demo.py

import json

import werkzeug
import tensorflow as tf

from config import config, parseArgs, configPDF
from extract_feature import get_img_feat, build_model
from main import setSession, loadWeights, setSavers
from model import MACnet
from preprocess import Preprocesser
import warnings

def predict(image, question):
  parseArgs()
  configPDF()
  with open(config.configFile(), "a+") as outFile:
      json.dump(vars(config), outFile)

  if config.gpus != "":
      config.gpusNum = len(config.gpus.split(","))
      os.environ["CUDA_VISIBLE_DEVICES"] = config.gpus
  tf.reset_default_graph()
  tf.Graph().as_default()
  tf.logging.set_verbosity(tf.logging.ERROR)
  cnn_model = build_model()
  imageData = get_img_feat(cnn_model, image)

  preprocessor = Preprocesser()
  qData, embeddings, answerDict = preprocessor.preprocessData(question)
  model = MACnet(embeddings, answerDict)
  init = tf.global_variables_initializer()

  savers = setSavers(model)
  saver, emaSaver = savers["saver"], savers["emaSaver"]
  sessionConfig = setSession()

  data = {'data': qData, 'image': imageData}

  with tf.Session(config=sessionConfig) as sess:
    sess.graph.finalize()

    epoch = loadWeights(sess, saver, init)
    emaSaver.restore(sess, config.weightsFile(epoch))

    evalRes = model.runBatch(sess, data['data'], data['image'], False)
    answer = None

    if evalRes in ['top', 'bottom']:
        answer = 'The caption at the %s side of the object.' % evalRes
    elif evalRes in ['True', 'False']:
        answer = 'There is at least one title object in this image.'
    else:
        answer = 'This image contain %s specific object(s).' % evalRes

  return answer