def test_generate_captions():
'''
Wherein we test test_mod.generate_captions. Since you may
use pre-computed weights from any source only print the
generated sentence to stdout and check that it is nonempty
'''
encoded_img=ei.encodings(ei.model_gen(),test_img)
caption=generate_captions(sd,model,encoded_img,beam_size=3)
def report():
print('The model generated the caption: '+caption)
atexit.register(report)
assert(len(caption)>0)
def test_encodings(): ''' We test encode_image.encodings(), which is just a wrapper for VGG16. This proceeds in the same fashion as in test_vgg16(). The truth value for the encodings differ slightly, since the encodings() wrapper uses another preprocessing routine defined in imagenet_utils. ''' model=model_gen() preds=encodings(model=model,path=os.path.join(image_folder,test_image)) encoding_pred_slice=np.array([0,0,0,0,6.943786,0,0,0,0,0.495808,0,0,1.2099614,0,0,1.9550853,0,2.5830698,0,0,1.5520213,6.7467823,0.30691597,0.6208435,0,3.8465405,3.7862554,2.3970299,0,0,3.5254822,2.7294598,0,0,2.7226853,0,3.2338202,2.3976898,0,6.3592043,0,2.7090664,0,0,10.004378,0,0,0,3.0425727,2.0538316,1.8156273,0.15581878,2.3381875,0.88823074,0,0,0,0,0,0,0.036334604,0,0,0,3.5556676,0.29299664,0,0,0,0,0,0,1.0033253,0,0,0,0.96017045,0,5.8062425,0,4.4312,0,0,0,0,0,0,0,2.7901797,0.5715834,0.76234996,1.7294867,1.2244358,0,0,0,0,0,0,0,]) assert(preds.shape==(4096,)) assert(len(preds[preds!=0])==1351) np.testing.assert_allclose(preds[:100],encoding_pred_slice,err_msg='encoding wrapper yielded wrong encoding for test image!')
def text(img): t1= time.time() encode = ei.model_gen() weight = 'Output/Weights.h5' sd = SceneDesc.scenedesc() model = sd.create_model(ret_model = True) model.load_weights(weight) image_path = img encoded_images = ei.encodings(encode, image_path) image_captions = tm.generate_captions(sd, model, encoded_images, beam_size=3) engine = pyttsx.init() print (image_captions) engine.say( str(image_captions)) engine.runAndWait()
def text(img):
t1 = time.time()
encode = ei.model_gen()
weight = 'Output/Weights.h5'
sd = SceneDesc.scenedesc()
model = sd.create_model(ret_model=True)
model.load_weights(weight)
image_path = img
encoded_images = ei.encodings(encode, image_path)
image_captions = tm.generate_captions(sd,
model,
encoded_images,
beam_size=3)
engine = pyttsx.init()
print '\nCaption Generated for the above Image is=\n', image_captions
def text(caption):
encode = ei.model_gen()
weight = 'Output/Weights.h5'
decoder = decode_caption.decoder()
model = decoder.model_gen(ret_model=True)
model.load_weights(weight)
bleu_sum = 0.0
count = 0
cc = SmoothingFunction()
test_imgs_id = open("Flickr8K_Text/Flickr_8k.testImages.txt").read().split(
'\n')[:-1]
for img_id in test_imgs_id:
image_path = "Flickr8K_Data/" + img_id
encoded_images = ei.encodings(encode, image_path)
image_captions = generate_captions(decoder,
model,
encoded_images,
beam_size=3)
print(image_captions)
# bleuscore-4
bleus = []
image_captions = image_captions.split()
for true_sentense in caption[img_id]:
true_sentense = true_sentense.split()
try:
bleu = sentence_bleu([true_sentense],
image_captions,
weights=(0.25, 0.25, 0.25, 0.25),
smoothing_function=cc.method4)
bleus.append(bleu)
except:
pass
if len(bleus) > 0:
print(np.mean(bleus))
bleu_sum += np.mean(bleus)
count += 1
print(bleu_sum / count)
