Initial commit

Author: bigaidream

.gitignore

@@ -1,41 +1,10 @@
-# Compiled Lua sources
-luac.out
-
-# luarocks build files
-*.src.rock
-*.zip
-*.tar.gz
-
-# Object files
-*.o
-*.os
-*.ko
-*.obj
-*.elf
-
-# Precompiled Headers
-*.gch
-*.pch
-
-# Libraries
-*.lib
-*.a
-*.la
-*.lo
-*.def
-*.exp
-
-# Shared objects (inc. Windows DLLs)
-*.dll
-*.so
-*.so.*
-*.dylib
-
-# Executables
-*.exe
-*.out
-*.app
-*.i*86
-*.x86_64
-*.hex
-
+mnist_lr/mnist/logs/*.log
+mnist_minibatch/mnist/logs/*.log
+mnist_lr/dqn/mnist.t7/
+mnist_minibatch/dqn/mnist.t7/
+tmp/
+mnist_minibatch/dqn/logs/*.log
+mnist_minibatch/dqn/logs/*.pdf
+mnist_lr/mnist/mnist.t7/
+mnist_minibatch/mnist/mnist.t7/
+mnist_minibatch/dqn/logs/paint.py

LICENSE

@@ -1,6 +1,6 @@
 The MIT License (MIT)
 
-Copyright (c) 2016 bigAIdream projects
+Copyright (c) 2016 Jie Fu, Zichuan Lin, Miao Liu, Nicholas Leonard, Jiashi Feng, Tat-Seng Chua projects
 
 Permission is hereby granted, free of charge, to any person obtaining a copy
 of this software and associated documentation files (the "Software"), to deal

README.md

@@ -1,2 +1,57 @@
-# qan
-Deep Q-Networks for Accelerating the Training of Deep Neural Networks
+# Deep Q-Networks for Accelerating the Training of Deep Neural Networks
+
+Source code to the paper [Deep Q-Networks for Accelerating the Training of Deep Neural Networks](https://arxiv.org/abs/1606.01467)
+
+## Reproduce our results on MNIST
+
+### Dependencies
+We are using Torch. The DQN component is mostly modified from [DeepMind Atari DQN](https://github.com/kuz/DeepMind-Atari-Deep-Q-Learner). 
+
+You might need to run `install_dependencies.sh` first. 
+
+### Tuning learning rates on MNIST
+```bash
+cd mnist_lr/;
+cd mnist;
+th train-on-mnist.lua; #get regression filter, save in ../save/
+./run_gpu; #Start tune learning rate using dqn
+#To get the test curve, run following command
+cd mnist_lr/dqn/logs;
+python paint_lr_episode.py;
+python paint_lr_vs.py;
+```
+
+### Tuning mini-batch selection on MNIST 
+```bash
+cd mnist_minibatch;
+cd mnist;
+th train-on-mnist.lua; #get regression filter, save in ../save/
+./run_gpu; #Start select mini-batch using dqn
+#To get the test curve, run following command
+cd mnist_minibatch/dqn/logs;
+python paint_mini_episode.py;
+python paint_mini_vs.py;
+```
+
+### Different Settings
+1. GPU device can be set in `run_gpu` where `gpu=0`
+2. Learning rate can be set in `/ataricifar/dqn/cnnGameEnv.lua`, in the `step` function. 
+3. When to stop doing regression is in `/ataricifar/dqn/cnnGameEnv/lua`, in line 250
+
+## TODO
+1. Experiments on CIFAR
+2. Transfer learning
+
+## Citation
+```
+@article{dqn-accelerate-dnn,
+  title={Deep Q-Networks for Accelerating the Training of Deep Neural Networks},
+  author={Fu, Jie and Lin, Zichuan and Liu, Miao and Leonard, Nicholas and Feng, Jiashi and Chua, Tat-Seng},
+  journal={arXiv preprint arXiv:1606.01467},
+  year={2016}
+}
+```
+
+## Contact
+
+If you have any problems or suggestions, please contact me: jie.fu A~_~T u.nus.edu~~cation~~

cifar_lr/cifar.torch/LICENSE

@@ -0,0 +1,22 @@
+The MIT License (MIT)
+
+Copyright (c) 2015 Sergey Zagoruyko
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+

cifar_lr/cifar.torch/README.md

@@ -0,0 +1,33 @@
+# cifar.torch
+
+The code achieves 92.45% accuracy on CIFAR-10 just with horizontal reflections.
+
+Corresponding blog post: http://torch.ch/blog/2015/07/30/cifar.html
+
+Accuracies:
+
+ | No flips | Flips
+--- | --- | ---
+VGG+BN+Dropout | 91.3% | 92.45%
+NIN+BN+Dropout | 90.4% | 91.9%
+
+Would be nice to add other architectures, PRs are welcome!
+
+Data preprocessing:
+
+```bash
+OMP_NUM_THREADS=2 th -i provider.lua
+```
+
+```lua
+provider = Provider()
+provider:normalize()
+torch.save('provider.t7',provider)
+```
+Takes about 30 seconds and saves 1400 Mb file.
+
+Training:
+
+```bash
+CUDA_VISIBLE_DEVICES=0 th train.lua --model vgg_bn_drop -s logs/vgg
+```

cifar_lr/cifar.torch/example_classify.lua

@@ -0,0 +1,81 @@
+require 'image'
+require 'cudnn'
+require 'cunn'
+local tablex = require 'pl.tablex'
+
+if #arg < 2 then
+  io.stderr:write('Usage: th example_classify.lua [MODEL] [FILE]...\n')
+  os.exit(1)
+end
+for _, f in ipairs(arg) do
+  if not paths.filep(f) then
+    io.stderr:write('file not found: ' .. f .. '\n')
+    os.exit(1)
+  end
+end
+
+local model_path = arg[1]
+local image_paths = tablex.sub(arg, 2, -1)
+
+-- loads the normalization parameters
+require 'provider'
+local provider = torch.load 'provider.t7'
+
+local function normalize(imgRGB)
+
+  -- preprocess trainSet
+  local normalization = nn.SpatialContrastiveNormalization(1, image.gaussian1D(7)):float()
+
+  -- rgb -> yuv
+  local yuv = image.rgb2yuv(imgRGB)
+  -- normalize y locally:
+  yuv[1] = normalization(yuv[{{1}}])
+
+  -- normalize u globally:
+  local mean_u = provider.trainData.mean_u
+  local std_u = provider.trainData.std_u
+  yuv:select(1,2):add(-mean_u)
+  yuv:select(1,2):div(std_u)
+  -- normalize v globally:
+  local mean_v = provider.trainData.mean_v
+  local std_v = provider.trainData.std_v
+  yuv:select(1,3):add(-mean_v)
+  yuv:select(1,3):div(std_v)
+
+  return yuv
+end
+
+local model = torch.load(model_path)
+model:add(nn.SoftMax():cuda())
+model:evaluate()
+
+-- model definition should set numInputDims
+-- hacking around it for the moment
+local view = model:findModules('nn.View')
+if #view > 0 then
+  view[1].numInputDims = 3
+end
+
+local cls = {'airplane', 'automobile', 'bird', 'cat',
+             'deer', 'dog', 'frog', 'horse', 'ship', 'truck'}
+
+for _, img_path in ipairs(image_paths) do
+  -- load image
+  local img = image.load(img_path, 3, 'float'):mul(255)
+
+  -- resize it to 32x32
+  img = image.scale(img, 32, 32)
+  -- normalize
+  img = normalize(img)
+  -- make it batch mode (for BatchNormalization)
+  img = img:view(1, 3, 32, 32)
+
+  -- get probabilities
+  local output = model:forward(img:cuda()):squeeze()
+
+  -- display
+  print('Probabilities for '..img_path)
+  for cl_id, cl in ipairs(cls) do
+    print(string.format('%-10s: %-05.2f%%', cl, output[cl_id] * 100))
+  end
+end