Models: object detection export inference graph error when RPN_only

[YuxianMeng]

When I set first_stage_only = true in my config file, it trains well. However, when I try to use export_inference_graph.py to export the trained model for inference, there is a bug:

Traceback (most recent call last):
File "export_inference_graph.py", line 101, in
tf.app.run()
File "/home/linhb/anaconda2/envs/tensorflow/lib/python2.7/site-packages/tensorflow/python/platform/app.py", line 48, in run
_sys.exit(main(_sys.argv[:1] + flags_passthrough))
File "export_inference_graph.py", line 97, in main
FLAGS.export_as_saved_model)
File "/home/linhb/horse/models/object_detection/exporter.py", line 339, in export_inference_graph
export_as_saved_model)
File "/home/linhb/horse/models/object_detection/exporter.py", line 310, in _export_inference_graph
outputs = _add_output_tensor_nodes(postprocessed_tensors)
File "/home/linhb/horse/models/object_detection/exporter.py", line 184, in _add_output_tensor_nodes
classes = postprocessed_tensors.get('detection_classes') + label_id_offset
TypeError: unsupported operand type(s) for +: 'NoneType' and 'int'

That's due to the output detection_classes is None when using first_stage_only. I change the code in exporter.py, line184, from,
classes = postprocessed_tensors.get('detection_classes')+label_id_offset
to,

classes = postprocessed_tensors.get('detection_classes') 
if classes: 
    classes += label_id_offset 
else:
    classes = tf.constant(1, tf.int32, scores.shape)

then the program can run correctly.

[PierreBoyeau]

Hey guys,

Any news on this subject? Personnaly, I also face issues when I try to run eval.py.
I get the following error

Traceback (most recent call last):
  File "object_detection/eval.py", line 161, in <module>
    tf.app.run()
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/platform/app.py", line 48, in run
    _sys.exit(main(_sys.argv[:1] + flags_passthrough))
  File "object_detection/eval.py", line 157, in main
    FLAGS.checkpoint_dir, FLAGS.eval_dir)
  File "/home/ubuntu/models-tf/object_detection/evaluator.py", line 132, in evaluate
    ignore_groundtruth=eval_config.ignore_groundtruth)
  File "/home/ubuntu/models-tf/object_detection/evaluator.py", line 70, in _extract_prediction_tensors
    tf.squeeze(detections['detection_classes'], axis=0) +
KeyError: 'detection_classes'
Traceback (most recent call last):
  File "object_detection/eval.py", line 161, in <module>
    tf.app.run()
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/platform/app.py", line 48, in run
    _sys.exit(main(_sys.argv[:1] + flags_passthrough))
  File "object_detection/eval.py", line 157, in main
    FLAGS.checkpoint_dir, FLAGS.eval_dir)
  File "/home/ubuntu/models-tf/object_detection/evaluator.py", line 132, in evaluate
    ignore_groundtruth=eval_config.ignore_groundtruth)
  File "/home/ubuntu/models-tf/object_detection/evaluator.py", line 70, in _extract_prediction_tensors
    tf.squeeze(detections['detection_classes'], axis=0) +
KeyError: 'detection_classes'

I made it work with the same tweak @shzygmyx used.

[slsd123]

Thanks for the above fix. I had to make one small change because I couldn't use a tensor as a python boolean. Here's what I replaced line 184 with (in case anyone else encounters this issue):

  if classes is not None:
    classes += label_id_offset
  else:
    classes = tf.constant(1, tf.int32, scores.shape)

[Abhijit-2592]

@slsd123 @tombstone @shzygmyx after your fix I am still running into a new error

File "export_inference_graph.py", line 119, in
tf.app.run()
File "/opt/conda/lib/python3.5/site-packages/tensorflow/python/platform/app.py", line 48, in run
_sys.exit(main(_sys.argv[:1] + flags_passthrough))
File "export_inference_graph.py", line 115, in main FLAGS.output_directory, input_shape) File "/code/dent_obj_detection/TF/object_detection/exporter.py", line 434, in export_inference_graph input_shape, optimize_graph, output_collection_name)
File "/code/dent_obj_detection/TF/object_detection/exporter.py", line 362, in _export_inference_graph
output_collection_name)
File "/code/dent_obj_detection/TF/object_detection/exporter.py", line 240, in _add_output_tensor_nodes
classes = tf.constant(1, tf.int32, scores.shape)
File "/opt/conda/lib/python3.5/site-packages/tensorflow/python/framework/constant_op.py", line 208, in constant
value, dtype=dtype, shape=shape, verify_shape=verify_shape))
File "/opt/conda/lib/python3.5/site-packages/tensorflow/python/framework/tensor_util.py", line 380, in make_tensor_proto
if shape is not None and np.prod(shape, dtype=np.int64) == 0:
File "/opt/conda/lib/python3.5/site-packages/numpy/core/fromnumeric.py", line 2518, in prod
out=out, **kwargs)
File "/opt/conda/lib/python3.5/site-packages/numpy/core/_methods.py", line 35, in _prod
return umr_prod(a, axis, dtype, out, keepdims)
TypeError: __int__ returned non-int (type NoneType)

Any idea how to fix this? The error I think is related to the tensorflow version.

Update:
The above error is fixed by providing values for the input_shape argument in export_inference_graph.py

[skye]

Is this still an issue or can I close?

[Abhijit-2592]

@skye yes, the issue still persists! Check my update on the previous comment. Even after the fix mentioned above, it works only if we give arguments for input_shape

[dsuess]

The problem is that your constant needs to have known shape at creation time. The way around it is to dynamically reshape it as a tensor:

  if classes:
    classes += label_id_offset
  else:
    one = tf.constant(1, dtype=tf.int32, shape=(1, 1), name='classes_dummy')
    classes = tf.tile(one, tf.shape(scores))

I am going to submit a pull request if that works for you as well.

[Harshini-Gadige]

Automatically closing due to lack of recent activity. Please update the issue when new information becomes available, and we will reopen the issue. Thanks!

[danyfang]

Thanks for the above fix. I had to make one small change because I couldn't use a tensor as a python boolean. Here's what I replaced line 184 with (in case anyone else encounters this issue):

  if classes is not None:
    classes += label_id_offset
  else:
    classes = tf.constant(1, tf.int32, scores.shape)

This fix is the correct one, if you use if classes you will encounter problems when you export the whole detection graph.