Pfthb

I have a strange issue, the DNN.predict method is quite slower when I load my model's weight than when I train with the fit method. I've also noted that when I run a prediction over a batch of images, it's getting faster and faster to predict.

Here is my code

class Reseau(object):

def init(self, img_size, lr=-1, activation=" "):
tf.logging.set_verbosity(tf.logging.ERROR)
self.lr = lr
self.activation = activation
self.img_size = img_size
self.alreadySaved = 0

def setting(self, X, Y, test_x, test_y, nbEpoch):
tflearn.init_graph(num_cores=32, gpu_memory_fraction=1)
with tf.device("/device:GPU:0"):
convnet = input_data(shape=[None, self.img_size, self.img_size, 3], name='input')

 convnet = conv_2d(convnet, 32, 5, activation=self.activation)
 convnet = max_pool_2d(convnet, 5)

 convnet = conv_2d(convnet, 64, 5, activation=self.activation)
 convnet = max_pool_2d(convnet, 5)

 convnet = conv_2d(convnet, 128, 5, activation=self.activation)
 convnet = max_pool_2d(convnet, 5)

 convnet = conv_2d(convnet, 64, 5, activation=self.activation)
 convnet = max_pool_2d(convnet, 5)

 convnet = conv_2d(convnet, 32, 5, activation=self.activation)
 convnet = max_pool_2d(convnet, 5)

 convnet = flatten(convnet)

 convnet = fully_connected(convnet, 1024, activation=self.activation, name='last')
 convnet = fully_connected(convnet, 1024, activation=self.activation, name='last')
 convnet = fully_connected(convnet, 1024, activation=self.activation, name='last')
 convnet = fully_connected(convnet, 1024, activation=self.activation, name='last')
 convnet = dropout(convnet, 0.8)

 convnet = fully_connected(convnet, 2, activation='softmax')
 convnet = regression(convnet, optimizer='adam', learning_rate=self.lr, loss='categorical_crossentropy', name='targets')

 self.model = tflearn.DNN(convnet, tensorboard_dir='log')

 if self.alreadySaved == 0:
 self.model.fit('input': X, 'targets': Y, n_epoch=nbEpoch, validation_set=('input': test_x, 'targets': test_y), snapshot_step=500, show_metric=True, run_id="model")
 self.model.save("./model")
 else:
 self.model.load("./model", weights_only=True)
 return self.model

def predire(self, img, label):
image = array(img).reshape(1, self.img_size,self.img_size,3)
model_out = self.model.predict(image)
rep = 0

 if np.argmax(model_out) == np.argmax(label): rep = 1
 else: rep = 0

 return rep

Here is a part of my main

reseau.setting(X, Y, test_x, test_y, NB_EPOCH)

X = np.array([i[0] for i in test]).reshape(-1,IMG_SIZE,IMG_SIZE,3)
Y = [i[1] for i in test]
cpt = 0

vrai = 0
start_time = time.time()

for i in range(20):
 cpt = 0

 vrai = 0
 start_time = time.time()
 for img in tqdm(X):
 prediction = reseau.predire(img, Y[cpt])
 cpt += 1
 if prediction == 1:
 vrai += 1

As you can see, I predict the same batch of images 20 times. The first time is always slower than the other ones (without fitting, I predict 82 images the first and then 340 a second, with fitting, it's 255 images the first time and 340 a seconde then).

I'm really out of idea to fix this.