GeezerVisonThe World Wide Web, restored to 800 pixels of glory. |
| Home|About|Gaming cabinet|Archives|How it works|Starter sites |
GeezerVison is showing github.com — DenseNet/train.lua at master · Lite-Java/DenseNet Expand file tree/ Copy pathtrain.luaMore file actions 210 lines (167 loc) · 6.65 KB / Copy pathtrain.luaFile metadata and controls210 lines (167 loc) · 6.65 KB 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 -- -- 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. -- -- The training loop and learning rate schedule -- -- Code modified for DenseNet (https://arxiv.org/abs/1608.06993) by Gao Huang. -- local optim = require 'optim' local M = {} local Trainer = torch.class('resnet.Trainer', M) function Trainer:__init(model, criterion, opt, optimState) self.model = model self.criterion = criterion self.optimState = optimState or { learningRate = opt.LR, learningRateDecay = 0.0, momentum = opt.momentum, nesterov = true, dampening = 0.0, weightDecay = opt.weightDecay, } self.opt = opt self.params, self.gradParams = model:getParameters() end function Trainer:train(epoch, dataloader) -- Trains the model for a single epoch ------for LR------ if self.opt.lrShape == 'multistep' then self.optimState.learningRate = self:learningRate(epoch) end ------for LR------ local timer = torch.Timer() local dataTimer = torch.Timer() local function feval() return self.criterion.output, self.gradParams end local trainSize = dataloader:size() local top1Sum, top5Sum, lossSum = 0.0, 0.0, 0.0 local N = 0 print('=> Training epoch # ' .. epoch) -- set the batch norm to training mode self.model:training() for n, sample in dataloader:run() do ------for LR------ if self.opt.lrShape == 'cosine' then self.optimState.learningRate = self:learningRateCosine(epoch, n, trainSize) end ------for LR------ local dataTime = dataTimer:time().real -- Copy input and target to the GPU self:copyInputs(sample) local output = self.model:forward(self.input):float() local batchSize = output:size(1) local loss = self.criterion:forward(self.model.output, self.target) self.model:zeroGradParameters() self.criterion:backward(self.model.output, self.target) self.model:backward(self.input, self.criterion.gradInput) optim.sgd(feval, self.params, self.optimState) local top1, top5 = self:computeScore(output, sample.target, 1) top1Sum = top1Sum + top1*batchSize top5Sum = top5Sum + top5*batchSize lossSum = lossSum + loss*batchSize N = N + batchSize print((' | Epoch: [%d][%d/%d] Time %.3f Data %.3f Err %1.3f top1 %7.2f top5 %7.2f lr %.4f'):format( epoch, n, trainSize, timer:time().real, dataTime, loss, top1, top5, self.optimState.learningRate)) -- check that the storage didn't get changed due to an unfortunate getParameters call assert(self.params:storage() == self.model:parameters()[1]:storage()) timer:reset() dataTimer:reset() end return top1Sum / N, top5Sum / N, lossSum / N end function Trainer:test(epoch, dataloader) -- Computes the top-1 and top-5 err on the validation set local timer = torch.Timer() local dataTimer = torch.Timer() local size = dataloader:size() local nCrops = self.opt.tenCrop and 10 or 1 local top1Sum, top5Sum = 0.0, 0.0 local N = 0 self.model:evaluate() for n, sample in dataloader:run() do local dataTime = dataTimer:time().real -- Copy input and target to the GPU self:copyInputs(sample) local output = self.model:forward(self.input):float() local batchSize = output:size(1) / nCrops local loss = self.criterion:forward(self.model.output, self.target) local top1, top5 = self:computeScore(output, sample.target, nCrops) top1Sum = top1Sum + top1*batchSize top5Sum = top5Sum + top5*batchSize N = N + batchSize print((' | Test: [%d][%d/%d] Time %.3f Data %.3f top1 %7.3f (%7.3f) top5 %7.3f (%7.3f)'):format( epoch, n, size, timer:time().real, dataTime, top1, top1Sum / N, top5, top5Sum / N)) timer:reset() dataTimer:reset() end self.model:training() print((' * Finished epoch # %d top1: %7.3f top5: %7.3f\n'):format( epoch, top1Sum / N, top5Sum / N)) return top1Sum / N, top5Sum / N end function Trainer:computeScore(output, target, nCrops) if nCrops > 1 then -- Sum over crops output = output:view(output:size(1) / nCrops, nCrops, output:size(2)) --:exp() :sum(2):squeeze(2) end -- Coputes the top1 and top5 error rate local batchSize = output:size(1) local _ , predictions = output:float():topk(5, 2, true, true) -- descending -- Find which predictions match the target local correct = predictions:eq( target:long():view(batchSize, 1):expandAs(predictions)) -- Top-1 score local top1 = 1.0 - (correct:narrow(2, 1, 1):sum() / batchSize) -- Top-5 score, if there are at least 5 classes local len = math.min(5, correct:size(2)) local top5 = 1.0 - (correct:narrow(2, 1, len):sum() / batchSize) return top1 * 100, top5 * 100 end local function getCudaTensorType(tensorType) if tensorType == 'torch.CudaHalfTensor' then return cutorch.createCudaHostHalfTensor() elseif tensorType == 'torch.CudaDoubleTensor' then return cutorch.createCudaHostDoubleTensor() else return cutorch.createCudaHostTensor() end end function Trainer:copyInputs(sample) -- Copies the input to a CUDA tensor, if using 1 GPU, or to pinned memory, -- if using DataParallelTable. The target is always copied to a CUDA tensor self.input = self.input or (self.opt.nGPU == 1 and torch[self.opt.tensorType:match('torch.(%a+)')]() or getCudaTensorType(self.opt.tensorType)) self.target = self.target or (torch.CudaLongTensor and torch.CudaLongTensor()) self.input:resize(sample.input:size()):copy(sample.input) self.target:resize(sample.target:size()):copy(sample.target) end function Trainer:learningRate(epoch) -- Training schedule local decay = 0 if self.opt.dataset == 'imagenet' then decay = math.floor((epoch - 1) / 30) elseif self.opt.dataset == 'cifar10' then decay = epoch >= 0.75*self.opt.nEpochs and 2 or epoch >= 0.5*self.opt.nEpochs and 1 or 0 elseif self.opt.dataset == 'cifar100' then decay = epoch >= 0.75*self.opt.nEpochs and 2 or epoch >= 0.5*self.opt.nEpochs and 1 or 0 end return self.opt.LR * math.pow(0.1, decay) end ------for LR------ function Trainer:learningRateCosine(epoch, iter, nBatches) local nEpochs = self.opt.nEpochs local T_total = nEpochs * nBatches local T_cur = ((epoch-1) % nEpochs) * nBatches + iter return 0.5 * self.opt.LR * (1 + torch.cos(math.pi * T_cur / T_total)) end ------for LR------ return M.Trainer |
| You are visitor 01402520 — GeezerVison does not store the pages you look up. Best viewed with Netscape Navigator 4.7 or Internet Explorer 6 at 800×600. No cookies required. |