Lesson 8 in-class


(Kent) #21

batcher.py.zip can’t be downloaded.


(RENJITH MADHAVAN) #22

batcher.py.zip
is not downloading

python3) renjith@dlp20db:~/datascience/kaggle/input/fastai2$ wget http://www.platform.ai/part2/lesson1/batcher.py.zip
–2017-02-27 20:53:28-- http://www.platform.ai/part2/lesson1/batcher.py.zip
Resolving www.platform.ai (www.platform.ai)… 67.205.12.187
Connecting to www.platform.ai (www.platform.ai)|67.205.12.187|:80… connected.
HTTP request sent, awaiting response… 500 Internal Server Error
2017-02-27 20:53:29 ERROR 500: Internal Server Error.


(RENJITH MADHAVAN) #23

Tried again and it worked.


(Roy) #24

wget http://www.platform.ai/part2/lesson1/DCGAN.ipynb
wget http://www.platform.ai/part2/lesson1/batcher.py.zip
wget http://www.platform.ai/part2/lesson1/dcgan.zip
wget http://www.platform.ai/part2/lesson1/neural-style.ipynb
wget http://www.platform.ai/part2/lesson1/torch_utils.zip
wget http://www.platform.ai/part2/lesson1/utils2.zip
wget http://www.platform.ai/part2/lesson1/vgg16_avg.zip
wget http://www.platform.ai/part2/lesson1/wgan-pytorch.ipynb


(Benedikt S) #25

@jeremy @rachel : Can you create a mailing list / resend a email to the mailing list? I haven’t received the first email about class information only the international fellowship email.


(Thundering Typhoons) #26

How many weeks is this course? Is it the same as deeplearning1 (7 weeks)?


(Rachel Thomas) #27

@benediktschifferer I’m having trouble with the internet, but there is a link in the first Lesson 8 Discussion post


(Rachel Thomas) #28

@thunderingtyphoons Yes


(Kaushik Patnaik) #29

Any thoughts around convolutional layers are being used for sequences also. Work from Deepmind around wavenet, pixelCNN, bytenet etc has outperformed LSTM based methods.


(sai kiran) #30

What is ordered data for which CNNs are used for? Any example?


(Rachel Thomas) #31

@benediktschifferer Here’s the intro email: http://us14.campaign-archive1.com/?u=11aeb8f4b5b91fd0da704eb1a&id=4a942ce016


(Kishore P. V.) #32

I got this random question in mind. How does having duplicates in training data affect the model created?
This question was as a result of tinkering with data augmentation: realized that after a few epochs there is likely duplicate data created, if the image data generator is used to augment data.


(Even Oldridge) #33

The steps to avoid overfitting were super helpful. I did run into issues regarding underfitting and I’d love to see a similar list for that.

Train longer and increase model complexity seem to be the main methods, but are there others?


(Thundering Typhoons) #34

Is it possible to post the slides?


(sai kiran) #35

When you augment the data, it is not the exact image you are passing. You are streching, zooming, move it to the right, left etc.
So it’s basically different pictures.


#36

Where is the link to the youtube stream?


(Rachel Thomas) #37

@superguy Are you one of the remote/international fellows? We are sharing it as needed


#38

I’m in the class, but it’s not that easy for me to see some of the smaller text so thought I could watch it in class at the same time


(RENJITH MADHAVAN) #39

Is opencv something used commonly. It took a lot of manual effort for me to configure opencv in python 3.6


(Kishore P. V.) #40

I understand that.For instance if you use just a few data augmentation parameters, like say just changing the width of the image. Running for a few epochs can end up in creating a sample that has the same width shift. This is less of a problem if more augmentation parameters are used.