hwalsuklee / tensorflow-mnist-cnn Goto Github PK

I think

should be written in

I use this code on my own data. I restore model(saver.restore(sess, myModelPath)),and then test.
if I set test batch size is 1,the result will be [ 0. 0. 0. 0. 0.].
if I set test batch size is 5,the result will be [ 1. 0. 0. 0. 1.].
Why is it different

If I run exactly your code, Tensorflow complains about variables being already defined with, e.g., the following error:

ValueError: Variable conv1/weights already exists, disallowed. Did you mean to set reuse=True or reuse=tf.AUTO_REUSE in VarScope?

Do variables have to be reused or not? How can you set it globally?
[I'm with Tensorflow 1.3, Python 3.6 on Linux]

Thanks for your code! And I have some questions, glad for your help.
With the Ubuntu 18.04 and the tensorflow 0.12.0, when I run
python mnist_cnn_test.py --model-dir model/model01_99.61 --batch-size 5000 --use-ensemble False
It don't work and return NotFoundError (see above for traceback): Key fc3/BatchNorm/moving_mean not found in checkpoint [[Node: save/RestoreV2_9 = RestoreV2[dtypes=[DT_FLOAT], _device="/job:localhost/replica:0/task:0/cpu:0"](_recv_save/Const_0, save/RestoreV2_9/tensor_names, save/RestoreV2_9/shape_and_slices)]]

And when I run 'python mnist_cnn_test.py --model-dir model/model01_99.61 --batch-size 5000 --use-ensemble True'
The "accuracy“ only 0.092, All models are like this.
Looking forward to your reply.

I ran python mnist_cnn_train.py in terminal and the terminal returned

Optimization` Finished!
test accuracy for the stored model: 0.9932

Training logs are saved in "logs/train" , however, there is no trained model saved as "model/model01_99.61/model.ckpt" or in other directories. When I run python mnist_cnn_test.py --model-dir model/model01_99.61 --batch-size 5000 --use-ensemble False and it returns an error message:

NotFoundError (see above for traceback): Key fc3/BatchNorm/beta not found in checkpoint
	 [[Node: save/RestoreV2 = RestoreV2[dtypes=[DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT], _device="/job:localhost/replica:0/task:0/device:CPU:0"](_arg_save/Const_0_0, save/RestoreV2/tensor_names, save/RestoreV2/shape_and_slices)]]

Where is the problem?

Hello,

Going through the code of your project, I think the parameter is_training is not taken into account for the CNN model in file mnist_cnn_train.py.

I've seen that the cnn_model.CNN function takes "is_training" argument with default equals True which prevent the code to crash.

In mnist_cnn_train, you define the is_training placeholder but you don't use it when calling the cnn_model.CNN function. You use it in the training and testing loops of the same file so I assume this is not an intended behavior.

I've not tested it yet, but I think the is_training entry of the feed_dict is just ignored and this cause dropout to be applied during the testing loop (same goes for batch normalization). This bug could be the cause of the issue #1

When the model is created "y = cnn_model.CNN(x) ", the is_training variable is not passed. Thus in the testing section when performing y_final = sess.run(y, feed_dict={x: batch_xs, y_: batch_ys, is_training: False}), the is_training: False has no affect. This will impact your accuracy.

If you use the mnist_cnn_train.py test function, the model is initialized with the is_training parameter and will give a result of approximately .5 % higher.

I changed y = cnn_model.CNN(x, is_training=is_training) and now the accuracy percents match for both modules.

Just as a side note: tf.scalar_summar is deprecated in Tensor 1.4