First, we will add the code to download all user-labeled images from the production server:

    import tensorflow as tf 
    import os 
    import json 
    import random 
    import requests 
    import shutil 
    from scipy.misc import imread, imsave 
    from datetime import datetime 
    from tqdm import tqdm 
 
    import nets, models, datasets 
 
 
    def ensure_folder_exists(folder_path): 
    if not os.path.exists(folder_path): 
        os.mkdir(folder_path) 
    return folder_path 
 
 
    def download_user_data(url, user_dir, train_ratio=0.8): 
    response = requests.get("%s/user-labels" % url) 
    data = json.loads(response.text) 
 
    if not os.path.exists(user_dir): 
        os.mkdir(user_dir) 
    user_dir = ensure_folder_exists(user_dir) 
    train_folder = ensure_folder_exists(os.path.join(user_dir,   
    "trainval")) 
    test_folder = ensure_folder_exists(os.path.join(user_dir,   
    "test")) 
 
    train_file = open(os.path.join(user_dir, 'trainval.txt'), 'w') 
    test_file = open(os.path.join(user_dir, 'test.txt'), 'w') 
 
    for image in data: 
        is_train = random.random() < train_ratio 
        image_url = image["url"] 
        file_name = image_url.split("/")[-1] 
        label = image["label"] 
        name = image["name"] 
 
        if is_train: 
          target_folder =  
          ensure_folder_exists(os.path.join(train_folder, name)) 
        else: 
          target_folder =   
          ensure_folder_exists(os.path.join(test_folder, name)) 
 
        target_file = os.path.join(target_folder, file_name) +   
        ".jpg" 
 
        if not os.path.exists(target_file): 
            response = requests.get("%s%s" % (url, image_url)) 
            temp_file_path = "/tmp/%s" % file_name 
            with open(temp_file_path, 'wb') as f: 
                for chunk in response: 
                    f.write(chunk) 
 
            image = imread(temp_file_path) 
            imsave(target_file, image) 
            os.remove(temp_file_path) 
            print("Save file: %s" % target_file) 
 
        label_path = "%s %s
" % (label, target_file) 
        if is_train: 
            train_file.write(label_path) 
        else: 
            test_file.write(label_path) 

In download_user_data, we call the /user-labels endpoint to get the list of user-labeled images. The JSON has the following format:

   [ 
    { 
     "id": 1,  
     "label": 0,  
     "name": "Abyssinian",  
     "url": "/uploads/2017-05-23_14-56-45_Abyssinian-cat.jpeg" 
    },  
    { 
     "id": 2,  
      "label": 32,  
      "name": "Siamese",  
     "url": "/uploads/2017-05-23_14-57-33_fat-Siamese-cat.jpeg" 
    } 
   ] 

In this JSON, label is the label that the user has chosen, and URL is the link to download the image from. For every image, we will download it into the tmp folder and use imread and imsave from scipy to make sure that the image is in JPEG format. We also create a trainval.txt and test.txt file, as in the training dataset.