TensorFlow bug（六）——Dimensions of inputs should match: shape[0] = [1,600,750,3] vs. shape[2] = [1,600

青旅半醒 2021-12-11 01:53 298阅读 0赞

# 问题描述： #

在用R-FCN目标检测API训练我自己的数据时， 我对原代码做了些改动，原来的rfcn\_resnet101\_coco.config文件中，**batch\_size**的值为1即一次处理1张图片，我改成了别的数（2、8、16、32...）都会出现如下的问题：

> InvalidArgumentError (see above for traceback): ConcatOp : Dimensions of inputs should match: shape\[0\] = \[1,600,750,3\] vs. shape\[2\] = \[1,600,692,3\]  
>      \[\[node concat (defined at /home/ubuntu/luanbo/VOCMaker/VOCMaker/object\_detection/legacy/trainer.py:190)  = ConcatV2\[N=64, T=DT\_FLOAT, Tidx=DT\_INT32, \_device="/job:localhost/replica:0/task:0/device:GPU:0"\](Preprocessor/sub, Preprocessor\_1/sub, Preprocessor\_2/sub, Preprocessor\_3/sub, Preprocessor\_4/sub, Preprocessor\_5/sub, Preprocessor\_6/sub, Preprocessor\_7/sub, Preprocessor\_8/sub, Preprocessor\_9/sub, Preprocessor\_10/sub, Preprocessor\_11/sub, Preprocessor\_12/sub, Preprocessor\_13/sub, Preprocessor\_14/sub, Preprocessor\_15/sub, Preprocessor\_16/sub, Preprocessor\_17/sub, Preprocessor\_18/sub, Preprocessor\_19/sub, Preprocessor\_20/sub, Preprocessor\_21/sub, Preprocessor\_22/sub, Preprocessor\_23/sub, Preprocessor\_24/sub, Preprocessor\_25/sub, Preprocessor\_26/sub, Preprocessor\_27/sub, Preprocessor\_28/sub, Preprocessor\_29/sub, Preprocessor\_30/sub, Preprocessor\_31/sub, Preprocessor\_32/sub, Preprocessor\_33/sub, Preprocessor\_34/sub, Preprocessor\_35/sub, Preprocessor\_36/sub, Preprocessor\_37/sub, Preprocessor\_38/sub, Preprocessor\_39/sub, Preprocessor\_40/sub, Preprocessor\_41/sub, Preprocessor\_42/sub, Preprocessor\_43/sub, Preprocessor\_44/sub, Preprocessor\_45/sub, Preprocessor\_46/sub, Preprocessor\_47/sub, Preprocessor\_48/sub, Preprocessor\_49/sub, Preprocessor\_50/sub, Preprocessor\_51/sub, Preprocessor\_52/sub, Preprocessor\_53/sub, Preprocessor\_54/sub, Preprocessor\_55/sub, Preprocessor\_56/sub, Preprocessor\_57/sub, Preprocessor\_58/sub, Preprocessor\_59/sub, Preprocessor\_60/sub, Preprocessor\_61/sub, Preprocessor\_62/sub, Preprocessor\_63/sub, Loss/RPNLoss/Loss/huber\_loss/assert\_broadcastable/is\_valid\_shape/has\_valid\_nonscalar\_shape/has\_invalid\_dims/x)\]\]  
>      \[\[\{ \{node assert\_equal\_3/y/\_3989\}\} = \_Recv\[client\_terminated=false, recv\_device="/job:localhost/replica:0/task:0/device:CPU:0", send\_device="/job:localhost/replica:0/task:0/device:GPU:0", send\_device\_incarnation=1, tensor\_name="edge\_23268\_assert\_equal\_3/y", tensor\_type=DT\_INT32, \_device="/job:localhost/replica:0/task:0/device:CPU:0"\]()\]\]

# 解决办法：  #

>       将batch\_size**改回了1就可以了**。

# 原因分析： #

怀疑是train.py文件中的'num\_clones'和'worker\_replicas'这两个参数的问题，但我将它们改成跟batch\_size相同大小后，又出了别的问题（这里忘记截图了，后面有机会补上），改回原来的1就好了。不解，暂时先这样，后面搞明白再回来填坑。

这里有几个类似问题的解释，但不知道跟我这个问题是否相关：

[关于bidirectional\_dynamic\_rnn出现 Dimensions of inputs should match问题][bidirectional_dynamic_rnn_ Dimensions of inputs should match]

[InvalidArgumentError (see above for traceback): Assign requires shapes of both tensors to match][InvalidArgumentError _see above for traceback_ Assign requires shapes of both tensors to match]

[bidirectional_dynamic_rnn_ Dimensions of inputs should match]: https://blog.csdn.net/junjie20082008/article/details/85762266
[InvalidArgumentError _see above for traceback_ Assign requires shapes of both tensors to match]: https://blog.csdn.net/yangfengling1023/article/details/81315612