The $300 credit is direct from google and is not restricted to fast.ai - https://cloud.google.com/free/
Current fast.ai GCP setup guide http://course-v3.fast.ai/start_gcp.html
1 Like
Note that this is outdated - please use the setup guide from the course site.
In general Resnet model was trained on two sizes 224 and 299 images of course you can use any size
1 Like
Can someone tell me what does data.c from the notebook mean? Number of classes?
Check out the lesson video and see if you can find this section - then tell us what you learn! 
3 Likes
Hi @johnhartquist,
I’m thinking about a music chord recognition application. Do you have some advice for me how to deal with that problem, please ?
I am looking for a data set about this but I think this field is quite lacking of resources compare to computer vision.
This is for my own passion. I really like to play guitar but just impossible to recognize which chord I should use when I listen to a song. I have to search for it on the internet and sometime their chord is bad. If there are such an application, it is very useful for some amateur guitarist, or to improve learning music by ear.
Thank you in advance 
5 Likes
Just wanted to know if a Kaggle kernel could be used to run Jupyter and the fastai library required for this course?
It can’t.
Have you given your credit card info and registered with GCP?
Very interesting question ! I saw that you had a very good result in your classification x-ray images using transfer learning from imagenet. I tried the same approach with parasite images and had good results too. Very curious why pretrained model from imagenet work so well (even with very different distribution of images) in the next parts
1 Like
Yes, I did. I already had an account, but not GCP because I tested some of the speech APIs. It is linked to my credit card.
Please move setup questions to the platform specific topics.
1 Like
Many thanks for the clarification, Jeremy
Incredible course content! Thrilled to be part of the community.
1 Like
Thanks!
@rachel I think this question got buried before it reached the 6 required votes during class, but would love to hear your answer to this question! Thank you.
Point taken; thanks.
It should be fine, unless there is some explicit contract you’ve signed to the contrary. Although I’m not sure the law in this area has been fully tested.
1 Like
I have been playing with Kaggle’s Plants Seedlings Classification challenge. When preparing the submission, I faced two challenges with the ImageDataBunch class (both were solved with work arounds, but probably there are more elegant ways):
-
Some of its factory methods have a valid_pct argument to specify the porcentual size of the validation set. Just before submission, I want to train the model with the entire dataset (without validation set). However, if valid_pct is set to 0, an exception is thrown. The work around around was to set it to a very low value like 0.001.
-
In order to create the submission file, the test set filenames are required. However, the ImageDataBunch doesn’t have a property to access the test dataset (unlike train_ds / valid_ds). Additionally, I couldn’t find a simple way to get the files names. In this case, the work around was to access the dataset through the dataloader and then its x property (data.test_dl.dataset.x). I am not sure if there is some other straightforward way or if some properties are missing (like data.test_ds.filenames).
9 Likes
In the lecture it was mentioned that, there will be information provided for creating custom dataset using Google images. Has anyone tried to do it ?
