calculatedcontent / big_deep_simple_mlp Goto Github PK

create setup.sh

• download all files
• install infimnist
• setup ENV vars

See line #3:

https://github.com/CalculatedContent/big_deep_simple_mlp/blob/master/infimnist.py#L3

I would like to have a simple, native, modern MLP (in parallel to tensorflow) because it is easier to get inside the code and monitor the training, gradients, jacobian, etc.

So I tried to install sklearn 0.18 using pip and run the neural network MLP

note: can't use anaconda for local packages
sklearn 0.18dev is needed for the sklearn MLP

but it is not easy to install on the Mac OSX

http://stackoverflow.com/questions/33568244/upgrade-to-dev-version-of-scikit-learn-on-anaconda

My approach is

1. pip uninstall sklearn
2. build sklearn 0.18dev from source
3. install sklearn 0.18dev
4. test on command line, and in notebook

if it fails, reinstall sklearn 0.17 (stable) with pip

1. pip uninstall sklearn ; pip install -U sklearn

I can build and install from source on the Mac , but it fails to link something at runtime

sklearn Library not loaded: libmkl_intel_lp64.dylib

a clue from Google...but I have not explored and tried yet

https://groups.google.com/a/continuum.io/forum/#!topic/anaconda/F8Q-8xyvrks

"it turned out to be a trivial linking issue.
Don't ask me why but something seems to have messed up my paths.
I fixed the issue on my box with some brute force.
Symbolically linking all anaconda libraries into my homebrew library folder did the trick.
for lib in ~/anaconda/lib/*; do ln -s $lib /usr/local/lib/$(basename $lib); done"

integrate unit tests into notebook

Trying to install sklearn 0.18 using pip and run the neural network MLP

note: can't use anaconda for local packages
sklearn 0.18dev is needed for the sklearn MLP
but it is not easy to install

http://stackoverflow.com/questions/33568244/upgrade-to-dev-version-of-scikit-learn-on-anaconda

I can build and install from source on the Mac , but it fails to link something at runtime

sklearn Library not loaded: libmkl_intel_lp64.dylib

a clues from Google i have not explored and tried yet

https://groups.google.com/a/continuum.io/forum/#!topic/anaconda/F8Q-8xyvrks

"it turned out to be a trivial linking issue.
Don't ask me why but something seems to have messed up my paths.
I fixed the issue on my box with some brute force.
Symbolically linking all anaconda libraries into my homebrew library folder did the trick.
for lib in ~/anaconda/lib/*; do ln -s $lib /usr/local/lib/$(basename $lib); done"

it is soo slow...can it be sped up ? perhaps by pre-generating the epoch files ?

TODO:

remove input_data.py

merge alignmnist.ipynb and alignmnist.py using notebook read/write

=> must create mnist.py , with same behavior, and pre-download files

i

Keras just blogged about their ImageDataGenerator class, which is similar in spirit to InfiMNIST and AlignMNIST

http://blog.keras.io/building-powerful-image-classification-models-using-very-little-data.html

I have no idea how it works or how it compares. It would be interesting to see how it compares

the MLPs being run so far are very simple

a modern implementation can be found in

https://github.com/terryum/TensorFlow_Exercises/blob/master/3b_MLP_MNIST_Modern_160517.ipynb

it seems like a next step would be to test these implementations and try to get as close to the best accuracy possible

just

trX, trY, teX, teY = alignmnist.next_epoch()

logging shows the increments are off...need to add logging and debug carefully