cnn_alexnet

Experiment on AlexNet (Krizhevsky, Alex, Ilya Sutskever, and Geoffrey E. Hinton. "Imagenet classification with deep convolutional neural networks." Advances in neural information processing systems. 2012.)

Dataset: wget https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz tar zxvf cifar-10-python.tar.gz wget https://www.dropbox.com/s/8ei5c8dqlro501o/alexnet_place.npy

    Download and decompress cifar 10 dataset (163MB)
    Download pre-trained AlexNet by imagenet (233MB)

    10 classes: airplance, automobile,
    bird, cat, deer, dog, frog, horse,
    ship, truck

    50000 traning images: 40000 for training and 10000 for validation
    10000 testing images

Training: CUDA_VISIBLE_DEVICES=0 python train.py

    (0 means the GPU's id, since tensorflow use all gpu by default)
    --model: use which model to test (deafult model/model_best)
    --batch_size: batch size for training (default 64)
    --lr: learning rate for loss (default 0.00001)
    --test: test or train (default train)

    time: 3mins / epo

Testing: CUDA_VISIBLE_DEVICES=0 python train.py --test

    --model means use which model to test (deafult model_best)

    time: 45 secs

Environment: Software: on Debian GNU/Linux testing (stretch)

1. Python 2.7
2. tensorflow 0.10.0
3. numpy 1.12.0

4. cv2, IPython, cPickle

   Hardware: NVIDIA Tesla GPU K80 (4 core and 10GB memory)

   tensorflow installment:

   Ubuntu/Linux 64-bit, GPU enabled, Python 2.7

   export TF_BINARY_URL=https://storage.googleapis.com/tensorflow/linux/gpu/tensorflow-0.10.0-cp27-none-linux_x86_64.whl pip install --upgrade $TF_BINARY_URL

Description: I have tried on oxford 17 first, but I think it is too small to prove the robustness. Therefore, for quick demo, I choice cifar10 dataset eventually. It's an appropiate size for proving the alexnet. And I use alexnet model which is pretrained by Imagenet for faster converaging and easy fine-tuning. Training code would produce model after epo ending, then we can use testing code to check the accuracy. Training code would print epo and loss message on the screen.

Epo Accuracy

1 0.21 2 0.533738057325 3 0.662320859873 4 0.711186305732 5 0.728503184713 6 0.748606687898 7 0.765525477707 8 0.765824044586 9 0.777368630573 11 0.79090366242 12 0.796476910828 14 0.797969745223 16 0.799064490446 17 0.801652070064

In my environment, this model achieves accuracy 0.8 in an hour. (3min/epo)

code: train.py: training and testing model.py: model structure of AlexNet data.py: Dataset for random sampling and data format