Added example script to train sketch-based image retrieval and deep s…

…hape matching network presented in ECCV18 paper: Deep Shape Matching.

cnntrain/train_network.m

@@ -70,7 +70,7 @@
 end
 
 for epoch=start+1:opts.numEpochs
-
+  
   % Set the random seed based on the epoch and opts.randomSeed.
   % This is important for reproducibility, including when training
   % is restarted from a checkpoint.
@@ -137,7 +137,7 @@
 
   % With multiple GPUs, return one copy
   if isa(net, 'Composite'), net = net{1} ; end
-
+     
 end
 
 % -------------------------------------------------------------------------

examples/test_cnnsketch2imageretrieval.m

@@ -24,7 +24,6 @@
 % Choose ECCV18 fine-tuned CNN network
 network_file = fullfile(data_root, 'networks', 'retrieval-SfM-30k', 'retrievalSfM30k-edgemac-vgg.mat');
 
-%% COMING SOON
 % After running the training script train_cnnsketch2imageretrieval.m you can evaluate fine-tuned network
 % network_file = fullfile(data_root, 'networks', 'exp', 'vgg_edgefilter_mac_test', 'net-epoch-20');
 

examples/train_cnnsketch2imageretrieval.m

@@ -0,0 +1,60 @@
+% TRAIN_CNNSKETCH2IMAGERETRIEVAL Code to train the methods presented in the paper:
+% F. Radenovic, G. Tolias, O. Chum, Deep Shape Matching, ECCV 2018
+%
+% Note: The method has been re-coded since our ECCV 2018 paper and minor differences in performance might appear.
+%
+% Authors: F. Radenovic, G. Tolias, O. Chum. 2018. 
+
+clear;
+
+%-------------------------------------------------------------------------------
+% Set data folder
+%-------------------------------------------------------------------------------
+
+% Set data folder, change if you have downloaded the data somewhere else
+data_root = fullfile(get_root_cnnimageretrieval(), 'data');
+% Check, and, if necessary, download train data (db with edgemaps), and pre-trained imagenet networks
+download_train_sketch(data_root);
+
+
+%-------------------------------------------------------------------------------
+% Reproduce training from ECCV18 paper: Deep Shape Matching ...
+%-------------------------------------------------------------------------------
+
+% Set architecture and initialization parameters
+opts.init.model = 'VGG'; % (ALEX | VGG | GOOGLENET | RESNET101)
+opts.init.modelDir = fullfile(data_root, 'networks', 'imagenet');
+opts.init.method = 'edgefilter_mac';
+opts.init.objectiveType = {'contrastiveloss', 0.7};
+opts.init.errorType = {'batchmap'};
+opts.init.averageImageScale = 0;
+opts.init.imageChannels = 1;
+
+% Set train parameters
+opts.train.dbPath = fullfile(data_root, 'train', 'dbs', 'retrieval-SfM-30k-edgemap.mat');
+opts.train.batchSize = 20;
+opts.train.numSubBatches = 4;
+opts.train.numEpochs = 20;
+opts.train.learningRate = 0.001 .* exp(-(0:99)*0.1);
+opts.train.numNegative = 5;
+opts.train.numRemine = 3;
+opts.train.gpus = [1];
+
+opts.train.augment.jitterFlip = true;
+opts.train.augment.jitterQueryBinarize = true;
+
+% Trial name (to name a save directory)
+trialName = 'test';
+
+% Export directory expDir named after model, method and trialName
+opts.init.method = [opts.init.method, '_', trialName];
+opts.train.expDir = fullfile(data_root, 'networks', 'exp', [lower(opts.init.model) '_' lower(opts.init.method)]);
+if ~exist(opts.train.expDir); mkdir(opts.train.expDir); end % create folder if its not there
+
+% Load opts by respecting added opts
+opts = load_opts_train(opts);
+
+% Initialize and train the network
+fprintf('>> Experiment folder is set to %s\n', opts.train.expDir);
+net = init_network(opts.init);
+[net, state, stats] = train_network(net, @(o,i,n,b,s,m,e) get_batch(o,i,n,b,s,m,e), opts.train);