Refinement experiments

Machine	Command
[email protected]	Mixed layers: 3 → 5
[email protected]	Combine lang WE + Hid (--mix-we)
[email protected]	Mixed size (Before mSRU): 4096
[email protected]	CUDA_VISIBLE_DEVICES=1 nohup python -u train.py --data /mnt/referit_data --backend vgg16 --save-folder unc_vgg16_sru_lowres_weights --lr 1e-5 > unc_vgg16_sru_lowres.log &
T1	CUDA_VISIBLE_DEVICES=0 nohup python -u train.py --data ../referit_data --backend dpn92 --lstm --high-res --save-folder unc_dpn92_lstm_highres_weights --lr 1e-5 > unc_dpn92_lstm_highres.log &
T2	CUDA_VISIBLE_DEVICES=1 nohup python -u train.py --data ../referit_data --backend dpn92 --high-res --save-folder unc_dpn92_sru_highres_weights --lr 1e-5 > unc_dpn92_sru_highres.log &
T3	CUDA_VISIBLE_DEVICES=2 nohup python -u train.py --data ../referit_data --backend vgg16 --lstm --high-res --save-folder unc_vgg16_lstm_highres_weights --lr 1e-5 > unc_vgg16_lstm_highres.log &
T4	CUDA_VISIBLE_DEVICES=3 nohup python -u train.py --data ../referit_data --backend vgg16 --high-res --save-folder unc_vgg16_sru_highres_weights --lr 1e-5 > unc_vgg16_sru_highres.log &