TensorFlow network not training?

Question

I had a very strange time with TensorFlow in recent days, and I can’t think about what is going wrong.

I created this network: link . This is a copy of the TensorFlow MNIST example.

Basically, what I was doing was changing the network to use 28x28x1 images (grayscale MNIST) up to 600x800x1 images (I took some images myself, a webcam with a relatively clean background and one object in different places).

What I wanted to do was play with that CNN and display the x-location of the object in the image on it. Thus, one single output neuron.

However, no matter what I tried, the network always outputs 1.0. Even (when you look at my testing section at the end of the code), when I feed everything, all zeros or all random numbers on the network.

Of course, since I only have 21 labeled training and 7 labeled test images, I expected the performance to be very poor (since images with a resolution of 800x600 pixels are huge for neural networks, and finding an object is not easy).

... but I have no idea why the network always outputs 1.0, even if it is fed nonsense. Any ideas?

Answer 1

, , , softmax (L.104 , ), softmax. , tf.nn.softmax() op (y_conv) batch_size x 1, 1.0 , tf.argmax(), 0 , . tf.argmax() y_train ( batch_size x 1) 0 , "" 100%.

, (i) , , (ii) y_train , .