diff --git a/README.md b/README.md index a17a7cf8f31341aa49b9ca30a27d1551b326766b..d4d76ddc1f535b6b16d4ec1a290221293bd86310 100644 --- a/README.md +++ b/README.md @@ -64,6 +64,7 @@ $ pip install -r requirements.txt * [如何训练U-Net](./turtorial/finetune_unet.md) * [如何训练ICNet](./turtorial/finetune_icnet.md) * [如何训练PSPNet](./turtorial/finetune_pspnet.md) +* [如何训练HRNet](./turtorial/finetune_hrnet.md) ### 预测部署 diff --git a/configs/hrnet_w18_pet.yaml b/configs/hrnet_w18_pet.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b1bfb9215e7f204444613fd9f6c78eba9c1c1432 --- /dev/null +++ b/configs/hrnet_w18_pet.yaml @@ -0,0 +1,49 @@ +TRAIN_CROP_SIZE: (512, 512) # (width, height), for unpadding rangescaling and stepscaling +EVAL_CROP_SIZE: (512, 512) # (width, height), for unpadding rangescaling and stepscaling +AUG: + AUG_METHOD: "unpadding" # choice unpadding rangescaling and stepscaling + FIX_RESIZE_SIZE: (512, 512) # (width, height), for unpadding + + INF_RESIZE_VALUE: 500 # for rangescaling + MAX_RESIZE_VALUE: 600 # for rangescaling + MIN_RESIZE_VALUE: 400 # for rangescaling + + MAX_SCALE_FACTOR: 1.25 # for stepscaling + MIN_SCALE_FACTOR: 0.75 # for stepscaling + SCALE_STEP_SIZE: 0.25 # for stepscaling + MIRROR: True +BATCH_SIZE: 4 +DATASET: + DATA_DIR: "./dataset/mini_pet/" + IMAGE_TYPE: "rgb" # choice rgb or rgba + NUM_CLASSES: 3 + TEST_FILE_LIST: "./dataset/mini_pet/file_list/test_list.txt" + TRAIN_FILE_LIST: "./dataset/mini_pet/file_list/train_list.txt" + VAL_FILE_LIST: "./dataset/mini_pet/file_list/val_list.txt" + VIS_FILE_LIST: "./dataset/mini_pet/file_list/test_list.txt" + IGNORE_INDEX: 255 + SEPARATOR: " " +FREEZE: + MODEL_FILENAME: "__model__" + PARAMS_FILENAME: "__params__" +MODEL: + MODEL_NAME: "hrnet" + DEFAULT_NORM_TYPE: "bn" + HRNET: + STAGE2: + NUM_CHANNELS: [18, 36] + STAGE3: + NUM_CHANNELS: [18, 36, 72] + STAGE4: + NUM_CHANNELS: [18, 36, 72, 144] +TRAIN: + PRETRAINED_MODEL_DIR: "./pretrained_model/hrnet_w18_bn_cityscapes/" + MODEL_SAVE_DIR: "./saved_model/hrnet_w18_bn_pet/" + SNAPSHOT_EPOCH: 10 +TEST: + TEST_MODEL: "./saved_model/hrnet_w18_bn_pet/final" +SOLVER: + NUM_EPOCHS: 100 + LR: 0.005 + LR_POLICY: "poly" + OPTIMIZER: "sgd" diff --git a/docs/configs/model_group.md b/docs/configs/model_group.md index c787c077103b87b2a3a9c798dda89f2770cdcb40..e11b769de7d8aabbd14583e6666045de6cfc5b42 100644 --- a/docs/configs/model_group.md +++ b/docs/configs/model_group.md @@ -5,10 +5,11 @@ MODEL Group存放所有和模型相关的配置，该Group还包含三个子Grou * [DeepLabv3p](./model_deeplabv3p_group.md) * [UNet](./model_unet_group.md) * [ICNet](./model_icnet_group.md) +* [HRNet](./model_hrnet_group.md) ## `MODEL_NAME` -所选模型，支持`deeplabv3p` `unet` `icnet`三种模型 +所选模型，支持`deeplabv3p` `unet` `icnet` `hrnet`四种模型 ### 默认值 diff --git a/docs/configs/model_hrnet_group.md b/docs/configs/model_hrnet_group.md new file mode 100644 index 0000000000000000000000000000000000000000..055a591a7e99b178a073e496d42035f307a538ed --- /dev/null +++ b/docs/configs/model_hrnet_group.md @@ -0,0 +1,69 @@ +# cfg.MODEL.HRNET + +MODEL.HRNET 子Group存放所有和HRNet模型相关的配置 + +## `STAGE2.NUM_MODULES` + +HRNet在第二阶段执行modularized block（multi-resolution parallel convolution + multi-resolution fusion)的重复次数 + +### 默认值 + +1 + +
+
+ +## `STAGE2.NUM_CHANNELS` + +HRNet在第二阶段各个分支的通道数 + +### 默认值 + +[40, 80] + +
+
+ +## `STAGE3.NUM_MODULES` + +HRNet在第三阶段执行modularized block的重复次数 + +### 默认值 + +4 + +
+
+ +## `STAGE3.NUM_CHANNELS` + +HRNet在第三阶段各个分支的通道数 + +### 默认值 + +[40, 80, 160] + +
+
+ +## `STAGE4.NUM_MODULES` + +HRNet在第四阶段执行modularized block的重复次数 + +### 默认值 + +3 + +
+
+ +## `STAGE4.NUM_CHANNELS` + +HRNet在第四阶段各个分支的通道数 + +### 默认值 + +[40, 80, 160, 320] + +
+
\ No newline at end of file diff --git a/docs/model_zoo.md b/docs/model_zoo.md index a591542cc31379b3df75829173d9cf63a4ae69c1..7e625db73a5ae185b8db00e8dd6f04e26d4e11e5 100644 --- a/docs/model_zoo.md +++ b/docs/model_zoo.md @@ -22,6 +22,16 @@ PaddleSeg对所有内置的分割模型都提供了公开数据集下的预训 | Xception65 | ImageNet | [Xception65_pretrained.tgz](https://paddleseg.bj.bcebos.com/models/Xception65_pretrained.tgz) | 80.32%/94.47% | | Xception71 | ImageNet | coming soon | -- | +| 模型 | 数据集合 | 下载地址 | Accuray Top1/5 Error | +|---|---|---|---| +| HRNet_W18 | ImageNet | [hrnet_w18_imagenet.tar](https://paddleseg.bj.bcebos.com/models/hrnet_w18_imagenet.tar) | 76.92%/93.39% | +| HRNet_W30 | ImageNet | [hrnet_w30_imagenet.tar](https://paddleseg.bj.bcebos.com/models/hrnet_w30_imagenet.tar) | 78.04%/94.02% | +| HRNet_W32 | ImageNet | [hrnet_w32_imagenet.tar](https://paddleseg.bj.bcebos.com/models/hrnet_w32_imagenet.tar) | 78.28%/94.24% | +| HRNet_W40 | ImageNet | [hrnet_w40_imagenet.tar](https://paddleseg.bj.bcebos.com/models/hrnet_w40_imagenet.tar) | 78.77%/94.47% | +| HRNet_W44 | ImageNet | [hrnet_w44_imagenet.tar](https://paddleseg.bj.bcebos.com/models/hrnet_w44_imagenet.tar) | 79.00%/94.51% | +| HRNet_W48 | ImageNet | [hrnet_w48_imagenet.tar](https://paddleseg.bj.bcebos.com/models/hrnet_w48_imagenet.tar) | 78.95%/94.42% | +| HRNet_W64 | ImageNet | [hrnet_w64_imagenet.tar](https://paddleseg.bj.bcebos.com/models/hrnet_w64_imagenet.tar) | 79.30%/94.61% | + ## COCO预训练模型 数据集为COCO实例分割数据集合转换成的语义分割数据集合 @@ -46,3 +56,4 @@ train数据集合为Cityscapes训练集合，测试为Cityscapes的验证集合 | ICNet/bn | Cityscapes |[icnet_cityscapes.tgz](https://paddleseg.bj.bcebos.com/models/icnet_cityscapes.tar.gz) |16|false| 0.6831 | | PSPNet/bn | Cityscapes |[pspnet50_cityscapes.tgz](https://paddleseg.bj.bcebos.com/models/pspnet50_cityscapes.tgz) |16|false| 0.7013 | | PSPNet/bn | Cityscapes |[pspnet101_cityscapes.tgz](https://paddleseg.bj.bcebos.com/models/pspnet101_cityscapes.tgz) |16|false| 0.7734 | +| HRNet_W18/bn | Cityscapes |[hrnet_w18_bn_cityscapes.tgz](https://paddleseg.bj.bcebos.com/models/hrnet_w18_bn_cityscapes.tgz) | 4 | false | 0.7936 | diff --git a/pdseg/models/model_builder.py b/pdseg/models/model_builder.py index 12805164832422189a8b33be1d8b24b983767dcc..56a959a9c20e1322ee5e906008e8bdcf392bd044 100644 --- a/pdseg/models/model_builder.py +++ b/pdseg/models/model_builder.py @@ -112,6 +112,7 @@ def softmax(logit): logit = fluid.layers.transpose(logit, [0, 3, 1, 2]) return logit + def sigmoid_to_softmax(logit): """ one channel to two channel @@ -143,19 +144,23 @@ def build_model(main_prog, start_prog, phase=ModelPhase.TRAIN): # 在导出模型的时候，增加图像标准化预处理,减小预测部署时图像的处理流程 # 预测部署时只须对输入图像增加batch_size维度即可 if ModelPhase.is_predict(phase): - origin_image = fluid.layers.data(name='image', - shape=[ -1, 1, 1, cfg.DATASET.DATA_DIM], - dtype='float32', - append_batch_size=False) + origin_image = fluid.layers.data( + name='image', + shape=[-1, 1, 1, cfg.DATASET.DATA_DIM], + dtype='float32', + append_batch_size=False) image = fluid.layers.transpose(origin_image, [0, 3, 1, 2]) origin_shape = fluid.layers.shape(image)[-2:] mean = np.array(cfg.MEAN).reshape(1, len(cfg.MEAN), 1, 1) mean = fluid.layers.assign(mean.astype('float32')) std = np.array(cfg.STD).reshape(1, len(cfg.STD), 1, 1) std = fluid.layers.assign(std.astype('float32')) - image = (image/255 - mean)/std - image = fluid.layers.resize_bilinear(image, - out_shape=[height, width], align_corners=False, align_mode=0) + image = fluid.layers.resize_bilinear( + image, + out_shape=[height, width], + align_corners=False, + align_mode=0) + image = (image / 255 - mean) / std else: image = fluid.layers.data( name='image', shape=image_shape, dtype='float32') @@ -180,15 +185,20 @@ def build_model(main_prog, start_prog, phase=ModelPhase.TRAIN): loss_type = list(loss_type) # dice_loss或bce_loss只适用两类分割中 - if class_num > 2 and (("dice_loss" in loss_type) or ("bce_loss" in loss_type)): - raise Exception("dice loss and bce loss is only applicable to binary classfication") - + if class_num > 2 and (("dice_loss" in loss_type) or + ("bce_loss" in loss_type)): + raise Exception( + "dice loss and bce loss is only applicable to binary classfication" + ) + # 在两类分割情况下，当loss函数选择dice_loss或bce_loss的时候，最后logit输出通道数设置为1 if ("dice_loss" in loss_type) or ("bce_loss" in loss_type): class_num = 1 if "softmax_loss" in loss_type: - raise Exception("softmax loss can not combine with dice loss or bce loss") - + raise Exception( + "softmax loss can not combine with dice loss or bce loss" + ) + logits = model_func(image, class_num) # 根据选择的loss函数计算相应的损失函数 @@ -196,9 +206,9 @@ def build_model(main_prog, start_prog, phase=ModelPhase.TRAIN): loss_valid = False avg_loss_list = [] valid_loss = [] - if "softmax_loss" in loss_type: - avg_loss_list.append(multi_softmax_with_loss(logits, - label, mask,class_num)) + if "softmax_loss" in loss_type: + avg_loss_list.append( + multi_softmax_with_loss(logits, label, mask, class_num)) loss_valid = True valid_loss.append("softmax_loss") if "dice_loss" in loss_type: @@ -210,13 +220,17 @@ def build_model(main_prog, start_prog, phase=ModelPhase.TRAIN): loss_valid = True valid_loss.append("bce_loss") if not loss_valid: - raise Exception("SOLVER.LOSS: {} is set wrong. it should " - "include one of (softmax_loss, bce_loss, dice_loss) at least" - " example: ['softmax_loss'], ['dice_loss'], ['bce_loss', 'dice_loss']".format(cfg.SOLVER.LOSS)) - + raise Exception( + "SOLVER.LOSS: {} is set wrong. it should " + "include one of (softmax_loss, bce_loss, dice_loss) at least" + " example: ['softmax_loss'], ['dice_loss'], ['bce_loss', 'dice_loss']" + .format(cfg.SOLVER.LOSS)) + invalid_loss = [x for x in loss_type if x not in valid_loss] if len(invalid_loss) > 0: - print("Warning: the loss {} you set is invalid. it will not be included in loss computed.".format(invalid_loss)) + print( + "Warning: the loss {} you set is invalid. it will not be included in loss computed." + .format(invalid_loss)) avg_loss = 0 for i in range(0, len(avg_loss_list)): @@ -238,7 +252,11 @@ def build_model(main_prog, start_prog, phase=ModelPhase.TRAIN): logit = sigmoid_to_softmax(logit) else: logit = softmax(logit) - logit = fluid.layers.resize_bilinear(logit, out_shape=origin_shape, align_corners=False, align_mode=0) + logit = fluid.layers.resize_bilinear( + logit, + out_shape=origin_shape, + align_corners=False, + align_mode=0) logit = fluid.layers.transpose(logit, [0, 2, 3, 1]) logit = fluid.layers.argmax(logit, axis=3) return origin_image, logit diff --git a/pdseg/models/modeling/hrnet.py b/pdseg/models/modeling/hrnet.py index 36ca5eaed70720943a666cb5690416eed8077bfd..741834e157105b233403772f2672ed60aafc488f 100644 --- a/pdseg/models/modeling/hrnet.py +++ b/pdseg/models/modeling/hrnet.py @@ -146,7 +146,7 @@ def layer1(input, name=None): name=name + '_' + str(i + 1)) return conv -def highResolutionNet(input, num_classes): +def high_resolution_net(input, num_classes): channels_2 = cfg.MODEL.HRNET.STAGE2.NUM_CHANNELS channels_3 = cfg.MODEL.HRNET.STAGE3.NUM_CHANNELS @@ -198,7 +198,7 @@ def highResolutionNet(input, num_classes): def hrnet(input, num_classes): - logit = highResolutionNet(input, num_classes) + logit = high_resolution_net(input, num_classes) return logit if __name__ == '__main__': diff --git a/pretrained_model/download_model.py b/pretrained_model/download_model.py index 6d7c265f6514ee3d16aa8e010ba9b071031ef07b..12b01472457bd25e22005141b21bb9d3014bf4fe 100644 --- a/pretrained_model/download_model.py +++ b/pretrained_model/download_model.py @@ -37,6 +37,20 @@ model_urls = { "https://paddleseg.bj.bcebos.com/models/Xception41_pretrained.tgz", "xception65_imagenet": "https://paddleseg.bj.bcebos.com/models/Xception65_pretrained.tgz", + "hrnet_w18_bn_imagenet": + "https://paddleseg.bj.bcebos.com/models/hrnet_w18_imagenet.tar", + "hrnet_w30_bn_imagenet": + "https://paddleseg.bj.bcebos.com/models/hrnet_w30_imagenet.tar", + "hrnet_w32_bn_imagenet": + "https://paddleseg.bj.bcebos.com/models/hrnet_w32_imagenet.tar" , + "hrnet_w40_bn_imagenet": + "https://paddleseg.bj.bcebos.com/models/hrnet_w40_imagenet.tar", + "hrnet_w44_bn_imagenet": + "https://paddleseg.bj.bcebos.com/models/hrnet_w44_imagenet.tar", + "hrnet_w48_bn_imagenet": + "https://paddleseg.bj.bcebos.com/models/hrnet_w48_imagenet.tar", + "hrnet_w64_bn_imagenet": + "https://paddleseg.bj.bcebos.com/models/hrnet_w64_imagenet.tar", # COCO pretrained "deeplabv3p_mobilenetv2-1-0_bn_coco": @@ -65,6 +79,8 @@ model_urls = { "https://paddleseg.bj.bcebos.com/models/pspnet50_cityscapes.tgz", "pspnet101_bn_cityscapes": "https://paddleseg.bj.bcebos.com/models/pspnet101_cityscapes.tgz", + "hrnet_w18_bn_cityscapes": + "https://paddleseg.bj.bcebos.com/models/hrnet_w18_bn_cityscapes.tgz", } if __name__ == "__main__": diff --git a/turtorial/finetune_hrnet.md b/turtorial/finetune_hrnet.md new file mode 100644 index 0000000000000000000000000000000000000000..f7feb9ddafd909fa829cf5f3e3d1c66c82505f57 --- /dev/null +++ b/turtorial/finetune_hrnet.md @@ -0,0 +1,131 @@ +# HRNet模型训练教程 + +* 本教程旨在介绍如何通过使用PaddleSeg提供的 ***`HRNet`*** 预训练模型在自定义数据集上进行训练。 + +* 在阅读本教程前，请确保您已经了解过PaddleSeg的[快速入门](../README.md#快速入门)和[基础功能](../README.md#基础功能)等章节，以便对PaddleSeg有一定的了解 + +* 本教程的所有命令都基于PaddleSeg主目录进行执行 + +## 一. 准备待训练数据 + +我们提前准备好了一份数据集，通过以下代码进行下载 + +```shell +python dataset/download_pet.py +``` + +## 二. 下载预训练模型 + +关于PaddleSeg支持的所有预训练模型的列表，我们可以从[模型组合](#模型组合)中查看我们所需模型的名字和配置 + +接着下载对应的预训练模型 + +```shell +python pretrained_model/download_model.py hrnet_w18_bn_cityscapes +``` + +## 三. 准备配置 + +接着我们需要确定相关配置，从本教程的角度，配置分为三部分： + +* 数据集 + * 训练集主目录 + * 训练集文件列表 + * 测试集文件列表 + * 评估集文件列表 +* 预训练模型 + * 预训练模型名称 + * 预训练模型各阶段通道数设置 + * 预训练模型的Normalization类型 + * 预训练模型路径 +* 其他 + * 学习率 + * Batch大小 + * ... + +在三者中，预训练模型的配置尤为重要，如果模型配置错误，会导致预训练的参数没有加载，进而影响收敛速度。预训练模型相关的配置如第二步所展示。 + +数据集的配置和数据路径有关，在本教程中，数据存放在`dataset/mini_pet`中 + +其他配置则根据数据集和机器环境的情况进行调节，最终我们保存一个如下内容的yaml配置文件，存放路径为**configs/hrnet_w18_pet.yaml** + +```yaml +# 数据集配置 +DATASET: + DATA_DIR: "./dataset/mini_pet/" + NUM_CLASSES: 3 + TEST_FILE_LIST: "./dataset/mini_pet/file_list/test_list.txt" + TRAIN_FILE_LIST: "./dataset/mini_pet/file_list/train_list.txt" + VAL_FILE_LIST: "./dataset/mini_pet/file_list/val_list.txt" + VIS_FILE_LIST: "./dataset/mini_pet/file_list/test_list.txt" + +# 预训练模型配置 +MODEL: + MODEL_NAME: "hrnet" + DEFAULT_NORM_TYPE: "bn" + HRNET: + STAGE2: + NUM_CHANNELS: [18, 36] + STAGE3: + NUM_CHANNELS: [18, 36, 72] + STAGE4: + NUM_CHANNELS: [18, 36, 72, 144] + +# 其他配置 +TRAIN_CROP_SIZE: (512, 512) +EVAL_CROP_SIZE: (512, 512) +AUG: + AUG_METHOD: "unpadding" + FIX_RESIZE_SIZE: (512, 512) +BATCH_SIZE: 4 +TRAIN: + PRETRAINED_MODEL_DIR: "./pretrained_model/hrnet_w18_bn_cityscapes/" + MODEL_SAVE_DIR: "./saved_model/hrnet_w18_bn_pet/" + SNAPSHOT_EPOCH: 10 +TEST: + TEST_MODEL: "./saved_model/hrnet_w18_bn_pet/final" +SOLVER: + NUM_EPOCHS: 100 + LR: 0.005 + LR_POLICY: "poly" + OPTIMIZER: "sgd" +``` + +## 四. 配置/数据校验 + +在开始训练和评估之前，我们还需要对配置和数据进行一次校验，确保数据和配置是正确的。使用下述命令启动校验流程 + +```shell +python pdseg/check.py --cfg ./configs/hrnet_w18_pet.yaml +``` + + +## 五. 开始训练 + +校验通过后，使用下述命令启动训练 + +```shell +python pdseg/train.py --use_gpu --cfg ./configs/hrnet_w18_pet.yaml +``` + +## 六. 进行评估 + +模型训练完成，使用下述命令启动评估 + +```shell +python pdseg/eval.py --use_gpu --cfg ./configs/hrnet_w18_pet.yaml +``` + +## 模型组合 + +|预训练模型名称|BackBone|Norm Type|数据集|配置| +|-|-|-|-|-| +|hrnet_w18_bn_cityscapes|-|bn| ImageNet | MODEL.MODEL_NAME: hrnet
MODEL.HRNET.STAGE2.NUM_CHANNELS: [18, 36]
MODEL.HRNET.STAGE3.NUM_CHANNELS: [18, 36, 72]
MODEL.HRNET.STAGE4.NUM_CHANNELS: [18, 36, 72, 144]
MODEL.DEFAULT_NORM_TYPE: bn| +| hrnet_w18_bn_imagenet |-|bn| ImageNet | MODEL.MODEL_NAME: hrnet
MODEL.HRNET.STAGE2.NUM_CHANNELS: [18, 36]
MODEL.HRNET.STAGE3.NUM_CHANNELS: [18, 36, 72]
MODEL.HRNET.STAGE4.NUM_CHANNELS: [18, 36, 72, 144]
MODEL.DEFAULT_NORM_TYPE: bn | +| hrnet_w30_bn_imagenet |-|bn| ImageNet | MODEL.MODEL_NAME: hrnet
MODEL.HRNET.STAGE2.NUM_CHANNELS: [30, 60]
MODEL.HRNET.STAGE3.NUM_CHANNELS: [30, 60, 120]
MODEL.HRNET.STAGE4.NUM_CHANNELS: [30, 60, 120, 240]
MODEL.DEFAULT_NORM_TYPE: bn | +| hrnet_w32_bn_imagenet |-|bn| ImageNet | MODEL.MODEL_NAME: hrnet
MODEL.HRNET.STAGE2.NUM_CHANNELS: [32, 64]
MODEL.HRNET.STAGE3.NUM_CHANNELS: [32, 64, 128]
MODEL.HRNET.STAGE4.NUM_CHANNELS: [32, 64, 128, 256]
MODEL.DEFAULT_NORM_TYPE: bn | +| hrnet_w40_bn_imagenet |-|bn| ImageNet | MODEL.MODEL_NAME: hrnet
MODEL.HRNET.STAGE2.NUM_CHANNELS: [40, 80]
MODEL.HRNET.STAGE3.NUM_CHANNELS: [40, 80, 160]
MODEL.HRNET.STAGE4.NUM_CHANNELS: [40, 80, 160, 320]
MODEL.DEFAULT_NORM_TYPE: bn | +| hrnet_w44_bn_imagenet |-|bn| ImageNet | MODEL.MODEL_NAME: hrnet
MODEL.HRNET.STAGE2.NUM_CHANNELS: [44, 88]
MODEL.HRNET.STAGE3.NUM_CHANNELS: [44, 88, 176]
MODEL.HRNET.STAGE4.NUM_CHANNELS: [44, 88, 176, 352]
MODEL.DEFAULT_NORM_TYPE: bn | +| hrnet_w48_bn_imagenet |-|bn| ImageNet | MODEL.MODEL_NAME: hrnet
MODEL.HRNET.STAGE2.NUM_CHANNELS: [48, 96]
MODEL.HRNET.STAGE3.NUM_CHANNELS: [48, 96, 192]
MODEL.HRNET.STAGE4.NUM_CHANNELS: [48, 96, 192, 384]
MODEL.DEFAULT_NORM_TYPE: bn | +| hrnet_w64_bn_imagenet |-|bn| ImageNet | MODEL.MODEL_NAME: hrnet
MODEL.HRNET.STAGE2.NUM_CHANNELS: [64, 128]
MODEL.HRNET.STAGE3.NUM_CHANNELS: [64, 128, 256]
MODEL.HRNET.STAGE4.NUM_CHANNELS: [64, 128, 256, 512]
MODEL.DEFAULT_NORM_TYPE: bn | +