问题
I am trying to train a model, at first I had dataset of 5000 images and training worked fine, Now I have added couple of more images and now my dataset contains 6,423 images. I am using python 3.6.1 on Ubuntu 18.04, my tensorflow version is 1.15 & numpy version is 1.16 (had same versions before and it worked fine). Now when I use:
python model_main.py --logtostderr --pipeline_config_path=training/faster_rcnn_resnet50_coco.config --model_dir=training
It starts settings up for couple of minutes and after these lines:
INFO:tensorflow:Saving checkpoints for 0 into training/model.ckpt.
I1123 10:26:21.548237 140482563244160 basic_session_run_hooks.py:606] Saving checkpoints for 0 into training/model.ckpt.
2019-11-23 10:28:30.801453: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcublas.so.10.0
I get following erros:
2019-11-23 10:08:38.843259: W tensorflow/core/framework/op_kernel.cc:1651] OP_REQUIRES failed at lookup_table_op.cc:788 : Not found: Resource localhost/_3_hash_table_2/N10tensorflow6lookup15LookupInterfaceE does not exist.
2019-11-23 10:08:38.843323: W tensorflow/core/framework/op_kernel.cc:1651] OP_REQUIRES failed at lookup_table_op.cc:788 : Not found: Resource localhost/_1_hash_table_1/N10tensorflow6lookup15LookupInterfaceE does not exist.
2019-11-23 10:08:38.843345: W tensorflow/core/framework/op_kernel.cc:1651] OP_REQUIRES failed at lookup_table_op.cc:788 : Not found: Resource localhost/_2_hash_table/N10tensorflow6lookup15LookupInterfaceE does not exist.
2019-11-23 10:08:38.851405: W tensorflow/core/framework/op_kernel.cc:1651] OP_REQUIRES failed at lookup_table_op.cc:788 : Not found: Resource localhost/_3_hash_table_2/N10tensorflow6lookup15LookupInterfaceE does not exist.
2019-11-23 10:08:38.851488: W tensorflow/core/framework/op_kernel.cc:1651] OP_REQUIRES failed at lookup_table_op.cc:788 : Not found: Resource localhost/_1_hash_table_1/N10tensorflow6lookup15LookupInterfaceE does not exist.
2019-11-23 10:08:38.851512: W tensorflow/core/framework/op_kernel.cc:1651] OP_REQUIRES failed at lookup_table_op.cc:788 : Not found: Resource localhost/_2_hash_table/N10tensorflow6lookup15LookupInterfaceE does not exist.
2019-11-23 10:08:38.851807: W tensorflow/core/framework/op_kernel.cc:1651] OP_REQUIRES failed at lookup_table_op.cc:788 : Not found: Resource localhost/_1_hash_table_1/N10tensorflow6lookup15LookupInterfaceE does not exist.
2019-11-23 10:08:38.851848: W tensorflow/core/framework/op_kernel.cc:1651] OP_REQUIRES failed at lookup_table_op.cc:788 : Not found: Resource localhost/_2_hash_table/N10tensorflow6lookup15LookupInterfaceE does not exist.
2019-11-23 10:08:38.851899: W tensorflow/core/framework/op_kernel.cc:1651] OP_REQUIRES failed at lookup_table_op.cc:788 : Not found: Resource localhost/_3_hash_table_2/N10tensorflow6lookup15LookupInterfaceE does not exist.
Traceback (most recent call last):
File "/usr/local/lib/python3.6/site-packages/tensorflow_core/python/client/session.py", line 1365, in _do_call
return fn(*args)
File "/usr/local/lib/python3.6/site-packages/tensorflow_core/python/client/session.py", line 1350, in _run_fn
target_list, run_metadata)
File "/usr/local/lib/python3.6/site-packages/tensorflow_core/python/client/session.py", line 1443, in _call_tf_sessionrun
run_metadata)
tensorflow.python.framework.errors_impl.InvalidArgumentError: 2 root error(s) found.
(0) Invalid argument: Cannot add tensor to the batch: number of elements does not match. Shapes are: [tensor]: [585,1024,3], [batch]: [600,799,3]
[[{{node IteratorGetNext}}]]
[[ToAbsoluteCoordinates_118/Assert/AssertGuard/Assert/data_0/_5709]]
(1) Invalid argument: Cannot add tensor to the batch: number of elements does not match. Shapes are: [tensor]: [585,1024,3], [batch]: [600,799,3]
[[{{node IteratorGetNext}}]]
0 successful operations.
0 derived errors ignored.
and training stops.
回答1:
It seems that the new images you've added have a resolution of 585x1024, which differs from the size that's expected by the model i.e. 600x799.
If so, then the solution is to resize these new images accordingly.
回答2:
Changing the batch_size to 1 fixed this issue for me.
回答3:
If you need batch size > 1, you can resize the images to a uniform size with the right image_resizer in the config, one of the ones defined in the image_resizer protobuf file, which I assume is what is used to parse that part of the config.
For example (stolen from here):
image_resizer {
fixed_shape_resizer {
height: 600
width: 800
}
}
This seems to fix the problem for me.
来源：https://stackoverflow.com/questions/59006696/cannot-add-tensor-to-the-batch-number-of-elements-does-not-match-shapes-are