I trained an object detection model on ML Engine and exported it by invoking:
python object_detection/export_inference_graph.py \
--input_type encoded_image_string_tensor ....
Then I successfully tested prediction locally by invoking:
gcloud ml-engine local predict --model-dir ../saved_model --json-instances=inputs.json --runtime-version=1.2
where inputs.json contains:
{"b64": "base64 encoded png image"}
When I try to create a model version on ML Engine using the following command:
gcloud ml-engine versions create ${YOUR_VERSION} --model ${YOUR_MODEL} --origin=${YOUR_GCS_BUCKET}/saved_model --runtime-version=1.2
it fails with the following message:
ERROR: (gcloud.ml-engine.versions.create) Bad model detected with error: "Error loading the model: Could not load model. "
Does ML Engine NOT support model versions of input_type=encoded_image_string_tensor and how can I obtain more details on the error?
Creating a model version on ml-engine using an exported model with input_type=image_tensor works fine.
Can you verify that you're exporting the model with tensorflow 1.2?
gcloud ml-engine local predict doesn't have the --runtime-version flag, so if you have TF 1.3 installed and exported your model with that, then local predict would work using TF 1.3, but there may be incompatibilities in the model when trying to use TF 1.2 on the service.