[slim] weird result with parameter is_training

[argman]

I am trying to train a inception_resnet_v2 on another dataset, when I test the accuracy of the model.

with slim.arg_scope(inception_resnet_v2.inception_resnet_v2_arg_scope()):
        logits, _ = inception_resnet_v2.inception_resnet_v2(images,
                                                             num_classes=dataset.num_classes,
                                                             is_training=True)

setting is_training with True give better result on validation set, but from the code, i should have set this to False

[argman]

@sguada I think this problem maybe with batch_norm during training and testing, there should be a better way to solve this

[argman]

I find its strange in _get_variables_to_train, I print out all the variable names, but in batch_norm, only beta like 'InceptionV3/Mixed_6b/Branch_2/Conv2d_0a_1x1/BatchNorm/beta:0' is included

[argman]

It's really confusig when I try fine-tuning, can someone explain this?

[sguada]

When is_training=True then batch_norm uses the statistics of the batch, when it is False it uses the moving average of the statistics computed during training.

When doing fine-tuning sometimes it is better not update the moving average of the statistics, since you are not going to run for long. So try training with is_training=False and testing with also is_training=False.

[argman]

@sguada , after I update my tf and cuda,cudnn, the problem goes away, so a suggestion maybe document your tf version and cuda, cudnn version in slim. Tf is now being developped by many and a bit complicated to debug.

And my tensorflow version is ed87884e50e1a50f7dc7b36dc7a7ff225442bee0， cuda 8.0, cudnn 5.1

[argman]

@sguada , in the slim implemention, you apply moving avg to all the model parameters(maybe just need for inception models ?), but in another implemention of resnet under tensorflow/models, they just apply to batch norm parameters.

[gunan]

Automatically closing due to lack of recent activity. Please update the issue when new information becomes available, and we will reopen the issue. Thanks!

[ilkarman]

I still have the same issue, 2 years later