未验证 提交 f88af205 编写于 作者: C cc 提交者: GitHub

Combine amp and qat (#33484)

* Combine amp and qat
* add unit test
上级 0905deec
......@@ -141,7 +141,7 @@ static inline std::shared_ptr<imperative::VarBase> CastToFP32(
}
static inline framework::proto::VarType::Type GetPromoteType(
const NameVarBaseMap& ins) {
const std::string& op_type, const NameVarBaseMap& ins) {
auto dst_type = framework::proto::VarType::FP16;
for (const auto& pair : ins) {
for (const auto& var : pair.second) {
......@@ -151,6 +151,18 @@ static inline framework::proto::VarType::Type GetPromoteType(
}
}
}
// NOTE(juncai): moving_average_abs_max_scale only consider the
// dtype of input(X)
if (op_type == "moving_average_abs_max_scale") {
for (const auto& pair : ins) {
if (pair.first == "X" &&
pair.second.front()->DataType() == framework::proto::VarType::FP16) {
dst_type = framework::proto::VarType::FP16;
}
}
}
return dst_type;
}
......@@ -183,7 +195,8 @@ NameVarBaseMap AutoCastInputs(const std::string& op_type,
}
return new_ins;
} else {
auto dst_type = GetPromoteType(ins);
auto dst_type = GetPromoteType(op_type, ins);
// NOTE(zhiqiu): if the op has op fp16 kernel, fall back to fp32.
if (dst_type == framework::proto::VarType::FP16 &&
AmpOperators::Instance().GetMutableUnsupportedFp16Ops()->count(
......
......@@ -25,18 +25,19 @@ __global__ void FindAbsMaxKernel(const T* in, const int n, T* out) {
int bid = threadIdx.x + blockIdx.x * blockDim.x;
int tid = threadIdx.x;
extern __shared__ T shared_max_data[];
extern __shared__ char* shared_max_data_tmp[];
auto shared_max_data = reinterpret_cast<T*>(shared_max_data_tmp);
if (gridDim.x > 1) {
shared_max_data[tid] = T(0);
for (int i = bid; i < n; i += blockDim.x * gridDim.x) {
T tmp = fabs(in[i]);
T tmp = abs(in[i]);
if (tmp > shared_max_data[tid]) {
shared_max_data[tid] = tmp;
}
}
} else {
if (bid < n) {
shared_max_data[tid] = fabs(in[bid]);
shared_max_data[tid] = abs(in[bid]);
} else {
shared_max_data[tid] = T(0);
}
......@@ -73,6 +74,8 @@ struct FindAbsMaxFunctor<platform::CUDADeviceContext, T> {
};
template struct FindAbsMaxFunctor<platform::CUDADeviceContext, float>;
template struct FindAbsMaxFunctor<platform::CUDADeviceContext,
paddle::platform::float16>;
template <typename T>
__global__ void FindChannelAbsMaxKernelQuantAxis0(const T* in, const int n,
......@@ -213,13 +216,16 @@ __global__ void ClipAndQuantDequantKernel(const T* in, const T* scale,
int tid = threadIdx.x;
T s = scale[0];
T inv_s = inverse(s);
T bin_cnt_t = static_cast<T>(bin_cnt);
for (int i = bid; i < n; i += blockDim.x * gridDim.x) {
T x = in[i];
T v = x > s ? s : x;
v = v < -s ? -s : v;
v = bin_cnt * inv_s * v;
out[i] = round(v) * s / bin_cnt;
x = x > s ? s : x;
x = x < -s ? -s : x;
x = (bin_cnt_t / s) * x;
x = static_cast<T>(round(static_cast<float>(x)));
out[i] = (x * s) / bin_cnt_t;
}
}
......@@ -261,9 +267,6 @@ struct ClipAndFakeQuantDequantFunctor<platform::CUDADeviceContext, T> {
}
};
template struct ClipAndFakeQuantDequantFunctor<platform::CUDADeviceContext,
float>;
// ChannelClipAndQuantKernel for quant_axis is 0
template <typename T>
__global__ void ChannelClipAndQuantKernelQuantAxis0(const T* in, const T* scale,
......@@ -423,8 +426,10 @@ struct FindMovingAverageAbsMaxFunctor<platform::CUDADeviceContext, T> {
memory::Copy(platform::CPUPlace(), &scale, gpu_place, cur_scale, sizeof(T),
ctx.stream());
ctx.Wait();
state = rate * state + 1;
accum = rate * accum + scale;
T rate_t = static_cast<T>(rate);
state = rate_t * state + static_cast<T>(1.0);
accum = rate_t * accum + scale;
scale = accum / state;
memory::Copy(gpu_place, out_accum->mutable_data<T>(gpu_place),
......@@ -527,10 +532,12 @@ template struct ChannelClipFakeQuantDequantFunctor<platform::CUDADeviceContext,
namespace ops = paddle::operators;
using CUDA = paddle::platform::CUDADeviceContext;
using float16 = paddle::platform::float16;
REGISTER_OP_CUDA_KERNEL(fake_quantize_abs_max,
ops::FakeQuantizeAbsMaxKernel<CUDA, float>);
REGISTER_OP_CUDA_KERNEL(fake_quantize_dequantize_abs_max,
ops::FakeQuantizeDequantizeAbsMaxKernel<CUDA, float>);
ops::FakeQuantizeDequantizeAbsMaxKernel<CUDA, float>,
ops::FakeQuantizeDequantizeAbsMaxKernel<CUDA, float16>);
REGISTER_OP_CUDA_KERNEL(fake_channel_wise_quantize_abs_max,
ops::FakeChannelWiseQuantizeAbsMaxKernel<CUDA, float>);
REGISTER_OP_CUDA_KERNEL(fake_quantize_range_abs_max,
......@@ -539,12 +546,15 @@ REGISTER_OP_CUDA_KERNEL(
fake_quantize_moving_average_abs_max,
ops::FakeQuantizeMovingAverageAbsMaxKernel<CUDA, float>);
REGISTER_OP_CUDA_KERNEL(moving_average_abs_max_scale,
ops::MovingAverageAbsMaxScaleKernel<CUDA, float>);
ops::MovingAverageAbsMaxScaleKernel<CUDA, float>,
ops::MovingAverageAbsMaxScaleKernel<CUDA, float16>);
REGISTER_OP_CUDA_KERNEL(
fake_quantize_dequantize_moving_average_abs_max,
ops::FakeQuantizeDequantizeMovingAverageAbsMaxKernel<CUDA, float>);
ops::FakeQuantizeDequantizeMovingAverageAbsMaxKernel<CUDA, float>,
ops::FakeQuantizeDequantizeMovingAverageAbsMaxKernel<CUDA, float16>);
REGISTER_OP_CUDA_KERNEL(stright_throuth_estimator_grad,
ops::StrightThroughEstimatorGradKernel<CUDA, float>);
ops::StrightThroughEstimatorGradKernel<CUDA, float>,
ops::StrightThroughEstimatorGradKernel<CUDA, float16>);
REGISTER_OP_CUDA_KERNEL(
fake_channel_wise_quantize_dequantize_abs_max,
ops::FakeChannelWiseQuantizeDequantizeAbsMaxKernel<CUDA, float>);
......@@ -127,6 +127,7 @@ if(WIN32)
list(REMOVE_ITEM TEST_OPS test_post_training_quantization_lstm_model)
list(REMOVE_ITEM TEST_OPS test_weight_quantization_mobilenetv1)
list(REMOVE_ITEM TEST_OPS test_quantize_transpiler_v2)
list(REMOVE_ITEM TEST_OPS test_imperative_qat_amp)
endif()
if(LINUX AND WITH_MKLDNN)
......
# copyright (c) 2018 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 print_function
import os
import numpy as np
import random
import shutil
import time
import unittest
import logging
import paddle
import paddle.fluid as fluid
from paddle.fluid.contrib.slim.quantization import ImperativeQuantAware
from paddle.fluid.log_helper import get_logger
from paddle.dataset.common import download
from imperative_test_utils import fix_model_dict, ImperativeLenet
os.environ["CPU_NUM"] = "1"
if paddle.is_compiled_with_cuda():
fluid.set_flags({"FLAGS_cudnn_deterministic": True})
_logger = get_logger(
__name__, logging.INFO, fmt='%(asctime)s-%(levelname)s: %(message)s')
class TestImperativeQatAmp(unittest.TestCase):
"""
Test the combination of qat and amp.
"""
@classmethod
def setUpClass(cls):
timestamp = time.strftime('%Y-%m-%d-%H-%M-%S', time.localtime())
cls.root_path = os.path.join(os.getcwd(),
"imperative_qat_amp_" + timestamp)
cls.save_path = os.path.join(cls.root_path, "model")
cls.download_path = 'dygraph_int8/download'
cls.cache_folder = os.path.expanduser('~/.cache/paddle/dataset/' +
cls.download_path)
cls.lenet_url = "https://paddle-inference-dist.cdn.bcebos.com/int8/unittest_model_data/lenet_pretrained.tar.gz"
cls.lenet_md5 = "953b802fb73b52fae42896e3c24f0afb"
seed = 1
np.random.seed(seed)
paddle.static.default_main_program().random_seed = seed
paddle.static.default_startup_program().random_seed = seed
@classmethod
def tearDownClass(cls):
try:
shutil.rmtree(cls.root_path)
except Exception as e:
print("Failed to delete {} due to {}".format(cls.root_path, str(e)))
def cache_unzipping(self, target_folder, zip_path):
if not os.path.exists(target_folder):
cmd = 'mkdir {0} && tar xf {1} -C {0}'.format(target_folder,
zip_path)
os.system(cmd)
def download_model(self, data_url, data_md5, folder_name):
download(data_url, self.download_path, data_md5)
file_name = data_url.split('/')[-1]
zip_path = os.path.join(self.cache_folder, file_name)
print('Data is downloaded at {0}'.format(zip_path))
data_cache_folder = os.path.join(self.cache_folder, folder_name)
self.cache_unzipping(data_cache_folder, zip_path)
return data_cache_folder
def set_vars(self):
self.qat = ImperativeQuantAware()
self.train_batch_num = 30
self.train_batch_size = 32
self.test_batch_num = 100
self.test_batch_size = 32
self.eval_acc_top1 = 0.99
def model_train(self, model, batch_num=-1, batch_size=32, use_amp=False):
model.train()
train_reader = paddle.batch(
paddle.dataset.mnist.train(), batch_size=batch_size)
adam = paddle.optimizer.Adam(
learning_rate=0.001, parameters=model.parameters())
scaler = paddle.amp.GradScaler(init_loss_scaling=500)
for batch_id, data in enumerate(train_reader()):
x_data = np.array([x[0].reshape(1, 28, 28)
for x in data]).astype('float32')
y_data = np.array(
[x[1] for x in data]).astype('int64').reshape(-1, 1)
img = paddle.to_tensor(x_data)
label = paddle.to_tensor(y_data)
if use_amp:
with paddle.amp.auto_cast():
out = model(img)
acc = fluid.layers.accuracy(out, label)
loss = fluid.layers.cross_entropy(out, label)
avg_loss = fluid.layers.mean(loss)
scaled_loss = scaler.scale(avg_loss)
scaled_loss.backward()
scaler.minimize(adam, scaled_loss)
adam.clear_gradients()
else:
out = model(img)
acc = fluid.layers.accuracy(out, label)
loss = fluid.layers.cross_entropy(out, label)
avg_loss = fluid.layers.mean(loss)
avg_loss.backward()
adam.minimize(avg_loss)
model.clear_gradients()
if batch_id % 100 == 0:
_logger.info("Train | step {}: loss = {:}, acc= {:}".format(
batch_id, avg_loss.numpy(), acc.numpy()))
if batch_num > 0 and batch_id + 1 >= batch_num:
break
def model_test(self, model, batch_num=-1, batch_size=32, use_amp=False):
model.eval()
test_reader = paddle.batch(
paddle.dataset.mnist.test(), batch_size=batch_size)
acc_top1_list = []
for batch_id, data in enumerate(test_reader()):
x_data = np.array([x[0].reshape(1, 28, 28)
for x in data]).astype('float32')
y_data = np.array(
[x[1] for x in data]).astype('int64').reshape(-1, 1)
img = paddle.to_tensor(x_data)
label = paddle.to_tensor(y_data)
with paddle.amp.auto_cast(use_amp):
out = model(img)
acc_top1 = fluid.layers.accuracy(input=out, label=label, k=1)
acc_top5 = fluid.layers.accuracy(input=out, label=label, k=5)
acc_top1_list.append(float(acc_top1.numpy()))
if batch_id % 100 == 0:
_logger.info("Test | At step {}: acc1 = {:}, acc5 = {:}".format(
batch_id, acc_top1.numpy(), acc_top5.numpy()))
if batch_num > 0 and batch_id + 1 >= batch_num:
break
acc_top1 = sum(acc_top1_list) / len(acc_top1_list)
return acc_top1
def test_ptq(self):
start_time = time.time()
self.set_vars()
params_path = self.download_model(self.lenet_url, self.lenet_md5,
"lenet")
params_path += "/lenet_pretrained/lenet.pdparams"
with fluid.dygraph.guard():
model = ImperativeLenet()
model_state_dict = paddle.load(params_path)
model.set_state_dict(model_state_dict)
_logger.info("Test fp32 model")
fp32_acc_top1 = self.model_test(model, self.test_batch_num,
self.test_batch_size)
self.qat.quantize(model)
use_amp = True
self.model_train(model, self.train_batch_num, self.train_batch_size,
use_amp)
_logger.info("Test int8 model")
int8_acc_top1 = self.model_test(model, self.test_batch_num,
self.test_batch_size, use_amp)
_logger.info('fp32_acc_top1: %f, int8_acc_top1: %f' %
(fp32_acc_top1, int8_acc_top1))
self.assertTrue(
int8_acc_top1 > fp32_acc_top1 - 0.01,
msg='fp32_acc_top1: %f, int8_acc_top1: %f' %
(fp32_acc_top1, int8_acc_top1))
input_spec = [
paddle.static.InputSpec(
shape=[None, 1, 28, 28], dtype='float32')
]
paddle.jit.save(layer=model, path=self.save_path, input_spec=input_spec)
print('Quantized model saved in {%s}' % self.save_path)
end_time = time.time()
print("total time: %ss" % (end_time - start_time))
if __name__ == '__main__':
unittest.main()
......@@ -29,6 +29,8 @@ WHITE_LIST = {
'matmul',
'matmul_v2',
'mul',
'fake_quantize_dequantize_abs_max',
'fake_quantize_dequantize_moving_average_abs_max',
}
# The set of ops that support fp16 calculation and are considered numerically-
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册