Text: /home/c-nrong/VQA/draw/Json/question_answers_jan.json /home/c-nrong/VQA/draw/Json/question_answers_genome.json
Images: /home/c-nrong/VQA/VG_100K_all/ /home/c-nrong/VQA/VG_100K/ /home/c-nrong/VQA/VG_100K_2/
There are three main directories: VQA/ fixtestset/ fixtestset_2/
About “VQA/” the number of training samples and test samples are changed in each file folder, though the sum of the two number is fixed.
However, since the number of test samples are not fixed, the results of those experiments are uncomparable.
That why I created “fixtestset/” and “fixtestset2/”
In each file folder, there is a “READMED.md” whch shows descriptions/details for the file folder.
About “fixtestset_2/” copy from ../fixtestset/ the difference is that when train the model, this dir will be initialized by /home/c-nrong/fixtestset/Jan_HieCoAttenVQA_3by10/save/train_vgg_Alternating/model_id0_iter0.t7
conclusion 1. “fixtestset_2” is useless. 2. most of the file folders in “VQA/” are useless. 3. focus on “fixtestset/”
by default: training is random. DataLoaderDish.lua:ix = split_ix[torch.random(max_index)]
To make different experiments comparable, I change it to “order”. that is: DataLoaderDish.lua:ix = split_ix[ri]
HieCoAttenVQA/ : for testing(for example:sometime I may wanna know the dimension of a variable) train: 345049 test:317564
HieCoAttenVQA2/: use to generate genome english attenmaps (github: VQA_en)
Jan_HieCoAttenVQA/: (check “_readme”) (github: VQA_jan) v1_all_1dividedby196(done) v1_all_1(done) v2_onlyA(done)Jan_HieCoAttenVQA2/: a copy of “3”. (only to run “without attenmaps”) th train.lua -use_english_attenmaps 0 res1: train randomly. (done) res2: train according to order.(DataLoaderDish.lua:ix = split_ix[torch.random(max_index)], ix = split_ix[ri]) (done)Jan_HieCoAttenVQA3/: a copy of “3”. (2016.11.02-14:53) v2_addone: attenmaps=attenmaps + 1 (changed prepro/prepro_atten_new.py) (running) [screen: addone]Jan_HieCoAttenVQA4//: a copy of “3”. v2_onlyA_order [screen:onlyA_order]Jan_HieCoAttenVQA5//: a copy of “5” v2_addone_order[screen:addone_order][8-12: order] 8. Jan_HieCoAttenVQA6/: a copy of “6” (trainset and testset number are changed)) train=train*(1/10) train: 35146 test:780214 9. Jan_HieCoAttenVQA7/: a copy of “8” (trainset and testset number are changed)) train=train*(3/10) train:103042 test:677405 10. Jan_HieCoAttenVQA8/: a copy of “8” (trainset and testset number are changed)) train=train*(5/10) train:170422 test:574596 11. Jan_HieCoAttenVQA9/: a copy of “8” (trainset and testset number are changed)) train=train*(7/10) train:237258 test:471787 12. Jan_HieCoAttenVQA10/: a copy of “8” (trainset and testset number are changed)) train=train+test(1:trainum/2)
There are two types of attention: “alternating”(default) and “parallel”.
Fix the test set, and change the number of training set(6 values in total): Jan_HieCoAttenVQA_${i}by10/ , Jan_HieCoAttenVQA_${i}by10_without/ where i={1,3,5,7,10,15}
And since the intialization does influence the performance of the model, each experiment will be run five times.
So, finally: Alternating “./fixtestset/1/” “./fixtestset/2/” “./fixtestset/3/” “./fixtestset/4/” “./fixtestset/5/” Parallel “./fixtestset/experiments/Parallel/1/” “./fixtestset/experiments/Parallel/2/” “./fixtestset/experiments/Parallel/3/” “./fixtestset/experiments/Parallel/4/” “./fixtestset/experiments/Parallel/5/”
In “fixtest/” directory, there are three “Analysis” file folders: 1. /home/c-nrong/fixtestset/Analysis/ 2. /home/c-nrong/fixtestset/experiments/Analysis 3. /home/c-nrong/fixtestset/experiments/Parallel/Analysis(copy from 2./home/c-nrong/fixtestset/experiments/Analysis)
If you just want to check the result for one experiment, use “1”. If you want to check the results for five experiments(since the initialization does influence the perfomance, for each setting, the experiment will be run five times), use “2”.
