Keras - my model always returns accuracy “0” and loss “nan”

Question

I have the following Keras code that tries to train a combination of an image and a vector of real number features:

from keras.layers import (Conv2D, Flatten, Lambda, Dense, concatenate,
                         Dropout, Input )
from keras.models import Model
import pandas as pd
import numpy as np
import cv2
import os

def label_img(img):
    word_label = img.split('.')[-3]
    if word_label == 'r':
        return 1
    elif word_label == 'i':
        return 0

train_directory = '/train'
images = []
y = []

dataset = pd.read_csv('features.csv')

dataset = dataset[[ 'first_value',
                    'second_value']]

features = dataset.iloc[:,0:2].values

for root, dirs, files in os.walk(train_directory):
    for file in files:
        image = cv2.imread(root + '/' + file)
        image = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
        image = cv2.resize(image,(512,512),interpolation=cv2.INTER_AREA)
        images.append(image)
        label = label_img(file)
        y.append(label)

images = np.asarray(images)
images = images.reshape((-1,512,512,1))

image_input = Input(shape=(512,512,1))
aux_input = Input(shape=(2,))

input_layer = Conv2D(32,(5,5), activation='relu')(image_input)
cov1 = Conv2D(24,(5,5), activation='relu', subsample=(2,2))(input_layer)
cov2 = Conv2D(36,(5,5), activation='relu', subsample=(2,2))(cov1)
cov3 = Conv2D(48,(5,5), activation='relu', subsample=(2,2))(cov2)
cov4 = Conv2D(64,(5,5), activation='relu')(cov3)
cov5 = Conv2D(64,(3,3), activation='relu')(cov4)
dropout = Dropout(0.5)(cov5)
flatten = Flatten()(dropout)

# Here we add in the feature vectors 
merge = concatenate([flatten, aux_input])

d1 = Dense(100, activation='elu')(merge)
d2 = Dense(50, activation='elu')(d1)
d3 = Dense(10, activation='elu')(d2)
out = Dense(1)(d3)

model = Model(inputs=[image_input, aux_input], outputs=[out])
model.compile(loss='binary_crossentropy', optimizer='adam',metrics=['acc'])
model.fit([images, features], y, epochs=50)

However, I'm getting something like the following:

Epoch 2/50
5/5 [==============================] - 0s - loss: nan - acc: 0.0000e+00
Epoch 3/50
5/5 [==============================] - 0s - loss: nan - acc: 0.0000e+00
Epoch 4/50
5/5 [==============================] - 0s - loss: nan - acc: 0.0000e+00
Epoch 5/50
...
...
...

Why is that? How can I solve the issue?

I would like to mention that my images are black background images with foreground showing an edge of the original image. I'm not sure if that's an issue at all, as there is not much data in the image?

Thanks.

Answer 1

To be able to use crossentropy correctly, you need to ensure, that the outputs of your network are between 0 and 1. Therefore you should use sigmoid or softmax activation in the last layer.

Elu or linear activations allow negative values, which will cause infinite loss, when using cross-entropy. On the other hand if output values are larger than one the same as before can happen and we can not ensure that the cross-entropy loss stays positive, therefore minimaization doesn't make sense anymore at this point.

This is caused by the definition of the cross-entropy:

-(y_true*log(y_pred)+(1-y_true)*log(1-y_pred))

The logarithm is not defined for values being less than zero.