Add inception-resnet-v2 model

checkpoints.lua

@@ -25,7 +25,7 @@ function checkpoint.latest(opt)
    return latest, optimState
 end
 
-function checkpoint.save(epoch, model, optimState, bestModel)
+function checkpoint.save(epoch, model, optimState, bestModel, opt)
    -- Don't save the DataParallelTable for easier loading on other machines
    if torch.type(model) == 'nn.DataParallelTable' then
       model = model:get(1)
@@ -34,16 +34,16 @@ function checkpoint.save(epoch, model, optimState, bestModel)
    local modelFile = 'model_' .. epoch .. '.t7'
    local optimFile = 'optimState_' .. epoch .. '.t7'
 
-   torch.save(modelFile, model)
-   torch.save(optimFile, optimState)
-   torch.save('latest.t7', {
+   torch.save(paths.concat(opt.resume, modelFile), model)
+   torch.save(paths.concat(opt.resume, optimFile), optimState)
+   torch.save(paths.concat(opt.resume, 'latest.t7'), {
       epoch = epoch,
       modelFile = modelFile,
       optimFile = optimFile,
    })
 
    if bestModel then
-      torch.save('model_best.t7', model)
+      torch.save(paths.concat(opt.resume, 'model_best.t7'), model)
    end
 end
 

dataloader.lua

@@ -39,7 +39,11 @@ function DataLoader:__init(dataset, opt, split)
       end
       torch.setnumthreads(1)
       _G.dataset = dataset
-      _G.preprocess = dataset:preprocess()
+      if opt.netType == 'inception-resnet-v2' or opt.netType == 'inception-resnet-v2-aux' then 
+        _G.preprocess = dataset:preprocess(328, 299)
+      else
+        _G.preprocess = dataset:preprocess(256, 224)
+      end  
       return dataset:size()
    end
 

datasets/create-imagenet-lmdb.lua

@@ -0,0 +1,189 @@
+--[[
+  A script to conver images to lmdb dataset
+
+  References
+  1. https://github.com/facebook/fb.resnet.torch/blob/master/datasets/init.lua
+  2. https://github.com/eladhoffer/ImageNet-Training/blob/master/CreateLMDBs.lua
+]]--
+
+local ffi = require 'ffi'
+local image = require 'image'
+
+-- Define local functions
+local function isvalid(opt, cachePath)
+   local imageInfo = torch.load(cachePath)
+   if imageInfo.basedir and imageInfo.basedir ~= opt.data then
+      return false
+   end
+   return true
+end
+
+local function split(dataset)
+  local tokens = {}
+  for word in string.gmatch(dataset, '([^-]+)') do
+      table.insert(tokens, word)
+  end
+  assert(tokens[2] == 'lmdb', string.format('opt.dataset should be <datset>-lmdb form; opt.dataset = %s', dataset))
+  return tokens[1]
+end
+
+local function _loadImage(path)
+   local ok, input = pcall(function()
+      return image.load(path, 3, 'float')
+   end)
+
+   -- Sometimes image.load fails because the file extension does not match the
+   -- image format. In that case, use image.decompress on a ByteTensor.
+   if not ok then
+      local f = io.open(path, 'r')
+      assert(f, 'Error reading: ' .. tostring(path))
+      local data = f:read('*a')
+      f:close()
+
+      local b = torch.ByteTensor(string.len(data))
+      ffi.copy(b:data(), data, b:size(1))
+
+      input = image.decompress(b, 3, 'float')
+   end
+
+   return input
+end
+
+local function getItem(basedir, imageClass, imagePath, i)
+   local path = ffi.string(imagePath[i]:data())
+   local image = _loadImage(paths.concat(basedir, path))
+   local class = imageClass[i]
+
+   return {
+      input = image,
+      target = class,
+   }
+end
+
+
+
+-- Init opt 
+local opt = {}
+opt.gen = 'gen'
+opt.dataset = 'imagenet-lmdb'
+opt.data = '/media/data1/image/ilsvrc15/ILSVRC2015/Data/CLS-LOC'
+
+opt.data_lmdb = '/media/data1/image'
+opt.shuffle = true
+--print(opt)
+
+-- Load imageInfo
+local cachePath = paths.concat(opt.gen, split(opt.dataset) .. '.t7')
+if not paths.filep(cachePath) or not isvalid(opt, cachePath) then
+   paths.mkdir('gen')
+
+   local script = paths.dofile(split(opt.dataset) .. '-gen.lua')
+   script.exec(opt, cachePath)
+end
+local imageInfo = torch.load(cachePath)
+--print(imageInfo)
+
+-- Create LMDB
+local lmdb = require 'lmdb'
+
+local train_env = lmdb.env{
+    Path = paths.concat(opt.data_lmdb, 'train_lmdb'),
+    Name = 'train_lmdb'
+}
+
+local val_env= lmdb.env{
+    Path = paths.concat(opt.data_lmdb, 'val_lmdb'),
+    Name = 'val_lmdb'
+}
+
+local path = ffi.string(imageInfo.train.imagePath[1]:data())
+--local image = self:_loadImage(paths.concat(self.dir, path))
+--local class = self.imageInfo.imageClass[i]
+print(path)
+
+local n_images = (#imageInfo.train.imagePath)[1]
+print(string.format("n_image: %d", n_images))
+
+local idxs
+if opt.shuffle then
+  idxs = torch.randperm(n_images)
+else 
+  idxs = torch.range(1, n_images)
+end
+print(string.format("opt.shuffle: %s, idxs[1]: %d", opt.shuffle, idxs[1]))
+
+local basedir = paths.concat(imageInfo.basedir, 'train')
+--local item = getItem(basedir, imageInfo.train.imageClass, imageInfo.train.imagePath, idxs[1])
+--print(item.target)
+----print(item.input)
+--print(#item.input)
+--print(item.input[1][1][1])
+
+train_env:open()
+local txn = train_env:txn()
+local cursor = txn:cursor()
+for i = 1, 1000 do --n_images do
+    local item = getItem(basedir, imageInfo.train.imageClass, imageInfo.train.imagePath, idxs[i])
+
+    cursor:put(string.format("%07d", i), item, lmdb.C.MDB_NODUPDATA)
+    if i % 100 == 0 then
+        txn:commit()
+        print(train_env:stat())
+        collectgarbage()
+        txn = train_env:txn()
+        cursor = txn:cursor()
+    end
+    xlua.progress(i, n_images)
+end
+txn:commit()
+train_env:close()
+
+--[[
+local sys = require 'sys'
+local n_test = 5000
+sys.tic()
+-------Read-------
+train_env:open()
+print(train_env:stat()) -- Current status
+local reader = train_env:txn(true) --Read-only transaction
+--local y = torch.Tensor(10,3,256,256)
+local y = {}
+
+local idxs = torch.randperm(n_test)
+for i=1,n_test do
+  local item = reader:get(string.format("%07d", idxs[i]))
+  if i % 1000 == 0 then
+    print(string.format('%d: %d', i, idxs[i]))
+    print(#item.input)
+  end 
+  --print(item)
+  --print(#item.input)
+  --print(item.input[1][1][1])
+end
+reader:abort()
+train_env:close()
+print(sys.toc())
+collectgarbage()
+
+sys.tic()
+-------Read-------
+train_env:open()
+print(train_env:stat()) -- Current status
+local reader = train_env:txn(true) --Read-only transaction
+--local y = torch.Tensor(10,3,256,256)
+local y = {}
+
+for i=1,n_test do
+  local item = reader:get(string.format("%07d", i))
+  if i % 1000 == 0 then
+    print(string.format('%d: %d', i, i))
+    print(#item.input)
+  end
+  --print(item)
+  --print(#item.input)
+  --print(item.input[1][1][1])
+end
+reader:abort()
+train_env:close()
+print(sys.toc())
+]]--

datasets/imagenet-lmdb.lua

@@ -0,0 +1,87 @@
+--
+--  Copyright (c) 2016, Facebook, Inc.
+--  All rights reserved.
+--
+--  This source code is licensed under the BSD-style license found in the
+--  LICENSE file in the root directory of this source tree. An additional grant
+--  of patent rights can be found in the PATENTS file in the same directory.
+--
+--  ImageNet dataset loader
+--
+
+local image = require 'image'
+local paths = require 'paths'
+local t = require 'datasets/transforms'
+local lmdb = require 'lmdb'
+
+local M = {}
+local ImagenetLMDBDataset = torch.class('resnet.ImagenetLMDBDataset', M)
+
+function ImagenetLMDBDataset:__init(imageInfo, opt, split)
+   self.imageInfo = imageInfo[split]
+   self.opt = opt
+   self.split = split
+   --self.dir = paths.concat(opt.data, split)
+   --assert(paths.dirp(self.dir), 'directory does not exist: ' .. self.dir)
+
+   self.env = lmdb.env{
+       Path = paths.concat(opt.data, string.format('%s_lmdb', split)),
+       Name = string.format('%s_lmdb', split)
+   }
+   assert(env:open(), 'directory does not exist: ' .. string.format('%s_lmdb', split))
+   self.stat = env:stat() -- Current status
+
+   self.reader = env:txn(true) --Read-only transaction
+   self.idxs = torch.randperm(n_images)
+   assert(self.imageInfo.imageClass:size(1) == #self.idxs, string.format('Something wrong with lmdb. The lmdb db should have %d number of items, but it has %d', self.imageInfo.imageClass:size(1), #self.idxs))
+end
+
+function ImagenetLMDBDataset:get(i)
+   local item = reader:get(string.format("%07d", self.idxs[i]))
+   return item
+end
+
+function ImagenetLMDBDataset:size()
+   return self.imageInfo.imageClass:size(1)
+end
+
+-- Computed from random subset of ImageNet training images
+local meanstd = {
+   mean = { 0.485, 0.456, 0.406 },
+   std = { 0.229, 0.224, 0.225 },
+}
+local pca = {
+   eigval = torch.Tensor{ 0.2175, 0.0188, 0.0045 },
+   eigvec = torch.Tensor{
+      { -0.5675,  0.7192,  0.4009 },
+      { -0.5808, -0.0045, -0.8140 },
+      { -0.5836, -0.6948,  0.4203 },
+   },
+}
+
+function ImagenetLMDBDataset:preprocess()
+   if self.split == 'train' then
+      return t.Compose{
+         t.RandomSizedCrop(224),
+         t.ColorJitter({
+            brightness = 0.4,
+            contrast = 0.4,
+            saturation = 0.4,
+         }),
+         t.Lighting(0.1, pca.eigval, pca.eigvec),
+         t.ColorNormalize(meanstd),
+         t.HorizontalFlip(0.5),
+      }
+   elseif self.split == 'val' then
+      local Crop = self.opt.tenCrop and t.TenCrop or t.CenterCrop
+      return t.Compose{
+         t.Scale(256),
+         t.ColorNormalize(meanstd),
+         Crop(224),
+      }
+   else
+      error('invalid split: ' .. self.split)
+   end
+end
+
+return M.ImagenetLMDBDataset

datasets/imagenet.lua

@@ -77,10 +77,13 @@ local pca = {
    },
 }
 
-function ImagenetDataset:preprocess()
+function ImagenetDataset:preprocess(minSize, cropSize) 
+   -- minSize : 256, cropSize : 224 for resnet
+   -- minSize : 328, cropSize : 299 for inception-resnet-v2 
+
    if self.split == 'train' then
       return t.Compose{
-         t.RandomSizedCrop(224),
+         t.RandomSizedCrop(cropSize),
          t.ColorJitter({
             brightness = 0.4,
             contrast = 0.4,
@@ -93,9 +96,9 @@ function ImagenetDataset:preprocess()
    elseif self.split == 'val' then
       local Crop = self.opt.tenCrop and t.TenCrop or t.CenterCrop
       return t.Compose{
-         t.Scale(256),
+         t.Scale(minSize),
          t.ColorNormalize(meanstd),
-         Crop(224),
+         Crop(cropSize),
       }
    else
       error('invalid split: ' .. self.split)