tfjs: Error: TensorList shape mismatch: Shapes -1 and 3 must match

I have the same exact error when training the model using the object detection API. all works fine right to the point where I convert the model to TFJS. after loading an image I got the error: TensorList shape mismatch: Shapes -1 and 3 must match

I’ve been working for a week. This error appears to be from open source applications that were not in previous versions. I even tried from old models. They are working. But the one I just created doesn’t work.

This error continues. How do we fix this?

tfjs-core: ^3.0.0
tfjs-converter: ^3.0.0
TF 2.4.1
Python 3.6.9

This issue is resolved in https://github.com/tensorflow/tfjs/pull/4657 , which is released in 3.1.0, please try the latest version.

I am having the exact same problem. I am trying to use a model I trained using the TF2 Obj Detection API. I did transfer learning on top of the MobileNet V2 SSD model, exported to saved graph using exporter_main_v2.py. I used the tensorflowjs_converter.

I get the same message from just trying to feed in a Zero Tensor, made with this code:

    const zeroTensor = tf.zeros([1, 300, 300, 3], 'int32');
    // Warmup the model.
    const result = await this.model.executeAsync(zeroTensor) as tf.Tensor[];

When I pause code execution in the browser and inspect, the loaded graph model has the following input signature.

signature:
  inputs:
    input_tensor:0:
    dtype: "DT_UINT8"
    name: "input_tensor:0"
    tensorShape:
      dim: Array(4)
      0: {size: "1"}
      1: {size: "-1"}
      2: {size: "-1"}
      3: {size: "3"}

I also checked the TF Save Model and it had the following signature:

!saved_model_cli show --dir {model_export_dir}saved_model --all
The given SavedModel SignatureDef contains the following input(s):
    inputs['input_tensor'] tensor_info:
        dtype: DT_UINT8
        shape: (1, -1, -1, 3)
        name: serving_default_input_tensor:0
  The given SavedModel SignatureDef contains the following output(s):
    outputs['detection_anchor_indices'] tensor_info:
        dtype: DT_FLOAT
        shape: (1, 100)
        name: StatefulPartitionedCall:0
    outputs['detection_boxes'] tensor_info:
        dtype: DT_FLOAT
        shape: (1, 100, 4)
        name: StatefulPartitionedCall:1
    outputs['detection_classes'] tensor_info:
        dtype: DT_FLOAT
        shape: (1, 100)
        name: StatefulPartitionedCall:2
    outputs['detection_multiclass_scores'] tensor_info:
        dtype: DT_FLOAT
        shape: (1, 100, 2)
        name: StatefulPartitionedCall:3
    outputs['detection_scores'] tensor_info:
        dtype: DT_FLOAT
        shape: (1, 100)
        name: StatefulPartitionedCall:4
    outputs['num_detections'] tensor_info:
        dtype: DT_FLOAT
        shape: (1)
        name: StatefulPartitionedCall:5
    outputs['raw_detection_boxes'] tensor_info:
        dtype: DT_FLOAT
        shape: (1, 1917, 4)
        name: StatefulPartitionedCall:6
    outputs['raw_detection_scores'] tensor_info:
        dtype: DT_FLOAT
        shape: (1, 1917, 2)
        name: StatefulPartitionedCall:7
  Method name is: tensorflow/serving/predict

I also checked and the Save Model from the TF2 Object Detection zoo has that same saved model: http://download.tensorflow.org/models/object_detection/tf2/20200711/ssd_mobilenet_v2_320x320_coco17_tpu-8.tar.gz

Does the input signature for the model need to be explicitly set to 1,300,300,3 somewhere?

Is there a better way to run a Custom TF2 Object Detection model in TFjs?

I tried too many times the problem persists. How am I going to do this? Can you help me?