HTTP vs. gRPC and output handling

[bacalfa]

How do I extract the desired result in the response from client.infer() when using HTTP and gRPC protocols? They seem to return different data structures.

image = ...
inputs = [
    tritonclient.InferInput("image", image.shape, "FP32"),
]
inputs[0].set_data_from_numpy(image.astype(np.float32))

model_name = "git"
url = "triton:8000"  # HTTP
# url = "triton:8001"  # gRPC
model_version = "1"
output_name = "generated_caption"
output = tritonclient.InferRequestedOutput(output_name)

with tritonclient.InferenceServerClient(url, verbose=False) as client:
     response = client.infer(model_name, model_version=model_version, inputs=inputs, outputs=[output])

# How do I handle variable `response` depending on the protocol? I want to retrieve the `"generated_caption"` value

Here's the model's config.pbtxt:

name: "git"
backend: "python"
input [
  {
    name: "image"
    data_type: TYPE_FP32
    dims: [-1, -1, -1]
  }
]
output [
  {
    name: "generated_caption"
    data_type: TYPE_STRING
    dims: [-1]
  }
]

I'm using the Python backend, and in the execute() method, I have the following:

inference_response = pb_utils.InferenceResponse(
    output_tensors=[
        pb_utils.Tensor(
            "generated_caption",
            np.array([output.encode("UTF-8")], dtype=np.bytes_),
        )
    ]
)

[bacalfa]

As I suspected, it shouldn't matter which protocol. I realized I can just do the following:

generated_caption = response.as_numpy(output_name)[0].decode("UTF-8")

Perfect!