使用PaddleRS中的变化检测模型报错

[Programrookie33]

Thanks for your bug report. To help us better solve the issue, please provide the following information：

1. PaddleRS version: (please specify the branch as well，e.g. PaddleRS release/1.0)
2. PaddlePaddle version: (e.g. PaddlePaddle 2.3.0)
3. Operation system: (e.g. Linux/Windows/MacOS)
4. Python version: (e.g. Python3.7/8)
5. CUDA/cuDNN version: (e.g. CUDA10.2/cuDNN 7.6.5)
6. Full codes: (if you modify any original code，please show the comparison of the codes before and after)
7. Detailed error information and releated running log: (if you used multi-gpus，the log can be found in log/worklog.0 by default)
8. Steps to reproduce the problem：
9. Additional context: (add any other context about the problem)

---

欢迎您反馈PaddleRS使用问题。辛苦您提供以下信息，以方便我们快速定位和解决问题：

1. PaddleRS版本：不是很确定具体版本，别人项目里缓存的版本，日期应该是1.0正式版以后的版本
2. PaddlePaddle版本：2.4.0
3. 操作系统信息：Linux
4. Python版本号：Python3.7.4
5. CUDA/cuDNN版本：（ 如CUDA10.2/cuDNN 7.6.5等）
6. 完整的代码：

导入PaddleRS库

import paddlers as pdrs
import paddlers.transforms as T
import os.path as osp
import paddle

train_transforms = T.Compose([
# 读取数据
T.DecodeImg(),
# 以50%的概率实施随机水平翻转
#T.RandomHorizontalFlip(prob=0.2),
# 以50%的概率实施随机垂直翻转
#T.RandomVerticalFlip(prob=0.3),
# 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.RandomFlipOrRotate(
probs = [0.1, 0.1], # p=0.3 to flip the image，p=0.2 to rotate the image，p=0.5 to keep the image unchanged.
probsf = [0.1, 0.1, 0, 0, 0], # p=0.3 and p=0.25 to perform horizontal and vertical flipping; probility of no-flipping is 0.45.
probsr = [0, 0.2, 0]), # p=0.65 to rotate the image by 180°; probility of no-rotation is 0.35.
T.ReloadMask(),
# 挑选训练过程中需要用到的数据，并按照指定顺序排列
T.ArrangeChangeDetector('train')
])
eval_transforms = T.Compose([
T.DecodeImg(),
# 在验证阶段，输入原始尺寸影像，对输入影像仅进行归一化处理
# 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(),
# 挑选验证过程中需要用到的数据，并按照指定顺序排列
T.ArrangeChangeDetector('eval')

])

处理后数据集所在目录

DATA_DIR = "/home/aistudio/work/Landslide_CD/"

train_dataset = pdrs.datasets.CDDataset(
data_dir=DATA_DIR,
file_list=osp.join(DATA_DIR, 'train.txt'),
transforms=train_transforms,
label_list=None,
# 使用4个辅助进程加载数据
num_workers=4,
shuffle=True,
# 将取值为{0,255}的标签自动归一化到{0,1}
binarize_labels=True
)
val_dataset = pdrs.datasets.CDDataset(
data_dir=DATA_DIR,
file_list=osp.join(DATA_DIR, 'val.txt'),
transforms=eval_transforms,
label_list=None,
num_workers=0,
shuffle=False,
binarize_labels=True
)
#模型
model = pdrs.tasks.cd.FCCDN()
#学习率
lr_scheduler = paddle.optimizer.lr.StepDecay(
0.001,
step_size=5000,
# 学习率衰减系数，这里指定每次减半
gamma=0.5
)
#优化器
optimizer = paddle.optimizer.Lamb(
learning_rate=lr_scheduler,
parameters=model.net.parameters()
)

执行模型训练

model.train(
num_epochs=50,
train_dataset=train_dataset,
train_batch_size=20,
eval_dataset=val_dataset,
optimizer=optimizer,
save_interval_epochs=1,
# 每多少次迭代记录一次日志
log_interval_steps=10,
save_dir='/home/aistudio/work/output/fccdn',
#pretrain_weights='/home/aistudio/work/output/cdnet/best_model/model.pdparams',
# 是否使用early stopping策略，当精度不再改善时提前终止训练
early_stop=True,
# 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
)

8. 详细的错误信息与相关log：

---

ValueError Traceback (most recent call last)
/tmp/ipykernel_64988/2465525996.py in
93 use_vdl=True,
94 # 指定从某个检查点继续训练
---> 95 resume_checkpoint=None
96 )

~/work/PaddleRS/paddlers/tasks/change_detector.py in train(self, num_epochs, train_dataset, train_batch_size, eval_dataset, optimizer, save_interval_epochs, log_interval_steps, save_dir, pretrain_weights, learning_rate, lr_decay_power, early_stop, early_stop_patience, use_vdl, resume_checkpoint)
333 early_stop=early_stop,
334 early_stop_patience=early_stop_patience,
--> 335 use_vdl=use_vdl)
336
337 def quant_aware_train(self,

~/work/PaddleRS/paddlers/tasks/base.py in train_loop(self, num_epochs, train_dataset, train_batch_size, eval_dataset, save_interval_epochs, log_interval_steps, save_dir, ema, early_stop, early_stop_patience, use_vdl)
372 outputs = self.train_step(step, data, ddp_net)
373 else:
--> 374 outputs = self.train_step(step, data, self.net)
375
376 scheduler_step(self.optimizer, outputs['loss'])

~/work/PaddleRS/paddlers/tasks/base.py in train_step(self, step, data, net)
663
664 def train_step(self, step, data, net):
--> 665 outputs = self.run(net, data, mode='train')
666
667 loss = outputs['loss']

~/work/PaddleRS/paddlers/tasks/change_detector.py in run(self, net, inputs, mode)
113
114 def run(self, net, inputs, mode):
--> 115 net_out = net(inputs[0], inputs[1])
116 logit = net_out[0]
117 outputs = OrderedDict()

/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/dygraph/layers.py in call(self, *inputs, **kwargs)
928 return self.forward(*inputs, **kwargs)
929 else:
--> 930 return self._dygraph_call_func(*inputs, **kwargs)
931
932 def forward(self, *inputs, **kwargs):

/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/dygraph/layers.py in _dygraph_call_func(self, *inputs, **kwargs)
913 outputs = self.forward(*inputs, **kwargs)
914 else:
--> 915 outputs = self.forward(*inputs, **kwargs)
916
917 for forward_post_hook in self._forward_post_hooks.values():

~/work/PaddleRS/paddlers/rs_models/cd/fccdn.py in forward(self, t1, t2)
453 y2 = self.block4(e3_2)
454
--> 455 y1 = self.center(y1)
456 y2 = self.center(y2)
457 c = self.df4(y1, y2)

/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/dygraph/layers.py in call(self, *inputs, **kwargs)
928 return self.forward(*inputs, **kwargs)
929 else:
--> 930 return self._dygraph_call_func(*inputs, **kwargs)
931
932 def forward(self, *inputs, **kwargs):

/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/dygraph/layers.py in _dygraph_call_func(self, *inputs, **kwargs)
913 outputs = self.forward(*inputs, **kwargs)
914 else:
--> 915 outputs = self.forward(*inputs, **kwargs)
916
917 for forward_post_hook in self._forward_post_hooks.values():

~/work/PaddleRS/paddlers/rs_models/cd/fccdn.py in forward(self, x)
143 nl = self.nl3(d3)
144 d3 = self.upsample_x2(paddle.multiply(d3, nl)) ##2C,H/2,W/2
--> 145 d2 = self.conv_d2(e2 + d3) # C,H/2,W/2
146 nl = self.nl2(d2)
147 d2 = self.upsample_x2(paddle.multiply(d2, nl)) # C,H,W

/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/dygraph/math_op_patch.py in impl(self, other_var)
297 else:
298 math_op = getattr(_C_ops, op_type)
--> 299 return math_op(self, other_var, 'axis', axis)
300
301 comment = OpProtoHolder.instance().get_op_proto(op_type).comment

ValueError: (InvalidArgument) Broadcast dimension mismatch. Operands could not be broadcast together with the shape of X = [20, 128, 7, 7] and the shape of Y = [20, 128, 6, 6]. Received [7] in X is not equal to [6] in Y at i:2.
[Hint: Expected x_dims_array[i] == y_dims_array[i] || x_dims_array[i] <= 1 || y_dims_array[i] <= 1 == true, but received x_dims_array[i] == y_dims_array[i] || x_dims_array[i] <= 1 || y_dims_array[i] <= 1:0 != true:1.] (at /paddle/paddle/phi/kernels/funcs/common_shape.h:84)
[operator < elementwise_add > error]
9. 问题复现步骤：使用官方提供的训练脚本
11. 其他内容: 使用changestar、FCCDN、DSIFN模型均出现类似问题，报错信息几乎一致

[Bobholamovic]

看起来是一个尺寸不匹配的问题，建议在训练阶段加入随机裁剪，在验证阶段加入resize，以保证输入图像具有固定的尺寸。受到U-Net型结构上下采样时的舍入误差影响，输入网络的图像长、宽最好是4或者8的整数倍（与具体的模型有关）。

[Programrookie33]

是的，我已经在代码中加入了随机裁剪和resize，但是还是报了相同的错误

[Bobholamovic]

请问现在的输入尺寸是多少呀？

[Programrookie33]

224*224

[Bobholamovic]

请问是通过数据预处理算子实现的吗？还是预先裁剪好图片呀？如果是前者的话，请贴一下代码～

[Programrookie33]

train_transforms = T.Compose([
# 读取数据
T.DecodeImg(),
# 以50%的概率实施随机水平翻转
#T.RandomHorizontalFlip(prob=0.2),
# 以50%的概率实施随机垂直翻转
#T.RandomVerticalFlip(prob=0.3),
# 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.RandomCrop(
# 裁剪区域将被缩放到224x224
crop_size=224,
# 裁剪区域的横纵比在0.5-2之间变动
aspect_ratio=[0.5, 2.0],
# 裁剪区域相对原始影像长宽比例在一定范围内变动，最小不低于原始长宽的1/5
scaling=[0.2, 1.0]),
T.RandomFlipOrRotate(
probs = [0.1, 0.1], # p=0.3 to flip the image，p=0.2 to rotate the image，p=0.5 to keep the image unchanged.
probsf = [0.1, 0.1, 0, 0, 0], # p=0.3 and p=0.25 to perform horizontal and vertical flipping; probility of no-flipping is 0.45.
probsr = [0, 0.2, 0]), # p=0.65 to rotate the image by 180°; probility of no-rotation is 0.35.
T.ReloadMask(),
# 挑选训练过程中需要用到的数据，并按照指定顺序排列
T.ArrangeChangeDetector('train')
])
eval_transforms = T.Compose([
T.DecodeImg(),
# 在验证阶段，输入原始尺寸影像，对输入影像仅进行归一化处理
# 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.Resize(224),
T.ReloadMask(),
# 挑选验证过程中需要用到的数据，并按照指定顺序排列
T.ArrangeChangeDetector('eval')
])
图片预先裁剪好的，但是代码中也添加了随机裁剪

[Bobholamovic]

有点儿奇怪……建议在~/work/PaddleRS/paddlers/rs_models/cd/fccdn.py的第135行的forward方法中打印x、e1、e2、e3、d3等张量的形状（例如print(x.shape)），我们看看是否符合预期～

[Programrookie33]

fccdn模型是这样的
x.shape=[20, 64, 14, 14]
e1.shape=[20, 64, 14, 14]
e2.shape=[20, 128, 7, 7]
e3.shape=[20, 256, 3, 3]
d3.shape=[20, 128, 3, 3]

[Bobholamovic]

我阅读了一下源码，FCCDN这个模型可能需要输入尺寸是64的整数倍（因为encoder下采样16倍，而NLFPN又会下采样4倍）。建议尝试将输入尺寸调整为192或256～

[Programrookie33]

好的，可以运行了！非常感谢您的耐心回答！