Skip to content

P
PaddleSlim

PaddlePaddle / PaddleSlim
大约 2 年 前同步成功

通知 51
Star 1434
Fork 344
P
PaddleSlim
“6dd2a578dbdba2737e7250d2e91cd0c262840837”上不存在“tests/script/git@gitcode.net:taosdata/tdengine.git”
提交 bb0f8fbb 编写于 作者: W wanghaoshuang
浏览文件
操作

Merge branch 'develop' of https://github.com/PaddlePaddle/PaddleSlim into base

上级 ae093d57 2cd6b44d
隐藏空白更改
内联 并排
Showing with 402 addition and 178 deletion
demo/nas/block_sa_nas_mobilenetv2.py
浏览文件 @ bb0f8fbb
...@@ -16,13 +16,6 @@ import imagenet_reader ...@@ -16,13 +16,6 @@ import imagenet_reader
_logger = get_logger(__name__, level=logging.INFO) _logger = get_logger(__name__, level=logging.INFO)
reduce_rate = 0.85
init_temperature = 10.24
max_flops = 321208544
server_address = ""
port = 8979
retain_epoch = 5
def create_data_loader(image_shape): def create_data_loader(image_shape):
data_shape = [None] + image_shape data_shape = [None] + image_shape
...@@ -71,17 +64,13 @@ def search_mobilenetv2_block(config, args, image_size): ...@@ -71,17 +64,13 @@ def search_mobilenetv2_block(config, args, image_size):
if args.is_server: if args.is_server:
sa_nas = SANAS( sa_nas = SANAS(
config, config,
server_addr=("", port), server_addr=(args.server_address, args.port),
init_temperature=init_temperature,
reduce_rate=reduce_rate,
search_steps=args.search_steps, search_steps=args.search_steps,
is_server=True) is_server=True)
else: else:
sa_nas = SANAS( sa_nas = SANAS(
config, config,
server_addr=(server_address, port), server_addr=(args.server_address, args.port),
init_temperature=init_temperature,
reduce_rate=reduce_rate,
search_steps=args.search_steps, search_steps=args.search_steps,
is_server=False) is_server=False)
...@@ -140,7 +129,7 @@ def search_mobilenetv2_block(config, args, image_size): ...@@ -140,7 +129,7 @@ def search_mobilenetv2_block(config, args, image_size):
current_flops = flops(train_program) current_flops = flops(train_program)
print('step: {}, current_flops: {}'.format(step, current_flops)) print('step: {}, current_flops: {}'.format(step, current_flops))
if current_flops > max_flops: if current_flops > int(321208544):
continue continue
place = fluid.CUDAPlace(0) if args.use_gpu else fluid.CPUPlace() place = fluid.CUDAPlace(0) if args.use_gpu else fluid.CPUPlace()
...@@ -178,7 +167,7 @@ def search_mobilenetv2_block(config, args, image_size): ...@@ -178,7 +167,7 @@ def search_mobilenetv2_block(config, args, image_size):
train_compiled_program = fluid.CompiledProgram( train_compiled_program = fluid.CompiledProgram(
train_program).with_data_parallel( train_program).with_data_parallel(
loss_name=avg_cost.name, build_strategy=build_strategy) loss_name=avg_cost.name, build_strategy=build_strategy)
for epoch_id in range(retain_epoch): for epoch_id in range(args.retain_epoch):
for batch_id, data in enumerate(train_loader()): for batch_id, data in enumerate(train_loader()):
fetches = [avg_cost.name] fetches = [avg_cost.name]
s_time = time.time() s_time = time.time()
...@@ -243,6 +232,11 @@ if __name__ == '__main__': ...@@ -243,6 +232,11 @@ if __name__ == '__main__':
type=int, type=int,
default=100, default=100,
help='controller server number.') help='controller server number.')
parser.add_argument(
'--server_address', type=str, default="", help='server ip.')
parser.add_argument('--port', type=int, default=8881, help='server port')
parser.add_argument(
'--retain_epoch', type=int, default=5, help='epoch for each token.')
parser.add_argument('--lr', type=float, default=0.1, help='learning rate.') parser.add_argument('--lr', type=float, default=0.1, help='learning rate.')
args = parser.parse_args() args = parser.parse_args()
print(args) print(args)
...@@ -257,7 +251,7 @@ if __name__ == '__main__': ...@@ -257,7 +251,7 @@ if __name__ == '__main__':
args.data)) args.data))
# block mask means block number, 1 mean downsample, 0 means the size of feature map don't change after this block # block mask means block number, 1 mean downsample, 0 means the size of feature map don't change after this block
config_info = {'block_mask': [0, 1, 1, 1, 1, 0, 1, 0]} config_info = {'block_mask': [0, 1, 1, 1, 0]}
config = [('MobileNetV2BlockSpace', config_info)] config = [('MobileNetV2BlockSpace', config_info)]
search_mobilenetv2_block(config, args, image_size) search_mobilenetv2_block(config, args, image_size)
demo/nas/sa_nas_mobilenetv2.py
浏览文件 @ bb0f8fbb
...@@ -18,13 +18,6 @@ import imagenet_reader ...@@ -18,13 +18,6 @@ import imagenet_reader
_logger = get_logger(__name__, level=logging.INFO) _logger = get_logger(__name__, level=logging.INFO)
reduce_rate = 0.85
init_temperature = 10.24
max_flops = 321208544
server_address = ""
port = 8989
retain_epoch = 5
def create_data_loader(image_shape): def create_data_loader(image_shape):
data_shape = [None] + image_shape data_shape = [None] + image_shape
...@@ -66,18 +59,14 @@ def search_mobilenetv2(config, args, image_size, is_server=True): ...@@ -66,18 +59,14 @@ def search_mobilenetv2(config, args, image_size, is_server=True):
### start a server and a client ### start a server and a client
sa_nas = SANAS( sa_nas = SANAS(
config, config,
server_addr=("", port), server_addr=(args.server_address, args.port),
init_temperature=init_temperature,
reduce_rate=reduce_rate,
search_steps=args.search_steps, search_steps=args.search_steps,
is_server=True) is_server=True)
else: else:
### start a client ### start a client
sa_nas = SANAS( sa_nas = SANAS(
config, config,
server_addr=(server_address, port), server_addr=(args.server_address, args.port),
init_temperature=init_temperature,
reduce_rate=reduce_rate,
search_steps=args.search_steps, search_steps=args.search_steps,
is_server=False) is_server=False)
...@@ -93,7 +82,7 @@ def search_mobilenetv2(config, args, image_size, is_server=True): ...@@ -93,7 +82,7 @@ def search_mobilenetv2(config, args, image_size, is_server=True):
current_flops = flops(train_program) current_flops = flops(train_program)
print('step: {}, current_flops: {}'.format(step, current_flops)) print('step: {}, current_flops: {}'.format(step, current_flops))
if current_flops > max_flops: if current_flops > int(321208544):
continue continue
test_loader, test_avg_cost, test_acc_top1, test_acc_top5 = build_program( test_loader, test_avg_cost, test_acc_top1, test_acc_top5 = build_program(
...@@ -139,7 +128,7 @@ def search_mobilenetv2(config, args, image_size, is_server=True): ...@@ -139,7 +128,7 @@ def search_mobilenetv2(config, args, image_size, is_server=True):
train_compiled_program = fluid.CompiledProgram( train_compiled_program = fluid.CompiledProgram(
train_program).with_data_parallel( train_program).with_data_parallel(
loss_name=avg_cost.name, build_strategy=build_strategy) loss_name=avg_cost.name, build_strategy=build_strategy)
for epoch_id in range(retain_epoch): for epoch_id in range(args.retain_epoch):
for batch_id, data in enumerate(train_loader()): for batch_id, data in enumerate(train_loader()):
fetches = [avg_cost.name] fetches = [avg_cost.name]
s_time = time.time() s_time = time.time()
...@@ -179,7 +168,7 @@ def search_mobilenetv2(config, args, image_size, is_server=True): ...@@ -179,7 +168,7 @@ def search_mobilenetv2(config, args, image_size, is_server=True):
def test_search_result(tokens, image_size, args, config): def test_search_result(tokens, image_size, args, config):
sa_nas = SANAS( sa_nas = SANAS(
config, config,
server_addr=("", 8887), server_addr=(args.server_address, args.port),
init_temperature=args.init_temperature, init_temperature=args.init_temperature,
reduce_rate=args.reduce_rate, reduce_rate=args.reduce_rate,
search_steps=args.search_steps, search_steps=args.search_steps,
...@@ -234,7 +223,7 @@ def test_search_result(tokens, image_size, args, config): ...@@ -234,7 +223,7 @@ def test_search_result(tokens, image_size, args, config):
train_compiled_program = fluid.CompiledProgram( train_compiled_program = fluid.CompiledProgram(
train_program).with_data_parallel( train_program).with_data_parallel(
loss_name=avg_cost.name, build_strategy=build_strategy) loss_name=avg_cost.name, build_strategy=build_strategy)
for epoch_id in range(retain_epoch): for epoch_id in range(args.retain_epoch):
for batch_id, data in enumerate(train_loader()): for batch_id, data in enumerate(train_loader()):
fetches = [avg_cost.name] fetches = [avg_cost.name]
s_time = time.time() s_time = time.time()
...@@ -298,6 +287,11 @@ if __name__ == '__main__': ...@@ -298,6 +287,11 @@ if __name__ == '__main__':
type=int, type=int,
default=100, default=100,
help='controller server number.') help='controller server number.')
parser.add_argument(
'--server_address', type=str, default="", help='server ip.')
parser.add_argument('--port', type=int, default=8881, help='server port')
parser.add_argument(
'--retain_epoch', type=int, default=5, help='epoch for each token.')
parser.add_argument('--lr', type=float, default=0.1, help='learning rate.') parser.add_argument('--lr', type=float, default=0.1, help='learning rate.')
args = parser.parse_args() args = parser.parse_args()
print(args) print(args)
......
demo/pantheon/run_student.py
浏览文件 @ bb0f8fbb
...@@ -80,8 +80,8 @@ def run(args): ...@@ -80,8 +80,8 @@ def run(args):
student.start() student.start()
if args.test_send_recv: if args.test_send_recv:
for t in xrange(2): for t in range(2):
for i in xrange(3): for i in range(3):
print(student.recv(t)) print(student.recv(t))
student.send("message from student!") student.send("message from student!")
......
demo/prune/README.md
浏览文件 @ bb0f8fbb
...@@ -17,7 +17,20 @@ ...@@ -17,7 +17,20 @@
1). 根据分类模型中[ImageNet数据准备文档](https://github.com/PaddlePaddle/models/tree/develop/PaddleCV/image_classification#%E6%95%B0%E6%8D%AE%E5%87%86%E5%A4%87)下载数据到`PaddleSlim/demo/data/ILSVRC2012`路径下。 1). 根据分类模型中[ImageNet数据准备文档](https://github.com/PaddlePaddle/models/tree/develop/PaddleCV/image_classification#%E6%95%B0%E6%8D%AE%E5%87%86%E5%A4%87)下载数据到`PaddleSlim/demo/data/ILSVRC2012`路径下。
2). 使用`train.py`脚本时，指定`--data`选项为`imagenet`. 2). 使用`train.py`脚本时，指定`--data`选项为`imagenet`.
## 2. 启动剪裁任务 ## 2. 下载预训练模型
如果使用`ImageNet`数据，建议在预训练模型的基础上进行剪裁，请从[分类库](https://github.com/PaddlePaddle/models/tree/develop/PaddleCV/image_classification#%E5%B7%B2%E5%8F%91%E5%B8%83%E6%A8%A1%E5%9E%8B%E5%8F%8A%E5%85%B6%E6%80%A7%E8%83%BD)中下载合适的预训练模型。
这里以`MobileNetV1`为例，下载并解压预训练模型到当前路径：
```
wget http://paddle-imagenet-models-name.bj.bcebos.com/MobileNetV1_pretrained.tar
tar -xf MobileNetV1_pretrained.tar
```
使用`train.py`脚本时，指定`--pretrained_model`加载预训练模型。
## 3. 启动剪裁任务
通过以下命令启动裁剪任务： 通过以下命令启动裁剪任务：
...@@ -25,8 +38,8 @@ ...@@ -25,8 +38,8 @@
export CUDA_VISIBLE_DEVICES=0 export CUDA_VISIBLE_DEVICES=0
python train.py \ python train.py \
--model "MobileNet" \ --model "MobileNet" \
--pruned_ratio 0.33 \ --pruned_ratio 0.31 \
--data "imagenet" --data "mnist"
``` ```
其中，`model`用于指定待裁剪的模型。`pruned_ratio`用于指定各个卷积层通道数被裁剪的比例。`data`选项用于指定使用的数据集。 其中，`model`用于指定待裁剪的模型。`pruned_ratio`用于指定各个卷积层通道数被裁剪的比例。`data`选项用于指定使用的数据集。
...@@ -35,7 +48,7 @@ python train.py \ ...@@ -35,7 +48,7 @@ python train.py \
在本示例中，会在日志中输出剪裁前后的`FLOPs`，并且每训练一轮就会保存一个模型到文件系统。 在本示例中，会在日志中输出剪裁前后的`FLOPs`，并且每训练一轮就会保存一个模型到文件系统。
## 3. 加载和评估模型 ## 4. 加载和评估模型
本节介绍如何加载训练过程中保存的模型。 本节介绍如何加载训练过程中保存的模型。
...@@ -43,14 +56,14 @@ python train.py \ ...@@ -43,14 +56,14 @@ python train.py \
``` ```
python eval.py \ python eval.py \
--model "mobilenet" \ --model "MobileNet" \
--data "mnist" \ --data "mnist" \
--model_path "./models/0" --model_path "./models/0"
``` ```
在脚本`eval.py`中，使用`paddleslim.prune.load_model`接口加载剪裁得到的模型。 在脚本`eval.py`中，使用`paddleslim.prune.load_model`接口加载剪裁得到的模型。
## 4. 接口介绍 ## 5. 接口介绍
该示例使用了`paddleslim.Pruner`工具类，用户接口使用介绍请参考：[API文档](https://paddlepaddle.github.io/PaddleSlim/api/prune_api/) 该示例使用了`paddleslim.Pruner`工具类，用户接口使用介绍请参考：[API文档](https://paddlepaddle.github.io/PaddleSlim/api/prune_api/)
......
demo/prune/eval.py
浏览文件 @ bb0f8fbb
...@@ -68,7 +68,7 @@ def eval(args): ...@@ -68,7 +68,7 @@ def eval(args):
val_feeder = feeder = fluid.DataFeeder( val_feeder = feeder = fluid.DataFeeder(
[image, label], place, program=val_program) [image, label], place, program=val_program)
load_model(val_program, "./model/mobilenetv1_prune_50") load_model(exe, val_program, args.model_path)
batch_id = 0 batch_id = 0
acc_top1_ns = [] acc_top1_ns = []
......
demo/prune/train.py
浏览文件 @ bb0f8fbb
...@@ -8,6 +8,7 @@ import math ...@@ -8,6 +8,7 @@ import math
import time import time
import numpy as np import numpy as np
import paddle.fluid as fluid import paddle.fluid as fluid
sys.path.append("../../")
from paddleslim.prune import Pruner, save_model from paddleslim.prune import Pruner, save_model
from paddleslim.common import get_logger from paddleslim.common import get_logger
from paddleslim.analysis import flops from paddleslim.analysis import flops
...@@ -37,6 +38,7 @@ add_arg('log_period', int, 10, "Log period in batches.") ...@@ -37,6 +38,7 @@ add_arg('log_period', int, 10, "Log period in batches.")
add_arg('test_period', int, 10, "Test period in epoches.") add_arg('test_period', int, 10, "Test period in epoches.")
add_arg('model_path', str, "./models", "The path to save model.") add_arg('model_path', str, "./models", "The path to save model.")
add_arg('pruned_ratio', float, None, "The ratios to be pruned.") add_arg('pruned_ratio', float, None, "The ratios to be pruned.")
add_arg('criterion', str, "l1_norm", "The prune criterion to be used, support l1_norm and batch_norm_scale.")
# yapf: enable # yapf: enable
model_list = models.__all__ model_list = models.__all__
...@@ -136,6 +138,8 @@ def compress(args): ...@@ -136,6 +138,8 @@ def compress(args):
return os.path.exists( return os.path.exists(
os.path.join(args.pretrained_model, var.name)) os.path.join(args.pretrained_model, var.name))
_logger.info("Load pretrained model from {}".format(
args.pretrained_model))
fluid.io.load_vars(exe, args.pretrained_model, predicate=if_exist) fluid.io.load_vars(exe, args.pretrained_model, predicate=if_exist)
val_reader = paddle.batch(val_reader, batch_size=args.batch_size) val_reader = paddle.batch(val_reader, batch_size=args.batch_size)
...@@ -200,10 +204,12 @@ def compress(args): ...@@ -200,10 +204,12 @@ def compress(args):
end_time - start_time)) end_time - start_time))
batch_id += 1 batch_id += 1
test(0, val_program)
params = get_pruned_params(args, fluid.default_main_program()) params = get_pruned_params(args, fluid.default_main_program())
_logger.info("FLOPs before pruning: {}".format( _logger.info("FLOPs before pruning: {}".format(
flops(fluid.default_main_program()))) flops(fluid.default_main_program())))
pruner = Pruner() pruner = Pruner(args.criterion)
pruned_val_program, _, _ = pruner.prune( pruned_val_program, _, _ = pruner.prune(
val_program, val_program,
fluid.global_scope(), fluid.global_scope(),
......
docs/en/model_zoo_en.md
浏览文件 @ bb0f8fbb
# Model Zoo # Model Zoo
## 1. 图象分类 ## 1. Image Classification
数据集：ImageNet1000类 Dataset：ImageNet1000
### 1.1 量化 ### 1.1 Quantization
| 模型 | 压缩方法 | Top-1/Top-5 Acc | 模型体积（MB） | 下载 | | Model | Method | Top-1/Top-5 Acc | Model Size（MB） | TensorRT latency(V100, ms) | Download |
|:--:|:---:|:--:|:--:|:--:|:--:|
|MobileNetV1|-|70.99%/89.68%| 17 | -| [model](http://paddle-imagenet-models-name.bj.bcebos.com/MobileNetV1_pretrained.tar) |
|MobileNetV1|quant_post|70.18%/89.25% (-0.81%/-0.43%)| 4.4 | - | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/MobileNetV1_quant_post.tar) |
|MobileNetV1|quant_aware|70.60%/89.57% (-0.39%/-0.11%)| 4.4 | -| [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/MobileNetV1_quant_aware.tar) |
| MobileNetV2 | - |72.15%/90.65%| 15 | - | [model](https://paddle-imagenet-models-name.bj.bcebos.com/MobileNetV2_pretrained.tar) |
| MobileNetV2 | quant_post | 71.15%/90.11% (-1%/-0.54%)| 4.0 | - | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/MobileNetV2_quant_post.tar) |
| MobileNetV2 | quant_aware |72.05%/90.63% (-0.1%/-0.02%)| 4.0 | - | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/MobileNetV2_quant_aware.tar) |
|ResNet50|-|76.50%/93.00%| 99 | 2.71 | [model](http://paddle-imagenet-models-name.bj.bcebos.com/ResNet50_pretrained.tar) |
|ResNet50|quant_post|76.33%/93.02% (-0.17%/+0.02%)| 25.1| 1.19 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/ResNet50_quant_post.tar) |
|ResNet50|quant_aware| 76.48%/93.11% (-0.02%/+0.11%)| 25.1 | 1.17 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/ResNet50_quant_awre.tar) |
PaddleLite latency(ms)
| Device | Model | Method | armv7 Thread 1 | armv7 Thread 2 | armv7 Thread 4 | armv8 Thread 1 | armv8 Thread 2 | armv8 Thread 4 |
| ------- | ----------- | ------------- | -------------- | -------------- | -------------- | -------------- | -------------- | -------------- |
| Qualcomm 835 | MobileNetV1 | FP32 baseline | 96.1942 | 53.2058 | 32.4468 | 88.4955 | 47.95 | 27.5189 |
| Qualcomm 835 | MobileNetV1 | quant_aware | 60.8186 | 32.1931 | 16.4275 | 56.4311 | 29.5446 | 15.1053 |
| Qualcomm 835 | MobileNetV1 | quant_post | 60.5615 | 32.4016 | 16.6596 | 56.5266 | 29.7178 | 15.1459 |
| Qualcomm 835 | MobileNetV2 | FP32 baseline | 65.715 | 38.1346 | 25.155 | 61.3593 | 36.2038 | 22.849 |
| Qualcomm 835 | MobileNetV2 | quant_aware | 48.3655 | 30.2021 | 21.9303 | 46.1487 | 27.3146 | 18.3053 |
| Qualcomm 835 | MobileNetV2 | quant_post | 48.3495 | 30.3069 | 22.1506 | 45.8715 | 27.4105 | 18.2223 |
| Qualcomm 835 | ResNet50 | FP32 baseline | 526.811 | 319.6486 | 205.8345 | 506.1138 | 335.1584 | 214.8936 |
| Qualcomm 835 | ResNet50 | quant_aware | 475.4538 | 256.8672 | 139.699 | 461.7344 | 247.9506 | 145.9847 |
| Qualcomm 835 | ResNet50 | quant_post | 476.0507 | 256.5963 | 139.7266 | 461.9176 | 248.3795 | 149.353 |
| Qualcomm 855 | MobileNetV1 | FP32 baseline | 33.5086 | 19.5773 | 11.7534 | 31.3474 | 18.5382 | 10.0811 |
| Qualcomm 855 | MobileNetV1 | quant_aware | 36.7067 | 21.628 | 11.0372 | 14.0238 | 8.199 | 4.2588 |
| Qualcomm 855 | MobileNetV1 | quant_post | 37.0498 | 21.7081 | 11.0779 | 14.0947 | 8.1926 | 4.2934 |
| Qualcomm 855 | MobileNetV2 | FP32 baseline | 25.0396 | 15.2862 | 9.6609 | 22.909 | 14.1797 | 8.8325 |
| Qualcomm 855 | MobileNetV2 | quant_aware | 28.1583 | 18.3317 | 11.8103 | 16.9158 | 11.1606 | 7.4148 |
| Qualcomm 855 | MobileNetV2 | quant_post | 28.1631 | 18.3917 | 11.8333 | 16.9399 | 11.1772 | 7.4176 |
| Qualcomm 855 | ResNet50 | FP32 baseline | 185.3705 | 113.0825 | 87.0741 | 177.7367 | 110.0433 | 74.4114 |
| Qualcomm 855 | ResNet50 | quant_aware | 327.6883 | 202.4536 | 106.243 | 243.5621 | 150.0542 | 78.4205 |
| Qualcomm 855 | ResNet50 | quant_post | 328.2683 | 201.9937 | 106.744 | 242.6397 | 150.0338 | 79.8659 |
| Kirin 970 | MobileNetV1 | FP32 baseline | 101.2455 | 56.4053 | 35.6484 | 94.8985 | 51.7251 | 31.9511 |
| Kirin 970 | MobileNetV1 | quant_aware | 62.5012 | 32.1863 | 16.6018 | 57.7477 | 29.2116 | 15.0703 |
| Kirin 970 | MobileNetV1 | quant_post | 62.4412 | 32.2585 | 16.6215 | 57.825 | 29.2573 | 15.1206 |
| Kirin 970 | MobileNetV2 | FP32 baseline | 70.4176 | 42.0795 | 25.1939 | 68.9597 | 39.2145 | 22.6617 |
| Kirin 970 | MobileNetV2 | quant_aware | 52.9961 | 31.5323 | 22.1447 | 49.4858 | 28.0856 | 18.7287 |
| Kirin 970 | MobileNetV2 | quant_post | 53.0961 | 31.7987 | 21.8334 | 49.383 | 28.2358 | 18.3642 |
| Kirin 970 | ResNet50 | FP32 baseline | 586.8943 | 344.0858 | 228.2293 | 573.3344 | 351.4332 | 225.8006 |
| Kirin 970 | ResNet50 | quant_aware | 488.361 | 260.1697 | 142.416 | 479.5668 | 249.8485 | 138.1742 |
| Kirin 970 | ResNet50 | quant_post | 489.6188 | 258.3279 | 142.6063 | 480.0064 | 249.5339 | 138.5284 |
### 1.2 Pruning
| Model | Method | Top-1/Top-5 Acc | Model Size（MB） | GFLOPs | Download |
|:--:|:---:|:--:|:--:|:--:|:--:|
| MobileNetV1 | Baseline | 70.99%/89.68% | 17 | 1.11 | [model](http://paddle-imagenet-models-name.bj.bcebos.com/MobileNetV1_pretrained.tar) |
| MobileNetV1 | uniform -50% | 69.4%/88.66% (-1.59%/-1.02%) | 9 | 0.56 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/MobileNetV1_uniform-50.tar) |
| MobileNetV1 | sensitive -30% | 70.4%/89.3% (-0.59%/-0.38%) | 12 | 0.74 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/MobileNetV1_sensitive-30.tar) |
| MobileNetV1 | sensitive -50% | 69.8% / 88.9% (-1.19%/-0.78%) | 9 | 0.56 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/MobileNetV1_sensitive-50.tar) |
| MobileNetV2 | - | 72.15%/90.65% | 15 | 0.59 | [model](https://paddle-imagenet-models-name.bj.bcebos.com/MobileNetV2_pretrained.tar) |
| MobileNetV2 | uniform -50% | 65.79%/86.11% (-6.35%/-4.47%) | 11 | 0.296 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/MobileNetV2_uniform-50.tar) |
| ResNet34 | - | 72.15%/90.65% | 84 | 7.36 | [model](https://paddle-imagenet-models-name.bj.bcebos.com/ResNet34_pretrained.tar) |
| ResNet34 | uniform -50% | 70.99%/89.95% (-1.36%/-0.87%) | 41 | 3.67 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/ResNet34_uniform-50.tar) |
| ResNet34 | auto -55.05% | 70.24%/89.63% (-2.04%/-1.06%) | 33 | 3.31 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/ResNet34_auto-55.tar) |
### 1.3 Distillation
| Model | Method | Top-1/Top-5 Acc | Model Size（MB） | Download |
|:--:|:---:|:--:|:--:|:--:| |:--:|:---:|:--:|:--:|:--:|
|MobileNetV1|-|70.99%/89.68%| xx | [下载链接]() | | MobileNetV1 | student | 70.99%/89.68% | 17 | [model](http://paddle-imagenet-models-name.bj.bcebos.com/MobileNetV1_pretrained.tar) |
|MobileNetV1|quant_post|xx%/xx%| xx | [下载链接]() | |ResNet50_vd|teacher|79.12%/94.44%| 99 | [model](https://paddle-imagenet-models-name.bj.bcebos.com/ResNet50_vd_pretrained.tar) |
|MobileNetV1|quant_aware|xx%/xx%| xx | [下载链接]() | |MobileNetV1|ResNet50_vd<sup>[1](#trans1)</sup> distill|72.77%/90.68% (+1.78%/+1.00%)| 17 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/MobileNetV1_distilled.tar) |
| MobileNetV2 | - |72.15%/90.65%| xx | [下载链接]() | | MobileNetV2 | student | 72.15%/90.65% | 15 | [model](https://paddle-imagenet-models-name.bj.bcebos.com/MobileNetV2_pretrained.tar) |
| MobileNetV2 | quant_post |xx%/xx%| xx | [下载链接]() | | MobileNetV2 | ResNet50_vd distill | 74.28%/91.53% (+2.13%/+0.88%) | 15 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/MobileNetV2_distilled.tar) |
| MobileNetV2 | quant_aware |xx%/xx%| xx | [下载链接]() | | ResNet50 | student | 76.50%/93.00% | 99 | [model](http://paddle-imagenet-models-name.bj.bcebos.com/ResNet50_pretrained.tar) |
|ResNet50|-|76.50%/93.00%| xx | [下载链接]() | |ResNet101|teacher|77.56%/93.64%| 173 | [model](http://paddle-imagenet-models-name.bj.bcebos.com/ResNet101_pretrained.tar) |
|ResNet50|quant_post|xx%/xx%| xx | [下载链接]() | | ResNet50 | ResNet101 distill | 77.29%/93.65% (+0.79%/+0.65%) | 99 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/ResNet50_distilled.tar) |
|ResNet50|quant_aware|xx%/xx%| xx | [下载链接]() |
!!! note "Note"
<a name="trans1">[1]</a>：The `_vd` suffix indicates that the pre-trained model uses Mixup. Please refer to the detailed introduction: [mixup: Beyond Empirical Risk Minimization](https://arxiv.org/abs/1710.09412)
### 1.2 剪裁
### 1.4 NAS
| 模型 | 压缩方法 | Top-1/Top-5 Acc | 模型体积（MB） | GFLOPs | 下载 | | Model | Method | Top-1/Top-5 Acc | Volume（MB） | GFLOPs | Download |
|:--:|:---:|:--:|:--:|:--:|:--:| |:--:|:---:|:--:|:--:|:--:|:--:|
| MobileNetV1 | Baseline | 70.99%/89.68% | 17 | 1.11 | [下载链接](http://paddle-imagenet-models-name.bj.bcebos.com/MobileNetV1_pretrained.tar) | | MobileNetV2 | - | 72.15%/90.65% | 15 | 0.59 | [model](https://paddle-imagenet-models-name.bj.bcebos.com/MobileNetV2_pretrained.tar) |
| MobileNetV1 | uniform -50% | 69.4%/88.66% (-1.59%/-1.02%) | 9 | 0.56 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/MobileNetV1_uniform-50.tar) | | MobileNetV2 | SANAS | 71.518%/90.208% (-0.632%/-0.442%) | 14 | 0.295 | [model](https://paddlemodels.cdn.bcebos.com/PaddleSlim/MobileNetV2_sanas.tar) |
| MobileNetV1 | sensitive -30% | 70.4%/89.3% (-0.59%/-0.38%) | 12 | 0.74 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/MobileNetV1_sensitive-30.tar) |
| MobileNetV1 | sensitive -50% | 69.8% / 88.9% (-1.19%/-0.78%) | 9 | 0.56 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/MobileNetV1_sensitive-50.tar) |
| MobileNetV2 | - | 72.15%/90.65% | 15 | 0.59 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/MobileNetV2_pretrained.tar) |
| MobileNetV2 | uniform -50% | 65.79%/86.11% (-6.35%/-4.47%) | 11 | 0.296 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/MobileNetV2_uniform-50.tar) |
| ResNet34 | - | 72.15%/90.65% | 84 | 7.36 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/ResNet34_pretrained.tar) |
| ResNet34 | uniform -50% | 70.99%/89.95% (-1.36%/-0.87%) | 41 | 3.67 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/ResNet34_uniform-50.tar) |
| ResNet34 | auto -55.05% | 70.24%/89.63% (-2.04%/-1.06%) | 33 | 3.31 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/ResNet34_auto-55.tar) |
## 2. Object Detection
### 2.1 Quantization
Dataset： COCO 2017
### 1.3 蒸馏 | Model | Method | Dataset | Image/GPU | Input 608 Box AP | Input 416 Box AP | Input 320 Box AP | Model Size（MB） | TensorRT latency(V100, ms) | Download |
| :----------------------------: | :---------: | :----: | :-------: | :------------: | :------------: | :------------: | :------------: | :----------: |:----------: |
| MobileNet-V1-YOLOv3 | - | COCO | 8 | 29.3 | 29.3 | 27.1 | 95 | - | [model](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1.tar) |
| MobileNet-V1-YOLOv3 | quant_post | COCO | 8 | 27.9 (-1.4)| 28.0 (-1.3) | 26.0 (-1.0) | 25 | - | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_quant_post.tar) |
| MobileNet-V1-YOLOv3 | quant_aware | COCO | 8 | 28.1 (-1.2)| 28.2 (-1.1) | 25.8 (-1.2) | 26.3 | - | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenet_coco_quant_aware.tar) |
| R34-YOLOv3 | - | COCO | 8 | 36.2 | 34.3 | 31.4 | 162 | - | [model](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r34.tar) |
| R34-YOLOv3 | quant_post | COCO | 8 | 35.7 (-0.5) | - | - | 42.7 | - | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r34_coco_quant_post.tar) |
| R34-YOLOv3 | quant_aware | COCO | 8 | 35.2 (-1.0) | 33.3 (-1.0) | 30.3 (-1.1)| 44 | - | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r34_coco_quant_aware.tar) |
| R50-dcn-YOLOv3 obj365_pretrain | - | COCO | 8 | 41.4 | - | - | 177 | 18.56 |[model](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r50vd_dcn_obj365_pretrained_coco.tar) |
| R50-dcn-YOLOv3 obj365_pretrain | quant_aware | COCO | 8 | 40.6 (-0.8) | 37.5 | 34.1 | 66 | 14.64 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r50vd_dcn_obj365_pretrained_coco_quant_aware.tar) |
| 模型 | 压缩方法 | Top-1/Top-5 Acc | 模型体积（MB） | 下载 |
|:--:|:---:|:--:|:--:|:--:|
| MobileNetV1 | student | 70.99%/89.68% | 17 | [下载链接](http://paddle-imagenet-models-name.bj.bcebos.com/MobileNetV1_pretrained.tar) |
|ResNet50_vd|teacher|79.12%/94.44%| 99 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/ResNet50_vd_pretrained.tar) |
|MobileNetV1|ResNet50_vd<sup>[1](#trans1)</sup> distill|72.77%/90.68% (+1.78%/+1.00%)| 17 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/MobileNetV1_distilled.tar) |
| MobileNetV2 | student | 72.15%/90.65% | 15 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/MobileNetV2_pretrained.tar) |
| MobileNetV2 | ResNet50_vd distill | 74.28%/91.53% (+2.13%/+0.88%) | 15 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/MobileNetV2_distilled.tar) |
| ResNet50 | student | 76.50%/93.00% | 99 | [下载链接](http://paddle-imagenet-models-name.bj.bcebos.com/ResNet50_pretrained.tar) |
|ResNet101|teacher|77.56%/93.64%| 173 | [下载链接](http://paddle-imagenet-models-name.bj.bcebos.com/ResNet101_pretrained.tar) |
| ResNet50 | ResNet101 distill | 77.29%/93.65% (+0.79%/+0.65%) | 99 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/ResNet50_distilled.tar) |
!!! note "Note"
<a name="trans1">[1]</a>：带_vd后缀代表该预训练模型使用了Mixup，Mixup相关介绍参考[mixup: Beyond Empirical Risk Minimization](https://arxiv.org/abs/1710.09412) Dataset：WIDER-FACE
## 2. 目标检测 | Model | Method | Image/GPU | Input Size | Easy/Medium/Hard | Model Size（MB） | Download |
| :------------: | :---------: | :-------: | :--------: | :-----------------------------: | :--------------: | :----------------------------------------------------------: |
| BlazeFace | - | 8 | 640 | 91.5/89.2/79.7 | 815 | [model](https://paddlemodels.bj.bcebos.com/object_detection/blazeface_original.tar) |
| BlazeFace | quant_post | 8 | 640 | 87.8/85.1/74.9 (-3.7/-4.1/-4.8) | 228 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/blazeface_origin_quant_post.tar) |
| BlazeFace | quant_aware | 8 | 640 | 90.5/87.9/77.6 (-1.0/-1.3/-2.1) | 228 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/blazeface_origin_quant_aware.tar) |
| BlazeFace-Lite | - | 8 | 640 | 90.9/88.5/78.1 | 711 | [model](https://paddlemodels.bj.bcebos.com/object_detection/blazeface_lite.tar) |
| BlazeFace-Lite | quant_post | 8 | 640 | 89.4/86.7/75.7 (-1.5/-1.8/-2.4) | 211 | [model]((https://paddlemodels.bj.bcebos.com/PaddleSlim/blazeface_lite_quant_post.tar)) |
| BlazeFace-Lite | quant_aware | 8 | 640 | 89.7/87.3/77.0 (-1.2/-1.2/-1.1) | 211 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/blazeface_lite_quant_aware.tar) |
| BlazeFace-NAS | - | 8 | 640 | 83.7/80.7/65.8 | 244 | [model](https://paddlemodels.bj.bcebos.com/object_detection/blazeface_nas.tar) |
| BlazeFace-NAS | quant_post | 8 | 640 | 81.6/78.3/63.6 (-2.1/-2.4/-2.2) | 71 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/blazeface_nas_quant_post.tar) |
| BlazeFace-NAS | quant_aware | 8 | 640 | 83.1/79.7/64.2 (-0.6/-1.0/-1.6) | 71 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/blazeface_nas_quant_aware.tar) |
### 2.1 量化 ### 2.2 Pruning
数据集： COCO 2017 Dataset：Pasacl VOC & COCO 2017
| 模型 | 压缩方法 | 数据集 | Image/GPU | 输入608 Box AP | 输入416 Box AP | 输入320 Box AP | 模型体积（MB） | 下载 | | Model | Method | Dataset | Image/GPU | Input 608 Box AP | Input 416 Box AP | Input 320 Box AP | Model Size(MB) | GFLOPs (608*608) | Download |
| :----------------------------: | :---------: | :----: | :-------: | :------------: | :------------: | :------------: | :------------: | :----------: | | :----------------------------: | :---------------: | :--------: | :-------: | :--------------: | :--------------: | :--------------: | :------------: | :--------------: | :----------------------------------------------------------: |
| MobileNet-V1-YOLOv3 | - | COCO | 8 | 29.3 | 29.3 | 27.1 | xx | [下载链接]() | | MobileNet-V1-YOLOv3 | Baseline | Pascal VOC | 8 | 76.2 | 76.7 | 75.3 | 94 | 40.49 | [model](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1_voc.tar) |
| MobileNet-V1-YOLOv3 | quant_post | COCO | 8 | xx | xx | xx | xx | [下载链接]() | | MobileNet-V1-YOLOv3 | sensitive -52.88% | Pascal VOC | 8 | 77.6 (+1.4) | 77.7 (1.0) | 75.5 (+0.2) | 31 | 19.08 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenet_v1_voc_prune.tar) |
| MobileNet-V1-YOLOv3 | quant_aware | COCO | 8 | xx | xx | xx | xx | [下载链接]() | | MobileNet-V1-YOLOv3 | - | COCO | 8 | 29.3 | 29.3 | 27.0 | 95 | 41.35 | [model](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1.tar) |
| R50-dcn-YOLOv3 obj365_pretrain | - | COCO | 8 | 41.4 | xx | xx | xx | [下载链接]() | | MobileNet-V1-YOLOv3 | sensitive -51.77% | COCO | 8 | 26.0 (-3.3) | 25.1 (-4.2) | 22.6 (-4.4) | 32 | 19.94 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenet_v1_prune.tar) |
| R50-dcn-YOLOv3 obj365_pretrain | quant_post | COCO | 8 | xx | xx | xx | xx | [下载链接]() | | R50-dcn-YOLOv3 | - | COCO | 8 | 39.1 | - | - | 177 | 89.60 | [model](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r50vd_dcn.tar) |
| R50-dcn-YOLOv3 obj365_pretrain | quant_aware | COCO | 8 | xx | xx | xx | xx | [下载链接]() | | R50-dcn-YOLOv3 | sensitive -9.37% | COCO | 8 | 39.3 (+0.2) | - | - | 150 | 81.20 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r50vd_dcn_prune.tar) |
| R50-dcn-YOLOv3 | sensitive -24.68% | COCO | 8 | 37.3 (-1.8) | - | - | 113 | 67.48 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r50vd_dcn_prune578.tar) |
| R50-dcn-YOLOv3 obj365_pretrain | - | COCO | 8 | 41.4 | - | - | 177 | 89.60 | [model](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r50vd_dcn_obj365_pretrained_coco.tar) |
| R50-dcn-YOLOv3 obj365_pretrain | sensitive -9.37% | COCO | 8 | 40.5 (-0.9) | - | - | 150 | 81.20 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r50vd_dcn_obj365_pretrained_coco_prune.tar) |
| R50-dcn-YOLOv3 obj365_pretrain | sensitive -24.68% | COCO | 8 | 37.8 (-3.3) | - | - | 113 | 67.48 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r50vd_dcn_obj365_pretrained_coco_prune578.tar) |
### 2.3 Distillation
Dataset：Pasacl VOC & COCO 2017
数据集：WIDER-FACE
| Model | Method | Dataset | Image/GPU | Input 608 Box AP | Input 416 Box AP | Input 320 Box AP | Model Size（MB） | Download |
| :-----------------: | :---------------------: | :--------: | :-------: | :--------------: | :--------------: | :--------------: | :--------------: | :----------------------------------------------------------: |
| MobileNet-V1-YOLOv3 | - | Pascal VOC | 8 | 76.2 | 76.7 | 75.3 | 94 | [model](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1_voc.tar) |
| ResNet34-YOLOv3 | - | Pascal VOC | 8 | 82.6 | 81.9 | 80.1 | 162 | [model](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r34_voc.tar) |
| MobileNet-V1-YOLOv3 | ResNet34-YOLOv3 distill | Pascal VOC | 8 | 79.0 (+2.8) | 78.2 (+1.5) | 75.5 (+0.2) | 94 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_voc_distilled.tar) |
| MobileNet-V1-YOLOv3 | - | COCO | 8 | 29.3 | 29.3 | 27.0 | 95 | [model](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1.tar) |
| ResNet34-YOLOv3 | - | COCO | 8 | 36.2 | 34.3 | 31.4 | 163 | [model](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r34.tar) |
| MobileNet-V1-YOLOv3 | ResNet34-YOLOv3 distill | COCO | 8 | 31.4 (+2.1) | 30.0 (+0.7) | 27.1 (+0.1) | 95 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_distilled.tar) |
| 模型 | 压缩方法 | Image/GPU | 输入尺寸 | Easy/Medium/Hard | 模型体积（MB） | 下载 | ### 2.4 NAS
| :------------: | :---------: | :-------: | :------: | :---------------: | :------------: | :----------: |
| BlazeFace | - | 8 | 640 | 0.915/0.892/0.797 | xx | [下载链接]() |
| BlazeFace | quant_post | 8 | 640 | xx/xx/xx | xx | [下载链接]() |
| BlazeFace | quant_aware | 8 | 640 | xx/xx/xx | xx | [下载链接]() |
| BlazeFace-Lite | - | 8 | 640 | 0.909/0.885/0.781 | xx | [下载链接]() |
| BlazeFace-Lite | quant_post | 8 | 640 | xx/xx/xx | xx | [下载链接]() |
| BlazeFace-Lite | quant_aware | 8 | 640 | xx/xx/xx | xx | [下载链接]() |
| BlazeFace-NAS | - | 8 | 640 | 0.837/0.807/0.658 | xx | [下载链接]() |
| BlazeFace-NAS | quant_post | 8 | 640 | xx/xx/xx | xx | [下载链接]() |
| BlazeFace-NAS | quant_aware | 8 | 640 | xx/xx/xx | xx | [下载链接]() |
### 2.2 剪裁 Dataset: WIDER-FACE
| Model | Method | Image/GPU | Input size | Easy/Medium/Hard | volume（KB） | latency（ms）| Download |
| :------------: | :---------: | :-------: | :------: | :-----------------------------: | :------------: | :------------: | :----------------------------------------------------------: |
| BlazeFace | - | 8 | 640 | 91.5/89.2/79.7 | 815 | 71.862 | [model](https://paddlemodels.bj.bcebos.com/object_detection/blazeface_original.tar) |
| BlazeFace-NAS | - | 8 | 640 | 83.7/80.7/65.8 | 244 | 21.117 |[model](https://paddlemodels.bj.bcebos.com/object_detection/blazeface_nas.tar) |
| BlazeFace-NAS1 | SANAS | 8 | 640 | 87.0/83.7/68.5 | 389 | 22.558 | [model](https://paddlemodels.bj.bcebos.com/object_detection/blazeface_nas2.tar) |
!!! note "Note"
数据集：Pasacl VOC & COCO 2017 <a name="trans1">[1]</a>: latency is based on latency_855.txt, the file is test on 855 by PaddleLite。
| 模型 | 压缩方法 | 数据集 | Image/GPU | 输入608 Box AP | 输入416 Box AP | 输入320 Box AP | 模型体积(MB) | GFLOPs (608*608) | 下载 |
| :----------------------------: | :---------------: | :--------: | :-------: | :------------: | :------------: | :------------: | :----------: | :--------------: | :----------------------------------------------------------: |
| MobileNet-V1-YOLOv3 | Baseline | Pascal VOC | 8 | 76.2 | 76.7 | 75.3 | 94 | 40.49 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1_voc.tar) |
| MobileNet-V1-YOLOv3 | sensitive -52.88% | Pascal VOC | 8 | 77.6 (+1.4) | 77.7 (1.0) | 75.5 (+0.2) | 31 | 19.08 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenet_v1_voc_prune.tar) |
| MobileNet-V1-YOLOv3 | - | COCO | 8 | 29.3 | 29.3 | 27.0 | 95 | 41.35 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1.tar) |
| MobileNet-V1-YOLOv3 | sensitive -51.77% | COCO | 8 | 26.0 (-3.3) | 25.1 (-4.2) | 22.6 (-4.4) | 32 | 19.94 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenet_v1_prune.tar) |
| R50-dcn-YOLOv3 | - | COCO | 8 | 39.1 | - | - | 177 | 89.60 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r50vd_dcn.tar) |
| R50-dcn-YOLOv3 | sensitive -9.37% | COCO | 8 | 39.3 (+0.2) | - | - | 150 | 81.20 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r50vd_dcn_prune.tar) |
| R50-dcn-YOLOv3 | sensitive -24.68% | COCO | 8 | 37.3 (-1.8) | - | - | 113 | 67.48 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r50vd_dcn_prune578.tar) |
| R50-dcn-YOLOv3 obj365_pretrain | - | COCO | 8 | 41.4 | - | - | 177 | 89.60 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r50vd_dcn_obj365_pretrained_coco.tar) |
| R50-dcn-YOLOv3 obj365_pretrain | sensitive -9.37% | COCO | 8 | 40.5 (-0.9) | - | - | 150 | 81.20 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r50vd_dcn_obj365_pretrained_coco_prune.tar) |
| R50-dcn-YOLOv3 obj365_pretrain | sensitive -24.68% | COCO | 8 | 37.8 (-3.3) | - | - | 113 | 67.48 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r50vd_dcn_obj365_pretrained_coco_prune578.tar) |
### 2.3 蒸馏 ## 3. Image Segmentation
Dataset：Cityscapes
数据集：Pasacl VOC & COCO 2017 ### 3.1 Quantization
| Model | Method | mIoU | Model Size（MB） | Download |
| :--------------------: | :---------: | :-----------: | :--------------: | :----------------------------------------------------------: |
| DeepLabv3+/MobileNetv1 | - | 63.26 | 6.6 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/deeplabv3_mobilenetv1.tar ) |
| DeepLabv3+/MobileNetv1 | quant_post | 58.63 (-4.63) | 1.8 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/deeplabv3_mobilenetv1_2049x1025_quant_post.tar) |
| DeepLabv3+/MobileNetv1 | quant_aware | 62.03 (-1.23) | 1.8 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/deeplabv3_mobilenetv1_2049x1025_quant_aware.tar) |
| DeepLabv3+/MobileNetv2 | - | 69.81 | 7.4 | [model](https://paddleseg.bj.bcebos.com/models/mobilenet_cityscapes.tgz) |
| DeepLabv3+/MobileNetv2 | quant_post | 67.59 (-2.22) | 2.1 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/deeplabv3_mobilenetv2_2049x1025_quant_post.tar) |
| DeepLabv3+/MobileNetv2 | quant_aware | 68.33 (-1.48) | 2.1 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/deeplabv3_mobilenetv2_2049x1025_quant_aware.tar) |
| 模型 | 压缩方法 | 数据集 | Image/GPU | 输入608 Box AP | 输入416 Box AP | 输入320 Box AP | 模型体积（MB） | 下载 | Image segmentation model PaddleLite latency (ms), input size 769x769
| :-----------------: | :---------------------: | :--------: | :-------: | :------------: | :------------: | :------------: | :------------: | :----------------------------------------------------------: |
| MobileNet-V1-YOLOv3 | - | Pascal VOC | 8 | 76.2 | 76.7 | 75.3 | 94 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1_voc.tar) |
| ResNet34-YOLOv3 | - | Pascal VOC | 8 | 82.6 | 81.9 | 80.1 | 162 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r34_voc.tar) |
| MobileNet-V1-YOLOv3 | ResNet34-YOLOv3 distill | Pascal VOC | 8 | 79.0 (+2.8) | 78.2 (+1.5) | 75.5 (+0.2) | 94 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_voc_distilled.tar) |
| MobileNet-V1-YOLOv3 | - | COCO | 8 | 29.3 | 29.3 | 27.0 | 95 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1.tar) |
| ResNet34-YOLOv3 | - | COCO | 8 | 36.2 | 34.3 | 31.4 | 163 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r34.tar) |
| MobileNet-V1-YOLOv3 | ResNet34-YOLOv3 distill | COCO | 8 | 31.4 (+2.1) | 30.0 (+0.7) | 27.1 (+0.1) | 95 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_distilled.tar) |
| Device | Model | Method | armv7 Thread 1 | armv7 Thread 2 | armv7 Thread 4 | armv8 Thread 1 | armv8 Thread 2 | armv8 Thread 4 |
| ------------ | ---------------------- | ------------- | -------------- | -------------- | -------------- | -------------- | -------------- | -------------- |
| Qualcomm 835 | Deeplabv3- MobileNetV1 | FP32 baseline | 1227.9894 | 734.1922 | 527.9592 | 1109.96 | 699.3818 | 479.0818 |
| Qualcomm 835 | Deeplabv3- MobileNetV1 | quant_aware | 848.6544 | 512.785 | 382.9915 | 752.3573 | 455.0901 | 307.8808 |
| Qualcomm 835 | Deeplabv3- MobileNetV1 | quant_post | 840.2323 | 510.103 | 371.9315 | 748.9401 | 452.1745 | 309.2084 |
| Qualcomm 835 | Deeplabv3-MobileNetV2 | FP32 baseline | 1282.8126 | 793.2064 | 653.6538 | 1193.9908 | 737.1827 | 593.4522 |
| Qualcomm 835 | Deeplabv3-MobileNetV2 | quant_aware | 976.0495 | 659.0541 | 513.4279 | 892.1468 | 582.9847 | 484.7512 |
| Qualcomm 835 | Deeplabv3-MobileNetV2 | quant_post | 981.44 | 658.4969 | 538.6166 | 885.3273 | 586.1284 | 484.0018 |
| Qualcomm 855 | Deeplabv3- MobileNetV1 | FP32 baseline | 568.8748 | 339.8578 | 278.6316 | 420.6031 | 281.3197 | 217.5222 |
| Qualcomm 855 | Deeplabv3- MobileNetV1 | quant_aware | 608.7578 | 347.2087 | 260.653 | 241.2394 | 177.3456 | 143.9178 |
| Qualcomm 855 | Deeplabv3- MobileNetV1 | quant_post | 609.0142 | 347.3784 | 259.9825 | 239.4103 | 180.1894 | 139.9178 |
| Qualcomm 855 | Deeplabv3-MobileNetV2 | FP32 baseline | 639.4425 | 390.1851 | 322.7014 | 477.7667 | 339.7411 | 262.2847 |
| Qualcomm 855 | Deeplabv3-MobileNetV2 | quant_aware | 703.7275 | 497.689 | 417.1296 | 394.3586 | 300.2503 | 239.9204 |
| Qualcomm 855 | Deeplabv3-MobileNetV2 | quant_post | 705.7589 | 474.4076 | 427.2951 | 394.8352 | 297.4035 | 264.6724 |
| Kirin 970 | Deeplabv3- MobileNetV1 | FP32 baseline | 1682.1792 | 1437.9774 | 1181.0246 | 1261.6739 | 1068.6537 | 690.8225 |
| Kirin 970 | Deeplabv3- MobileNetV1 | quant_aware | 1062.3394 | 1248.1014 | 878.3157 | 774.6356 | 710.6277 | 528.5376 |
| Kirin 970 | Deeplabv3- MobileNetV1 | quant_post | 1109.1917 | 1339.6218 | 866.3587 | 771.5164 | 716.5255 | 500.6497 |
| Kirin 970 | Deeplabv3-MobileNetV2 | FP32 baseline | 1771.1301 | 1746.0569 | 1222.4805 | 1448.9739 | 1192.4491 | 760.606 |
| Kirin 970 | Deeplabv3-MobileNetV2 | quant_aware | 1320.2905 | 921.4522 | 676.0732 | 1145.8801 | 821.5685 | 590.1713 |
| Kirin 970 | Deeplabv3-MobileNetV2 | quant_post | 1320.386 | 918.5328 | 672.2481 | 1020.753 | 820.094 | 591.4114 |
## 3. 图像分割
数据集：Cityscapes
### 3.1 量化
| 模型 | 压缩方法 | mIoU | 模型体积（MB） | 下载 |
| :--------------------: | :---------: | :---: | :------------: | :----------: |
| DeepLabv3+/MobileNetv1 | - | 63.26 | xx | [下载链接]() |
| DeepLabv3+/MobileNetv1 | quant_post | xx | xx | [下载链接]() |
| DeepLabv3+/MobileNetv1 | quant_aware | xx | xx | [下载链接]() |
| DeepLabv3+/MobileNetv2 | - | 69.81 | xx | [下载链接]() |
| DeepLabv3+/MobileNetv2 | quant_post | xx | xx | [下载链接]() |
| DeepLabv3+/MobileNetv2 | quant_aware | xx | xx | [下载链接]() |
### 3.2 剪裁 ### 3.2 Pruning
| 模型 | 压缩方法 | mIoU | 模型体积（MB） | GFLOPs | 下载 | | Model | Method | mIoU | Model Size（MB） | GFLOPs | Download |
| :-------: | :---------------: | :-----------: | :------------: | :----: | :----------------------------------------------------------: | | :-------: | :---------------: | :-----------: | :--------------: | :----: | :----------------------------------------------------------: |
| fast-scnn | baseline | 69.64 | 11 | 14.41 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/fast_scnn_cityscape.tar) | | fast-scnn | baseline | 69.64 | 11 | 14.41 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/fast_scnn_cityscape.tar) |
| fast-scnn | uniform -17.07% | 69.58 (-0.06) | 8.5 | 11.95 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/fast_scnn_cityscape_uniform-17.tar) | | fast-scnn | uniform -17.07% | 69.58 (-0.06) | 8.5 | 11.95 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/fast_scnn_cityscape_uniform-17.tar) |
| fast-scnn | sensitive -47.60% | 66.68 (-2.96) | 5.7 | 7.55 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/fast_scnn_cityscape_sensitive-47.tar) | | fast-scnn | sensitive -47.60% | 66.68 (-2.96) | 5.7 | 7.55 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/fast_scnn_cityscape_sensitive-47.tar) |
docs/zh_cn/api_cn/nas_api.rst
浏览文件 @ bb0f8fbb
...@@ -128,7 +128,7 @@ SANAS（Simulated Annealing Neural Architecture Search）是基于模拟退火 ...@@ -128,7 +128,7 @@ SANAS（Simulated Annealing Neural Architecture Search）是基于模拟退火
- **tokens(list):** - 一组tokens。tokens的长度和范围取决于搜索空间。 - **tokens(list):** - 一组tokens。tokens的长度和范围取决于搜索空间。
**返回：** **返回：**
根据传入的token得到一个模型结构实例。 根据传入的token得到一个模型结构实例列表。
**示例代码：** **示例代码：**
...@@ -153,8 +153,10 @@ SANAS（Simulated Annealing Neural Architecture Search）是基于模拟退火 ...@@ -153,8 +153,10 @@ SANAS（Simulated Annealing Neural Architecture Search）是基于模拟退火
**示例代码：** **示例代码：**
.. code-block:: python .. code-block:: python
import paddle.fluid as fluid import paddle.fluid as fluid
from paddleslim.nas import SANAS from paddleslim.nas import SANAS
config = [('MobileNetV2Space')] config = [('MobileNetV2Space')]
sanas = SANAS(configs=config) sanas = SANAS(configs=config)
print(sanas.current_info()) print(sanas.current_info())
docs/zh_cn/api_cn/pantheon_api.md
浏览文件 @ bb0f8fbb
# 多进程蒸馏 # 大规模可扩展知识蒸馏框架 Pantheon
## Teacher ## Teacher
...@@ -100,7 +100,8 @@ pantheon.Teacher.start\_knowledge\_service(feed\_list, schema, program, reader\_ ...@@ -100,7 +100,8 @@ pantheon.Teacher.start\_knowledge\_service(feed\_list, schema, program, reader\_
- **times (int):** The maximum repeated serving times, default 1. Whenever - **times (int):** The maximum repeated serving times, default 1. Whenever
the public method **get\_knowledge\_generator()** in **Student** the public method **get\_knowledge\_generator()** in **Student**
object called once, the serving times will be added one, object called once, the serving times will be added one,
until reaching the maximum and ending the service. until reaching the maximum and ending the service. Only
valid in online mode, and will be ignored in offline mode.
**Return:** None **Return:** None
......
docs/zh_cn/api_cn/prune_api.rst
浏览文件 @ bb0f8fbb
...@@ -378,7 +378,7 @@ load_sensitivities ...@@ -378,7 +378,7 @@ load_sensitivities
} }
} }
sensitivities_file = "sensitive_api_demo.data" sensitivities_file = "sensitive_api_demo.data"
with open(sensitivities_file, 'w') as f: with open(sensitivities_file, 'wb') as f:
pickle.dump(sen, f) pickle.dump(sen, f)
sensitivities = load_sensitivities(sensitivities_file) sensitivities = load_sensitivities(sensitivities_file)
print(sensitivities) print(sensitivities)
......
docs/zh_cn/model_zoo.md
浏览文件 @ bb0f8fbb
# 模型库 # 模型库
## 1. 图象分类 ## 1. 图像分类
数据集：ImageNet1000类 数据集：ImageNet1000类
...@@ -16,7 +16,7 @@ ...@@ -16,7 +16,7 @@
| MobileNetV2 | quant_aware |72.05%/90.63% (-0.1%/-0.02%)| 4.0 | - | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/MobileNetV2_quant_aware.tar) | | MobileNetV2 | quant_aware |72.05%/90.63% (-0.1%/-0.02%)| 4.0 | - | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/MobileNetV2_quant_aware.tar) |
|ResNet50|-|76.50%/93.00%| 99 | 2.71 | [下载链接](http://paddle-imagenet-models-name.bj.bcebos.com/ResNet50_pretrained.tar) | |ResNet50|-|76.50%/93.00%| 99 | 2.71 | [下载链接](http://paddle-imagenet-models-name.bj.bcebos.com/ResNet50_pretrained.tar) |
|ResNet50|quant_post|76.33%/93.02% (-0.17%/+0.02%)| 25.1| 1.19 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/ResNet50_quant_post.tar) | |ResNet50|quant_post|76.33%/93.02% (-0.17%/+0.02%)| 25.1| 1.19 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/ResNet50_quant_post.tar) |
|ResNet50|quant_aware| 76.48%/93.11% (-0.02%/+0.11%)| 25.1 | 1.17 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/ResNet50_quant_awre.tar) | |ResNet50|quant_aware| 76.48%/93.11% (-0.02%/+0.11%)| 25.1 | 1.17 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/ResNet50_quant_awre.tar) |
分类模型Lite时延(ms) 分类模型Lite时延(ms)
...@@ -89,6 +89,12 @@ ...@@ -89,6 +89,12 @@
<a name="trans1">[1]</a>：带_vd后缀代表该预训练模型使用了Mixup，Mixup相关介绍参考[mixup: Beyond Empirical Risk Minimization](https://arxiv.org/abs/1710.09412) <a name="trans1">[1]</a>：带_vd后缀代表该预训练模型使用了Mixup，Mixup相关介绍参考[mixup: Beyond Empirical Risk Minimization](https://arxiv.org/abs/1710.09412)
### 1.4 搜索
| 模型 | 压缩方法 | Top-1/Top-5 Acc | 模型体积（MB） | GFLOPs | 下载 |
|:--:|:---:|:--:|:--:|:--:|:--:|
| MobileNetV2 | - | 72.15%/90.65% | 15 | 0.59 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/MobileNetV2_pretrained.tar) |
| MobileNetV2 | SANAS | 71.518%/90.208% (-0.632%/-0.442%) | 14 | 0.295 | [下载链接](https://paddlemodels.cdn.bcebos.com/PaddleSlim/MobileNetV2_sanas.tar) |
## 2. 目标检测 ## 2. 目标检测
...@@ -99,8 +105,8 @@ ...@@ -99,8 +105,8 @@
| 模型 | 压缩方法 | 数据集 | Image/GPU | 输入608 Box AP | 输入416 Box AP | 输入320 Box AP | 模型体积（MB） | TensorRT时延(V100, ms) | 下载 | | 模型 | 压缩方法 | 数据集 | Image/GPU | 输入608 Box AP | 输入416 Box AP | 输入320 Box AP | 模型体积（MB） | TensorRT时延(V100, ms) | 下载 |
| :----------------------------: | :---------: | :----: | :-------: | :------------: | :------------: | :------------: | :------------: | :----------: |:----------: | | :----------------------------: | :---------: | :----: | :-------: | :------------: | :------------: | :------------: | :------------: | :----------: |:----------: |
| MobileNet-V1-YOLOv3 | - | COCO | 8 | 29.3 | 29.3 | 27.1 | 95 | - | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1.tar) | | MobileNet-V1-YOLOv3 | - | COCO | 8 | 29.3 | 29.3 | 27.1 | 95 | - | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1.tar) |
| MobileNet-V1-YOLOv3 | quant_post | COCO | 8 | 27.9 (-1.4)| 28.0 (-1.3) | 26.0 (-1.0) | 25 | - | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_quant_post.tar) | | MobileNet-V1-YOLOv3 | quant_post | COCO | 8 | 27.9 (-1.4)| 28.0 (-1.3) | 26.0 (-1.0) | 25 | - | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_quant_post.tar) |
| MobileNet-V1-YOLOv3 | quant_aware | COCO | 8 | 28.1 (-1.2)| 28.2 (-1.1) | 25.8 (-1.2) | 26.3 | - | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenet_coco_quant_aware.tar) | | MobileNet-V1-YOLOv3 | quant_aware | COCO | 8 | 28.1 (-1.2)| 28.2 (-1.1) | 25.8 (-1.2) | 26.3 | - | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenet_coco_quant_aware.tar) |
| R34-YOLOv3 | - | COCO | 8 | 36.2 | 34.3 | 31.4 | 162 | - | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r34.tar) | | R34-YOLOv3 | - | COCO | 8 | 36.2 | 34.3 | 31.4 | 162 | - | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r34.tar) |
| R34-YOLOv3 | quant_post | COCO | 8 | 35.7 (-0.5) | - | - | 42.7 | - | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r34_coco_quant_post.tar) | | R34-YOLOv3 | quant_post | COCO | 8 | 35.7 (-0.5) | - | - | 42.7 | - | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r34_coco_quant_post.tar) |
| R34-YOLOv3 | quant_aware | COCO | 8 | 35.2 (-1.0) | 33.3 (-1.0) | 30.3 (-1.1)| 44 | - | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r34_coco_quant_aware.tar) | | R34-YOLOv3 | quant_aware | COCO | 8 | 35.2 (-1.0) | 33.3 (-1.0) | 30.3 (-1.1)| 44 | - | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r34_coco_quant_aware.tar) |
...@@ -157,6 +163,20 @@ ...@@ -157,6 +163,20 @@
| MobileNet-V1-YOLOv3 | ResNet34-YOLOv3 distill | COCO | 8 | 31.4 (+2.1) | 30.0 (+0.7) | 27.1 (+0.1) | 95 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_distilled.tar) | | MobileNet-V1-YOLOv3 | ResNet34-YOLOv3 distill | COCO | 8 | 31.4 (+2.1) | 30.0 (+0.7) | 27.1 (+0.1) | 95 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_distilled.tar) |
### 2.4 搜索
数据集：WIDER-FACE
| 模型 | 压缩方法 | Image/GPU | 输入尺寸 | Easy/Medium/Hard | 模型体积（KB） | 硬件延时（ms）| 下载 |
| :------------: | :---------: | :-------: | :------: | :-----------------------------: | :------------: | :------------: | :----------------------------------------------------------: |
| BlazeFace | - | 8 | 640 | 91.5/89.2/79.7 | 815 | 71.862 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/blazeface_original.tar) |
| BlazeFace-NAS | - | 8 | 640 | 83.7/80.7/65.8 | 244 | 21.117 |[下载链接](https://paddlemodels.bj.bcebos.com/object_detection/blazeface_nas.tar) |
| BlazeFace-NAS1 | SANAS | 8 | 640 | 87.0/83.7/68.5 | 389 | 22.558 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/blazeface_nas2.tar) |
!!! note "Note"
<a name="trans1">[1]</a>: 硬件延时时间是利用提供的硬件延时表得到的，硬件延时表是在855芯片上基于PaddleLite测试的结果。
## 3. 图像分割 ## 3. 图像分割
数据集：Cityscapes 数据集：Cityscapes
......
paddleslim/__init__.py
浏览文件 @ bb0f8fbb
...@@ -19,4 +19,5 @@ from paddleslim import nas ...@@ -19,4 +19,5 @@ from paddleslim import nas
from paddleslim import analysis from paddleslim import analysis
from paddleslim import dist from paddleslim import dist
from paddleslim import quant from paddleslim import quant
__all__ = ['models', 'prune', 'nas', 'analysis', 'dist', 'quant'] from paddleslim import pantheon
__all__ = ['models', 'prune', 'nas', 'analysis', 'dist', 'quant', 'pantheon']
paddleslim/nas/sa_nas.py
浏览文件 @ bb0f8fbb
...@@ -190,7 +190,10 @@ class SANAS(object): ...@@ -190,7 +190,10 @@ class SANAS(object):
self._iter = 0 self._iter = 0
def _get_host_ip(self): def _get_host_ip(self):
return socket.gethostbyname(socket.gethostname()) if os.name == 'posix':
return socket.gethostbyname('localhost')
else:
return socket.gethostbyname(socket.gethostname())
def tokens2arch(self, tokens): def tokens2arch(self, tokens):
""" """
......
paddleslim/pantheon/README.md
浏览文件 @ bb0f8fbb
...@@ -13,7 +13,7 @@ The illustration below shows an application of Pantheon, where the sudent model ...@@ -13,7 +13,7 @@ The illustration below shows an application of Pantheon, where the sudent model
## Prerequisites ## Prerequisites
- Python 2.7.x or 3.x - Python 2.7.x or 3.x
- PaddlePaddle >= 1.6.0 - PaddlePaddle >= 1.7.0
## APIs ## APIs
......
paddleslim/pantheon/student.py
浏览文件 @ bb0f8fbb
...@@ -158,7 +158,7 @@ class Student(object): ...@@ -158,7 +158,7 @@ class Student(object):
if end_recved: if end_recved:
break break
with open(in_path, 'r') as fin: with open(in_path, 'rb') as fin:
# get knowledge desc # get knowledge desc
desc = pickle.load(fin) desc = pickle.load(fin)
out_queue.put(desc) out_queue.put(desc)
...@@ -222,7 +222,7 @@ class Student(object): ...@@ -222,7 +222,7 @@ class Student(object):
self._started = True self._started = True
def _merge_knowledge(self, knowledge): def _merge_knowledge(self, knowledge):
for k, tensors in knowledge.items(): for k, tensors in list(knowledge.items()):
if len(tensors) == 0: if len(tensors) == 0:
del knowledge[k] del knowledge[k]
elif len(tensors) == 1: elif len(tensors) == 1:
...@@ -308,7 +308,7 @@ class Student(object): ...@@ -308,7 +308,7 @@ class Student(object):
print("Knowledge merging strategy: {}".format( print("Knowledge merging strategy: {}".format(
self._merge_strategy)) self._merge_strategy))
print("Knowledge description after merging:") print("Knowledge description after merging:")
for schema, desc in knowledge_desc.items(): for schema, desc in list(knowledge_desc.items()):
print("{}: {}".format(schema, desc)) print("{}: {}".format(schema, desc))
self._knowledge_desc = knowledge_desc self._knowledge_desc = knowledge_desc
...@@ -426,13 +426,13 @@ class Student(object): ...@@ -426,13 +426,13 @@ class Student(object):
end_received = [0] * len(queues) end_received = [0] * len(queues)
while True: while True:
knowledge = OrderedDict( knowledge = OrderedDict(
[(k, []) for k, v in self._knowledge_desc.items()]) [(k, []) for k, v in list(self._knowledge_desc.items())])
for idx, receiver in enumerate(data_receivers): for idx, receiver in enumerate(data_receivers):
if not end_received[idx]: if not end_received[idx]:
batch_samples = receiver.next( batch_samples = receiver.next(
) if six.PY2 else receiver.__next__() ) if six.PY2 else receiver.__next__()
if not isinstance(batch_samples, EndSignal): if not isinstance(batch_samples, EndSignal):
for k, v in batch_samples.items(): for k, v in list(batch_samples.items()):
knowledge[k].append(v) knowledge[k].append(v)
else: else:
end_received[idx] = 1 end_received[idx] = 1
......
paddleslim/pantheon/teacher.py
浏览文件 @ bb0f8fbb
...@@ -151,7 +151,7 @@ class Teacher(object): ...@@ -151,7 +151,7 @@ class Teacher(object):
self._t2s_queue = None self._t2s_queue = None
self._cmd_queue = None self._cmd_queue = None
self._out_file = open(self._out_path, "w") if self._out_path else None self._out_file = open(self._out_path, "wb") if self._out_path else None
if self._out_file: if self._out_file:
return return
...@@ -231,7 +231,7 @@ class Teacher(object): ...@@ -231,7 +231,7 @@ class Teacher(object):
"The knowledge data should be a dict or OrderedDict!") "The knowledge data should be a dict or OrderedDict!")
knowledge_desc = {} knowledge_desc = {}
for name, value in knowledge.items(): for name, value in list(knowledge.items()):
knowledge_desc[name] = { knowledge_desc[name] = {
"shape": [-1] + list(value.shape[1:]), "shape": [-1] + list(value.shape[1:]),
"dtype": str(value.dtype), "dtype": str(value.dtype),
...@@ -294,7 +294,8 @@ class Teacher(object): ...@@ -294,7 +294,8 @@ class Teacher(object):
times (int): The maximum repeated serving times. Default 1. Whenever times (int): The maximum repeated serving times. Default 1. Whenever
the public method 'get_knowledge_generator()' in Student the public method 'get_knowledge_generator()' in Student
object called once, the serving times will be added one, object called once, the serving times will be added one,
until reaching the maximum and ending the service. until reaching the maximum and ending the service. Only
valid in online mode, and will be ignored in offline mode.
""" """
if not self._started: if not self._started:
raise ValueError("The method start() should be called first!") raise ValueError("The method start() should be called first!")
...@@ -339,9 +340,12 @@ class Teacher(object): ...@@ -339,9 +340,12 @@ class Teacher(object):
if not times > 0: if not times > 0:
raise ValueError("Repeated serving times should be positive!") raise ValueError("Repeated serving times should be positive!")
self._times = times self._times = times
if self._times > 1 and self._out_file:
self._times = 1
print("WARNING: args 'times' will be ignored in offline mode")
desc = {} desc = {}
for name, var in schema.items(): for name, var in list(schema.items()):
if not isinstance(var, fluid.framework.Variable): if not isinstance(var, fluid.framework.Variable):
raise ValueError( raise ValueError(
"The member of schema must be fluid Variable.") "The member of schema must be fluid Variable.")
...@@ -412,10 +416,14 @@ class Teacher(object): ...@@ -412,10 +416,14 @@ class Teacher(object):
else: else:
if self._knowledge_queue: if self._knowledge_queue:
self._knowledge_queue.put(EndSignal()) self._knowledge_queue.put(EndSignal())
# should close file in child thread to wait for all
# writing finished
if self._out_file:
self._out_file.close()
# Asynchronous output # Asynchronous output
out_buf_queue = Queue.Queue(self._buf_size) out_buf_queue = Queue.Queue(self._buf_size)
schema_keys, schema_vars = zip(*self._schema.items()) schema_keys, schema_vars = zip(*list(self._schema.items()))
out_thread = Thread(target=writer, args=(out_buf_queue, schema_keys)) out_thread = Thread(target=writer, args=(out_buf_queue, schema_keys))
out_thread.daemon = True out_thread.daemon = True
out_thread.start() out_thread.start()
...@@ -424,8 +432,9 @@ class Teacher(object): ...@@ -424,8 +432,9 @@ class Teacher(object):
self._program).with_data_parallel() self._program).with_data_parallel()
print("Knowledge description {}".format(self._knowledge_desc)) print("Knowledge description {}".format(self._knowledge_desc))
print(time.strftime('%Y-%m-%d %H:%M:%S', time.localtime(time.time())) + print(
" Teacher begins to serve ...") time.strftime('%Y-%m-%d %H:%M:%S', time.localtime(time.time())) +
" Teacher begins to serve ...")
# For offline dump, write the knowledge description to the head of file # For offline dump, write the knowledge description to the head of file
if self._out_file: if self._out_file:
self._out_file.write(pickle.dumps(self._knowledge_desc)) self._out_file.write(pickle.dumps(self._knowledge_desc))
...@@ -491,11 +500,10 @@ class Teacher(object): ...@@ -491,11 +500,10 @@ class Teacher(object):
if self._knowledge_queue: if self._knowledge_queue:
self._knowledge_queue.join() self._knowledge_queue.join()
print(time.strftime('%Y-%m-%d %H:%M:%S', time.localtime(time.time())) + print(
" Teacher ends serving.") time.strftime('%Y-%m-%d %H:%M:%S', time.localtime(time.time())) +
" Teacher ends serving.")
def __del__(self): def __del__(self):
if self._manager: if self._manager:
self._manager.shutdown() self._manager.shutdown()
if self._out_file:
self._out_file.close()
paddleslim/prune/pruner.py
浏览文件 @ bb0f8fbb
...@@ -13,6 +13,7 @@ ...@@ -13,6 +13,7 @@
# limitations under the License. # limitations under the License.
import logging import logging
import sys
import numpy as np import numpy as np
import paddle.fluid as fluid import paddle.fluid as fluid
import copy import copy
...@@ -79,8 +80,8 @@ class Pruner(): ...@@ -79,8 +80,8 @@ class Pruner():
pruned_num = int(round(param_v.shape()[0] * ratio)) pruned_num = int(round(param_v.shape()[0] * ratio))
pruned_idx = [0] * pruned_num pruned_idx = [0] * pruned_num
else: else:
param_t = np.array(scope.find_var(param).get_tensor()) pruned_idx = self._cal_pruned_idx(
pruned_idx = self._cal_pruned_idx(param_t, ratio, axis=0) graph, scope, param, ratio, axis=0)
param = graph.var(param) param = graph.var(param)
conv_op = param.outputs()[0] conv_op = param.outputs()[0]
walker = conv2d_walker( walker = conv2d_walker(
...@@ -130,7 +131,7 @@ class Pruner(): ...@@ -130,7 +131,7 @@ class Pruner():
graph.infer_shape() graph.infer_shape()
return graph.program, param_backup, param_shape_backup return graph.program, param_backup, param_shape_backup
def _cal_pruned_idx(self, param, ratio, axis): def _cal_pruned_idx(self, graph, scope, param, ratio, axis):
""" """
Calculate the index to be pruned on axis by given pruning ratio. Calculate the index to be pruned on axis by given pruning ratio.
...@@ -145,11 +146,26 @@ class Pruner(): ...@@ -145,11 +146,26 @@ class Pruner():
Returns: Returns:
list<int>: The indexes to be pruned on axis. list<int>: The indexes to be pruned on axis.
""" """
prune_num = int(round(param.shape[axis] * ratio))
reduce_dims = [i for i in range(len(param.shape)) if i != axis]
if self.criterion == 'l1_norm': if self.criterion == 'l1_norm':
criterions = np.sum(np.abs(param), axis=tuple(reduce_dims)) param_t = np.array(scope.find_var(param).get_tensor())
pruned_idx = criterions.argsort()[:prune_num] prune_num = int(round(param_t.shape[axis] * ratio))
reduce_dims = [i for i in range(len(param_t.shape)) if i != axis]
criterions = np.sum(np.abs(param_t), axis=tuple(reduce_dims))
pruned_idx = criterions.argsort()[:prune_num]
elif self.criterion == "batch_norm_scale":
param_var = graph.var(param)
conv_op = param_var.outputs()[0]
conv_output = conv_op.outputs("Output")[0]
bn_op = conv_output.outputs()[0]
if bn_op is not None:
bn_scale_param = bn_op.inputs("Scale")[0].name()
bn_scale_np = np.array(
scope.find_var(bn_scale_param).get_tensor())
prune_num = int(round(bn_scale_np.shape[axis] * ratio))
pruned_idx = np.abs(bn_scale_np).argsort()[:prune_num]
else:
raise SystemExit(
"Can't find BatchNorm op after Conv op in Network.")
return pruned_idx return pruned_idx
def _prune_tensor(self, tensor, pruned_idx, pruned_axis, lazy=False): def _prune_tensor(self, tensor, pruned_idx, pruned_axis, lazy=False):
......
tests/test_slim_prune.py 0 → 100644
浏览文件 @ bb0f8fbb
# 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 sys
sys.path.append("../")
import unittest
import paddle.fluid as fluid
from paddleslim.prune import Pruner
from layers import conv_bn_layer
class TestPrune(unittest.TestCase):
def test_prune(self):
main_program = fluid.Program()
startup_program = fluid.Program()
# X X O X O
# conv1-->conv2-->sum1-->conv3-->conv4-->sum2-->conv5-->conv6
# | ^ | ^
# |____________| |____________________|
#
# X: prune output channels
# O: prune input channels
with fluid.program_guard(main_program, startup_program):
input = fluid.data(name="image", shape=[None, 3, 16, 16])
conv1 = conv_bn_layer(input, 8, 3, "conv1")
conv2 = conv_bn_layer(conv1, 8, 3, "conv2")
sum1 = conv1 + conv2
conv3 = conv_bn_layer(sum1, 8, 3, "conv3")
conv4 = conv_bn_layer(conv3, 8, 3, "conv4")
sum2 = conv4 + sum1
conv5 = conv_bn_layer(sum2, 8, 3, "conv5")
conv6 = conv_bn_layer(conv5, 8, 3, "conv6")
shapes = {}
for param in main_program.global_block().all_parameters():
shapes[param.name] = param.shape
place = fluid.CPUPlace()
exe = fluid.Executor(place)
scope = fluid.Scope()
exe.run(startup_program, scope=scope)
criterion = 'batch_norm_scale'
pruner = Pruner(criterion)
main_program, _, _ = pruner.prune(
main_program,
scope,
params=["conv4_weights"],
ratios=[0.5],
place=place,
lazy=False,
only_graph=False,
param_backup=None,
param_shape_backup=None)
shapes = {
"conv1_weights": (4L, 3L, 3L, 3L),
"conv2_weights": (4L, 4L, 3L, 3L),
"conv3_weights": (8L, 4L, 3L, 3L),
"conv4_weights": (4L, 8L, 3L, 3L),
"conv5_weights": (8L, 4L, 3L, 3L),
"conv6_weights": (8L, 8L, 3L, 3L)
}
for param in main_program.global_block().all_parameters():
if "weights" in param.name:
print("param: {}; param shape: {}".format(param.name,
param.shape))
self.assertTrue(param.shape == shapes[param.name])
if __name__ == '__main__':
unittest.main()
Markdown is supported
0% 或 .
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 或