This is a nice opportunity to contribute to fastai as it seems one needs a test directory… A nice functionality would be to pass a pandas dataframe for test predictions.
thank you very much!
Does cleaning up over here means removing irrelevant images?
Why is this error coming ?
RuntimeError: DataLoader worker (pid 1100) is killed by signal: Bus error. Even on Pytorch forums, this issue hasn’t been resolved. How to set higher shared memory to avoid this error ?
1 Like
I have been lagging a bit due to my College exam’s . I am now only half way through the first video lecture. Is it okay, if binche and do stuff soon in upcoming weeks?
The University of San Francisco deserves much GLOBAL praise and admiration for hosting Jeremy’s course; it demonstrates great and genuinely progressive community spirit and scientific ethic based behavior in supporting the Fast AI course, content presentation and hosting logistics.
At the risk of embarrasing Jeremy etal, this is a cut and paste from another thread re The Master (Jeremy).
“Is there a general thread where I can sing Jeremy’s (and Rachael’s) praises? I think thus should be done from the global AI rooftops. He , virtually alone, has revived Ai to a fabulous state, translated it from ‘nerd-speak’, worked like a deamon in proving himself and his methods (non-confirmist, and for good reason), wrestled the AI from the strict realms if the academics, read heaps of papers and sorted them, cut through the in-bread jargon, prepared superbly understandable lessons and videos into THE premier AI course and presents it FREE for us unclean masses in the fabulous MOOC. WHAT A GREAT GUY! And, of course, he is Australian, AND from Melbourne. Thanks also to the Sanfrancisco University for supporting him in this Globally Disruptive Technological thrust. This will inspire global changes to many things, you ain’t seen nothin’ yet.
I’ll try to keep you informed if any significant progress on my little desktop Ilia.
Cheers for now,
Peter Kelly”
Well done to all; look as hard as you may and you willnot find anyone who could not gain from this specracular enterprise. Great stuff.
Cheers, p
12 Likes
It seems that the library is not using bcolz anymore, does anyone knows why?
edit: It seems the library is no longer storing activations to train just the classifier head (is this correct?)
When checking the length of fnames in Oxford IIIT Pet Dataset the len of it comes out to be 7390 but actual size of dataset is only 7349 as mentioned here http://www.robots.ox.ac.uk/~vgg/data/pets/
Is the len right ??
Yeah it does, there is no harm in doing so (assuming that when we speak of cleaning we are talking about doing so to the training set).
1 Like
you can use reflect to prevent black borders.
def pad(x, padding:int, mode='reflection'):
"Pad `x` with `padding` pixels. `mode` fills in space ('zeros','reflection','border')."
mode = _pad_mode_convert[mode]
return F.pad(x[None], (padding,)*4, mode=mode)[0]Learner.get_preds
It has always been in the lib. Learner.get_preds
1 Like
where is the link to join lesson 3?
3 Likes
Here it is:
https://www.youtube.com/watch?v=VPg2ZlRPiXI
Why is it that people generally discourage pytorch models for production?
1 Like
Part two was mentioned a couple times in lecture 3. Is there a approximate data for when it is supposed to be held?
Does .fit_one_cycle() cycle the learning rate over each epoch or just one “up” and “down” over the course of all the epochs specified in the fucntion?
1 Like
It’s just one up and one down over the course of all the epochs
Thanks, I figured based on the plots shown in lesson 3, but wanted to confirm since I sometimes fit a few epochs at a time.
is there any way to load image data from dataframe containing image ids file without labels?