search

maxpumperla / deep_learning_and_the_game_of_go

Code and other material for the book "Deep Learning and the Game of Go"
https://www.manning.com/books/deep-learning-and-the-game-of-go
987 stars 390 forks source link

Issue 12 v2 #18

Closed jeffhgs closed 5 years ago

jeffhgs commented 5 years ago

Tested on python 2.7.15 and python 3.6.8.

The code is based a different idea: don't reinvent the wheel that Keras invented to distribute the mnist dataset. The Keras dataset loads mnist from its s3 object, and it works in python 2 and python 3. Its file format turns out to also be a variation on pickle, but their file is created in a way that both python 2 and python 3 unpicklers grok by default. There are several minor differences, and I'm not sure which tricks are the important one. I copied the few lines of their code for dealing with the way they format their dictionaries and checked in their version of the file.

jeffhgs commented 5 years ago

I see this as a bugfix, so I have based on master and sent to master.

maxpumperla commented 5 years ago

makes sense, thanks. although, keep in mind that we're not reinventing the wheel as much as we're acting as if we're inventing it. that's a big difference. Chapter 5 is built up from first principles, no external libraries apart from numpy. and one of the killer features of chapter 6 is to see how incredibly convenient all this suddenly becomes with Keras. In particular, loading data is messy. Same goes for the Go data in later chapters. We could also just provide readers with fully prepared .npy data and leave out the full data module, but that's very misleading. Preparing the data is the hard part.