Network.py provides a more polished version of the Network class from backprop2.py. Improvements are:

• Better documentation.
• Saving and loading.
• List-like operators (splicing, concatonating).

(The list-like operators are convenient for pretraining work.)

This is intended for an experimental branch, but github won't let me propose a pull request to a non-existant branch.

As they slow down the sigmoid functions considerably and unnecessarily. The network's SGD method runs several times faster with this simple modification.

I added the Git Attributes which will normalize all line enddings in the binary of all text files. (The enddings are different for linux, windows, mac, etc.) It's kinda sorta pointless, but it's a nice perk Git offers, so why not have it? It also provides a bunch of other things not really needed for this repo.

I also added basic operating system junk files to the gitignore (because they were going to be committed)

with theano upgrade to 0.8 or above,the function downsample.max_pool_2d() has been removed. Should replace with pool.pool_2d.Otherwise, it will crash.

nabla_b and nabla_w are lists which store gradient, but their initial values are not used. So it isn't needed to call np.zero to create vector/matrix.

No intended behavior change. The vectorize() documentation points out that vectorize() is not fast, and indeed after this change the time to train one epoch is reduced from ~37.6s to ~10.5s on my laptop -- a 72% reduction.

I haven't read more than Chapter 1 yet (in fact, I haven't completed Chapter 1), so maybe something like this can be done in other places too.

Hi thanks for this great learning resource, I think your writing is amazing.

I'm a complete newbie in python but I think this PR makes this bit of code less verbose.

I know it's not a big change but still... There might be a good reason this wasn't done in the first place, in which case please disregard this PR.

These are the basic modifications needed for network3.py to be compatible with the latest Theano 0.9.0 (which pip will automatically install right now). Possibly a mention should also be added to the README about Theano moving towards a new GPU backend (more details at http://deeplearning.net/software/theano/tutorial/using_gpu.html).

I've update some of the code that uses deprecated theano functions

• Replaced theano.tensor.signal.downsample.max_pool_2d function with the newer theano.tensor.signal.pool.pool_2d function (Python throws a warning when using the earlier one)
• Replaced the deprecated theano.tensor.nnet.conv.conv2d with the newer Theano.tensor.nnet.conv2d function. Also, replaced the deprecated image_shape parameter with input_shape .

Also, removed the redundant cost_ij variable in network3.py.

"downsample module has been moved to the theano.tensor.signal.pool module." so i modify the theano.tensor.signal.downsample.max_pool_2d() to theano.tensor.signal.pool.pool_2d()

While reading the code base, I noticed that you had lots of class declarations of the forms

class Foo():
    ....

class Foo:
    ....

My sense is that you're not planning on making any classes to be extensible via inheritance, but the more Pythonic way of doing this is to inherit the base object:

class Foo(object):

This permits inheritance, which the old-style classes (eg, class Foo:) preclude.

In the process, my editor, which aggressively trims trailing whitespace, discovered lots of this in your source files -- hence the larger-than-expected diff. I'm happy amend this commit to include only the class declaration changes and leave the whitespace alone if you're particular about these things. Lemme know.

This isn't meant to be a PR, so sorry for making one, but I didn't find another way to notify anyone who might be interested. As you may know, the current repository is made to work with python 2.7, which is old and incompatible. I tried searching for one which was compatible with the latest version but no luck(Michal Daniel Dobrzanski isn't the one I was looking for), so I decided to make one: https://github.com/unexploredtest/neural-networks-and-deep-learning Hopefully somebody will find it useful.

I am working on your online book and projects using Python 3.5. Since I used a different and newer python version, I came across several problems and solved them by making some changes in network.py and mnist_loader files. Now the code is compatible with Python 3.5 The biggest problem is zip variables haven't len() attribute in 3.5. So I converted training_data, test_data to the list variables. I had to make a deep copy of original training and test data because when they are converted to list variables, they can't be reused.

I am working on your online book and projects using Python 3.5. Since I used a different and newer python version, I came across several problems and solved them by making some changes in network.py and mnist_loader files. Now the code is compatible with Python 3.5 I added just "import _pickle as cPickle" line because in 3.5 cPickle isn't defined and it should be used in standard library version.