Lesson 1 In-Class Discussion

(yinterian) #108

! pip install isoweek

(Kousik) #109

(Brian Muhia) #110

Add these lines to `requirements.txt` and do `pip install -r requirements.txt`

``````isoweek>=1.3.3
pandas_summary>=0.0.41
torchtext>=0.2.0
``````

Edited to fix pandas_summary. Thanks @TheLariat!

(yinterian) #111

Pre-prosessing image data is a very particular. Also, every model needs a slightly different pre-processing. That is why it is done by the library.

(Setu) #112

​conda install﻿ -c conda-forge opencv

(Ike Okonkwo) #113

cv2 refers to openCV… you can install via conda or build from source

(Amrit ) #114

pip install opencv python

(yinterian) #115

conda install opencv

(Davide Boschetto) #116

Hey there. Going out of memory on my 970. Any guidance on how to approach this with this fastai library?

(Pete Condon) #117

Are we going to cover cyclic learning rates in this course, or leave it for part 2?

(Anand Saha) #118

My intuition is that:

Traditionally in ML, we have been coming up with our own approximate function (the model) that we think would mimic the function we are trying to predict. The shortcoming is that the approximate function we come up, is based on the best of our knowledge and it does not have the freedom to change it’s form. It’s form is fixed. For e.g. if we think the model for a particular problem should be f(x) = w1x^2 + w2x+b, the form remains so and never changes.

In deep learning, by stacking weighted sum of inputs and non-linear activation functions, we let the network figure out by itself what form the approximate function should be - it does so by adjusting it’s own weights. By adding more layers and nodes, and by choosing the right activation function, we give the network enough wiggle room to approximate any function -
given the loss function.

Makes sense?

(Rajat) #119

@jeremy How about if we have three classes unlike in class example in which probability of cat is close to 0 and for dog to 1. Then how can we detect which probability belong to which of three classes.

(yinterian) #120

Lowering the mini-batch size.

(Daniel Freire) #121

Hey guys, how can I install cv2?

(Pete Condon) #122

huh, I guess that question got answered quick …

(yinterian) #123

Jeremy will show you how to do it.

(Kishore P. V.) #124

@jeremy In the lecture you mention that you prefer to build a model first and then analyze the results to clean the data. This approach sounds good for images (and maybe audio and text too).
What about data which is just a bunch of features? Is it better to do the reverse approach in this case too?

(Ching June Hao) #125

I get this error when I launched my jupyter notebook.

Notebook validation failed
The save operation succeeded, but the notebook does not appear to be valid. The validation error was:

``````Notebook Validation failed: {u'model_id': u'43a3aea4b93c41128be460941019c404', u'version_minor': 0, u'version_major': 2} is not valid under any of the given schemas:
{
"model_id": "43a3aea4b93c41128be460941019c404",
"version_minor": 0,
"version_major": 2
}
``````

What should I correct in order to get this right?

(Ike Okonkwo) #126

install via conda or build from source

(Davide Boschetto) #127

Good, thanks! Was asking since there’s no “clear” parameter this time around for this