soumith

==> processing options
options are:
table: 0x41910ed8
==> loading dataset
==> preprocessing data
==> Preprocessing total time: 0.28143811225891
==> construct ReLU model
==> define loss
model add to cuda
==> defining some tools

soumith

overfeat schedule
=================
LR=0.01,epochNumber=1,weightDecay=5e-4
LR=0.005,epochNumber=20,weightDecay=5e-4
LR=0.001,epochNumber=33,weightDecay=0
LR=0.0005,epochNumber=48,weightDecay=0
Run till epoch 76

Epoch logs
=============

soumith

~/code/lstm$ th main.lua
Loading ./data/ptb.train.txt, size of data = 929589
Loading ./data/ptb.valid.txt, size of data = 73760
Loading ./data/ptb.test.txt, size of data = 82430
Using 1-th gpu
Network parameters:
{
  layers : 2
  lr : 1
  max_max_epoch : 13

soumith

----------------------------------------------------------------------
-- CIFAR 8x8
opt.scale = 8 
opt.geometry = {3, opt.scale, opt.scale}
local input_sz = opt.geometry[1] * opt.geometry[2] * opt.geometry[3]
local numhid = 600 
model_D = nn.Sequential()
model_D:add(nn.Reshape(input_sz))
model_D:add(nn.Linear(input_sz, numhid))
model_D:add(nn.ReLU())

soumith

   model_G = nn.Sequential()
   model_G:add(nn.JoinTable(2, 2))
   model_G:add(cudnn.SpatialConvolutionUpsample(3+1, 64, 7, 7, 1, 1)):add(cudnn.ReLU(true))
   model_G:add(nn.SpatialBatchNormalization(64, nil, nil, false))
   model_G:add(cudnn.SpatialConvolutionUpsample(64, 368, 7, 7, 1, 4)):add(cudnn.ReLU(true))
   model_G:add(nn.SpatialBatchNormalization(368, nil, nil, false))
   model_G:add(nn.SpatialDropout(0.5))
   model_G:add(cudnn.SpatialConvolutionUpsample(368, 128, 7, 7, 1, 4)):add(cudnn.ReLU(true))
   model_G:add(nn.SpatialBatchNormalization(128, nil, nil, false))
   model_G:add(nn.FeatureLPPooling(2,2,2,true))

soumith

2015-03-28 22:57:04.475	Epoch: [1][10000/10000]	Time 0.304 DataTime 0.005 Err 3.9184 	
2015-03-28 22:57:04.476	==> Validation epoch # 1	
2015-03-28 22:58:07.329	json_stats: {"learningRate":0.25,"batchSize":128,"train_loss":4.7269560290813,"manualSeed":1,"msra_mul":0,"decay":0.5,"backend":"cudnn","epochSize":10000,"nDonkeys":16,"train_time":3231.9889090061,"test_accuracy":14.646,"weightDecay":0.0005,"epoch":1,"nEpochs":30,"model":"alexnetowtbn","momentum":0.9,"test_loss":4.6024402600098,"GPU":1,"retrain":"","train_accuracy":13.362265625,"test_time":62.852720022202,"best_accuracy":0,"bestAccuracy":0,"nGPU":4}	
--
2015-03-28 23:53:36.332	Epoch: [2][10000/10000]	Time 0.286 DataTime 0.006 Err 3.7128 	
2015-03-28 23:53:36.332	==> Validation epoch # 2	
2015-03-28 23:54:37.711	json_stats: {"learningRate":0.25,"batchSize":128,"train_loss":3.6877586714745,"manualSeed":1,"msra_mul":0,"decay":0.5,"backend":"cudnn","epochSize":10000,"nDonkeys":16,"train_time":3214.1851511002,"test_accuracy":16.416,"weightDecay":0.0005,"ep

soumith

--[[
   This file implements Batch Normalization as described in the paper:
   "Batch Normalization: Accelerating Deep Network Training
                         by Reducing Internal Covariate Shift"
                by Sergey Ioffe, Christian Szegedy

   This implementation is useful for inputs coming from convolution layers.
   For Non-convolutional layers, see BatchNormalization.lua

   The operation implemented is:

soumith

-- Torch Android demo script
-- Script: main.lua
-- Copyright (C) 2013 Soumith Chintala

require 'torch'
require 'cunn'
require 'nnx'
require 'dok'
require 'image'

soumith

-- multiple learning rates per network. Optimizes two copies of a model network and checks if the optimization steps (2) and (3) produce the same weights/parameters.

require 'torch'
require 'nn'
require 'optim'
torch.setdefaulttensortype('torch.FloatTensor')

-- (1) Define a model for this example.
local model = nn.Sequential()
model:add(nn.Linear(10,20))

soumith

[
   {
      "oC":32,
      "name":"conv",
      "strideH":2,
      "kW":3,
      "iC":3,
      "padW":0,
      "strideW":2,
      "padH":0,