Lesson 11 discussion and wiki

Part 2 (2019)
61

We are normalizing those after activation: if you look at the list of layers inside mods, those are conv-relus.

62

would LSUV also work with RNN?

1 Like
63

thanks!
so if we have a not conv model, i.e. simple FC, we should create modules that aggregate the linear layer with the nonlinearity on order for this to work as expected?

64

Try and tell us :slight_smile:

2 Likes
65

That would be the best yes, create a basic block with those, then replace the isinstance(o, ConvLayer) a few cells before by isinstance(o, YourBlock).

66

Is it not expensive to repeat those transforms every time we access one or more images in the ImageList? Or do we suppose items only need to be accessed (and transformed) once?

67
1 Like
68

You get them when you need to access them for training, only once each per epoch. We care less about the speed otherwise.

1 Like
69

Any gotchas to be aware of or heuristics for deciding on order of image transforms?

1 Like
70

Oh, I remember I had a very rough time trying to figure out where the error comes from when had a couple of gray-scale images in my RGB pipelineā€¦

1 Like
71

Look at the end result and see if the result makes sense :wink:

1 Like
72

Jeremy will go into much more detail on image augmentation/transformations later in this lesson.

4 Likes
73

We have learned from all the annoyed users and what they told us they couldnā€™t be doing :wink:

3 Likes
75

Will we see part 3 of the course? Would be great!

76

In case anyone else was wonderingā€¦

7 Likes
77

any precautions that is needed , if we are doing a segmentation ?

78

You just need to make sure you apply the same transforms for data augmentation to your inputs and targets.

79

And if you are better served with videos these are good.

2 Likes
80

where was ListContainer implemented ?

81

It might be worth mentioning that uniquefy is order-preserving ā€“ otherwise you could just do set(list(x)) or something similar.

3 Likes