OSSForks/pytorch
1
0
Fork 0
mirror of https://github.com/zebrajr/pytorch.git synced 2025-12-06 12:20:52 +01:00
Code Issues Actions 38 Packages Projects Releases Wiki Activity

[Quant][docs] Replace qconfig_dict with QConfigMapping in docs

Browse Source 
Summary: https://github.com/pytorch/pytorch/pull/78452 replaced
qconfig_dict with QConfigMapping as the default API for prepare_fx,
prepare_qat_fx, and convert_fx. We should update the docs to reflect
this change as well.

Test Plan:
```
cd docs
make html
cd build/html
python -m server.http
```

Reviewers: jerryzh168, vkuzo

Subscribers: jerryzh168, vkuzo

Pull Request resolved: https://github.com/pytorch/pytorch/pull/78533

Approved by: https://github.com/vkuzo
This commit is contained in:
Andrew Or 2022-05-31 10:37:37 -07:00 committed by PyTorch MergeBot
parent 43c09b5cef
commit e41389f84b
2 changed files with 14 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

22
docs/source/quantization.rst
View File

@ -296,7 +296,7 @@ by module basis. Specifically, for all quantization techniques, the user needs t
additional parameters) from functionals to module form (for example,
using ``torch.nn.ReLU`` instead of ``torch.nn.functional.relu``).
2. Specify which parts of the model need to be quantized either by assigning
``.qconfig`` attributes on submodules or by specifying ``qconfig_dict``.
``.qconfig`` attributes on submodules or by specifying ``qconfig_mapping``.
For example, setting ``model.conv1.qconfig = None`` means that the
``model.conv`` layer will not be quantized, and setting
``model.linear1.qconfig = custom_qconfig`` means that the quantization
@ -322,10 +322,11 @@ to do the following in addition:
(Prototype) FX Graph Mode Quantization
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
There are multiple quantization types in post training quantization (weight only, dynamic and static) and the configuration is done through `qconfig_dict` (an argument of the `prepare_fx` function).
There are multiple quantization types in post training quantization (weight only, dynamic and static) and the configuration is done through `qconfig_mapping` (an argument of the `prepare_fx` function).
API Example::
from torch.quantization import QConfigMapping
import torch.quantization.quantize_fx as quantize_fx
import copy
@ -338,9 +339,9 @@ API Example::
# we need to deepcopy if we still want to keep model_fp unchanged after quantization since quantization apis change the input model
model_to_quantize = copy.deepcopy(model_fp)
model_to_quantize.eval()
qconfig_dict = {"": torch.quantization.default_dynamic_qconfig}
qconfig_mapping = QConfigMapping().set_global(torch.quantization.default_dynamic_qconfig)
# prepare
model_prepared = quantize_fx.prepare_fx(model_to_quantize, qconfig_dict)
model_prepared = quantize_fx.prepare_fx(model_to_quantize, qconfig_mapping)
# no calibration needed when we only have dynamici/weight_only quantization
# quantize
model_quantized = quantize_fx.convert_fx(model_prepared)
@ -350,10 +351,10 @@ API Example::
#
model_to_quantize = copy.deepcopy(model_fp)
qconfig_dict = {"": torch.quantization.get_default_qconfig('qnnpack')}
qconfig_mapping = QConfigMapping().set_global(torch.quantization.get_default_qconfig('qnnpack'))
model_to_quantize.eval()
# prepare
model_prepared = quantize_fx.prepare_fx(model_to_quantize, qconfig_dict)
model_prepared = quantize_fx.prepare_fx(model_to_quantize, qconfig_mapping)
# calibrate (not shown)
# quantize
model_quantized = quantize_fx.convert_fx(model_prepared)
@ -363,10 +364,10 @@ API Example::
#
model_to_quantize = copy.deepcopy(model_fp)
qconfig_dict = {"": torch.quantization.get_default_qat_qconfig('qnnpack')}
qconfig_mapping = QConfigMapping().set_global(torch.quantization.get_default_qat_qconfig('qnnpack'))
model_to_quantize.train()
# prepare
model_prepared = quantize_fx.prepare_qat_fx(model_to_quantize, qconfig_dict)
model_prepared = quantize_fx.prepare_qat_fx(model_to_quantize, qconfig_mapping)
# training loop (not shown)
# quantize
model_quantized = quantize_fx.convert_fx(model_prepared)
@ -793,6 +794,7 @@ Example::
import torch
import torch.nn.quantized as nnq
from torch.quantization import QConfigMapping
import torch.quantization.quantize_fx
# original fp32 module to replace
@ -868,7 +870,7 @@ Example::
m = torch.nn.Sequential(CustomModule()).eval()
qconfig_dict = {'': torch.quantization.default_qconfig}
qconfig_mapping = QConfigMapping().set_global(torch.quantization.default_qconfig)
prepare_custom_config_dict = {
"float_to_observed_custom_module_class": {
"static": {
@ -884,7 +886,7 @@ Example::
}
}
mp = torch.quantization.quantize_fx.prepare_fx(
m, qconfig_dict, prepare_custom_config_dict=prepare_custom_config_dict)
m, qconfig_mapping, prepare_custom_config_dict=prepare_custom_config_dict)
# calibration (not shown)
mq = torch.quantization.quantize_fx.convert_fx(
mp, convert_custom_config_dict=convert_custom_config_dict)

2
torch/ao/quantization/quantize_fx.py
View File

@ -353,6 +353,8 @@ def prepare_fx(
* `qconfig_mapping` (required): mapping from model ops to qconfigs::
from torch.quantization import QConfigMapping
qconfig_mapping = QConfigMapping() \
.set_global(global_qconfig) \
.set_object_type(torch.nn.Linear, qconfig1) \