tensorflow/models · error · NotImplementedError

Video classification do not support image bytes input.

Error message

Video classification do not support image bytes input.

What it means

Capability stub in the video-classification export module: inference_from_image_bytes is declared to complete the export-module interface but the video pipeline requires decoded frame tensors, so any attempt to build a concrete function from image-byte (serialized single-image) input hits this unconditional refusal.

Solutions

1. Use input_type 'image_tensor' (decoded frames) for video classification serving.
2. Decode video/image bytes to tensors on the client side before sending.

When it happens

Trigger: Thrown at official/vision/serving/video_classification.py:112 when the library encounters an invalid state.

Common situations: See trigger scenarios.

AI-assisted analysis of tensorflow/models@e006f5f0d5 (2026-08-24). Data as JSON: /api/errors/dd61a9d60ff451bb. Report an issue: GitHub.

Appendix: source

Thrown at official/vision/serving/video_classification.py:112

            fn_output_signature={
                video_input.IMAGE_KEY: tf.string,
            })
        return self.serve(inputs[video_input.IMAGE_KEY], tf.zeros([1, 1]))

  @tf.function
  def inference_from_image_tensors(
      self, input_frames: tf.Tensor) -> Mapping[str, tf.Tensor]:
    return self.serve(input_frames, tf.zeros([1, 1]))

  @tf.function
  def inference_from_image_audio_tensors(
      self, input_frames: tf.Tensor,
      input_audio: tf.Tensor) -> Mapping[str, tf.Tensor]:
    return self.serve(input_frames, input_audio)

  @tf.function
  def inference_from_image_bytes(self, inputs: tf.Tensor):
    raise NotImplementedError(
        'Video classification do not support image bytes input.')

  def serve(self, input_frames: tf.Tensor, input_audio: tf.Tensor):
    """Cast image to float and run inference.

    Args:
      input_frames: uint8 Tensor of shape [batch_size, None, None, 3]
      input_audio: float32

    Returns:
      Tensor holding classification output logits.
    """
    with tf.device('cpu:0'):
      inputs = tf.map_fn(
          self._preprocess_image, (input_frames),
          fn_output_signature={
              'image': tf.float32,
          })

View on GitHub (pinned to e006f5f0d5)