_qconfig_mapping, qconfig_mapping) # type: ignore[arg-type]
convert_qconfig_map = generate_qconfig_map(model, modules_copy, model.graph, qconfig_mapping, node_name_to_scope)
# check the convert_qconfig_map generated and ensure that all the values either match what was set in prepare qconfig_map
# or are set to None in the convert_qconfig_map.
for k, v in qconfig_map.items():
assert k in convert_qconfig_map, 'Expected key {} in convert qconfig_map'.format(k)
if convert_qconfig_map[k] is not None:
assert qconfig_equals(v, convert_qconfig_map[k]), \
"Expected k {} to have the same value in prepare and convert QConfigMappings, " \
"but {} was updated to {}".format(k, v, convert_qconfig_map[k])
qconfig_map = convert_qconfig_map
custom_module_classes = get_custom_module_class_keys(convert_custom_config.observed_to_quantized_mapping)
custom_module_class_mapping = convert_custom_config.observed_to_quantized_mapping
if model._equalization_qconfig_map is not None:
# If we want to do equalization then do the following:
# Calculate the equalization scale, update the observers with the scaled
# inputs, and scale the weight
weight_eq_obs_dict = update_obs_for_equalization(model, modules)
convert_eq_obs(model, modules, weight_eq_obs_dict)
# always run weight observers in the top level forward method
# for dynamic quant ops or weight only quant ops
run_weight_observers(model)
graph_inputs: List[str] = []
for node in model.graph.nodes:
if node.op == 'placeholder':
graph_inputs.append(node.name)
# TODO: move this outside of this function
def replace_observer_with_quantize_dequantize_node(
model: torch.nn.Module,
graph: Graph,
node: Node,
modules: Dict[str, torch.nn.Module],
node_name_to_scope: Dict[str, Tuple[str, type]],
qconfig_map: Dict[str, QConfigAny]) -> None:
""" Replace activation_post_process module call node with quantize and
dequantize node
Before:
... -> observer_0(x) -> ...
After:
... -> torch.quantize_per_tensor(x, ...) -> x.dequantize() -> ...
"""
assert modules is not None
assert isinstance(node.target, str)
module_path, prefix = get_module_path_and_prefix(node, node_name_to_scope, qconfig_map)
observer_module = modules[node.target]
maybe_quantize_node_info = get_quantize_node_info(observer_module)
# Skip replacing observers to quant/dequant nodes if the qconfigs of all
# consumers and producers of this observer are None
skip_replacement = all([
has_none_qconfig(n, qconfig_map) for n in
list(node.args) + list(node.users.keys())])
if skip_replacement or maybe_quantize_node_info is None:
# didn't find correponding quantize op and info for the observer_module
# so we just remove the observer
with graph.inserting_before(node):
node.replace_all_uses_with(node.args[0])
graph.erase_node(node)
else:
# otherwise, we can convert the observer moduel call to quantize/dequantize node
node_type, quantize_op, qparams = maybe_quantize_node_info
# replace observer node with quant - dequant node
with graph.inserting_before(node):
input_node = node.args[0]
inputs = [input_node]
for key, value in qparams.items():
# TODO: we can add the information of whether a value needs to
# be registered as an attribute in qparams dict itself
if key in ['_scale_', '_zero_point_']:
# For scale and zero_point values we register them as buffers in the root module.
# TODO: maybe need more complex attr name here
qparam_node = create_getattr_from_value(model, graph, module_path + prefix + key, value)
inputs.append(qparam_node)
else:
# for qparams that are not scale/zero_point (like axis, dtype) we store them as literals in the graph.
inputs.append(value)
quantized_node = graph.create_node(node_type, quantize_op, tuple(inputs), {})
dequantized_node = graph.call_method("dequantize", args=(quantized_node,))
node.replace_all_uses_with(dequantized_node)
graph.erase_node(node)
# this is a temporary hack for custom module, we may want to implement
# this properly after the custom module class design is finalized
def replace_observer_with_dequantize_node(node: Node, graph: Graph):
call_custom_module_node = node.args[0]
assert isinstance(call_custom_module_node, Node), \
f"Expecting the for call custom module node to be a Node, but got {call_custom_module_node}"
node.replace_all_uses_with(call_custom_module_node)
graph.erase_node(node)
insert_dequantize_node(call_custom_module_node, graph)
# additional state to override inputs to be quantized, if specified
# by the user
placeholder_node_seen_cnt = 0
input_quantized_idxs: List[int] = prepare_custom_config.input_quantized_indexes
output_quantized_idxs: List[int] = prepare_custom_config.output_quantized_indexes
if backend_config_dict is None:
backend_config_dict = get_native_backend_config_dict()
root_module_to_quantized_reference_module = get_root_module_to_quantized_reference_module(backend_config_dict)
# convert tuples so that it can work with isinstance(module, tuple_of_classes)
root_module_classes = tuple(root_module_to_quantized_reference_module.keys())
qat_module_classes = get_qat_module_classes(backend_config_dict)
fused_module_classes = get_fused_module_classes(backend_config_dict)
statically_quantized_custom_module_nodes: Set[Node] = set()
for node in list(model.graph.nodes):
if node.op == 'placeholder':
cur_placeholder_node_idx = placeholder_node_seen_cnt
placeholder_node_seen_cnt += 1
if cur_placeholder_node_idx in input_quantized_idxs:
# Inputs are assumed to be quantized if the user specifid the
# input_quantized_idxs override.
# we need to dequantize the inputs since all operators took
# floating point inputs in reference quantized models
insert_dequantize_node(node, model.graph)
elif node.op == "output":
# If the argument is empty we don't need to do anything
if len(output_quantized_idxs) == 0:
continue
# Result are kept quantized if the user specified the
# output_quantized_idxs override.
# Remove the dequantize operator for the node in the end if any
return_node = node
output = node.args[0]
# outputs can be Node, list, tuple, dict, other cases are not supported yet
if isinstance(output, (list, tuple)):
for idx in output_quantized_idxs:
maybe_recursive_remove_dequantize(output[idx], return_node, model.graph)
elif isinstance(output, (Node, dict)):
# we treat dict as a single argument currently, but it can be extended
# to support {"key": dtype} after we change output_quantized_idxs to
# dict
if 0 in output_quantized_idxs:
maybe_recursive_remove_dequantize(output, return_node, model.graph)
else:
warnings.warn(f"Unsupported node type for output_quantized_idxs: {type(output)}")
elif node.op == "call_module":
if is_activation_post_process(modules[node.target]):
observed_node = node.args[0]
if observed_node in statically_quantized_custom_module_nodes:
replace_observer_with_dequantize_node(node, model.graph)
else:
replace_observer_with_quantize_dequantize_node(
model, model.graph, node, modules, node_name_to_scope,
qconfig_map)
elif is_observed_standalone_module(modules[node.target]):
convert_standalone_module(
node, modules, model, is_reference, backend_config_dict)
# below this point `type_before_parametrizations` is used
# instead of `type` to handle situations with fx quant + sparsity
elif type_before_parametrizations(modules[node.target]) in set(
root_module_classes).union(qat_module_classes).union(fused_module_classes):
# extra check for fused module classes to make sure they are fused module classes
# of target modules
if type_before_parametrizations(modules[node.target]) in fused_module_classes and \
type_before_parametrizations(modules[node.target][0]) not in root_module_classes:
continue
convert_weighted_module(
node, modules, observed_node_names, qconfig_map, backend_config_dict)
elif type_before_parametrizations(modules[node.target]) in custom_module_classes:
convert_custom_module(
node, model.graph, modules, custom_module_class_mapping,
statically_quantized_custom_module_nodes)
preserved_attributes = set(convert_custom_config.preserved_attributes)
model = QuantizedGraphModule(model, copy.deepcopy(model.graph), preserved_attributes)
# remove deadcode after converting observers to quant/dequant ops
model.graph.eliminate_dead_code()
model.recompile()
# TODO: maybe move this to quantize_fx.py
if not is_reference:
model = duplicate_dequantize_node(model)
model = duplicate_quantize_dynamic_node(model)
model = lower_to_fbgemm(model, qconfig_map, node_name_to_scope)
model = remove_quant_dequant_pairs(model)
model = remove_extra_dequantize(model)
# TODO: this looks hacky, we want to check why we need this and see if we can
# remove this
# removes qconfig and activation_post_process modules
if _remove_qconfig_flag:
_remove_qconfig(model)
return model
def get_matched_types(m):
if isinstance(m, tuple):
return tuple(map(get_matched_types, m))
if isinstance(m, torch.nn.Module):
return type_before_parametrizations(m)
return m
def __init__(self, batch_norm, relu):
assert type_before_parametrizations(batch_norm) == BatchNorm3d and type_before_parametrizations(relu) == ReLU, \
'Incorrect types for input modules{}{}'.format(
type_before_parametrizations(batch_norm), type_before_parametrizations(relu))
super().__init__(batch_norm, relu)
