mindspore-Ecosystem
/
mindspore
21
0
Fork 27
Code Issues Pull Requests 6 Projects Releases Wiki Activity

!19257 support FaceDetection for GPU 

Merge pull request !19257 from 周莉莉/master
This commit is contained in:
i-robot 2021-07-03 06:41:46 +00:00 committed by Gitee
parent 11b6b6b5a5 4ba02f5dff
commit 89d7803bb3
5 changed files with 158 additions and 32 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

94
model_zoo/research/cv/FaceDetection/README.md
View File

@ -69,8 +69,8 @@ We use about 13K images as training dataset and 3K as evaluating dataset in this
# [Environment Requirements](#contents)
- Hardware（Ascend, CPU）
- Prepare hardware environment with Ascend or CPU processor.
- Hardware（Ascend, CPU, GPU）
- Prepare hardware environment with Ascend, CPU or GPU processor.
- Framework
- [MindSpore](https://www.mindspore.cn/install/en)
- For more information, please check the resources below：
@ -96,7 +96,9 @@ The entire code structure is as following:
└─ moxing_adapter.py # Moxing adapter for ModelArts
├─ scripts
├─ run_standalone_train.sh # launch standalone training(1p) in ascend
├─ run_standalone_train_gpu.sh # launch standalone training(1p) in GPU
├─ run_distribute_train.sh # launch distributed training(8p) in ascend
├─ run_distribute_train_gpu.sh # launch distributed training(8p) in GPU
├─ run_eval.sh # launch evaluating in ascend
├─ run_infer_310.sh # launch inference on Ascend310
└─ run_export.sh # launch exporting air model
@ -129,6 +131,8 @@ The entire code structure is as following:
- Stand alone mode
# Training on Ascend
```bash
cd ./scripts
bash run_standalone_train.sh [PLATFORM] [MINDRECORD_FILE] [USE_DEVICE_ID]
@ -148,8 +152,21 @@ The entire code structure is as following:
bash run_standalone_train.sh CPU /home/train.mindrecord 0 /home/a.ckpt
```
# Training on GPU
```python
python ./train.py --config_path=./default_config.yaml
```
```bash
cd ./scripts
bash run_standalone_train_gpu.sh [CONFIG_PATH]
```
- Distribute mode (recommended)
# Distribute training on Ascend
```bash
cd ./scripts
bash run_distribute_train.sh [MINDRECORD_FILE] [RANK_TABLE]
@ -169,6 +186,19 @@ The entire code structure is as following:
bash run_distribute_train.sh /home/train.mindrecord ./rank_table_8p.json /home/a.ckpt
```
# Distribute training on GPU
```python
mpirun --allow-run-as-root -n $RANK_SIZE --output-filename log_output --merge-stderr-to-stdout \
python ./train.py \
--config_path=$CONFIG_PATH > train.log 2>&1 &
```
```bash
cd ./scripts
bash run_distribute_train_gpu.sh [CONFIG_PATH]
```
*Distribute mode doesn't support running on CPU*. You will get the loss value of each step as following in "./scripts/device0/output/[TIME]/[TIME].log" or "./scripts/device0/train.log":
```python
@ -278,6 +308,11 @@ cd ./scripts
bash run_eval.sh [PLATFORM] [MINDRECORD_FILE] [USE_DEVICE_ID] [PRETRAINED_BACKBONE]
```
```python
GPU platform
python eval.py [CONFIG_PATH]
```
for example:
```bash
@ -285,6 +320,11 @@ cd ./scripts
bash run_eval.sh Ascend /home/eval.mindrecord 0 /home/a.ckpt
```
```python
GPU platform
python eval.py --config_path=./default_config.yaml
```
You will get the result as following in "./scripts/device0/eval.log":
```python
@ -335,34 +375,34 @@ Saving ../../results/0-2441_61000/.._.._results_0-2441_61000_face_AP_0.7575.png
### Training Performance
| Parameters | Face Detection |
| -------------------------- | ----------------------------------------------------------- |
| Model Version | V1 |
| Resource | Ascend 910; CPU 2.60GHz, 192cores; Memory 755G; OS Euler2.8 |
| uploaded Date | 09/30/2020 (month/day/year) |
| MindSpore Version | 1.0.0 |
| Dataset | 13K images |
| Training Parameters | epoch=2500, batch_size=64, momentum=0.5 |
| Optimizer | Momentum |
| Loss Function | Softmax Cross Entropy, Sigmoid Cross Entropy, SmoothL1Loss |
| outputs | boxes and label |
| Speed | 1pc: 800~850 ms/step; 8pcs: 1000~1150 ms/step |
| Total time | 1pc: 120 hours; 8pcs: 18 hours |
| Checkpoint for Fine tuning | 37M (.ckpt file) |
| Parameters | Face Detection | Face Detection |
| -------------------------- | ----------------------------------------------------------- | ----------------------------------------------------------- |
| Model Version | V1 | V1 |
| Resource | Ascend 910; CPU 2.60GHz, 192cores; Memory 755G; OS Euler2.8 | GPU PCIE V100-32G |
| uploaded Date | 09/30/2020 (month/day/year) | 07/01/2021 (month/day/year) |
| MindSpore Version | 1.0.0 | 1.3.0 |
| Dataset | 13K images | 13K images |
| Training Parameters | epoch=2500, batch_size=64, momentum=0.5 | epoch=2500, batch_size=64, momentum=0.5 |
| Optimizer | Momentum | Momentum |
| Loss Function | Softmax Cross Entropy, Sigmoid Cross Entropy, SmoothL1Loss | Softmax Cross Entropy, Sigmoid Cross Entropy, SmoothL1Loss |
| outputs | boxes and label | boxes and label |
| Speed | 1pc: 800~850 ms/step; 8pcs: 1000~1150 ms/step | 1pc: 180fps; 4pcs: 385fps |
| Total time | 1pc: 120 hours; 8pcs: 18 hours | 4pcs: 23 hours |
| Checkpoint for Fine tuning | 37M (.ckpt file) | -- |
### Evaluation Performance
| Parameters | Face Detection |
| ------------------- | --------------------------- |
| Model Version | V1 |
| Resource | Ascend 910; OS Euler2.8 |
| Uploaded Date | 09/30/2020 (month/day/year) |
| MindSpore Version | 1.0.0 |
| Dataset | 3K images |
| batch_size | 1 |
| outputs | mAP |
| Accuracy | 8pcs: 76.0% |
| Model for inference | 37M (.ckpt file) |
| Parameters | Face Detection | Face Detection |
| ------------------- | --------------------------- | --------------------------- |
| Model Version | V1 | V1 |
| Resource | Ascend 910; OS Euler2.8 | GPU NV SMX2 V100-32G |
| Uploaded Date | 09/30/2020 (month/day/year) | 07/01/2021 (month/day/year) |
| MindSpore Version | 1.0.0 | 1.3.0 |
| Dataset | 3K images | 3K images |
| batch_size | 1 | 1 |
| outputs | mAP | mAP |
| Accuracy | 8pcs: 76.0% | 4pcs: 77.8% |
| Model for inference | 37M (.ckpt file) | -- |
### Inference Performance

10
model_zoo/research/cv/FaceDetection/default_config.yaml
View File

@ -13,9 +13,9 @@ modelarts_dataset_unzip_name: "face_detect_dataset"
# ==============================================================================
# train options
run_platform: "Ascend" # choices in ("Ascend", "CPU")
mindrecord_path: ""
pretrained: ""
run_platform: "Ascend" # choices in ("Ascend", "CPU", "GPU")
mindrecord_path: "" # change to your own path of train data
pretrained: "" # if you have pretrained backbone, set the path of pretrained_ckpt
use_loss_scale: True
# default options
@ -29,9 +29,9 @@ momentum: 0.5
max_epoch: 2500
log_interval: 10
ckpt_path: "../../output"
ckpt_path: "./output"
ckpt_interval: 1000
result_path: "../../results"
result_path: "./results"
input_shape: [768, 448]
jitter: 0.3

40
model_zoo/research/cv/FaceDetection/scripts/run_distribute_train_gpu.sh Normal file
View File

@ -0,0 +1,40 @@
#!/bin/bash
# Copyright 2021 Huawei Technologies Co., Ltd
#
# 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.
# ============================================================================
if [ $# != 1 ]
then
echo "Usage: bash run_distribute_train_gpu.sh [CONFIG_PATH]"
exit 1
fi
current_exec_path=$(pwd)
dirname_path=$(dirname "$(pwd)")
export PYTHONPATH=${dirname_path}:$PYTHONPATH
SCRIPT_NAME='train.py'
rm -rf ${current_exec_path}/train_parallel
ulimit -c unlimited
CONFIG_PATH=$1
echo 'start training'
mkdir ${current_exec_path}/train_parallel
cd ${current_exec_path}/train_parallel || exit
mpirun --allow-run-as-root -n $RANK_SIZE --output-filename log_output --merge-stderr-to-stdout \
python ${dirname_path}/${SCRIPT_NAME} --config_path=$CONFIG_PATH > train.log 2>&1 &

44
model_zoo/research/cv/FaceDetection/scripts/run_standalone_train_gpu.sh Normal file
View File

@ -0,0 +1,44 @@
#!/bin/bash
# Copyright 2021 Huawei Technologies Co., Ltd
#
# 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.
# ============================================================================
if [ $# != 1 ]
then
echo "Usage: bash run_standalone_train_gpu.sh [CONFIG_PATH]"
exit 1
fi
current_exec_path=$(pwd)
echo ${current_exec_path}
dirname_path=$(dirname "$(pwd)")
echo ${dirname_path}
export PYTHONPATH=${dirname_path}:$PYTHONPATH
export RANK_SIZE=1
SCRIPT_NAME='train.py'
ulimit -c unlimited
CONFIG_PATH=$1
export RANK_ID=0
rm -rf ${current_exec_path}/train
mkdir ${current_exec_path}/train
cd ${current_exec_path}/train || exit
python ${dirname_path}/${SCRIPT_NAME} --config_path=$CONFIG_PATH > train.log 2>&1 &

2
model_zoo/research/cv/FaceDetection/train.py
View File

@ -108,6 +108,8 @@ def run_train():
config.use_loss_scale = False
config.world_size = 1
config.local_rank = 0
if config.run_platform == "GPU":
config.use_loss_scale = False
if config.world_size != 8:
config.lr_steps = [i * 8 // config.world_size for i in config.lr_steps]
config.weight_decay = config.weight_decay if config.world_size != 1 else 0.