nnUnet代码分析一训练

nnUnet是做分割的一套完整代码，用在医学图像分析中较多，效果还很不错。

先看训练的代码 run_training.py

一般用法：nnUNet_train 2d nnUNetTrainerV2 TaskXXX_MYTASK FOLD --npz

2d代表2d Unet网络，nnUNetTrainerV2代表trainer，Task是任务id，

还有其他的参数。详见代码。

plans_file, output_folder_name, dataset_directory, batch_dice, stage, \

    trainer_class = get_default_configuration(network, task, network_trainer, plans_identifier)

根据网络、任务、trainer，计划生产一个trainer_class.

命令：nnUNet_train 2d  nnUNetTrainerV2 Task004_Hippocampus 1 --npz

输出结果如下：

###############################################
I am running the following nnUNet: 2d
My trainer class is:  
For that I will be using the following configuration:
num_classes:  2
modalities:  {0: 'MRI'}
use_mask_for_norm OrderedDict([(0, False)])
keep_only_largest_region None
min_region_size_per_class None
min_size_per_class None
normalization_schemes OrderedDict([(0, 'nonCT')])
stages...

stage:  0
{'batch_size': 366, 'num_pool_per_axis': [3, 3], 'patch_size': array([56, 40]), 'median_patient_size_in_voxels': array([36, 50, 35]), 'current_spacing': array([1., 1., 1.]), 'original_spacing': array([1., 1., 1.]), 'pool_op_kernel_sizes': [[2, 2], [2, 2], [2, 2]], 'conv_kernel_sizes': [[3, 3], [3, 3], [3, 3], [3, 3]], 'do_dummy_2D_data_aug': False}

I am using stage 0 from these plans
I am using batch dice + CE loss

I am using data from this folder:  /mnt/nnUNet_preprocessed/Task004_Hippocampus/nnUNetData_plans_v2.1_2D
###############################################

这里有一个batch dice 和sample dice，stage的概念不是很理解。

按照上面的提示，我们用的trainer是'nnunet.training.network_training.nnUNetTrainerV2.nnUNetTrainerV2'> 

继承自class nnUNetTrainer(NetworkTrainer):

先看：NetworkTrainer.py，nnUNetTrainer.py

do_split :5折验证

训练过程没有什么特别，sgd+poly_lr

loss是dice+ce

数据增强也非常简单，只有缩放和旋转

    def setup_DA_params(self):

        """

        - we increase roation angle from [-15, 15] to [-30, 30]

        - scale range is now (0.7, 1.4), was (0.85, 1.25)

        - we don't do elastic deformation anymore

有早停，patience=50.

下面先跑一下海马体分割的例子。

nnUNet_convert_decathlon_task -i /xxx/Task04_Hippocampus 可以把msd格式的数据转成nnUnet格式。

nnUNet_plan_and_preprocess -t 4 配置plan和preprocess是nnunet中的重要一环。后续再研究

接下去就是训练了：nnUNet_train 3d_fullres nnUNetTrainerV2 4 0

msd中的海马体分割数据集不到30M，经过预处理后生成两个nnUNetData_plans_v2.1_2D_stage0，nnUNetData_plans_v2.1_stage0 ，每个目录有166M，还有一个gt_segmentation目录 ，应该是label？

图像大小才30*50*30左右，，但是速度挺慢的。

k80一个epoch超过了6分钟

epoch:  0
2022-09-03 00:48:13.676541: train loss : -0.3330
2022-09-03 00:48:33.931392: validation loss: -0.7511
2022-09-03 00:48:33.935500: Average global foreground Dice: [0.8332, 0.8122]
2022-09-03 00:48:33.936098: (interpret this as an estimate for the Dice of the different classes. This is not exact.)
2022-09-03 00:48:35.234157: lr: 0.009991
2022-09-03 00:48:35.235549: This epoch took 360.665858 s

2022-09-03 00:48:35.236174: 
epoch:  1

2080ti一个epoch 30s，快了12倍。

epoch:  0
2022-09-03 00:57:25.939667: train loss : -0.3076
2022-09-03 00:57:28.172749: validation loss: -0.7510
2022-09-03 00:57:28.174898: Average global foreground Dice: [0.8301, 0.8239]
2022-09-03 00:57:28.175579: (interpret this as an estimate for the Dice of the different classes. This is not exact.)
2022-09-03 00:57:29.656255: lr: 0.009991
2022-09-03 00:57:29.657619: This epoch took 30.441361 s

2022-09-03 00:57:29.658239: 
epoch:  1

loss为什么是负的呢？

海马体分割有两个区域，2018年nnunet的成绩是0.90，0.89

10个epoch可以到0.89，0.877

33个epoch时得到： [0.9017, 0.8877]

2022-09-03 01:13:15.730579: 
epoch:  33
2022-09-03 01:13:40.102154: train loss : -0.8661
2022-09-03 01:13:41.597394: validation loss: -0.8594
2022-09-03 01:13:41.599368: Average global foreground Dice: [0.9017, 0.8877]
2022-09-03 01:13:41.599944: (interpret this as an estimate for the Dice of the different classes. This is not exact.)
2022-09-03 01:13:43.128609: lr: 0.009693
2022-09-03 01:13:43.135475: saving checkpoint...
2022-09-03 01:13:44.503204: done, saving took 1.37 seconds
2022-09-03 01:13:44.544393: This epoch took 28.813107 s

2022-09-03 01:13:44.545232: 

换成local loss后：

10 epoch 到 [0.8562, 0.8408]

33epochs 到[0.8769, 0.8605] 34 [0.8808, 0.8643]