s2anet问题：

[jackie8310]

用release2.2中s2anet_1x_spine.yml训练spine_coco数据集，配置文件如下：
BASE: [
'../datasets/spine_coco.yml',
'../runtime.yml',
'base/s2anet_optimizer_1x.yml',
'base/s2anet.yml',
'base/s2anet_reader.yml',
]

weights: output/s2anet_1x_spine/model_final

for 4 card

LearningRate:
base_lr: 0.00125

S2ANetHead:
anchor_strides: [8, 16, 32, 64, 128]
anchor_scales: [4]
anchor_ratios: [4.0]
anchor_assign: RBoxAssigner
stacked_convs: 2
feat_in: 256
feat_out: 256
num_classes: 9
align_conv_type: 'AlignConv' # AlignConv Conv
align_conv_size: 3
use_sigmoid_cls: True
reg_loss_weight: [1.0, 1.0, 1.0, 1.0, 1.05]
cls_loss_weight: [1.05, 1.0]
reg_loss_type: 'l1'

训练12epoch后：mAP(0.50, 11point) = 9.45%

[08/03 10:48:35] ppdet.engine INFO: Epoch: [11] [200/230] learning_rate: 0.000013 fam_cls_loss: 0.093969 fam_reg_loss: 1.341817 odm_cls_loss: 0.100388 odm_reg_loss: 0.080870 loss: 1.627664 eta: 0:00:15 batch_cost: 0.5041 data_cost: 0.0002 ips: 1.9838 images/s
[08/03 10:48:53] ppdet.utils.checkpoint INFO: Save checkpoint: output/s2anet_1x_spine
[08/03 10:48:54] ppdet.engine INFO: Eval iter: 0
[08/03 10:49:36] ppdet.metrics.metrics INFO: The bbox result is saved to bbox.json.
[08/03 10:49:36] ppdet.metrics.metrics INFO: Accumulating evaluatation results...
[08/03 10:49:36] ppdet.metrics.metrics INFO: mAP(0.50, 11point) = 9.45%
[08/03 10:49:36] ppdet.engine INFO: Total sample number: 57, averge FPS: 1.3343229141074622
[08/03 10:49:36] ppdet.engine INFO: Best test bbox ap is 0.094.

训练自己的数据集（通过roLabelImg 来标注旋转矩形框，并转成coco），训练24epoch后：mAP(0.50, 11point) = 3.1%。
问题：1、首先两个数据集没有问题，为何mAP很低？
2、 其中anchor_scales: [4]和 anchor_ratios: [1.0] 与anchor生成相关，如果自己数据集anchor分布很分散可能多个尺寸更好回归，当anchor_ratios设置多种比例时就会报错（如下），只能固定一种吗？该如何调整以提高mAP？
Traceback (most recent call last):
File "tools/train.py", line 139, in
main()
File "tools/train.py", line 135, in main
run(FLAGS, cfg)
File "tools/train.py", line 110, in run
trainer.train(FLAGS.eval)
File "/home/aistudio/PaddleDetection-release-2.2/ppdet/engine/trainer.py", line 357, in train
outputs = model(data)
File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/dygraph/layers.py", line 898, in call
outputs = self.forward(*inputs, **kwargs)
File "/home/aistudio/PaddleDetection-release-2.2/ppdet/modeling/architectures/meta_arch.py", line 26, in forward
out = self.get_loss()
File "/home/aistudio/PaddleDetection-release-2.2/ppdet/modeling/architectures/s2anet.py", line 97, in get_loss
loss = self._forward()
File "/home/aistudio/PaddleDetection-release-2.2/ppdet/modeling/architectures/s2anet.py", line 73, in _forward
self.s2anet_head(body_feats)
File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/dygraph/layers.py", line 898, in call
outputs = self.forward(*inputs, **kwargs)
File "/home/aistudio/PaddleDetection-release-2.2/ppdet/modeling/heads/s2anet_head.py", line 441, in forward
featmap_size, self.anchor_strides[feat_idx])
File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/dygraph/layers.py", line 898, in call
outputs = self.forward(*inputs, **kwargs)
File "/home/aistudio/PaddleDetection-release-2.2/ppdet/modeling/heads/s2anet_head.py", line 90, in forward
all_anchors = self.base_anchors[:, :] + shifts[:, :]
File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/dygraph/math_op_patch.py", line 250, in impl
return math_op(self, other_var, 'axis', axis)
ValueError: (InvalidArgument) Broadcast dimension mismatch. Operands could not be broadcast together with the shape of X = [3, 4] and the shape of Y = [13824, 4]. Received [3] in X is not equal to [13824] in Y at i:0.
[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/fluid/operators/elementwise/elementwise_op_function.h:169)
[operator < elementwise_add > error]

[lyuwenyu]

其中anchor_scales: [4]和 anchor_ratios: [1.0]
有尝试其他的比例嘛 也会出现这种错误？

[lyuwenyu]

建议用8卡的配置文件，4卡的话得修改optimizer的超参

[jackie8310]

@lyuwenyu ,用了最新的develop分支，单卡train，LR：0.00125，epoch:36，LOSS有下降，但mAP(0.50, 11point) = 4.72%
08/04 18:49:33] ppdet.engine INFO: Eval iter: 200
[08/04 18:49:45] ppdet.metrics.metrics INFO: The bbox result is saved to bbox.json.
[08/04 18:49:45] ppdet.metrics.metrics INFO: Accumulating evaluatation results...
[08/04 18:49:45] ppdet.metrics.metrics INFO: mAP(0.50, 11point) = 4.72%
修改ratios，如下：
anchor_strides: [8, 16, 32, 64, 128]
anchor_scales: [4]
anchor_ratios: [1.0, 16.0]
报错：
Traceback (most recent call last):
File "tools/train.py", line 139, in
main()
File "tools/train.py", line 135, in main
run(FLAGS, cfg)
File "tools/train.py", line 110, in run
trainer.train(FLAGS.eval)
File "/home/aistudio/PaddleDetection-develop/ppdet/engine/trainer.py", line 360, in train
outputs = model(data)
File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/dygraph/layers.py", line 898, in call
outputs = self.forward(*inputs, **kwargs)
File "/home/aistudio/PaddleDetection-develop/ppdet/modeling/architectures/meta_arch.py", line 26, in forward
out = self.get_loss()
File "/home/aistudio/PaddleDetection-develop/ppdet/modeling/architectures/s2anet.py", line 97, in get_loss
loss = self._forward()
File "/home/aistudio/PaddleDetection-develop/ppdet/modeling/architectures/s2anet.py", line 73, in _forward
self.s2anet_head(body_feats)
File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/dygraph/layers.py", line 898, in call
outputs = self.forward(*inputs, **kwargs)
File "/home/aistudio/PaddleDetection-develop/ppdet/modeling/heads/s2anet_head.py", line 441, in forward
featmap_size, self.anchor_strides[feat_idx])
File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/dygraph/layers.py", line 898, in call
outputs = self.forward(*inputs, **kwargs)
File "/home/aistudio/PaddleDetection-develop/ppdet/modeling/heads/s2anet_head.py", line 90, in forward
all_anchors = self.base_anchors[:, :] + shifts[:, :]
File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/dygraph/math_op_patch.py", line 250, in impl
return math_op(self, other_var, 'axis', axis)
ValueError: (InvalidArgument) Broadcast dimension mismatch. Operands could not be broadcast together with the shape of X = [2, 4] and the shape of Y = [16384, 4]. Received [2] in X is not equal to [16384] in Y at i:0.
[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/fluid/operators/elementwise/elementwise_op_function.h:169)
[operator < elementwise_add > error]

结论：只能设置单一ratios, 问题：数据集anchor分布不均衡，可能多个ratios更好拟合以达到更高mAP, 该如何设置呢？

[yumeng88]

我遇到和你相同的问题，单卡训练，边训练边评估，训练的mAP也只有3.5左右，如果是训练完成在评估(我测试了好几个数据集，其中有一个的结果可以达到4.1，其余的全是0)，目前还没有解决这个问题。

[lyuwenyu]

用develop试一下吧

[jackie8310]

@lyuwenyu , 是用的最新develop分支，单卡train，LR：0.00125，epoch:36，LOSS有下降，推荐spine数据集和自己的数据集mAP(0.50, 11point) = 4.72%，这是最高的情况。