TFJS-converter | Cannot convert saved_model to tfjs_graph_model

Environment

Python version

3.6

tensorflowjs version

1.0.1

OS version

MacOS Mojave 10.4.3

Error

While I am trying to convert faster_rcnn_inception_v2_coco saved model to Javascript format, I received an issue:

KeyError: <tf.Tensor 'SecondStagePostprocessor/BatchMultiClassNonMaxSuppresion/map/while/Switch_1:0' shape=<unkknown> dtype=float32>

Code to reproduce the bug / link to feature request

1. Download faster_rcnn_inception_v2_coco tar and extract it
2. Try to convert with tensorflowjs_converter --input_format=tf_saved_model --output_format=tfjs_graph_model <pathToExtractedTar>/saved_model ./output_folder

Would you be so kind to provide some help according this issue? Thank you!

[vpanichkin]

Hi, Google Team!

Thank you for such a quick response. Today I succeed in converting this model by using 0.8.x version and passing a frozen model with

output_node_names=detection_boxes,detection_scores,detection_classes,num_detections

I suppose, that, while I am providing output_node_names, I stay focused on a particular part of the graph and this part converts without any issue.

Is it possible to add output_node_names parameter to future releases?

[vpanichkin]

Hi, @pyu10055!

Solved an issue with older converter. Thanks for cooperating!