Problem Converting tensorflow garden model in format of Tensorflow Saved Model to a Web Model with tensorflowjs

General Discussion
, ,
#1

System information

Describe the current behavior
I tried to convert the model i trained with my own custom dataset, but when the conversion of the saved model hapened, it first output this repeated error.

2024-06-15 03:25:44.975342: E tensorflow/core/grappler/optimizers/meta_optimizer.cc:961] remapper failed: INVALID_ARGUMENT: MutableGraphView::MutableGraphView error: node 'StatefulPartitionedCall/StatefulPartitionedCall/fpn_1/activation_3/Relu' has missing fanin 'StatefulPartitionedCall/StatefulPartitionedCall/fpn_1/coarser_6/BiasAdd'.
2024-06-15 03:25:45.312683: E tensorflow/core/grappler/optimizers/dependency_optimizer.cc:772] Iteration = 0, topological sort failed with message: Non-existent input StatefulPartitionedCall/StatefulPartitionedCall/fpn_1/coarser_6/BiasAdd in node StatefulPartitionedCall/StatefulPartitionedCall/fpn_1/activation_3/Relu
2024-06-15 03:25:45.371166: E tensorflow/core/grappler/optimizers/dependency_optimizer.cc:772] Iteration = 1, topological sort failed with message: Non-existent input StatefulPartitionedCall/StatefulPartitionedCall/fpn_1/post_hoc_5/BiasAdd in node StatefulPartitionedCall/StatefulPartitionedCall/fpn_1/coarser_6/Conv2D
2024-06-15 03:25:45.745130: E tensorflow/core/grappler/optimizers/dependency_optimizer.cc:772] Iteration = 0, topological sort failed with message: Non-existent input StatefulPartitionedCall/StatefulPartitionedCall/fpn_1/post_hoc_5/BiasAdd in node StatefulPartitionedCall/StatefulPartitionedCall/fpn_1/coarser_6/Conv2D
2024-06-15 03:25:45.775964: E tensorflow/core/grappler/optimizers/dependency_optimizer.cc:772] Iteration = 1, topological sort failed with message: Non-existent input StatefulPartitionedCall/StatefulPartitionedCall/fpn_1/post_hoc_5/BiasAdd in node StatefulPartitionedCall/StatefulPartitionedCall/fpn_1/coarser_6/Conv2D

Describe the expected behavior
I would expect that my model would be converted normally in the same way i converted my model into tflite, just as in my notebook.

To reproduce the issue

Full logs

2024-06-15 03:25:44.975342: E tensorflow/core/grappler/optimizers/meta_optimizer.cc:961] remapper failed: INVALID_ARGUMENT: MutableGraphView::MutableGraphView error: node 'StatefulPartitionedCall/StatefulPartitionedCall/fpn_1/activation_3/Relu' has missing fanin 'StatefulPartitionedCall/StatefulPartitionedCall/fpn_1/coarser_6/BiasAdd'.
2024-06-15 03:25:45.312683: E tensorflow/core/grappler/optimizers/dependency_optimizer.cc:772] Iteration = 0, topological sort failed with message: Non-existent input StatefulPartitionedCall/StatefulPartitionedCall/fpn_1/coarser_6/BiasAdd in node StatefulPartitionedCall/StatefulPartitionedCall/fpn_1/activation_3/Relu
2024-06-15 03:25:45.371166: E tensorflow/core/grappler/optimizers/dependency_optimizer.cc:772] Iteration = 1, topological sort failed with message: Non-existent input StatefulPartitionedCall/StatefulPartitionedCall/fpn_1/post_hoc_5/BiasAdd in node StatefulPartitionedCall/StatefulPartitionedCall/fpn_1/coarser_6/Conv2D
2024-06-15 03:25:45.745130: E tensorflow/core/grappler/optimizers/dependency_optimizer.cc:772] Iteration = 0, topological sort failed with message: Non-existent input StatefulPartitionedCall/StatefulPartitionedCall/fpn_1/post_hoc_5/BiasAdd in node StatefulPartitionedCall/StatefulPartitionedCall/fpn_1/coarser_6/Conv2D
2024-06-15 03:25:45.775964: E tensorflow/core/grappler/optimizers/dependency_optimizer.cc:772] Iteration = 1, topological sort failed with message: Non-existent input StatefulPartitionedCall/StatefulPartitionedCall/fpn_1/post_hoc_5/BiasAdd in node StatefulPartitionedCall/StatefulPartitionedCall/fpn_1/coarser_6/Conv2D
Traceback (most recent call last):
  File "/usr/local/bin/tensorflowjs_converter", line 8, in <module>
    sys.exit(pip_main())
  File "/usr/local/lib/python3.10/dist-packages/tensorflowjs/converters/converter.py", line 959, in pip_main
    main([' '.join(sys.argv[1:])])
  File "/usr/local/lib/python3.10/dist-packages/tensorflowjs/converters/converter.py", line 963, in main
    convert(argv[0].split(' '))
  File "/usr/local/lib/python3.10/dist-packages/tensorflowjs/converters/converter.py", line 949, in convert
    _dispatch_converter(input_format, output_format, args, quantization_dtype_map,
  File "/usr/local/lib/python3.10/dist-packages/tensorflowjs/converters/converter.py", line 655, in _dispatch_converter
    tf_saved_model_conversion_v2.convert_tf_saved_model(
  File "/usr/local/lib/python3.10/dist-packages/tensorflowjs/converters/tf_saved_model_conversion_v2.py", line 982, in convert_tf_saved_model
    _convert_tf_saved_model(output_dir, saved_model_dir=saved_model_dir,
  File "/usr/local/lib/python3.10/dist-packages/tensorflowjs/converters/tf_saved_model_conversion_v2.py", line 857, in _convert_tf_saved_model
    optimized_graph = optimize_graph(frozen_graph, signature,
  File "/usr/local/lib/python3.10/dist-packages/tensorflowjs/converters/tf_saved_model_conversion_v2.py", line 188, in optimize_graph
    optimized_graph = _run_grappler(config, optimized_graph, graph, signature_def)
  File "/usr/local/lib/python3.10/dist-packages/tensorflowjs/converters/tf_saved_model_conversion_v2.py", line 129, in _run_grappler
    return tf_optimizer.OptimizeGraph(
  File "/usr/local/lib/python3.10/dist-packages/tensorflow/python/grappler/tf_optimizer.py", line 61, in OptimizeGraph
    out_graph = optimize_method(
tensorflow.python.framework.errors_impl.InvalidArgumentError: Non-existent input StatefulPartitionedCall/StatefulPartitionedCall/fpn_1/post_hoc_5/BiasAdd in node StatefulPartitionedCall/StatefulPartitionedCall/fpn_1/coarser_6/Conv2D
#3

Hi @Linubius ,
There could be 2 reasons according error logs:

  1. Since you are using older TensorFlow version 1.16.0 (assuming it is not typo error here), The TensorFlow.js (4.20.0) converter may not be compatible with the TensorFlow version you’re using. Please use the latest tensorflow and try converting.

  2. Grappler optimizer is unable to find certain input nodes (e.g., fpn_1/coarser_6/BiasAdd` fpn_1/post_hoc_5/BiasAdd)

    Try disabling the “Grappler optimizations” during the conversion by setting the appropriate TensorFlow options may help resolving the above issue IMO.

Thanks.

#5
  1. Yes, it was a typo. I meant to type 2.16 because this is also the version of the TFM package used.
  2. After did some digging, I decided to use the convert_tf_saved_model() function from the tf_saved_model_conversion_v2.py and i got this new error message ValueError: Unsupported Ops in the model before optimization Assert. It seems like the assert Ops isn’t supported “Yet” in tfjs. here is the related gist-file even if i disabled the Grappler Optimizations it resulted the same error of Non-existent input StatefulPartitionedCall/StatefulPartitionedCall/fpn_1/post_hoc_5/BiasAdd in node StatefulPartitionedCall/StatefulPartitionedCall/fpn_1/coarser_6/Conv2D
    Thanks for the response @Laxma_Reddy_Patlolla
#7

Hi @Linubius ,

I think you got answer from github that some of the TensorFLow ops not supported while in conversion.

Thanks.