[lisa-lab/pylearn2] b05eab: Bugfixes in SVHN and DenseDesignMatrixPyTables

[GitHub]

Branch: refs/heads/master
Home: https://github.com/lisa-lab/pylearn2
Commit: b05eabb8e8600dce1d625c9ffdde46c8f1b81caa
https://github.com/lisa-lab/pylearn2/commit/b05eabb8e8600dce1d625c9ffdde46c8f1b81caa
Author: Sandeep Bhat <sandee...@gmail.com>
Date: 2015-05-21 (Thu, 21 May 2015)

Changed paths:
M pylearn2/datasets/dense_design_matrix.py
M pylearn2/datasets/svhn.py
M pylearn2/devtools/tests/test_format.py
A pylearn2/scripts/datasets/make_svhn_pytables.py

Log Message:
-----------
Bugfixes in SVHN and DenseDesignMatrixPyTables

[fix] split_train_valid method in SVHN class is expected to return two
tuples, each with two values. But it was returning 4 values. This was
wrongly introduced during commit
af07b6cc076942ef4f80c84ec7c7148afb3999ad ("remove one_hot, first try...")
[fix] Added support to specify y_labels in DenseDesignMatrixPyTables
constructor.This should have been done as part of the following commits
af07b6cc076942ef4f80c84ec7c7148afb3999ad ("remove one_hot, first try...")
d6050106f137147ba976e472748ad77273218b55 ("Modified all super() params...")
[fix] Table shape for y (the labels) is now set to [sizes[which_set], 1].
Before it was being wrongly set to [sizes[which_set], 10]. Also y
labels for SVHN now range from [0,9], similar to the MNIST dataset
labels. These changes should have been done as part of commit
af07b6cc076942ef4f80c84ec7c7148afb3999ad ("remove one_hot, first try...")
[fix] Now the SVHN dataset class properly sets the filters attribute
before call to resize().
[chg] Removed hardcoded title specification in init_hdf5(). It can now
be specified by an optional argument. SVHN class now sets the optional
title argument.
[chg] If dataset y is of integer type then int type atom is used by
init_hdf5() for storing the labels in pytables.
[new] Added ./pylearn2/scripts/datasets/download_svhn.sh script for SVHN
dataset download to the $PYLEARN2_DATA_PATH/SVHN/format2 directory. The
script is closely based on download_cifar10.sh, but it uses a verbose
wget or curl.
[new] Added ./pylearn2/scripts/datasets/make_svhn_pytables.py which
converts the download MAT files for SVHN dataset to Pytables format.
[note] The above fixes have been tested using SVHN YAML files in
./pylearn2/scripts/papers/maxout as follows:
[x] Setup your theano/pylearn2 to use GPU and set config.floatX = 'float32'
[x] Set the PYLEARN2_DATA_PATH environment variable.
[x] Download the SVHN dataset MAT files by running
$ ./download_svhn.sh
[x] Generate the necessary Pytables files by running
$ python make_svhn_pytables.py
[x] Copy the generated dataset to an alternate path say /tmp/SVHN
[x] Set the SVHN_LOCAL_PATH environment variable to this path
$ export SVHN_LOCAL_PATH=/tmp/SVHN/
[x] Pre-process the dataset by running
$ python svhn_preprocessing.py
[x] Train on the processed dataset by running
$ THEANO_FLAGS="device=gpu,floatX=float32" python train.py svhn.yaml
[x] For the first 10 iterations I got the following test_y_misclass
test_y_misclass: 0.923491537571
test_y_misclass: 0.804302871227
test_y_misclass: 0.0927878990769
test_y_misclass: 0.0584590695798
test_y_misclass: 0.0481450334191
test_y_misclass: 0.0444889366627
test_y_misclass: 0.0405634306371
test_y_misclass: 0.0373691543937
test_y_misclass: 0.0350215472281
test_y_misclass: 0.0334051698446


Commit: 2516e631a808078d86a2ec1f7ebba5deda30483b
https://github.com/lisa-lab/pylearn2/commit/2516e631a808078d86a2ec1f7ebba5deda30483b
Author: Sandeep Bhat <sandee...@gmail.com>
Date: 2015-05-21 (Thu, 21 May 2015)

Changed paths:
M pylearn2/datasets/dense_design_matrix.py
M pylearn2/datasets/svhn.py
M pylearn2/scripts/datasets/make_svhn_pytables.py

Log Message:
-----------
SVHN/DenseDesignMatrixPyTables fixes pep8 compliant

[fix] Added check in init_hdf5() to ensure y_dtype is either
'float' or 'int' only.
[fix] split_train_valid() in svhn.py is now using the passed path
argument instead of the hardcoded SVHN.data_path.
[chg] Minor edits for pep8.


Commit: cf126221c1f3ba34661c499a627e7f0de4597f35
https://github.com/lisa-lab/pylearn2/commit/cf126221c1f3ba34661c499a627e7f0de4597f35
Author: Sandeep Bhat <sandee...@gmail.com>
Date: 2015-05-21 (Thu, 21 May 2015)

Changed paths:
M pylearn2/scripts/datasets/make_svhn_pytables.py

Log Message:
-----------
Skip make_svhn_pytables testing if data path is not found

[chg] make_svhn_pytables.py script now raises SkipTest to skip testing in the
absence of environment variable PYLEARN2_DATA_PATH


Commit: 16fe83292f8a155ee231f2d268e7de86b603c40d
https://github.com/lisa-lab/pylearn2/commit/16fe83292f8a155ee231f2d268e7de86b603c40d
Author: Sandeep Bhat <sandee...@gmail.com>
Date: 2015-05-21 (Thu, 21 May 2015)

Changed paths:
M pylearn2/scripts/datasets/make_svhn_pytables.py

Log Message:
-----------
Added make_svhn_pytables.py to whitelist_docstrings

[chg] Whitelisted make_svhn_pytables.py to skip docstrings testing on Travis CI.
[chg] Added check to ensure 'PYLEARN2_DATA_PATH' is defined as an environment
variable before conversion to pytables.


Commit: d7c3fa67ccce5f79dd4bbc7e767be3fc3bc61753
https://github.com/lisa-lab/pylearn2/commit/d7c3fa67ccce5f79dd4bbc7e767be3fc3bc61753
Author: Pascal Lamblin <lamb...@iro.umontreal.ca>
Date: 2015-06-03 (Wed, 03 Jun 2015)

Changed paths:
M pylearn2/datasets/dense_design_matrix.py
M pylearn2/datasets/svhn.py
M pylearn2/devtools/tests/test_format.py
A pylearn2/scripts/datasets/make_svhn_pytables.py

Log Message:
-----------
Merge pull request #1371 from sandeepkbhat/pytables_svhn_bugfixes

Bugfixes in SVHN and DenseDesignMatrixPyTables


Compare: https://github.com/lisa-lab/pylearn2/compare/393a75583e20...d7c3fa67ccce