- I'll verify the HTM using the task of ImageNet LSVRC-2014.
- Dataset
- Install nupic/pylearn2
- Simple task
- The CIFAR-10 dataset consists of 60000 32x32 colour images in 10 classes, with 6000 images per class.
- The CIFAR-10 dataset
- The MNIST database of handwritten digits.
- THE MNIST DATABASE
-
What is ILSVRC2013
-
How to get images from imageNet
-
Get wnid list
- You need to prepare csv file of wnid.
- I created the csv file by shaping LSVRC page.(http://image-net.org/challenges/LSVRC/2014/browse-synsets)
- Here is sample csv file (data/classification_categorys.csv).
-
Run get_image_from_imagenet.py
- This script to get the image from imagenet.
- Simply, this script get image url by wnid, and download image file.
- I have excluded the following file.
- flicr not found image file
- cannot open file
- Too small file
cd data python get_image_from_imagenet.py
- install
- tutorial
- When I execute make_dataset.py, the error has occurred.
- I edited train_example_path of pylearn2/scripts/tutorials/grbm_smd/make_dataset.py.
IOError: permission error creating /Library/Python/2.7/site-packages/pylearn2/scripts/tutorials/grbm_smd/cifar10_preprocessed_train.pkl
- environment values
export PYLEARN2_DATA_PATH=/Users/karino-t/data export PYLEARN2_VIEWER_COMMAND="open -Wn"
- architecture
- input
- (3, 224, 224)
- convolutional layer
-
- 96 kernel, (11, 11) max-pooling
-
- 256 kernel, (5,5,48) max-pooling
-
- 384 kernel, (3,3,256) all connect with second layer
-
- 384 kernel, (3,3,192)
-
- 384 kernel, (3,3,192) max-pooling
-
- full connected layer
-
- 2048+2048 sigmoid?, tanh?, other?
-
- 2048+2048
-
- softmax 1000
-
- reducing overfiting
- data augmentation
- dropout (full connected layer ?)
- datamodel : cifar10.py
- preprocess : make_dataset.py
- model : conv_sample.yaml