def requestGenerator(input_name, output_name, c, h, w, format, dtype, FLAGS): # Preprocess image into input data according to model requirements image_data = None with Image.open(FLAGS.image_filename) as img: image_data = preprocess(img, format, dtype, c, h, w, FLAGS.scaling) repeated_image_data = [image_data for _ in range(FLAGS.batch_size)] batched_image_data = np.stack(repeated_image_data, axis=0) # Set the input data inputs = [] if FLAGS.protocol.lower() == "grpc": inputs.append(grpcclient.InferInput(input_name, batched_image_data.shape, dtype)) inputs[0].set_data_from_numpy(batched_image_data) else: inputs.append(httpclient.InferInput(input_name, batched_image_data.shape, dtype)) inputs[0].set_data_from_numpy(batched_image_data, binary_data=False) outputs = [] if FLAGS.protocol.lower() == "grpc": outputs.append(grpcclient.InferRequestedOutput(output_name, class_count=FLAGS.classes)) else: outputs.append(httpclient.InferRequestedOutput( output_name, binary_data=False, class_count=FLAGS.classes)) yield inputs, outputs, FLAGS.model_name, FLAGS.model_version
def async_stream_send(triton_client, values, batch_size, sequence_id, model_name, model_version): count = 1 for value in values: # Create the tensor for INPUT value_data = np.full(shape=[batch_size, 1], fill_value=value, dtype=np.int32) inputs = [] inputs.append(grpcclient.InferInput('INPUT', value_data.shape, "INT32")) # Initialize the data inputs[0].set_data_from_numpy(value_data) outputs = [] outputs.append(grpcclient.InferRequestedOutput('OUTPUT')) # Issue the asynchronous sequence inference. triton_client.async_stream_infer(model_name=model_name, inputs=inputs, outputs=outputs, request_id='{}_{}'.format( sequence_id, count), sequence_id=sequence_id, sequence_start=(count == 1), sequence_end=(count == len(values))) count = count + 1
"input0_data", cudashm.get_raw_handle(shm_ip0_handle), 0, input_byte_size) triton_client.register_cuda_shared_memory( "input1_data", cudashm.get_raw_handle(shm_ip1_handle), 0, input_byte_size) # Set the parameters to use data from shared memory inputs = [] inputs.append(grpcclient.InferInput('INPUT0', [1, 16], "INT32")) inputs[-1].set_shared_memory("input0_data", input_byte_size) inputs.append(grpcclient.InferInput('INPUT1', [1, 16], "INT32")) inputs[-1].set_shared_memory("input1_data", input_byte_size) outputs = [] outputs.append(grpcclient.InferRequestedOutput('OUTPUT0')) outputs[-1].set_shared_memory("output0_data", output_byte_size) outputs.append(grpcclient.InferRequestedOutput('OUTPUT1')) outputs[-1].set_shared_memory("output1_data", output_byte_size) results = triton_client.infer(model_name=model_name, inputs=inputs, outputs=outputs) # Read results from the shared memory. output0 = results.get_output("OUTPUT0") if output0 is not None: output0_data = cudashm.get_contents_as_numpy( shm_op0_handle, utils.triton_to_np_dtype(output0.datatype), output0.shape)