GitLab

Fix for TF-models #7216: CIFAR-10 tutorial for multi-GPU fails because full shape isn't passed to prefetch_queue contributed by mattmann. (#7217)

Fix for #7225: CIFAR-10 eval fails with error TypeError: Input 'predictions' of 'InTopKV2' Op has type float16 that contributed by mattmann. (#7227)

* Bug fix of cifar10_eval.py

line 120: changed from eval_data = FLAGS.eval_data == 'test'
to: eval_data = FLAGS.eval_data
comment: the original code assigns 'true' to eval_data, when the script is being used to evaluate networks on the evaluation set, it DOES NOT load the evaluation set as intended, but actually loads the trainning set (line 105 of cifar10_input.py).

* Remove one line that was an artifact of issue.

* restored missing function

* missing import

* missing imports

* updated tutorial link

* recovered _print_download_progress func

* change default train with float16 instead of float32 accuracy

* test disable func call

* redundant function call, currently data is pulled automatically in

* optimized imports

* optimized imports

* TF cifar10 cnn tutorial: use tensorflow-datasets to load the data.

* Load cifar10 in memory.

* Fix imports

* More import fixes

the number of batches per epoch also depends on the number of gpus. it should be `num_batches_per_epoch = (cifar10.NUM_EXAMPLES_PER_EPOCH_FOR_TRAIN / FLAGS.batch_size / FLAGS.num_gpus)`

While the original code works fine in practice, it technically
allows gradient application and moving average update to happen
in any order. This causes the behavior to deviate from pure
mathematical specifications.

We can directly get `batch_size` by `images.get_shape()[0]` in `inference` method, since maybe we will not use `cifar.inputs` method to build the input.

The change you added did not work with python2, it does not have encoding in .log.  This fixes it. feel free to test on Python3.  I tested python2 and I am really sue this works fine for python3.  Good catch.  sorry for not having it fixed earlier.

change cPickle to pickle for python3
specify  encoding='latin1' when call pickle.load()

Updated cifar10_input.py to put the data augmentation into a name_scope named data_augmentation so the tensorboard graph is easier to read and more meaningful

tf.contrib.framework.get_or_create_global_step -> tf.train.get_or_create_global_step

CIFAR website is now using https.
Since urllib does not follow the http to https redirection and thus never retrieves any data, the data URL needs an update.

Trying to remove the +1.