Inference output determined by batch size about squeezenet HOT 9 CLOSED

Ahhh, my first guess is this is due to how epochs are calculated combined with how the looping over the data is done. When you use a batch size of 1, the # of epochs are perfectly calculated to run over exactly all your data once. However with a batch size > 1, then the number of epochs is the ceiling of dividing the size of your training set by the batch size. For example, with 42 test samples and a batch size of 5, the number of epochs would be ceil(42 / 5) = 9. But 9 * 5 is 45 samples, so you'd be evaluating 3 test samples twice....

It's a lazy way to calculate epochs and do the looping...
Really, the evaluation_loop should make better use of the batch queue (by setting allow_smaller_final_batch=True).

Glad you are making use of the model though! Just beware... this was quickly put together for a class homework and definitely is not polished.

from squeezenet.

Thanks! Evaluation loop aside, if I load the checkpoint train model directly and run inference directly, the logit outputs are wildly different for various batch sizes with same input. Just wondering why that would be?

Yes thanks was just playing around with squeezenet.

from squeezenet.

Hey, so it seems the issue is that batch_norm needs to be set to is_training=False during inference. You can do this by adjusting your squeezenet model as such:

def squeezenet_arg_scope(is_training=True): batch_norm_params = {'is_training': is_training, 'decay': 0.9, 'updates_collections': None} with slim.arg_scope([slim.conv2d], normalizer_fn=slim.batch_norm, normalizer_params=batch_norm_params) as sc: return sc

from squeezenet.

Ohhh... woops! Good call, thank you!

from squeezenet.

Finally had some time to include your fix. Thanks again.

from squeezenet.

Hi @levelvc @vonclites
I have followed similar steps
Converting my images to TFrecords and then trying to use the train_squeezenet script for the same
When i do the same I am getting the below log.txt error

log.txt

Please let me know what exactly is wrong and the workaround for it

Thanks

from squeezenet.

from squeezenet.

Looks like your batch size is zero.
…
On Oct 30, 2018, at 5:16 AM, Bhagyoday-Patil @.***> wrote: Hi @levelvc @vonclites I have followed similar steps Converting my images to TFrecords and then trying to use the train_squeezenet script for the same When i do the same I am getting the below log.txt error log.txt Please let me know what exactly is wrong and the workaround for it Thanks — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.

Sorry @levelvc slight mistake
Please find attachment logs.
logs.txt

I have used https://github.com/kwotsin/create_tfrecords to convert my images to tfrecords.
The same tfrecords when given as input to the train_squeezenet script is throwing that error. I have only 2 classes. I am not understanding what exactly is wrong. I have not made any changes in any of the scripts.
tensorflow version = 1.11.0
ubuntu version = 16.04.3
no GPU/CUDA

from squeezenet.

@vonclites @muthiyanbhushan Can you please help me with this error.
As descried above I am getting this error.
PFA log_1.7.1.txt

from squeezenet.