Merge pull request #1657 from baiyfbupt/dygraph
Add dygraph quantizationdependabot/pip/PPOCRLabel/requirements/lxml-4.6.2
Double_V
4 years ago
committed by
GitHub
commit
541c3170f8
@ -0,0 +1,118 @@
|
||||
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
|
||||
import os
|
||||
import sys
|
||||
|
||||
__dir__ = os.path.dirname(os.path.abspath(__file__))
|
||||
sys.path.append(__dir__)
|
||||
sys.path.append(os.path.abspath(os.path.join(__dir__, '..', '..', '..')))
|
||||
sys.path.append(
|
||||
os.path.abspath(os.path.join(__dir__, '..', '..', '..', 'tools')))
|
||||
|
||||
import argparse
|
||||
|
||||
import paddle
|
||||
from paddle.jit import to_static
|
||||
|
||||
from ppocr.modeling.architectures import build_model
|
||||
from ppocr.postprocess import build_post_process
|
||||
from ppocr.utils.save_load import init_model
|
||||
from ppocr.utils.logging import get_logger
|
||||
from tools.program import load_config, merge_config, ArgsParser
|
||||
from ppocr.metrics import build_metric
|
||||
import tools.program as program
|
||||
from paddleslim.dygraph.quant import QAT
|
||||
from ppocr.data import build_dataloader
|
||||
|
||||
|
||||
def main():
|
||||
############################################################################################################
|
||||
# 1. quantization configs
|
||||
############################################################################################################
|
||||
quant_config = {
|
||||
# weight preprocess type, default is None and no preprocessing is performed.
|
||||
'weight_preprocess_type': None,
|
||||
# activation preprocess type, default is None and no preprocessing is performed.
|
||||
'activation_preprocess_type': None,
|
||||
# weight quantize type, default is 'channel_wise_abs_max'
|
||||
'weight_quantize_type': 'channel_wise_abs_max',
|
||||
# activation quantize type, default is 'moving_average_abs_max'
|
||||
'activation_quantize_type': 'moving_average_abs_max',
|
||||
# weight quantize bit num, default is 8
|
||||
'weight_bits': 8,
|
||||
# activation quantize bit num, default is 8
|
||||
'activation_bits': 8,
|
||||
# data type after quantization, such as 'uint8', 'int8', etc. default is 'int8'
|
||||
'dtype': 'int8',
|
||||
# window size for 'range_abs_max' quantization. default is 10000
|
||||
'window_size': 10000,
|
||||
# The decay coefficient of moving average, default is 0.9
|
||||
'moving_rate': 0.9,
|
||||
# for dygraph quantization, layers of type in quantizable_layer_type will be quantized
|
||||
'quantizable_layer_type': ['Conv2D', 'Linear'],
|
||||
}
|
||||
FLAGS = ArgsParser().parse_args()
|
||||
config = load_config(FLAGS.config)
|
||||
merge_config(FLAGS.opt)
|
||||
logger = get_logger()
|
||||
# build post process
|
||||
|
||||
post_process_class = build_post_process(config['PostProcess'],
|
||||
config['Global'])
|
||||
|
||||
# build model
|
||||
# for rec algorithm
|
||||
if hasattr(post_process_class, 'character'):
|
||||
char_num = len(getattr(post_process_class, 'character'))
|
||||
config['Architecture']["Head"]['out_channels'] = char_num
|
||||
model = build_model(config['Architecture'])
|
||||
|
||||
# get QAT model
|
||||
quanter = QAT(config=quant_config)
|
||||
quanter.quantize(model)
|
||||
|
||||
init_model(config, model, logger)
|
||||
model.eval()
|
||||
|
||||
# build metric
|
||||
eval_class = build_metric(config['Metric'])
|
||||
|
||||
# build dataloader
|
||||
valid_dataloader = build_dataloader(config, 'Eval', device, logger)
|
||||
|
||||
# start eval
|
||||
metirc = program.eval(model, valid_dataloader, post_process_class,
|
||||
eval_class)
|
||||
logger.info('metric eval ***************')
|
||||
for k, v in metirc.items():
|
||||
logger.info('{}:{}'.format(k, v))
|
||||
|
||||
save_path = '{}/inference'.format(config['Global']['save_inference_dir'])
|
||||
infer_shape = [3, 32, 100] if config['Architecture'][
|
||||
'model_type'] != "det" else [3, 640, 640]
|
||||
|
||||
quanter.save_quantized_model(
|
||||
model,
|
||||
save_path,
|
||||
input_spec=[
|
||||
paddle.static.InputSpec(
|
||||
shape=[None] + infer_shape, dtype='float32')
|
||||
])
|
||||
logger.info('inference QAT model is saved to {}'.format(save_path))
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
config, device, logger, vdl_writer = program.preprocess()
|
||||
main()
|
||||
@ -0,0 +1,166 @@
|
||||
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
|
||||
from __future__ import absolute_import
|
||||
from __future__ import division
|
||||
from __future__ import print_function
|
||||
|
||||
import os
|
||||
import sys
|
||||
|
||||
__dir__ = os.path.dirname(os.path.abspath(__file__))
|
||||
sys.path.append(__dir__)
|
||||
sys.path.append(os.path.abspath(os.path.join(__dir__, '..', '..', '..')))
|
||||
sys.path.append(
|
||||
os.path.abspath(os.path.join(__dir__, '..', '..', '..', 'tools')))
|
||||
|
||||
import yaml
|
||||
import paddle
|
||||
import paddle.distributed as dist
|
||||
|
||||
paddle.seed(2)
|
||||
|
||||
from ppocr.data import build_dataloader
|
||||
from ppocr.modeling.architectures import build_model
|
||||
from ppocr.losses import build_loss
|
||||
from ppocr.optimizer import build_optimizer
|
||||
from ppocr.postprocess import build_post_process
|
||||
from ppocr.metrics import build_metric
|
||||
from ppocr.utils.save_load import init_model
|
||||
import tools.program as program
|
||||
from paddleslim.dygraph.quant import QAT
|
||||
|
||||
dist.get_world_size()
|
||||
|
||||
|
||||
class PACT(paddle.nn.Layer):
|
||||
def __init__(self):
|
||||
super(PACT, self).__init__()
|
||||
alpha_attr = paddle.ParamAttr(
|
||||
name=self.full_name() + ".pact",
|
||||
initializer=paddle.nn.initializer.Constant(value=20),
|
||||
learning_rate=1.0,
|
||||
regularizer=paddle.regularizer.L2Decay(2e-5))
|
||||
|
||||
self.alpha = self.create_parameter(
|
||||
shape=[1], attr=alpha_attr, dtype='float32')
|
||||
|
||||
def forward(self, x):
|
||||
out_left = paddle.nn.functional.relu(x - self.alpha)
|
||||
out_right = paddle.nn.functional.relu(-self.alpha - x)
|
||||
x = x - out_left + out_right
|
||||
return x
|
||||
|
||||
|
||||
quant_config = {
|
||||
# weight preprocess type, default is None and no preprocessing is performed.
|
||||
'weight_preprocess_type': None,
|
||||
# activation preprocess type, default is None and no preprocessing is performed.
|
||||
'activation_preprocess_type': None,
|
||||
# weight quantize type, default is 'channel_wise_abs_max'
|
||||
'weight_quantize_type': 'channel_wise_abs_max',
|
||||
# activation quantize type, default is 'moving_average_abs_max'
|
||||
'activation_quantize_type': 'moving_average_abs_max',
|
||||
# weight quantize bit num, default is 8
|
||||
'weight_bits': 8,
|
||||
# activation quantize bit num, default is 8
|
||||
'activation_bits': 8,
|
||||
# data type after quantization, such as 'uint8', 'int8', etc. default is 'int8'
|
||||
'dtype': 'int8',
|
||||
# window size for 'range_abs_max' quantization. default is 10000
|
||||
'window_size': 10000,
|
||||
# The decay coefficient of moving average, default is 0.9
|
||||
'moving_rate': 0.9,
|
||||
# for dygraph quantization, layers of type in quantizable_layer_type will be quantized
|
||||
'quantizable_layer_type': ['Conv2D', 'Linear'],
|
||||
}
|
||||
|
||||
|
||||
def main(config, device, logger, vdl_writer):
|
||||
# init dist environment
|
||||
if config['Global']['distributed']:
|
||||
dist.init_parallel_env()
|
||||
|
||||
global_config = config['Global']
|
||||
|
||||
# build dataloader
|
||||
train_dataloader = build_dataloader(config, 'Train', device, logger)
|
||||
if config['Eval']:
|
||||
valid_dataloader = build_dataloader(config, 'Eval', device, logger)
|
||||
else:
|
||||
valid_dataloader = None
|
||||
|
||||
# build post process
|
||||
post_process_class = build_post_process(config['PostProcess'],
|
||||
global_config)
|
||||
|
||||
# build model
|
||||
# for rec algorithm
|
||||
if hasattr(post_process_class, 'character'):
|
||||
char_num = len(getattr(post_process_class, 'character'))
|
||||
config['Architecture']["Head"]['out_channels'] = char_num
|
||||
model = build_model(config['Architecture'])
|
||||
|
||||
# prepare to quant
|
||||
quanter = QAT(config=quant_config, act_preprocess=PACT)
|
||||
quanter.quantize(model)
|
||||
|
||||
if config['Global']['distributed']:
|
||||
model = paddle.DataParallel(model)
|
||||
|
||||
# build loss
|
||||
loss_class = build_loss(config['Loss'])
|
||||
|
||||
# build optim
|
||||
optimizer, lr_scheduler = build_optimizer(
|
||||
config['Optimizer'],
|
||||
epochs=config['Global']['epoch_num'],
|
||||
step_each_epoch=len(train_dataloader),
|
||||
parameters=model.parameters())
|
||||
|
||||
# build metric
|
||||
eval_class = build_metric(config['Metric'])
|
||||
# load pretrain model
|
||||
pre_best_model_dict = init_model(config, model, logger, optimizer)
|
||||
|
||||
logger.info('train dataloader has {} iters, valid dataloader has {} iters'.
|
||||
format(len(train_dataloader), len(valid_dataloader)))
|
||||
# start train
|
||||
program.train(config, train_dataloader, valid_dataloader, device, model,
|
||||
loss_class, optimizer, lr_scheduler, post_process_class,
|
||||
eval_class, pre_best_model_dict, logger, vdl_writer)
|
||||
|
||||
|
||||
def test_reader(config, device, logger):
|
||||
loader = build_dataloader(config, 'Train', device, logger)
|
||||
import time
|
||||
starttime = time.time()
|
||||
count = 0
|
||||
try:
|
||||
for data in loader():
|
||||
count += 1
|
||||
if count % 1 == 0:
|
||||
batch_time = time.time() - starttime
|
||||
starttime = time.time()
|
||||
logger.info("reader: {}, {}, {}".format(
|
||||
count, len(data[0]), batch_time))
|
||||
except Exception as e:
|
||||
logger.info(e)
|
||||
logger.info("finish reader: {}, Success!".format(count))
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
config, device, logger, vdl_writer = program.preprocess(is_train=True)
|
||||
main(config, device, logger, vdl_writer)
|
||||
# test_reader(config, device, logger)
|
||||
Loading…
Reference in new issue