DOWNLOAD.md

Download Pre-trained Weights

CIRPLANT

Performance

('recall_top1_correct_composition', 0.20174599378139202),
('recall_top2_correct_composition', 0.3271944510882564),
('recall_top5_correct_composition', 0.5320975843099737),
('recall_top10_correct_composition', 0.691580961492466),
('recall_top50_correct_composition', 0.9306386032049749),
('recall_top100_correct_composition', 0.9681894283664195),
('recall_subset_top1_correct_composition', 0.4038985888543411),
('recall_subset_top2_correct_composition', 0.6477397751734034),
('recall_subset_top3_correct_composition', 0.8100932791198278)

How-To

Run with the same arguments as in training, but with the additional --validateonly and --load_from_checkpoint appended at the end.

python trainval_oscar.py --dataset cirr \
                         --usefeat nlvr-resnet152_w_empty --max_epochs 300 \
                         --model CIRPLANT-img --model_type 'bert' \
                         --model_name_or_path data/Oscar_pretrained_models/base-vg-labels/ep_107_1192087 \
                         --task_name cirr --gpus 1 \
                         --img_feature_dim 2054 --max_img_seq_length 1 \
                         --model_type bert --do_lower_case --max_seq_length 40 \
                         --learning_rate 1e-05 --loss_type xe --seed 88 \
                         --drop_out 0.3 --weight_decay 0.05 --warmup_steps 0 \
                         --loss st --batch_size 32 --num_batches 529 \
                         --pin_memory --num_workers_per_gpu 0 \
                         --comment input_your_comments \
                         --output saved_models/cirr_rc2_iccv_release_test \
                         --log_by recall_inset_top1_correct_composition \
                         --validateonly  --load_from_checkpoint $PATH_TO_CKPT

Results will be displayed and saved to the output directory.