diff --git a/README.md b/README.md index fd886ce..4a55317 100644 --- a/README.md +++ b/README.md @@ -64,30 +64,35 @@ PaddleRS具有以下五大特色：
  • ResNet50-vd
  • MobileNetV3
  • HRNet
    • +
  • ...
    • 语义分割
    目标检测
    图像复原
    变化检测
    @@ -114,6 +119,7 @@ PaddleRS具有以下五大特色：
  • ReduceDim
  • SelectBand
  • RandomSwap
    • +
  • ...
    • @@ -122,12 +128,15 @@ PaddleRS具有以下五大特色：
  • coco to mask
  • mask to shpfile
  • mask to geojson
    • +
  • ...
    • 数据预处理
    @@ -135,7 +144,7 @@ PaddleRS具有以下五大特色： - 遥感语义分割
    + 遥感图像分割
    @@ -147,7 +156,7 @@ PaddleRS具有以下五大特色： - 遥感影像超分
    + 遥感图像复原
    @@ -191,8 +200,9 @@ PaddleRS目录树中关键部分如下： * [智能标注工具EISeg](https://github.com/PaddlePaddle/PaddleSeg/tree/release/2.6/EISeg) * [遥感影像处理工具集](./docs/data/tools.md) * 组件介绍 - * [数据预处理/数据增强](./docs/intro/transforms.md) + * [数据集预处理脚本](./docs/intro/data_prep.md) * [模型库](./docs/intro/model_zoo.md) + * [数据变换算子](./docs/intro/transforms.md) * 模型训练 * [模型训练API说明](./docs/apis/train.md) * 模型部署 diff --git a/docs/apis/data.md b/docs/apis/data.md index ab8a523..6be2e32 100644 --- a/docs/apis/data.md +++ b/docs/apis/data.md @@ -86,6 +86,22 @@ ### 图像复原数据集`ResDataset` +`ResDataset`定义在：https://github.com/PaddlePaddle/PaddleRS/blob/develop/paddlers/datasets/res_dataset.py + +其初始化参数列表如下： + +|参数名称|类型|参数说明|默认值| +|-------|----|--------|-----| +|`data_dir`|`str`|数据集存放目录。|| +|`file_list`|`str`|file list路径。file list是一个文本文件，其中每一行包含一个样本的路径信息。`ResDataset`对file list的具体要求请参见下文。|| +|`transforms`|`paddlers.transforms.Compose`|对输入数据应用的数据变换算子。|| +|`num_workers`|`int` \| `str`|加载数据时使用的辅助进程数。若设置为`'auto'`，则按照如下规则确定使用进程数：当CPU核心数大于16时，使用8个数据读取辅助进程；否则，使用CPU核心数一半数量的辅助进程。|`'auto'`| +|`shuffle`|`bool`|是否随机打乱数据集中的样本。|`False`| +|`sr_factor`|`int` \| `None`|对于超分辨率重建任务，指定为超分辨率倍数；对于其它任务，指定为`None`。|`None`| + +`ResDataset`对file list的要求如下： + +- file list中的每一行应该包含2个以空格分隔的项，依次表示输入影像（例如超分辨率重建任务中的低分辨率影像）相对`data_dir`的路径以及目标影像（例如超分辨率重建任务中的高分辨率影像）相对`data_dir`的路径。 ### 图像分割数据集`SegDataset` diff --git a/docs/apis/infer.md b/docs/apis/infer.md index 2a5b62d..880b5cc 100644 --- a/docs/apis/infer.md +++ b/docs/apis/infer.md @@ -89,7 +89,28 @@ def predict(self, img_file, transforms=None): #### `BaseRestorer.predict()` +接口形式： + +```python +def predict(self, img_file, transforms=None): +``` + +输入参数： + +|参数名称|类型|参数说明|默认值| +|-------|----|--------|-----| +|`img_file`|`list[str\|np.ndarray]` \| `str` \| `np.ndarray`|输入影像数据（NumPy数组形式）或输入影像路径。若需要一次性预测一组影像，以列表包含这些影像的数据或路径（每幅影像对应列表中的一个元素）。|| +|`transforms`|`paddlers.transforms.Compose` \| `None`|对输入数据应用的数据变换算子。若为`None`，则使用训练器在验证阶段使用的数据变换算子。|`None`| + +返回格式： + +若`img_file`是一个字符串或NumPy数组，则返回对象为包含下列键值对的字典： +``` +{"res_map": 模型输出的复原或重建影像（以[h, w, c]格式排布）} +``` + +若`img_file`是一个列表，则返回对象为与`img_file`等长的列表，其中的每一项为一个字典（键值对如上所示），顺序对应`img_file`中的每个元素。 #### `BaseSegmenter.predict()` @@ -190,7 +211,7 @@ def predict(self, |参数名称|类型|参数说明|默认值| |-------|----|--------|-----| -|`img_file`|`list[str\|tuple\|np.ndarray]` \| `str` \| `tuple` \| `np.ndarray`|对于场景分类、目标检测和图像分割任务来说，该参数可为单一图像路径，或是解码后的、排列格式为[h, w, c]且具有float32类型的图像数据（表示为NumPy数组形式），或者是一组图像路径或np.ndarray对象构成的列表；对于变化检测任务来说，该参数可以为图像路径二元组（分别表示前后两个时相影像路径），或是解码后的两幅图像组成的二元组，或者是上述两种二元组之一构成的列表。|| +|`img_file`|`list[str\|tuple\|np.ndarray]` \| `str` \| `tuple` \| `np.ndarray`|对于场景分类、目标检测、图像复原和图像分割任务来说，该参数可为单一图像路径，或是解码后的、排列格式为[h, w, c]且具有float32类型的图像数据（表示为NumPy数组形式），或者是一组图像路径或np.ndarray对象构成的列表；对于变化检测任务来说，该参数可以为图像路径二元组（分别表示前后两个时相影像路径），或是解码后的两幅图像组成的二元组，或者是上述两种二元组之一构成的列表。|| |`topk`|`int`|场景分类模型预测时使用，表示选取模型输出概率大小排名前`topk`的类别作为最终结果。|`1`| |`transforms`|`paddlers.transforms.Compose`\|`None`|对输入数据应用的数据变换算子。若为`None`，则使用从`model.yml`中读取的算子。|`None`| |`warmup_iters`|`int`|预热轮数，用于评估模型推理以及前后处理速度。若大于1，将预先重复执行`warmup_iters`次推理，而后才开始正式的预测及其速度评估。|`0`| diff --git a/docs/apis/train.md b/docs/apis/train.md index 5ac48f3..d28aa94 100644 --- a/docs/apis/train.md +++ b/docs/apis/train.md @@ -1,6 +1,6 @@ # PaddleRS训练API说明 -**训练器**封装了模型训练、验证、量化以及动态图推理等逻辑，定义在`paddlers/tasks/`目录下的文件中。为了方便用户使用，PaddleRS为所有支持的模型均提供了继承自父类[`BaseModel`](https://github.com/PaddlePaddle/PaddleRS/blob/develop/paddlers/tasks/base.py)的训练器，并对外提供数个API。变化检测、场景分类、图像分割以及目标检测任务对应的训练器类型分别为`BaseChangeDetector`、`BaseClassifier`、`BaseDetector`和`BaseSegmenter`。本文档介绍训练器的初始化函数以及`train()`、`evaluate()` API。 +**训练器**封装了模型训练、验证、量化以及动态图推理等逻辑，定义在`paddlers/tasks/`目录下的文件中。为了方便用户使用，PaddleRS为所有支持的模型均提供了继承自父类[`BaseModel`](https://github.com/PaddlePaddle/PaddleRS/blob/develop/paddlers/tasks/base.py)的训练器，并对外提供数个API。变化检测、场景分类、目标检测、图像复原以及图像分割任务对应的训练器类型分别为`BaseChangeDetector`、`BaseClassifier`、`BaseDetector`、`BaseRestorer`和`BaseSegmenter`。本文档介绍训练器的初始化函数以及`train()`、`evaluate()` API。 ## 初始化训练器 @@ -10,27 +10,33 @@ - 一般支持设置`num_classes`、`use_mixed_loss`以及`in_channels`参数，分别表示模型输出类别数、是否使用预置的混合损失以及输入通道数。部分子类如`DSIFN`暂不支持对`in_channels`参数的设置。 - `use_mixed_loss`参将在未来被弃用，因此不建议使用。 +- 可通过`losses`参数指定模型训练时使用的损失函数。`losses`需为一个字典，其中`'types'`键和`'coef'`键对应的值为两个等长的列表，分别表示损失函数对象（一个可调用对象）和损失函数的权重。例如：`losses={'types': [LossType1(), LossType2()], 'coef': [1.0, 0.5]}`在训练过程中将等价于计算如下损失函数：`1.0*LossType1()(logits, labels)+0.5*LossType2()(logits, labels)`，其中`logits`和`labels`分别是模型输出和真值标签。 - 不同的子类支持与模型相关的输入参数，详情请参考[模型定义](https://github.com/PaddlePaddle/PaddleRS/blob/develop/paddlers/rs_models/cd)和[训练器定义](https://github.com/PaddlePaddle/PaddleRS/blob/develop/paddlers/tasks/change_detector.py)。 ### 初始化`BaseClassifier`子类对象 - 一般支持设置`num_classes`和`use_mixed_loss`参数，分别表示模型输出类别数以及是否使用预置的混合损失。 - `use_mixed_loss`参将在未来被弃用，因此不建议使用。 +- 可通过`losses`参数指定模型训练时使用的损失函数，传入实参需为`paddlers.models.clas_losses.CombinedLoss`类型对象。 - 不同的子类支持与模型相关的输入参数，详情请参考[模型定义](https://github.com/PaddlePaddle/PaddleRS/blob/develop/paddlers/rs_models/clas)和[训练器定义](https://github.com/PaddlePaddle/PaddleRS/blob/develop/paddlers/tasks/classifier.py)。 ### 初始化`BaseDetector`子类对象 - 一般支持设置`num_classes`和`backbone`参数，分别表示模型输出类别数以及所用的骨干网络类型。相比其它任务，目标检测任务的训练器支持设置的初始化参数较多，囊括网络结构、损失函数、后处理策略等方面。 +- 与分割、分类、变化检测等任务不同，检测任务不支持通过`losses`参数指定损失函数。不过对于部分训练器如`PPYOLO`，可通过`use_iou_loss`等参数定制损失函数。 - 不同的子类支持与模型相关的输入参数，详情请参考[模型定义](https://github.com/PaddlePaddle/PaddleRS/blob/develop/paddlers/rs_models/det)和[训练器定义](https://github.com/PaddlePaddle/PaddleRS/blob/develop/paddlers/tasks/object_detector.py)。 ### 初始化`BaseRestorer`子类对象 - +- 一般支持设置`sr_factor`参数，表示超分辨率倍数；对于不支持超分辨率重建任务的模型，`sr_factor`设置为`None`。 +- 可通过`losses`参数指定模型训练时使用的损失函数，传入实参需为可调用对象或字典。手动指定的`losses`与子类的`default_loss()`方法返回值必须具有相同的格式。 +- 不同的子类支持与模型相关的输入参数，详情请参考[模型定义](https://github.com/PaddlePaddle/PaddleRS/blob/develop/paddlers/rs_models/res)和[训练器定义](https://github.com/PaddlePaddle/PaddleRS/blob/develop/paddlers/tasks/restorer.py)。 ### 初始化`BaseSegmenter`子类对象 - 一般支持设置`in_channels`、`num_classes`以及`use_mixed_loss`参数，分别表示输入通道数、输出类别数以及是否使用预置的混合损失。部分模型如`FarSeg`暂不支持对`in_channels`参数的设置。 - `use_mixed_loss`参将在未来被弃用，因此不建议使用。 +- 可通过`losses`参数指定模型训练时使用的损失函数。`losses`需为一个字典，其中`'types'`键和`'coef'`键对应的值为两个等长的列表，分别表示损失函数对象（一个可调用对象）和损失函数的权重。例如：`losses={'types': [LossType1(), LossType2()], 'coef': [1.0, 0.5]}`在训练过程中将等价于计算如下损失函数：`1.0*LossType1()(logits, labels)+0.5*LossType2()(logits, labels)`，其中`logits`和`labels`分别是模型输出和真值标签。 - 不同的子类支持与模型相关的输入参数，详情请参考[模型定义](https://github.com/PaddlePaddle/PaddleRS/blob/develop/paddlers/rs_models/seg)和[训练器定义](https://github.com/PaddlePaddle/PaddleRS/blob/develop/paddlers/tasks/segmentor.py)。 ## `train()` @@ -176,6 +182,46 @@ def train(self, ### `BaseRestorer.train()` +接口形式： + +```python +def train(self, + num_epochs, + train_dataset, + train_batch_size=2, + eval_dataset=None, + optimizer=None, + save_interval_epochs=1, + log_interval_steps=2, + save_dir='output', + pretrain_weights='CITYSCAPES', + learning_rate=0.01, + lr_decay_power=0.9, + early_stop=False, + early_stop_patience=5, + use_vdl=True, + resume_checkpoint=None): +``` + +其中各参数的含义如下： + +|参数名称|类型|参数说明|默认值| +|-------|----|--------|-----| +|`num_epochs`|`int`|训练的epoch数目。|| +|`train_dataset`|`paddlers.datasets.ResDataset`|训练数据集。|| +|`train_batch_size`|`int`|训练时使用的batch size。|`2`| +|`eval_dataset`|`paddlers.datasets.ResDataset` \| `None`|验证数据集。|`None`| +|`optimizer`|`paddle.optimizer.Optimizer` \| `None`|训练时使用的优化器。若为`None`，则使用默认定义的优化器。|`None`| +|`save_interval_epochs`|`int`|训练时存储模型的间隔epoch数。|`1`| +|`log_interval_steps`|`int`|训练时打印日志的间隔step数（即迭代数）。|`2`| +|`save_dir`|`str`|存储模型的路径。|`'output'`| +|`pretrain_weights`|`str` \| `None`|预训练权重的名称/路径。若为`None`，则不适用预训练权重。|`'CITYSCAPES'`| +|`learning_rate`|`float`|训练时使用的学习率大小，适用于默认优化器。|`0.01`| +|`lr_decay_power`|`float`|学习率衰减系数，适用于默认优化器。|`0.9`| +|`early_stop`|`bool`|训练过程是否启用早停策略。|`False`| +|`early_stop_patience`|`int`|启用早停策略时的`patience`参数（参见[`EarlyStop`](https://github.com/PaddlePaddle/PaddleRS/blob/develop/paddlers/utils/utils.py)）。|`5`| +|`use_vdl`|`bool`|是否启用VisualDL日志。|`True`| +|`resume_checkpoint`|`str` \| `None`|检查点路径。PaddleRS支持从检查点（包含先前训练过程中存储的模型权重和优化器权重）继续训练，但需注意`resume_checkpoint`与`pretrain_weights`不得同时设置为`None`以外的值。|`None`| ### `BaseSegmenter.train()` @@ -280,7 +326,7 @@ def evaluate(self, eval_dataset, batch_size=1, return_details=False): ``` {"top1": top1准确率, - "top5": `top5准确率} + "top5": top5准确率} ``` ### `BaseDetector.evaluate()` @@ -320,6 +366,26 @@ def evaluate(self, ### `BaseRestorer.evaluate()` +接口形式： + +```python +def evaluate(self, eval_dataset, batch_size=1, return_details=False): +``` + +输入参数如下： + +|参数名称|类型|参数说明|默认值| +|-------|----|--------|-----| +|`eval_dataset`|`paddlers.datasets.ResDataset`|评估数据集。|| +|`batch_size`|`int`|评估时使用的batch size（多卡训练时，为所有设备合计batch size）。|`1`| +|`return_details`|`bool`|*当前版本请勿手动设置此参数。*|`False`| + +输出为一个`collections.OrderedDict`对象，包含如下键值对： + +``` +{"psnr": PSNR指标, + "ssim": SSIM指标} +``` ### `BaseSegmenter.evaluate()` diff --git a/docs/data/tools.md b/docs/data/tools.md index f434c49..035832d 100644 --- a/docs/data/tools.md +++ b/docs/data/tools.md @@ -8,8 +8,9 @@ PaddleRS在`tools`目录中提供了丰富的遥感影像处理工具，包括 - `match.py`：用于实现两幅影像的配准。 - `split.py`：用于对大幅面影像数据进行切片。 - `coco_tools/`：COCO工具合集，用于统计处理COCO格式标注文件。 +- `prepare_dataset/`：数据集预处理脚本合集。 -## 使用示例 +## 使用说明 首先请确保您已将PaddleRS下载到本地。进入`tools`目录： @@ -101,3 +102,24 @@ python split.py --image_path {输入影像路径} [--mask_path {真值标签路 - `json_Merge.py`: 将多个json文件合并为一个。 详细使用方法请参见[coco_tools使用说明](coco_tools.md)。 + +### prepare_dataset + +`prepare_dataset`目录中包含一系列数据预处理脚本，主要用于预处理已下载到本地的遥感开源数据集，使其符合PaddleRS训练、验证、测试的标准。 + +在执行脚本前，您可以通过`--help`选项获取帮助信息。例如： + +```shell +python prepare_dataset/prepare_levircd.py --help +``` + +以下列出了脚本中常见的命令行选项： + +- `--in_dataset_dir`：下载到本地的原始数据集所在路径。示例：`--in_dataset_dir downloads/LEVIR-CD`。 +- `--out_dataset_dir`：处理后的数据集存放路径。示例：`--out_dataset_dir data/levircd`。 +- `--crop_size`：对于支持影像裁块的数据集，指定切分的影像块大小。示例：`--crop_size 256`。 +- `--crop_stride`：对于支持影像裁块的数据集，指定切分时滑窗移动的步长。示例：`--crop_stride 256`。 +- `--seed`：随机种子。可用于固定随机数生成器产生的伪随机数序列，从而得到固定的数据集划分结果。示例：`--seed 1919810` +- `--ratios`：对于支持子集随机划分的数据集，指定需要划分的各个子集的样本比例。示例：`--ratios 0.7 0.2 0.1`。 + +您可以在[此文档](https://github.com/PaddlePaddle/PaddleRS/blob/develop/docs/intro/data_prep.md)中查看PaddleRS提供哪些数据集的预处理脚本。 diff --git a/docs/dev/dev_guide.md b/docs/dev/dev_guide.md index 9f678af..e4de181 100644 --- a/docs/dev/dev_guide.md +++ b/docs/dev/dev_guide.md @@ -70,6 +70,15 @@ Args: 4. 在全局变量`__all__`中添加新增训练器的类名。 +需要注意的是，对于图像复原任务，模型的前向、反向逻辑均实现在训练器定义中。对于GAN等需要用到多个网络的模型，训练器的编写请参照如下规范： +- 重写`build_net()`方法，使用`GANAdapter`维护所有网络。`GANAdapter`对象在构造时接受两个列表作为输入：第一个列表中包含所有的生成器，其中第一个元素为主生成器；第二个列表中包含所有的判别器。 +- 重写`default_loss()`方法，构建损失函数。若训练过程中需要用到多个损失函数，推荐以字典的形式组织。 +- 重写`default_optimizer()`方法，构建一个或多个优化器。当`build_net()`返回值的类型为`GANAdapter`时，`parameters`参数为一个字典。其中，`parameters['params_g']`是一个列表，顺序包含各个生成器的state dict；`parameters['params_d']`是一个列表，顺序包含各个判别器的state dict。若构建多个优化器，在返回时应使用`OptimizerAdapter`包装。 +- 重写`run_gan()`方法，该方法接受`net`、`inputs`、`mode`、和`gan_mode`四个参数，用于执行训练过程中的某一个子任务，例如生成器的前向计算、判别器的前向计算等等。 +- 重写`train_step()`方法，在其中编写模型训练过程中一次迭代的具体逻辑。通常的做法是反复调用`run_gan()`，每次调用时都根据需要构造不同的`inputs`、并使其工作在不同的`gan_mode`，并从每次返回的`outputs`字典中抽取有用的字段（如各项损失），汇总至最终结果。 + +GAN训练器的具体例子可以参考`ESRGAN`。 + ## 2 新增数据预处理/数据增强函数或算子 ### 2.1 新增数据预处理/数据增强函数 diff --git a/docs/images/whole_picture.png b/docs/images/whole_picture.png index dcd5945..0d10a23 100644 Binary files a/docs/images/whole_picture.png and b/docs/images/whole_picture.png differ diff --git a/docs/intro/data_prep.md b/docs/intro/data_prep.md new file mode 100644 index 0000000..6932e6b --- /dev/null +++ b/docs/intro/data_prep.md @@ -0,0 +1,9 @@ +# 数据集预处理脚本 + +## PaddleRS已支持的数据集预处理脚本列表 + +| 任务 | 数据集名称 | 数据集地址 | 预处理脚本 | +|-----|-----------|----------|----------| +| 变化检测 | LEVIR-CD | https://justchenhao.github.io/LEVIR/ | [prepare_levircd.py](https://github.com/PaddlePaddle/PaddleRS/blob/develop/tools/prepare_dataset/prepare_levircd.py) | +| 变化检测 | Season-varying | https://paperswithcode.com/dataset/cdd-dataset-season-varying | [prepare_svcd.py](https://github.com/PaddlePaddle/PaddleRS/blob/develop/tools/prepare_dataset/prepare_svcd.py) | +| 目标检测 | RSOD | https://github.com/RSIA-LIESMARS-WHU/RSOD-Dataset- | [prepare_rsod](https://github.com/PaddlePaddle/PaddleRS/blob/develop/tools/prepare_dataset/prepare_rsod.py) | diff --git a/docs/intro/transforms.md b/docs/intro/transforms.md index a86e63e..2478134 100644 --- a/docs/intro/transforms.md +++ b/docs/intro/transforms.md @@ -1,4 +1,4 @@ -# 数据预处理/数据增强 +# 数据变换算子 ## PaddleRS已支持的数据变换算子列表 diff --git a/examples/README.md b/examples/README.md index 14d3864..18f0105 100644 --- a/examples/README.md +++ b/examples/README.md @@ -8,9 +8,31 @@ PaddleRS提供从科学研究到产业应用的丰富示例，希望帮助遥感 ## 2 社区贡献案例 -[AI Studio](https://aistudio.baidu.com/aistudio/index)是基于百度深度学习平台飞桨的人工智能学习与实训社区，提供在线编程环境、免费GPU算力、海量开源算法和开放数据，帮助开发者快速创建和部署模型。您可以在AI Studio上探索PaddleRS的更多玩法： +### 2.1 基于PaddleRS的遥感解译平台 -[AI Studio上的PaddleRS相关项目](https://aistudio.baidu.com/aistudio/projectoverview/public?kw=PaddleRS) +#### 小桨神瞳 + +

    + +

    + +- 作者：白菜 +- 代码仓库：https://github.com/CrazyBoyM/webRS +- 演示视频：https://www.bilibili.com/video/BV1W14y1s7fs?vd_source=0de109a09b98176090b8aa3295a45bb6 + +#### 遥感图像智能解译平台 + +

    + +

    + +- 作者：HHU-河马海牛队 +- 代码仓库：https://github.com/terayco/Intelligent-RS-System +- 演示视频：https://www.bilibili.com/video/BV1eY4y1u7Eq/?vd_source=75a73fc15a4e8b25195728ee93a5b322 + +### 2.2 AI Studio上的PaddleRS相关项目 + +[AI Studio](https://aistudio.baidu.com/aistudio/index)是基于百度深度学习平台飞桨的人工智能学习与实训社区，提供在线编程环境、免费GPU算力、海量开源算法和开放数据，帮助开发者快速创建和部署模型。您可以[在AI Studio上探索PaddleRS的更多玩法](https://aistudio.baidu.com/aistudio/projectoverview/public?kw=PaddleRS)。 本文档收集了部分由开源爱好者贡献的精品项目： diff --git a/paddlers/rs_models/cd/bit.py b/paddlers/rs_models/cd/bit.py index 720969d..13f2290 100644 --- a/paddlers/rs_models/cd/bit.py +++ b/paddlers/rs_models/cd/bit.py @@ -56,7 +56,7 @@ class BIT(nn.Layer): Default: 2. enc_with_pos (bool, optional): Whether to add leanred positional embedding to the input feature sequence of the encoder. Default: True. - enc_depth (int, optional): Number of attention blocks used in the encoder. Default: 1 + enc_depth (int, optional): Number of attention blocks used in the encoder. Default: 1. enc_head_dim (int, optional): Embedding dimension of each encoder head. Default: 64. dec_depth (int, optional): Number of attention blocks used in the decoder. Default: 8. dec_head_dim (int, optional): Embedding dimension of each decoder head. Default: 8. diff --git a/paddlers/rs_models/seg/farseg.py b/paddlers/rs_models/seg/farseg.py index b9c6b95..7a5a62a 100644 --- a/paddlers/rs_models/seg/farseg.py +++ b/paddlers/rs_models/seg/farseg.py @@ -11,11 +11,10 @@ # 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. -""" -This code is based on https://github.com/Z-Zheng/FarSeg -Ths copyright of Z-Zheng/FarSeg is as follows: -Apache License [see LICENSE for details] -""" + +# This code is based on https://github.com/Z-Zheng/FarSeg +# The copyright of Z-Zheng/FarSeg is as follows: +# Apache License (see https://github.com/Z-Zheng/FarSeg/blob/master/LICENSE for details). import math @@ -164,7 +163,7 @@ class SceneRelation(nn.Layer): return refined_feats -class AssymetricDecoder(nn.Layer): +class AsymmetricDecoder(nn.Layer): def __init__(self, in_channels, out_channels, @@ -172,7 +171,7 @@ class AssymetricDecoder(nn.Layer): out_feat_output_stride=4, norm_fn=nn.BatchNorm2D, num_groups_gn=None): - super(AssymetricDecoder, self).__init__() + super(AsymmetricDecoder, self).__init__() if norm_fn == nn.BatchNorm2D: norm_fn_args = dict(num_features=out_channels) elif norm_fn == nn.GroupNorm: @@ -215,9 +214,12 @@ class AssymetricDecoder(nn.Layer): class ResNet50Encoder(nn.Layer): - def __init__(self, pretrained=True): + def __init__(self, in_ch=3, pretrained=True): super(ResNet50Encoder, self).__init__() self.resnet = resnet50(pretrained=pretrained) + if in_ch != 3: + self.resnet.conv1 = nn.Conv2D( + in_ch, 64, kernel_size=7, stride=2, padding=3, bias_attr=False) def forward(self, inputs): x = inputs @@ -234,25 +236,35 @@ class ResNet50Encoder(nn.Layer): class FarSeg(nn.Layer): """ - The FarSeg implementation based on PaddlePaddle. + The FarSeg implementation based on PaddlePaddle. + + The original article refers to + Zheng, Zhuo, et al. "Foreground-Aware Relation Network for Geospatial Object Segmentation in High Spatial Resolution + Remote Sensing Imagery" + (https://openaccess.thecvf.com/content_CVPR_2020/papers/Zheng_Foreground-Aware_Relation_Network_for_Geospatial_Object_Segmentation_in_High_Spatial_CVPR_2020_paper.pdf) - The original article refers to - Zheng, Zhuo, et al. "Foreground-Aware Relation Network for Geospatial Object - Segmentation in High Spatial Resolution Remote Sensing Imagery" - (https://openaccess.thecvf.com/content_CVPR_2020/papers/Zheng_Foreground-Aware_Relation_Network_for_Geospatial_Object_Segmentation_in_High_Spatial_CVPR_2020_paper.pdf) + Args: + in_channels (int, optional): Number of bands of the input images. Default: 3. + num_classes (int, optional): Number of target classes. Default: 16. + fpn_ch_list (list[int]|tuple[int], optional): Channel list of the FPN. Default: (256, 512, 1024, 2048). + mid_ch (int, optional): Output channels of the FPN. Default: 256. + out_ch (int, optional): Output channels of the decoder. Default: 128. + sr_ch_list (list[int]|tuple[int], optional): Channel list of the foreground-scene relation module. Default: (256, 256, 256, 256). + pretrained_encoder (bool, optional): Whether to use a pretrained encoder. Default: True. """ def __init__(self, + in_channels=3, num_classes=16, fpn_ch_list=(256, 512, 1024, 2048), mid_ch=256, out_ch=128, sr_ch_list=(256, 256, 256, 256), - encoder_pretrained=True): + pretrained_encoder=True): super(FarSeg, self).__init__() - self.en = ResNet50Encoder(encoder_pretrained) + self.en = ResNet50Encoder(in_channels, pretrained_encoder) self.fpn = FPN(in_channels_list=fpn_ch_list, out_channels=mid_ch) - self.decoder = AssymetricDecoder( + self.decoder = AsymmetricDecoder( in_channels=mid_ch, out_channels=out_ch) self.cls_pred_conv = nn.Conv2D(out_ch, num_classes, 1) self.upsample4x_op = nn.UpsamplingBilinear2D(scale_factor=4) @@ -273,5 +285,4 @@ class FarSeg(nn.Layer): final_feat = self.decoder(refined_fpn_feat_list) cls_pred = self.cls_pred_conv(final_feat) cls_pred = self.upsample4x_op(cls_pred) - cls_pred = F.softmax(cls_pred, axis=1) return [cls_pred] diff --git a/paddlers/tasks/change_detector.py b/paddlers/tasks/change_detector.py index 7a45172..f822b90 100644 --- a/paddlers/tasks/change_detector.py +++ b/paddlers/tasks/change_detector.py @@ -31,7 +31,7 @@ import paddlers.utils.logging as logging from paddlers.models import seg_losses from paddlers.transforms import Resize, decode_image from paddlers.utils import get_single_card_bs -from paddlers.utils.checkpoint import seg_pretrain_weights_dict +from paddlers.utils.checkpoint import cd_pretrain_weights_dict from .base import BaseModel from .utils import seg_metrics as metrics from .utils.infer_nets import InferCDNet @@ -275,7 +275,7 @@ class BaseChangeDetector(BaseModel): exit=True) if pretrain_weights is not None and resume_checkpoint is not None: logging.error( - "pretrain_weights and resume_checkpoint cannot be set simultaneously.", + "`pretrain_weights` and `resume_checkpoint` cannot be set simultaneously.", exit=True) self.labels = train_dataset.labels if self.losses is None: @@ -289,23 +289,30 @@ class BaseChangeDetector(BaseModel): else: self.optimizer = optimizer - if pretrain_weights is not None and not osp.exists(pretrain_weights): - if pretrain_weights not in seg_pretrain_weights_dict[ - self.model_name]: - logging.warning( - "Path of pretrain_weights('{}') does not exist!".format( - pretrain_weights)) - logging.warning("Pretrain_weights is forcibly set to '{}'. " - "If don't want to use pretrain weights, " - "set pretrain_weights to be None.".format( - seg_pretrain_weights_dict[self.model_name][ - 0])) - pretrain_weights = seg_pretrain_weights_dict[self.model_name][0] - elif pretrain_weights is not None and osp.exists(pretrain_weights): - if osp.splitext(pretrain_weights)[-1] != '.pdparams': - logging.error( - "Invalid pretrain weights. Please specify a '.pdparams' file.", - exit=True) + if pretrain_weights is not None: + if not osp.exists(pretrain_weights): + if self.model_name not in cd_pretrain_weights_dict: + logging.warning( + "Path of pretrained weights ('{}') does not exist!". + format(pretrain_weights)) + pretrain_weights = None + elif pretrain_weights not in cd_pretrain_weights_dict[ + self.model_name]: + logging.warning( + "Path of pretrained weights ('{}') does not exist!". + format(pretrain_weights)) + pretrain_weights = cd_pretrain_weights_dict[ + self.model_name][0] + logging.warning( + "`pretrain_weights` is forcibly set to '{}'. " + "If you don't want to use pretrained weights, " + "please set `pretrain_weights` to None.".format( + pretrain_weights)) + else: + if osp.splitext(pretrain_weights)[-1] != '.pdparams': + logging.error( + "Invalid pretrained weights. Please specify a .pdparams file.", + exit=True) pretrained_dir = osp.join(save_dir, 'pretrain') is_backbone_weights = pretrain_weights == 'IMAGENET' self.net_initialize( @@ -409,18 +416,18 @@ class BaseChangeDetector(BaseModel): key-value pairs: For binary change detection (number of classes == 2), the key-value pairs are like: - {"iou": `intersection over union for the change class`, - "f1": `F1 score for the change class`, - "oacc": `overall accuracy`, - "kappa": ` kappa coefficient`}. + {"iou": intersection over union for the change class, + "f1": F1 score for the change class, + "oacc": overall accuracy, + "kappa": kappa coefficient}. For multi-class change detection (number of classes > 2), the key-value pairs are like: - {"miou": `mean intersection over union`, - "category_iou": `category-wise mean intersection over union`, - "oacc": `overall accuracy`, - "category_acc": `category-wise accuracy`, - "kappa": ` kappa coefficient`, - "category_F1-score": `F1 score`}. + {"miou": mean intersection over union, + "category_iou": category-wise mean intersection over union, + "oacc": overall accuracy, + "category_acc": category-wise accuracy, + "kappa": kappa coefficient, + "category_F1-score": F1 score}. """ self._check_transforms(eval_dataset.transforms, 'eval') diff --git a/paddlers/tasks/classifier.py b/paddlers/tasks/classifier.py index 23ab154..83c20fb 100644 --- a/paddlers/tasks/classifier.py +++ b/paddlers/tasks/classifier.py @@ -246,7 +246,7 @@ class BaseClassifier(BaseModel): exit=True) if pretrain_weights is not None and resume_checkpoint is not None: logging.error( - "pretrain_weights and resume_checkpoint cannot be set simultaneously.", + "`pretrain_weights` and `resume_checkpoint` cannot be set simultaneously.", exit=True) self.labels = train_dataset.labels if self.losses is None: @@ -262,25 +262,32 @@ class BaseClassifier(BaseModel): else: self.optimizer = optimizer - if pretrain_weights is not None and not osp.exists(pretrain_weights): - if pretrain_weights not in cls_pretrain_weights_dict[ - self.model_name]: - logging.warning( - "Path of pretrain_weights('{}') does not exist!".format( - pretrain_weights)) - logging.warning("Pretrain_weights is forcibly set to '{}'. " - "If don't want to use pretrain weights, " - "set pretrain_weights to be None.".format( - cls_pretrain_weights_dict[self.model_name][ - 0])) - pretrain_weights = cls_pretrain_weights_dict[self.model_name][0] - elif pretrain_weights is not None and osp.exists(pretrain_weights): - if osp.splitext(pretrain_weights)[-1] != '.pdparams': - logging.error( - "Invalid pretrain weights. Please specify a '.pdparams' file.", - exit=True) + if pretrain_weights is not None: + if not osp.exists(pretrain_weights): + if self.model_name not in cls_pretrain_weights_dict: + logging.warning( + "Path of `pretrain_weights` ('{}') does not exist!". + format(pretrain_weights)) + pretrain_weights = None + elif pretrain_weights not in cls_pretrain_weights_dict[ + self.model_name]: + logging.warning( + "Path of `pretrain_weights` ('{}') does not exist!". + format(pretrain_weights)) + pretrain_weights = cls_pretrain_weights_dict[ + self.model_name][0] + logging.warning( + "`pretrain_weights` is forcibly set to '{}'. " + "If you don't want to use pretrained weights, " + "set `pretrain_weights` to None.".format( + pretrain_weights)) + else: + if osp.splitext(pretrain_weights)[-1] != '.pdparams': + logging.error( + "Invalid pretrained weights. Please specify a .pdparams file.", + exit=True) pretrained_dir = osp.join(save_dir, 'pretrain') - is_backbone_weights = False # pretrain_weights == 'IMAGENET' # TODO: this is backbone + is_backbone_weights = False self.net_initialize( pretrain_weights=pretrain_weights, save_dir=pretrained_dir, @@ -380,8 +387,8 @@ class BaseClassifier(BaseModel): Returns: If `return_details` is False, return collections.OrderedDict with key-value pairs: - {"top1": `acc of top1`, - "top5": `acc of top5`}. + {"top1": acc of top1, + "top5": acc of top5}. """ self._check_transforms(eval_dataset.transforms, 'eval') diff --git a/paddlers/tasks/object_detector.py b/paddlers/tasks/object_detector.py index 6531481..ca25213 100644 --- a/paddlers/tasks/object_detector.py +++ b/paddlers/tasks/object_detector.py @@ -274,7 +274,7 @@ class BaseDetector(BaseModel): exit=True) if pretrain_weights is not None and resume_checkpoint is not None: logging.error( - "pretrain_weights and resume_checkpoint cannot be set simultaneously.", + "`pretrain_weights` and `resume_checkpoint` cannot be set simultaneously.", exit=True) if train_dataset.__class__.__name__ == 'VOCDetDataset': train_dataset.data_fields = { @@ -323,23 +323,29 @@ class BaseDetector(BaseModel): self.optimizer = optimizer # Initiate weights - if pretrain_weights is not None and not osp.exists(pretrain_weights): - if pretrain_weights not in det_pretrain_weights_dict['_'.join( - [self.model_name, self.backbone_name])]: - logging.warning( - "Path of pretrain_weights('{}') does not exist!".format( - pretrain_weights)) - pretrain_weights = det_pretrain_weights_dict['_'.join( - [self.model_name, self.backbone_name])][0] - logging.warning("Pretrain_weights is forcibly set to '{}'. " - "If you don't want to use pretrain weights, " - "set pretrain_weights to be None.".format( - pretrain_weights)) - elif pretrain_weights is not None and osp.exists(pretrain_weights): - if osp.splitext(pretrain_weights)[-1] != '.pdparams': - logging.error( - "Invalid pretrain weights. Please specify a '.pdparams' file.", - exit=True) + if pretrain_weights is not None: + if not osp.exists(pretrain_weights): + key = '_'.join([self.model_name, self.backbone_name]) + if key not in det_pretrain_weights_dict: + logging.warning( + "Path of pretrained weights ('{}') does not exist!". + format(pretrain_weights)) + pretrain_weights = None + elif pretrain_weights not in det_pretrain_weights_dict[key]: + logging.warning( + "Path of pretrained weights ('{}') does not exist!". + format(pretrain_weights)) + pretrain_weights = det_pretrain_weights_dict[key][0] + logging.warning( + "`pretrain_weights` is forcibly set to '{}'. " + "If you don't want to use pretrained weights, " + "please set `pretrain_weights` to None.".format( + pretrain_weights)) + else: + if osp.splitext(pretrain_weights)[-1] != '.pdparams': + logging.error( + "Invalid pretrained weights. Please specify a .pdparams file.", + exit=True) pretrained_dir = osp.join(save_dir, 'pretrain') self.net_initialize( pretrain_weights=pretrain_weights, @@ -477,7 +483,7 @@ class BaseDetector(BaseModel): Returns: If `return_details` is False, return collections.OrderedDict with key-value pairs: - {"bbox_mmap":`mean average precision (0.50, 11point)`}. + {"bbox_mmap": mean average precision (0.50, 11point)}. """ if metric is None: diff --git a/paddlers/tasks/restorer.py b/paddlers/tasks/restorer.py index fe17f82..d9ce6ad 100644 --- a/paddlers/tasks/restorer.py +++ b/paddlers/tasks/restorer.py @@ -31,6 +31,7 @@ from paddlers.models import res_losses from paddlers.transforms import Resize, decode_image from paddlers.transforms.functions import calc_hr_shape from paddlers.utils import get_single_card_bs +from paddlers.utils.checkpoint import res_pretrain_weights_dict from .base import BaseModel from .utils.res_adapters import GANAdapter, OptimizerAdapter from .utils.infer_nets import InferResNet @@ -234,7 +235,7 @@ class BaseRestorer(BaseModel): exit=True) if pretrain_weights is not None and resume_checkpoint is not None: logging.error( - "pretrain_weights and resume_checkpoint cannot be set simultaneously.", + "`pretrain_weights` and `resume_checkpoint` cannot be set simultaneously.", exit=True) if self.losses is None: @@ -256,14 +257,30 @@ class BaseRestorer(BaseModel): else: self.optimizer = optimizer - if pretrain_weights is not None and not osp.exists(pretrain_weights): - logging.warning("Path of pretrain_weights('{}') does not exist!". - format(pretrain_weights)) - elif pretrain_weights is not None and osp.exists(pretrain_weights): - if osp.splitext(pretrain_weights)[-1] != '.pdparams': - logging.error( - "Invalid pretrain weights. Please specify a '.pdparams' file.", - exit=True) + if pretrain_weights is not None: + if not osp.exists(pretrain_weights): + if self.model_name not in res_pretrain_weights_dict: + logging.warning( + "Path of pretrained weights ('{}') does not exist!". + format(pretrain_weights)) + pretrain_weights = None + elif pretrain_weights not in res_pretrain_weights_dict[ + self.model_name]: + logging.warning( + "Path of pretrained weights ('{}') does not exist!". + format(pretrain_weights)) + pretrain_weights = res_pretrain_weights_dict[ + self.model_name][0] + logging.warning( + "`pretrain_weights` is forcibly set to '{}'. " + "If you don't want to use pretrained weights, " + "please set `pretrain_weights` to None.".format( + pretrain_weights)) + else: + if osp.splitext(pretrain_weights)[-1] != '.pdparams': + logging.error( + "Invalid pretrained weights. Please specify a .pdparams file.", + exit=True) pretrained_dir = osp.join(save_dir, 'pretrain') is_backbone_weights = pretrain_weights == 'IMAGENET' self.net_initialize( @@ -365,8 +382,8 @@ class BaseRestorer(BaseModel): Returns: If `return_details` is False, return collections.OrderedDict with key-value pairs: - {"psnr": `peak signal-to-noise ratio`, - "ssim": `structural similarity`}. + {"psnr": peak signal-to-noise ratio, + "ssim": structural similarity}. """ diff --git a/paddlers/tasks/segmenter.py b/paddlers/tasks/segmenter.py index b9c586f..83cbffa 100644 --- a/paddlers/tasks/segmenter.py +++ b/paddlers/tasks/segmenter.py @@ -267,7 +267,7 @@ class BaseSegmenter(BaseModel): exit=True) if pretrain_weights is not None and resume_checkpoint is not None: logging.error( - "pretrain_weights and resume_checkpoint cannot be set simultaneously.", + "`pretrain_weights` and `resume_checkpoint` cannot be set simultaneously.", exit=True) self.labels = train_dataset.labels if self.losses is None: @@ -281,23 +281,30 @@ class BaseSegmenter(BaseModel): else: self.optimizer = optimizer - if pretrain_weights is not None and not osp.exists(pretrain_weights): - if pretrain_weights not in seg_pretrain_weights_dict[ - self.model_name]: - logging.warning( - "Path of pretrain_weights('{}') does not exist!".format( - pretrain_weights)) - logging.warning("Pretrain_weights is forcibly set to '{}'. " - "If don't want to use pretrain weights, " - "set pretrain_weights to be None.".format( - seg_pretrain_weights_dict[self.model_name][ - 0])) - pretrain_weights = seg_pretrain_weights_dict[self.model_name][0] - elif pretrain_weights is not None and osp.exists(pretrain_weights): - if osp.splitext(pretrain_weights)[-1] != '.pdparams': - logging.error( - "Invalid pretrain weights. Please specify a '.pdparams' file.", - exit=True) + if pretrain_weights is not None: + if not osp.exists(pretrain_weights): + if self.model_name not in seg_pretrain_weights_dict: + logging.warning( + "Path of pretrained weights ('{}') does not exist!". + format(pretrain_weights)) + pretrain_weights = None + elif pretrain_weights not in seg_pretrain_weights_dict[ + self.model_name]: + logging.warning( + "Path of pretrained weights ('{}') does not exist!". + format(pretrain_weights)) + pretrain_weights = seg_pretrain_weights_dict[ + self.model_name][0] + logging.warning( + "`pretrain_weights` is forcibly set to '{}'. " + "If you don't want to use pretrained weights, " + "please set `pretrain_weights` to None.".format( + pretrain_weights)) + else: + if osp.splitext(pretrain_weights)[-1] != '.pdparams': + logging.error( + "Invalid pretrained weights. Please specify a .pdparams file.", + exit=True) pretrained_dir = osp.join(save_dir, 'pretrain') is_backbone_weights = pretrain_weights == 'IMAGENET' self.net_initialize( @@ -398,12 +405,12 @@ class BaseSegmenter(BaseModel): Returns: collections.OrderedDict with key-value pairs: - {"miou": `mean intersection over union`, - "category_iou": `category-wise mean intersection over union`, - "oacc": `overall accuracy`, - "category_acc": `category-wise accuracy`, - "kappa": ` kappa coefficient`, - "category_F1-score": `F1 score`}. + {"miou": mean intersection over union, + "category_iou": category-wise mean intersection over union, + "oacc": overall accuracy, + "category_acc": category-wise accuracy, + "kappa": kappa coefficient, + "category_F1-score": F1 score}. """ @@ -909,6 +916,7 @@ class BiSeNetV2(BaseSegmenter): class FarSeg(BaseSegmenter): def __init__(self, + in_channels=3, num_classes=2, use_mixed_loss=False, losses=None, @@ -918,4 +926,5 @@ class FarSeg(BaseSegmenter): num_classes=num_classes, use_mixed_loss=use_mixed_loss, losses=losses, + in_channels=in_channels, **params) diff --git a/paddlers/utils/checkpoint.py b/paddlers/utils/checkpoint.py index 82b5b12..029c4fc 100644 --- a/paddlers/utils/checkpoint.py +++ b/paddlers/utils/checkpoint.py @@ -21,20 +21,14 @@ import paddle from . import logging from .download import download_and_decompress +cd_pretrain_weights_dict = {} + cls_pretrain_weights_dict = { 'ResNet50_vd': ['IMAGENET'], 'MobileNetV3_small_x1_0': ['IMAGENET'], 'HRNet_W18_C': ['IMAGENET'], } -seg_pretrain_weights_dict = { - 'UNet': ['CITYSCAPES'], - 'DeepLabV3P': ['CITYSCAPES', 'PascalVOC', 'IMAGENET'], - 'FastSCNN': ['CITYSCAPES'], - 'HRNet': ['CITYSCAPES', 'PascalVOC'], - 'BiSeNetV2': ['CITYSCAPES'] -} - det_pretrain_weights_dict = { 'PicoDet_ESNet_s': ['COCO', 'IMAGENET'], 'PicoDet_ESNet_m': ['COCO', 'IMAGENET'], @@ -74,6 +68,16 @@ det_pretrain_weights_dict = { 'MaskRCNN_ResNet101_vd_fpn': ['COCO', 'IMAGENET'] } +res_pretrain_weights_dict = {} + +seg_pretrain_weights_dict = { + 'UNet': ['CITYSCAPES'], + 'DeepLabV3P': ['CITYSCAPES', 'PascalVOC', 'IMAGENET'], + 'FastSCNN': ['CITYSCAPES'], + 'HRNet': ['CITYSCAPES', 'PascalVOC'], + 'BiSeNetV2': ['CITYSCAPES'] +} + cityscapes_weights = { 'UNet_CITYSCAPES': 'https://bj.bcebos.com/paddleseg/dygraph/cityscapes/unet_cityscapes_1024x512_160k/model.pdparams', diff --git a/test_tipc/README.md b/test_tipc/README.md index 70fd203..934a83f 100644 --- a/test_tipc/README.md +++ b/test_tipc/README.md @@ -44,6 +44,7 @@ | 目标检测 | PP-YOLOv2 | 支持 | - | - | - | | 目标检测 | YOLOv3 | 支持 | - | - | - | | 图像分割 | DeepLab V3+ | 支持 | - | - | - | +| 图像分割 | FarSeg | 支持 | - | - | - | | 图像分割 | UNet | 支持 | - | - | - | ## 3 测试工具简介 diff --git a/test_tipc/configs/seg/farseg/farseg_rsseg.yaml b/test_tipc/configs/seg/farseg/farseg_rsseg.yaml new file mode 100644 index 0000000..fa6d97b --- /dev/null +++ b/test_tipc/configs/seg/farseg/farseg_rsseg.yaml @@ -0,0 +1,11 @@ +# Configurations of FarSeg with RSSeg dataset + +_base_: ../_base_/rsseg.yaml + +save_dir: ./test_tipc/output/seg/farseg/ + +model: !Node + type: FarSeg + args: + in_channels: 10 + num_classes: 5 \ No newline at end of file diff --git a/test_tipc/configs/seg/farseg/train_infer_python.txt b/test_tipc/configs/seg/farseg/train_infer_python.txt new file mode 100644 index 0000000..6619052 --- /dev/null +++ b/test_tipc/configs/seg/farseg/train_infer_python.txt @@ -0,0 +1,53 @@ +===========================train_params=========================== +model_name:seg:farseg +python:python +gpu_list:0|0,1 +use_gpu:null|null +--precision:null +--num_epochs:lite_train_lite_infer=3|lite_train_whole_infer=3|whole_train_whole_infer=20 +--save_dir:adaptive +--train_batch_size:lite_train_lite_infer=4|lite_train_whole_infer=4|whole_train_whole_infer=4 +--model_path:null +--config:lite_train_lite_infer=./test_tipc/configs/seg/farseg/farseg_rsseg.yaml|lite_train_whole_infer=./test_tipc/configs/seg/farseg/farseg_rsseg.yaml|whole_train_whole_infer=./test_tipc/configs/seg/farseg/farseg_rsseg.yaml +train_model_name:best_model +null:null +## +trainer:norm +norm_train:test_tipc/run_task.py train seg +pact_train:null +fpgm_train:null +distill_train:null +null:null +null:null +## +===========================eval_params=========================== +eval:null +null:null +## +===========================export_params=========================== +--save_dir:adaptive +--model_dir:adaptive +--fixed_input_shape:[-1,10,512,512] +norm_export:deploy/export/export_model.py +quant_export:null +fpgm_export:null +distill_export:null +export1:null +export2:null +===========================infer_params=========================== +infer_model:null +infer_export:null +infer_quant:False +inference:test_tipc/infer.py +--device:cpu|gpu +--enable_mkldnn:True +--cpu_threads:6 +--batch_size:1 +--use_trt:False +--precision:fp32 +--model_dir:null +--config:null +--save_log_path:null +--benchmark:True +--model_name:farseg +null:null \ No newline at end of file diff --git a/test_tipc/docs/test_train_inference_python.md b/test_tipc/docs/test_train_inference_python.md index 72a321b..b117cce 100644 --- a/test_tipc/docs/test_train_inference_python.md +++ b/test_tipc/docs/test_train_inference_python.md @@ -31,6 +31,7 @@ Linux GPU/CPU 基础训练推理测试的主程序为`test_train_inference_pytho | 目标检测 | PP-YOLOv2 | 正常训练 | 正常训练 | mAP=59.37% | | 目标检测 | YOLOv3 | 正常训练 | 正常训练 | mAP=47.33% | | 图像分割 | DeepLab V3+ | 正常训练 | 正常训练 | mIoU=56.05% | +| 图像分割 | FarSeg | 正常训练 | 正常训练 | mIoU=49.58% | | 图像分割 | UNet | 正常训练 | 正常训练 | mIoU=55.50% | *注：参考预测精度为whole_train_whole_infer模式下单卡训练汇报的精度数据。* @@ -61,6 +62,7 @@ Linux GPU/CPU 基础训练推理测试的主程序为`test_train_inference_pytho | 目标检测 | PP-YOLOv2 | 支持 | 支持 | 1 | | 目标检测 | YOLOv3 | 支持 | 支持 | 1 | | 图像分割 | DeepLab V3+ | 支持 | 支持 | 1 | +| 图像分割 | FarSeg | 支持 | 支持 | 1 | | 图像分割 | UNet | 支持 | 支持 | 1 | ## 2 测试流程 diff --git a/tests/deploy/test_predictor.py b/tests/deploy/test_predictor.py index 24db0ff..8675e65 100644 --- a/tests/deploy/test_predictor.py +++ b/tests/deploy/test_predictor.py @@ -105,7 +105,7 @@ class TestPredictor(CommonTest): dict_[key], expected_dict[key], rtol=1.e-4, atol=1.e-6) -# @TestPredictor.add_tests +@TestPredictor.add_tests class TestCDPredictor(TestPredictor): MODULE = pdrs.tasks.change_detector TRAINER_NAME_TO_EXPORT_OPTS = { @@ -177,7 +177,7 @@ class TestCDPredictor(TestPredictor): self.assertEqual(len(out_multi_array_t), num_inputs) -# @TestPredictor.add_tests +@TestPredictor.add_tests class TestClasPredictor(TestPredictor): MODULE = pdrs.tasks.classifier TRAINER_NAME_TO_EXPORT_OPTS = { @@ -242,7 +242,7 @@ class TestClasPredictor(TestPredictor): self.check_dict_equal(out_multi_array_p, out_multi_array_t) -# @TestPredictor.add_tests +@TestPredictor.add_tests class TestDetPredictor(TestPredictor): MODULE = pdrs.tasks.object_detector TRAINER_NAME_TO_EXPORT_OPTS = { @@ -355,7 +355,7 @@ class TestResPredictor(TestPredictor): self.assertEqual(len(out_multi_array_t), num_inputs) -# @TestPredictor.add_tests +@TestPredictor.add_tests class TestSegPredictor(TestPredictor): MODULE = pdrs.tasks.segmenter TRAINER_NAME_TO_EXPORT_OPTS = { diff --git a/tests/rs_models/test_cd_models.py b/tests/rs_models/test_cd_models.py index 3b3e683..6deed6b 100644 --- a/tests/rs_models/test_cd_models.py +++ b/tests/rs_models/test_cd_models.py @@ -21,7 +21,7 @@ __all__ = [ 'TestBITModel', 'TestCDNetModel', 'TestChangeStarModel', 'TestDSAMNetModel', 'TestDSIFNModel', 'TestFCEarlyFusionModel', 'TestFCSiamConcModel', 'TestFCSiamDiffModel', 'TestSNUNetModel', 'TestSTANetModel', - 'TestChangeFormerModel' + 'TestChangeFormerModel', 'TestFCCDNModel' ] @@ -32,8 +32,11 @@ class TestCDModel(TestModel): self.assertIsInstance(output, list) self.check_output_equal(len(output), len(target)) for o, t in zip(output, target): - o = o.numpy() - self.check_output_equal(o.shape, t.shape) + if isinstance(o, list): + self.check_output(o, t) + else: + o = o.numpy() + self.check_output_equal(o.shape, t.shape) def set_inputs(self): if self.EF_MODE == 'Concat': @@ -225,3 +228,27 @@ class TestChangeFormerModel(TestCDModel): dict(**base_spec, decoder_softmax=True), dict(**base_spec, embed_dim=56) ] # yapf: disable + + +class TestFCCDNModel(TestCDModel): + MODEL_CLASS = paddlers.rs_models.cd.FCCDN + + def set_specs(self): + self.specs = [ + dict(in_channels=3, num_classes=2), + dict(in_channels=8, num_classes=2), + dict(in_channels=3, num_classes=8), + dict(in_channels=3, num_classes=2, _phase='eval', _stop_grad=True) + ] # yapf: disable + + def set_targets(self): + b = self.DEFAULT_BATCH_SIZE + h = self.DEFAULT_HW[0] // 2 + w = self.DEFAULT_HW[1] // 2 + tar_c2 = [ + self.get_zeros_array(2), [self.get_zeros_array(1, b, h, w)] * 2 + ] + self.targets = [ + tar_c2, tar_c2, [self.get_zeros_array(8), tar_c2[1]], + [self.get_zeros_array(2)] + ] diff --git a/tests/rs_models/test_seg_models.py b/tests/rs_models/test_seg_models.py index 88fb6e1..156f311 100644 --- a/tests/rs_models/test_seg_models.py +++ b/tests/rs_models/test_seg_models.py @@ -25,8 +25,11 @@ class TestSegModel(TestModel): self.assertIsInstance(output, list) self.check_output_equal(len(output), len(target)) for o, t in zip(output, target): - o = o.numpy() - self.check_output_equal(o.shape, t.shape) + if isinstance(o, list): + self.check_output(o, t) + else: + o = o.numpy() + self.check_output_equal(o.shape, t.shape) def set_inputs(self): def _gen_data(specs): @@ -50,7 +53,8 @@ class TestFarSegModel(TestSegModel): def set_specs(self): self.specs = [ - dict(), dict(num_classes=20), dict(encoder_pretrained=False) + dict(), dict(num_classes=20), dict(pretrained_encoder=False), + dict(in_channels=10) ] def set_targets(self): diff --git a/tools/prepare_dataset/common.py b/tools/prepare_dataset/common.py index 09eec87..3848915 100644 --- a/tools/prepare_dataset/common.py +++ b/tools/prepare_dataset/common.py @@ -107,7 +107,7 @@ def crop_patches(crop_size, if max_workers < 0: raise ValueError("`max_workers` must be a non-negative integer!") - if subset is None: + if subsets is None: subsets = ('', ) if max_workers == 0: @@ -280,4 +280,4 @@ def random_split(samples, # Append remainder to the last split splits[-1].append(splits[ed_idx:]) - return splits \ No newline at end of file + return splits diff --git a/tutorials/train/README.md b/tutorials/train/README.md index 44e93a3..44c2491 100644 --- a/tutorials/train/README.md +++ b/tutorials/train/README.md @@ -27,6 +27,7 @@ |object_detection/ppyolov2.py | 目标检测 | PP-YOLOv2 | |object_detection/yolov3.py | 目标检测 | YOLOv3 | |semantic_segmentation/deeplabv3p.py | 图像分割 | DeepLab V3+ | +|semantic_segmentation/farseg.py | 图像分割 | FarSeg | |semantic_segmentation/unet.py | 图像分割 | UNet | ## 环境准备 diff --git a/tutorials/train/semantic_segmentation/deeplabv3p.py b/tutorials/train/semantic_segmentation/deeplabv3p.py index b3dbd50..4e0cb0b 100644 --- a/tutorials/train/semantic_segmentation/deeplabv3p.py +++ b/tutorials/train/semantic_segmentation/deeplabv3p.py @@ -71,7 +71,7 @@ eval_dataset = pdrs.datasets.SegDataset( # 目前已支持的模型请参考：https://github.com/PaddlePaddle/PaddleRS/blob/develop/docs/intro/model_zoo.md # 模型输入参数请参考：https://github.com/PaddlePaddle/PaddleRS/blob/develop/paddlers/tasks/segmenter.py model = pdrs.tasks.seg.DeepLabV3P( - input_channel=NUM_BANDS, + in_channels=NUM_BANDS, num_classes=len(train_dataset.labels), backbone='ResNet50_vd') diff --git a/tutorials/train/semantic_segmentation/farseg.py b/tutorials/train/semantic_segmentation/farseg.py new file mode 100644 index 0000000..f8561b5 --- /dev/null +++ b/tutorials/train/semantic_segmentation/farseg.py @@ -0,0 +1,94 @@ +#!/usr/bin/env python + +# 图像分割模型FarSeg训练示例脚本 +# 执行此脚本前，请确认已正确安装PaddleRS库 + +import paddlers as pdrs +from paddlers import transforms as T + +# 数据集存放目录 +DATA_DIR = './data/rsseg/' +# 训练集`file_list`文件路径 +TRAIN_FILE_LIST_PATH = './data/rsseg/train.txt' +# 验证集`file_list`文件路径 +EVAL_FILE_LIST_PATH = './data/rsseg/val.txt' +# 数据集类别信息文件路径 +LABEL_LIST_PATH = './data/rsseg/labels.txt' +# 实验目录，保存输出的模型权重和结果 +EXP_DIR = './output/farseg/' + +# 下载和解压多光谱地块分类数据集 +pdrs.utils.download_and_decompress( + 'https://paddlers.bj.bcebos.com/datasets/rsseg.zip', path='./data/') + +# 定义训练和验证时使用的数据变换（数据增强、预处理等） +# 使用Compose组合多种变换方式。Compose中包含的变换将按顺序串行执行 +# API说明：https://github.com/PaddlePaddle/PaddleRS/blob/develop/docs/apis/data.md +train_transforms = T.Compose([ + # 读取影像 + T.DecodeImg(), + # 选择前三个波段 + T.SelectBand([1, 2, 3]), + # 将影像缩放到512x512大小 + T.Resize(target_size=512), + # 以50%的概率实施随机水平翻转 + T.RandomHorizontalFlip(prob=0.5), + # 将数据归一化到[-1,1] + T.Normalize( + mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]), + T.ArrangeSegmenter('train') +]) + +eval_transforms = T.Compose([ + T.DecodeImg(), + # 验证阶段与训练阶段应当选择相同的波段 + T.SelectBand([1, 2, 3]), + T.Resize(target_size=512), + # 验证阶段与训练阶段的数据归一化方式必须相同 + T.Normalize( + mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]), + T.ReloadMask(), + T.ArrangeSegmenter('eval') +]) + +# 分别构建训练和验证所用的数据集 +train_dataset = pdrs.datasets.SegDataset( + data_dir=DATA_DIR, + file_list=TRAIN_FILE_LIST_PATH, + label_list=LABEL_LIST_PATH, + transforms=train_transforms, + num_workers=0, + shuffle=True) + +eval_dataset = pdrs.datasets.SegDataset( + data_dir=DATA_DIR, + file_list=EVAL_FILE_LIST_PATH, + label_list=LABEL_LIST_PATH, + transforms=eval_transforms, + num_workers=0, + shuffle=False) + +# 构建FarSeg模型 +# 目前已支持的模型请参考：https://github.com/PaddlePaddle/PaddleRS/blob/develop/docs/intro/model_zoo.md +# 模型输入参数请参考：https://github.com/PaddlePaddle/PaddleRS/blob/develop/paddlers/tasks/segmenter.py +model = pdrs.tasks.seg.FarSeg(num_classes=len(train_dataset.labels)) + +# 执行模型训练 +model.train( + num_epochs=10, + train_dataset=train_dataset, + train_batch_size=4, + eval_dataset=eval_dataset, + save_interval_epochs=5, + # 每多少次迭代记录一次日志 + log_interval_steps=4, + save_dir=EXP_DIR, + pretrain_weights=None, + # 初始学习率大小 + learning_rate=0.001, + # 是否使用early stopping策略，当精度不再改善时提前终止训练 + early_stop=False, + # 是否启用VisualDL日志功能 + use_vdl=True, + # 指定从某个检查点继续训练 + resume_checkpoint=None) diff --git a/tutorials/train/semantic_segmentation/unet.py b/tutorials/train/semantic_segmentation/unet.py index e1e8b82..1aee709 100644 --- a/tutorials/train/semantic_segmentation/unet.py +++ b/tutorials/train/semantic_segmentation/unet.py @@ -71,7 +71,7 @@ eval_dataset = pdrs.datasets.SegDataset( # 目前已支持的模型请参考：https://github.com/PaddlePaddle/PaddleRS/blob/develop/docs/intro/model_zoo.md # 模型输入参数请参考：https://github.com/PaddlePaddle/PaddleRS/blob/develop/paddlers/tasks/segmenter.py model = pdrs.tasks.seg.UNet( - input_channel=NUM_BANDS, num_classes=len(train_dataset.labels)) + in_channels=NUM_BANDS, num_classes=len(train_dataset.labels)) # 执行模型训练 model.train(