tfserve_vision21.py

# create serving function
def predict_apparel_model2_serving(img, img_dims=(32,32), label_map=class_names):
    
    sample_img_processed = (np.array([resize_image_array(img, 
                                                         img_size_dims=img_dims) 
                                                for img in np.stack([[img]]*3, 
                                                                    axis=-1)])) / 255.
    data = json.dumps({"signature_name": "serving_default", 
                       "instances": sample_img_processed.tolist()})
    
    HEADERS = {'content-type': 'application/json'}
    MODEL2_API_URL = 'http://localhost:8501/v1/models/fashion_model_serving/versions/2:predict'
    json_response = requests.post(MODEL2_API_URL, data=data, headers=HEADERS)
    prediction = json.loads(json_response.text)['predictions']
    prediction = np.argmax(np.array(prediction), axis=1)[0]
    return label_map[prediction]
    
# benchmark 10K requests
%%time

pred_labels = []
for img in tqdm(test_images):
    pred_label = predict_apparel_model2_serving(img)
    pred_labels.append(img)
len(pred_labels)