adamsolomou / sc-dnn Goto Github PK

Hi~, I am very grateful for the code you shared, it worked very well and helped me a lot.

But I found a small problem, which may cause the program to output unexpected results (NAN)

In SC-DNN/src/training/sc_train_creg.py

The variable gain is randomly initialized and used in subsequent calculations (in tf.div)

lo_matmul = {
    'upscale': True,
    'gain': tf.get_variable(name="G_matmul_2",
                            initializer=tf.random_uniform([n_classes],minval=0,maxval=16))
}

# ......

if(upscale):
    z_upscaled = tf.multiply(z,gain)
    new_scale = tf.div(curr_scale,gain)
else:
    z_upscaled = z
    new_scale = curr_scale

However, when the variable gain unfortunately gets 0 in the process of random initialization, there would be a division by zero operation in tf.div(curr_scale,gain). This is illegal and will eventually cause the program to output NaN

Maybe it's better if we added a small constant (1e-7) to tf.div(curr_scale,gain) to avoid division by zero? Just like: tf.div(curr_scale,gain + 1e-7) Or setting minval=1e-7 in the initialization of gain

Looking forward to receiving your reply~

Thanks in advance!