Huawei_Technology
/
mindspore
Not watched
1
0
Fork 0
Code
Releases 13 Wiki evaluate Activity
Issues 0 Pull Requests 0 Datasets Model Cloudbrain HPC
Browse Source

!15095 update unet README 

From: @zhao_ting_v
Reviewed-by: @guoqi1024,@c_34
Signed-off-by: @c_34
pull/15095/MERGE
mindspore-ci-bot Gitee 4 years ago
parent
d65b370bdf 1e0f164c23
commit
a094ff5b7d
4 changed files with 34 additions and 22 deletions
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. +16
    -12
      model_zoo/official/cv/unet/README.md
  2. +11
    -7
      model_zoo/official/cv/unet/README_CN.md
  3. +5
    -1
      model_zoo/official/cv/unet/src/config.py
  4. +2
    -2
      model_zoo/official/cv/unet/train.py

+ 16
- 12
model_zoo/official/cv/unet/README.md View File

@@ -28,7 +28,7 @@
## [Unet Description](#contents)
Unet Medical model for 2D image segmentation. This implementation is as described in the original paper [UNet: Convolutional Networks for Biomedical Image Segmentation](https://arxiv.org/abs/1505.04597). Unet, in the 2015 ISBI cell tracking competition, many of the best are obtained. In this paper, a network model for medical image segmentation is proposed, and a data enhancement method is proposed to effectively use the annotation data to solve the problem of insufficient annotation data in the medical field. A U-shaped network structure is also used to extract the context and location information.
Unet for 2D image segmentation. This implementation is as described in the original paper [UNet: Convolutional Networks for Biomedical Image Segmentation](https://arxiv.org/abs/1505.04597). Unet, in the 2015 ISBI cell tracking competition, many of the best are obtained. In this paper, a network model for medical image segmentation is proposed, and a data enhancement method is proposed to effectively use the annotation data to solve the problem of insufficient annotation data in the medical field. A U-shaped network structure is also used to extract the context and location information.
UNet++ is a neural architecture for semantic and instance segmentation with re-designed skip pathways and deep supervision.
@@ -71,7 +71,8 @@ After installing MindSpore via the official website, you can start training and
- Select the network and dataset to use
Refer to `src/config.py`. We support some parameter configurations for quick start. You can set `'model'` to `'unet_medical'`,`'unet_nested'` or `'unet_simple'` to select which net to use. We support `ISBI` and `Cell_nuclei` two dataset, you can set `'dataset'` to `'Cell_nuclei'` to use `Cell_nuclei` dataset, default is `ISBI`.
1. Select `cfg_unet` in `src/config.py`. We support unet and unet++, and we provide some parameter configurations for quick start.
2. If you want other parameters, please refer to `src/config.py`. You can set `'model'` to `'unet_nested'` or `'unet_simple'` to select which net to use. We support `ISBI` and `Cell_nuclei` two dataset, you can set `'dataset'` to `'Cell_nuclei'` to use `Cell_nuclei` dataset, default is `ISBI`.
- Run on Ascend
@@ -157,6 +158,7 @@ Parameters for both training and evaluation can be set in config.py
'name': 'Unet', # model name
'lr': 0.0001, # learning rate
'epochs': 400, # total training epochs when run 1p
'repeat': 400, # Repeat times pre one epoch
'distribute_epochs': 1600, # total training epochs when run 8p
'batchsize': 16, # training batch size
'cross_valid_ind': 1, # cross valid ind
@@ -185,6 +187,7 @@ Parameters for both training and evaluation can be set in config.py
'img_size': [96, 96], # image size
'lr': 3e-4, # learning rate
'epochs': 200, # total training epochs when run 1p
'repeat': 10, # Repeat times pre one epoch
'distribute_epochs': 1600, # total training epochs when run 8p
'batchsize': 16, # batch size
'num_classes': 2, # the number of classes in the dataset
@@ -206,6 +209,8 @@ Parameters for both training and evaluation can be set in config.py
'eval_interval': 1 # valuation interval when run_eval is True
```
*Note: total steps pre epoch is floor(epochs / repeat), because unet dataset usually is small, we repeat the dataset to avoid drop too many images when add batch size.*
## [Training Process](#contents)
### Training
@@ -269,7 +274,7 @@ You can add `run_eval` to start shell and set it True, if you want evaluation wh
Before running the command below, please check the checkpoint path used for evaluation. Please set the checkpoint path to be the absolute full path, e.g., "username/unet/ckpt_unet_medical_adam-48_600.ckpt".
```shell
python eval.py --data_url=/path/to/data/ --ckpt_path=/path/to/checkpoint/ > eval.log 2>&1 &
python eval.py --data_url=/path/to/data/ --ckpt_path=/path/to/unet.ckpt > eval.log 2>&1 &
OR
bash scripts/run_standalone_eval.sh [DATASET] [CHECKPOINT]
```
@@ -278,7 +283,7 @@ The above python command will run in the background. You can view the results th
```shell
# grep "Cross valid dice coeff is:" eval.log
============== Cross valid dice coeff is: {'dice_coeff': 0.9085704886070473}
============== Cross valid dice coeff is: {'dice_coeff': 0.9111}
```
## [Model Description](#contents)
@@ -292,17 +297,16 @@ The above python command will run in the background. You can view the results th
| Model Version | Unet |
| Resource | Ascend 910; CPU 2.60GHz, 192cores; Memory 755G; OS Euler2.8 |
| uploaded Date | 09/15/2020 (month/day/year) |
| MindSpore Version | 1.0.0 |
| MindSpore Version | 1.2.0 |
| Dataset | ISBI |
| Training Parameters | 1pc: epoch=400, total steps=600, batch_size = 16, lr=0.0001 |
| | 8pc: epoch=1600, total steps=300, batch_size = 16, lr=0.0001 |
| Optimizer | ADAM |
| Optimizer | Adam |
| Loss Function | Softmax Cross Entropy |
| outputs | probability |
| Loss | 0.22070312 |
| Speed | 1pc: 267 ms/step; 8pc: 280 ms/step; |
| Total time | 1pc: 2.67 mins; 8pc: 1.40 mins |
| Parameters (M) | 93M |
| Speed | 1pc: 267 ms/step |
| Total time | 1pc: 2.67 mins |
| Parameters (M) | 93M |
| Checkpoint for Fine tuning | 355.11M (.ckpt file) |
| Scripts | [unet script](https://gitee.com/mindspore/mindspore/tree/master/model_zoo/official/cv/unet) |
@@ -345,7 +349,7 @@ Set options `resume` to True in `config.py`, and set `resume_ckpt` to the path o
```python
'resume': True,
'resume_ckpt': 'ckpt_0/ckpt_unet_medical_adam_1-1_600.ckpt',
'resume_ckpt': 'ckpt_0/ckpt_unet_sample_adam_1-1_600.ckpt',
'transfer_training': False,
'filter_weight': ["final.weight"]
```
@@ -356,7 +360,7 @@ Do the same thing as resuming traing above. In addition, set `transfer_training`
```python
'resume': True,
'resume_ckpt': 'ckpt_0/ckpt_unet_medical_adam_1-1_600.ckpt',
'resume_ckpt': 'ckpt_0/ckpt_unet_sample_adam_1-1_600.ckpt',
'transfer_training': True,
'filter_weight': ["final.weight"]
```


+ 11
- 7
model_zoo/official/cv/unet/README_CN.md View File

@@ -31,7 +31,7 @@
## U-Net说明
U-Net医学模型基于二维图像分割。实现方式见论文[UNet：Convolutional Networks for Biomedical Image Segmentation](https://arxiv.org/abs/1505.04597)。在2015年ISBI细胞跟踪竞赛中，U-Net获得了许多最佳奖项。论文中提出了一种用于医学图像分割的网络模型和数据增强方法，有效利用标注数据来解决医学领域标注数据不足的问题。U型网络结构也用于提取上下文和位置信息。
U-Net模型基于二维图像分割。实现方式见论文[UNet：Convolutional Networks for Biomedical Image Segmentation](https://arxiv.org/abs/1505.04597)。在2015年ISBI细胞跟踪竞赛中，U-Net获得了许多最佳奖项。论文中提出了一种用于医学图像分割的网络模型和数据增强方法，有效利用标注数据来解决医学领域标注数据不足的问题。U型网络结构也用于提取上下文和位置信息。
UNet++是U-Net的增强版本，使用了新的跨层链接方式和深层监督，可以用于语义分割和实例分割。
@@ -75,7 +75,8 @@ UNet++是U-Net的增强版本，使用了新的跨层链接方式和深层监督
- 选择模型及数据集
我们在`src/config.py`预备了一些网络及数据集的参数配置用于快速体验。也可以通过设置`'model'` 为 `'unet_medical'`,`'unet_nested'` 或者 `'unet_simple'` 来选择使用什么网络结构。我们支持`ISBI` 和 `Cell_nuclei`两种数据集处理，默认使用`ISBI`，可以设置`'dataset'` 为 `'Cell_nuclei'`使用`Cell_nuclei`数据集。
1. 在`src/config.py`中选择相应的配置项赋给`cfg_unet`，现在支持unet和unet++，我们在`src/config.py`预备了一些网络及数据集的参数配置用于快速体验。
2. 如果使用其他的参数，也可以参考`src/config.py`通过设置`'model'` 为 `'unet_nested'` 或者 `'unet_simple'` 来选择使用什么网络结构。我们支持`ISBI` 和 `Cell_nuclei`两种数据集处理，默认使用`ISBI`，可以设置`'dataset'` 为 `'Cell_nuclei'`使用`Cell_nuclei`数据集。
- Ascend处理器环境运行
@@ -161,6 +162,7 @@ bash scripts/docker_start.sh unet:20.1.0 [DATA_DIR] [MODEL_DIR]
'name': 'Unet', # 模型名称
'lr': 0.0001, # 学习率
'epochs': 400, # 运行1p时的总训练轮次
'repeat': 400, # 每一遍epoch重复数据集的次数
'distribute_epochs': 1600, # 运行8p时的总训练轮次
'batchsize': 16, # 训练批次大小
'cross_valid_ind': 1, # 交叉验证指标
@@ -182,6 +184,7 @@ bash scripts/docker_start.sh unet:20.1.0 [DATA_DIR] [MODEL_DIR]
'img_size': [96, 96], # 输入图像大小
'lr': 3e-4, # 学习率
'epochs': 200, # 运行1p时的总训练轮次
'repeat': 10, # 每一遍epoch重复数据集的次数
'distribute_epochs': 1600, # 运行8p时的总训练轮次
'batchsize': 16, # 训练批次大小
'num_classes': 2, # 数据集类数
@@ -199,6 +202,8 @@ bash scripts/docker_start.sh unet:20.1.0 [DATA_DIR] [MODEL_DIR]
'filter_weight': ['final1.weight', 'final2.weight', 'final3.weight', 'final4.weight'] # 迁移学习过滤参数名
```
注意: 实际运行时的每epoch的step数为 floor(epochs / repeat)。这是因为unet的数据集一般都比较小，每一遍epoch重复数据集用来避免在加batch时丢掉过多的图片。
## 训练过程
### 用法
@@ -262,7 +267,7 @@ step: 300, loss is 0.18949677, fps is 57.63118508760329
在运行以下命令之前，请检查用于评估的检查点路径。将检查点路径设置为绝对全路径，如"username/unet/ckpt_unet_medical_adam-48_600.ckpt"。
```shell
python eval.py --data_url=/path/to/data/ --ckpt_path=/path/to/checkpoint/ > eval.log 2>&1 &
python eval.py --data_url=/path/to/data/ --ckpt_path=/path/to/unet.ckpt/ > eval.log 2>&1 &
OR
bash scripts/run_standalone_eval.sh [DATASET] [CHECKPOINT]
```
@@ -271,7 +276,7 @@ step: 300, loss is 0.18949677, fps is 57.63118508760329
```shell
# grep "Cross valid dice coeff is:" eval.log
============== Cross valid dice coeff is: {'dice_coeff': 0.9085704886070473}
============== Cross valid dice coeff is: {'dice_coeff': 0.9111}
```
## 模型描述
@@ -285,11 +290,10 @@ step: 300, loss is 0.18949677, fps is 57.63118508760329
| 模型版本 | U-Net |
| 资源 | Ascend 910；CPU 2.60GHz，192核；内存 755GB；系统 Euler2.8 |
| 上传日期 | 2020-9-15 |
| MindSpore版本 | 1.0.0 |
| MindSpore版本 | 1.2.0 |
| 数据集 | ISBI |
| 训练参数 | 1pc: epoch=400, total steps=600, batch_size = 16, lr=0.0001 |
| | 8pc: epoch=1600, total steps=300, batch_size = 16, lr=0.0001 |
| 优化器 | ADAM |
| 优化器 | Adam |
| 损失函数 | Softmax交叉熵 |
| 输出 | 概率 |
| 损失 | 0.22070312 |


+ 5
- 1
model_zoo/official/cv/unet/src/config.py View File

@@ -19,6 +19,7 @@ cfg_unet_medical = {
'img_size': [572, 572],
'lr': 0.0001,
'epochs': 400,
'repeat': 400,
'distribute_epochs': 1600,
'batchsize': 16,
'cross_valid_ind': 1,
@@ -44,6 +45,7 @@ cfg_unet_nested = {
'img_size': [576, 576],
'lr': 0.0001,
'epochs': 400,
'repeat': 400,
'distribute_epochs': 1600,
'batchsize': 16,
'cross_valid_ind': 1,
@@ -73,6 +75,7 @@ cfg_unet_nested_cell = {
'img_size': [96, 96],
'lr': 3e-4,
'epochs': 200,
'repeat': 10,
'distribute_epochs': 1600,
'batchsize': 16,
'cross_valid_ind': 1,
@@ -101,6 +104,7 @@ cfg_unet_simple = {
'img_size': [576, 576],
'lr': 0.0001,
'epochs': 400,
'repeat': 400,
'distribute_epochs': 1600,
'batchsize': 16,
'cross_valid_ind': 1,
@@ -120,7 +124,7 @@ cfg_unet_simple = {
'eval_resize': False
}
cfg_unet = cfg_unet_medical
cfg_unet = cfg_unet_simple
if not ('dataset' in cfg_unet and cfg_unet['dataset'] == 'Cell_nuclei') and cfg_unet['eval_resize']:
print("ISBI dataset not support resize to original image size when in evaluation.")
cfg_unet['eval_resize'] = False

+ 2
- 2
model_zoo/official/cv/unet/train.py View File

@@ -80,7 +80,7 @@ def train_net(args_opt,
else:
criterion = CrossEntropyWithLogits()
if 'dataset' in cfg and cfg['dataset'] == "Cell_nuclei":
repeat = 10
repeat = cfg['repeat']
dataset_sink_mode = True
per_print_times = 0
train_dataset = create_cell_nuclei_dataset(data_dir, cfg['img_size'], repeat, batch_size,
@@ -90,7 +90,7 @@ def train_net(args_opt,
eval_resize=cfg["eval_resize"], split=0.8,
python_multiprocessing=False)
else:
repeat = epochs
repeat = cfg['repeat']
dataset_sink_mode = False
per_print_times = 1
train_dataset, valid_dataset = create_dataset(data_dir, repeat, batch_size, True, cross_valid_ind,


Write Preview
Loading…
Cancel
Save