CHAPTER 2 Fastbook

Dear all,
I am running the Bing Search Image and even if I copied the Key I got this error: Access Denied, any ideas?

I checked also others topic from previous years, but I am not sure if it would be appropriate to share the code of the fastbook with them at the moment directly in the forum, I saw several useful posts.
Let me know if you have the same issue

Best to just use this category for stuff we’re studying in the course.

Note that we’ll be studying this notebook in the next lesson BTW (but skipping ahead is fine too!)

1 Like

Thank @jeremy, yes it is such an amazing book and course, I am too excited.

I was playing with Data Augmentation with Cat / Dog in the Chapter One.
Not sure what I am doing wrong. Is the operation too simple for our NN that it refuses to work?!

I am definetely doing something wrong

Why are all of them labeled “images” ( this is your y’s that you are predicting)
looks like there is only one image (not sure though, hope that helps)

you can check your classes/labels with - dls.train.vocab (see what those are)

1 Like

lookin at this image, you can see that you have only one class. :slight_smile:
you are missing something here:
bear_types = 'grizzly','black','teddy'

1 Like

awesome catch, @barnacl I had the get_y=parent_label and not label_func=is_cat.
Thank you so much


After Data Augmentation

It is raise the number of False Positive.
from 1 --> 5

Is it because we cheated basically in the NN?

You didn’t do a fair comparison, if you notice your second one runs a resnet18 :wink:


Thank @muellerzr,
Yes it was a huge difference with resnet34

Do you know where can I find more material to understand the benefit of Data Augmentation and its correlation with Confusion Matrix?

The error rate started higher and it drops quicker.
I will analyse also that curve of the error rate, doing a comparison.

Besides Jeremy teaching about data augmentation, there are a number of papers that discuss the idea and why it works. However here is a paper discussing it:

1 Like

A general understanding of why though can be summed up like this:

Our dataset will realistically always be limited. There is a limited number of samples, and these samples are built and shown in a particular fashion. But in the real world it won’t look this way, there are images that we will see that was never seen in our training data. How do we deal with this? NN’s are great at generalizing, but data augmentation can take us a step further. By augmenting our data, we provide the network a larger corpus of data to train on, as an image that’s rotated 45 degrees is quite different from it’s regular image, and in the case of RandomCrop, the particular cropped region could make certain features larger. The concept is further explored in Chapter 5 as well :slight_smile: (See the “presizing” section)

1 Like

You could also try this Visual GUI as a way to play around with image augmentations which was recently updated to v2 with more functionality to come.


This is nice visualization of different image augmentations


This is awesome work @amritv thank you so much, I will try for sure

1 Like

I have the same issue too

Dear @jeremy, I am still having the same issue. I recall that you mentioned there is a step by step tutorial available, am I right?

I personally:

  • Create Azure account
  • Create Cognitive Service account and resource
  • Open the 7days trial with Image Search
  • Copy paste the key

Byt still “Access Denied”

Please don’t at-mention me for questions unless it’s something specific to me.

You need to go to your Azure portal and get the key from there. That’s a different key from the 7-day-trial key which has already expired.