Tensorflow CNN模型不训练吗？ 持续的损失和准确性

Question

I have built a model using this as a base.

And the train portion from this code .

This model does not train and always gives the cost/loss output for every iteration.

I dont think it learns anything.

I have checked for the usual stuff like shuffled inputs. Making sure that each batch is new.

Any idea why ?

This is my code.

Output

Iter 1280, Minibatch Loss= 4.615120, Training Accuracy= 0.03125
Testing Accuracy: 0.0
Iter 2560, Minibatch Loss= 4.615120, Training Accuracy= 0.00000
Testing Accuracy: 0.0
Iter 3840, Minibatch Loss= 4.615120, Training Accuracy= 0.00000
Testing Accuracy: 0.015625
Iter 5120, Minibatch Loss= 4.615120, Training Accuracy= 0.00000
Testing Accuracy: 0.078125
Iter 6400, Minibatch Loss= 4.615120, Training Accuracy= 0.03125
Testing Accuracy: 0.0
Iter 7680, Minibatch Loss= 4.615120, Training Accuracy= 0.03125
Testing Accuracy: 0.015625
Iter 8960, Minibatch Loss= 4.615120, Training Accuracy= 0.00000
Testing Accuracy: 0.0
Iter 10240, Minibatch Loss= 4.615120, Training Accuracy= 0.00000
Testing Accuracy: 0.015625
Iter 11520, Minibatch Loss= 4.615120, Training Accuracy= 0.00000
Testing Accuracy: 0.0
Iter 12800, Minibatch Loss= 4.615120, Training Accuracy= 0.01562
Testing Accuracy: 0.03125
Iter 14080, Minibatch Loss= 4.615120, Training Accuracy= 0.01562
Testing Accuracy: 0.0
Iter 15360, Minibatch Loss= 4.615120, Training Accuracy= 0.01562
Testing Accuracy: 0.0

Answer 1

The code you started from is just a benchmark of the forward and backward pass and isn't designed to do training. You should start from an example that actually trains a model instead and ignore the benchmark code.

You might have an easier time starting from a completely working training example program instead of trying to combine two pieces.