09:30:01	 From Haider Fadil Ali Hussein Al-Saadi : how do you deal with boundaries where there are no values for the individual kernel entries?
09:30:17	 From Haider Fadil Ali Hussein Al-Saadi : nvm I see you wrote in padding :)
09:37:56	 From Troels Christian Petersen : By many parameters, one is talking about MILLIONS!
09:39:22	 From Troels Christian Petersen : Think about how many parameters are in your models, e.g. for the small project…
09:42:08	 From kristoffer : 26x26?
09:42:27	 From Joachim Kønigslieb : i agree with kristoffer! 
09:44:16	 From kristoffer : spaced out
09:44:22	 From Haider Fadil Ali Hussein Al-Saadi : 28x28?
09:44:23	 From Joachim Kønigslieb : 600
09:46:06	 From Troels Christian Petersen : Good answer, Joachim - I forgot the bias terms as well...
10:06:43	 From Andy Anker : Why is validation loss smaller than training loss? I would always expect the opposite.
10:41:22	 From Carl-Johannes Johnsen : It should be noted, that when training on GPUs, larger batch sizes reduces the training time, as larger chunks of data can be transferred to the GPU. The GPU likes fewer and larger chunks, when talking about performance! (of course up to a certain limit). However, a larger batch size might decrease the accuracy of the trained model.
10:44:52	 From Carl-Johannes Johnsen : To quantify my statement, on a large GPU with bach 128 each epoch takes 2.35 secs with acc of .9894. With batch 1024 each epoch takes 1.16 sec with acc of .9797.
10:47:42	 From kristoffer : Some more info on GPU on erda this would be dope!!
10:47:53	 From Emil Martiny : no
10:47:54	 From Emy Alerskans : no
10:47:54	 From Simone Vejlgaard : no
10:47:54	 From Edwin Vargas : no
10:47:55	 From Simon Hilbard : NO
10:47:55	 From Andy Anker : no
10:47:56	 From joachim : no
10:47:56	 From Bagne : no
10:47:56	 From Michael Haahr : no
10:47:57	 From Mikkel Langgaard Lauritzen : no
10:47:57	 From Haider Fadil Ali Hussein Al-Saadi : no
10:47:57	 From Rasmus Salmon : No
10:47:59	 From Aske R. : no
10:48:20	 From Carl-Johannes Johnsen : I am not sure about the "general" GPU machines, but I'll ask around! For the large machines, you need access. 
10:48:30	 From Carl-Johannes Johnsen : And my microphone broke yesterday :(
10:54:29	 From Aske R. : wasn't there a second part of presentations?
10:54:58	 From Haider Fadil Ali Hussein Al-Saadi : why is there such a huge dispartity between the growth of the parameters and growth of the epoch time? when i run 5000 parameters only, it takes 10 seconds but 20000 runs for 15 seconds? 
10:55:58	 From Haider Fadil Ali Hussein Al-Saadi : GPU
10:56:51	 From Dina Rapp  To  Troels Christian Petersen(privately) : How can we run TensorFlow in ERDA? When I try to import the library, ERDA says ‘no module named tensorflow’,  :)
10:56:55	 From Haider Fadil Ali Hussein Al-Saadi : 1080
10:57:10	 From Carl-Johannes Johnsen : A relatively big one :)
10:57:22	 From Haider Fadil Ali Hussein Al-Saadi : in my room yes
10:57:26	 From Jonathan : I have a question regarding regression in the small project:
I'm trying to transform/normalize/standardize (whatever gets my training data in the range [0;1]).
Should I transform the target values in the training data as well?
And how do I make sure that I transform the test data correctly so that the predictions corresponds as well as possible to the test set target values (which is not known)?
10:57:52	 From Carl-Johannes Johnsen : Tensorflow
10:58:18	 From Carl-Johannes Johnsen : Not the one I've been using at least!
11:00:42	 From Jonathan : it makes sense
11:00:49	 From Jonathan : thanks
11:01:05	 From Troels Christian Petersen  To  Dina Rapp(privately) : Did you get it to work now?
11:02:08	 From Haider Fadil Ali Hussein Al-Saadi : is there any particular reason the second convolution layer is twice the size of the first? i tried changing it so the second layer has exactly as many filters as the first and it led to a relatively large decrease in accuracy even when somewhat conserving the amount of parameters by increasing other layers
11:03:30	 From Carl-Johannes Johnsen : For GPU training, there is Google Colab, where you can change the runtime to "GPU" (a tesla k80 if i remember correctly), which is a lot faster than their CPU. You can use a max of 12 hours at a time. 
11:05:32	 From Haider Fadil Ali Hussein Al-Saadi : these networks seem optimized for grad student search :)
11:05:37	 From Carl-Johannes Johnsen : I stand corrected, I got a P100 in colab :O
11:06:32	 From Simon Hilbard : Does anyone else get the error: AttributeError: module 'tensorflow' has no attribute 'confusion_matrix'
when runing the secound block in the "inference" part of the CNN code ? 
11:07:07	 From Haider Fadil Ali Hussein Al-Saadi : replace the session and confusion with this maybe
11:07:08	 From Haider Fadil Ali Hussein Al-Saadi : confusion = tf.math.confusion_matrix(labels=np.argmax(test_labels, axis=1), predictions=predictions, num_classes=num_classes)
conf_matrix = confusion.numpy()
11:10:08	 From Simon Hilbard : That solve the problem 
11:10:51	 From Simon Hilbard : thank you 
11:11:01	 From Aske R. : I get an NotImplementedError: eval is not supported when eager execution is enabled, is .numpy() what you're looking for?
11:11:13	 From Aske R. : when trying Haiders solution :/
11:11:25	 From Haider Fadil Ali Hussein Al-Saadi : that's why you do confusion.numpy()?
11:11:46	 From Haider Fadil Ali Hussein Al-Saadi : the original in the notebook does .eval
11:11:54	 From Haider Fadil Ali Hussein Al-Saadi : which should give you that error
11:12:11	 From Haider Fadil Ali Hussein Al-Saadi : but .numpy() is exactly what the error suggests you should do
11:12:14	 From Haider Fadil Ali Hussein Al-Saadi : which is very odd
11:12:21	 From Aske R. : missed that last line
11:12:22	 From Aske R. : thanks
11:16:14	 From Carl-Johannes Johnsen : I just got back from the ERDA team, and no we currently don't have any other machines with GPUs, other then the ones that require admission. We are working on another solution, which we hoped would be ready to use in this course, but sadly didn't finish in time. I would suggest that you use Google Colab, as they offer free GPUs. I also read that Kaggle also offers GPUs, but only for 30 hours a week.
11:23:20	 From Haider Fadil Ali Hussein Al-Saadi : its funny how that there is a 5 that sort of looks like a 6 that many of the networks get wrong. its somewhat consistent
11:25:53	 From Aleksandar Topic : That is a good point. Do you have suggestions for how to overcome this?
11:27:40	 From Aleksandar Topic : For example, what would happen if we added more augmented representations of this exact image to the dataset?
11:27:59	 From Haider Fadil Ali Hussein Al-Saadi : ah
11:28:39	 From Haider Fadil Ali Hussein Al-Saadi : then that image would be easier to explain with our network
11:31:53	 From Aleksandar Topic : Yes, or rather the network would have more reason to adapt to the image, instead of seeing it as an outlier. Remember that we want a general network, not a specific one.
11:33:40	 From Haider Fadil Ali Hussein Al-Saadi : i think we got introduced to a method where you make copies of wrong predictions after each epoch, though i cant remember exactly when.
11:39:09	 From Andy Anker : Should this be understood as you predict on a combination of the raw data AND images made of the raw data? 
11:41:37	 From Aleksandar Topic : Yes so you expand only the training set to include these augmented images. It would not make sense to expand the testset, since we don't have this distribution.
11:42:17	 From Aleksandar Topic : So you actually increase the variance in your training data, which might needs to be reflected in the architecture.
11:44:20	 From Jonathan :  I'm getting an error when loading Keras
TypeError: expected bytes, Descriptor found
11:44:30	 From Jonathan : any ideas, anyone? :)
11:48:00	 From Runi : What does the Bias correction involve?