training scripts for flickr30k #9
Description
In order to reproduce the experiments asap, i use flickr30k datasets as the benchmark and utilize ddp to run the code.
the modified training script for flickr is as follows:
CUDA_VISIBLE_DEVICES=0,1,2,3 python3 train.py \ --data_name f30k --cnn_type resnet152 --wemb_type glove \ --margin 0.2 --max_violation --img_num_embeds 4 --txt_num_embeds 4 \ --img_attention --txt_attention --img_finetune --txt_finetune \ --mmd_weight 0.01 --unif_weight 0.01 \ --batch_size 200 --warm_epoch 0 --num_epochs 80 \ --optimizer adamw --lr_scheduler cosine --lr_step_size 30 --lr_step_gamma 0.1 \ --warm_img --finetune_lr_lower 1 \ --lr 1e-3 --txt_lr_scale 1 --img_pie_lr_scale 0.1 --txt_pie_lr_scale 0.1 \ --eval_on_gpu --sync_bn --amp \ --loss smooth_chamfer --eval_similarity smooth_chamfer --temperature 16 \ --txt_pooling rnn --arch slot --txt_attention_input wemb \ --spm_img_pos_enc_type none --spm_txt_pos_enc_type sine \ --spm_1x1 --spm_residual --spm_residual_norm --spm_residual_activation none \ --spm_activation gelu \ --spm_ff_mult 4 --spm_last_ln \ --img_res_pool max --img_res_first_fc \ --spm_input_dim 1024 --spm_query_dim 1024 \ --spm_depth 4 --spm_weight_sharing \ --remark coco_butd_bigru \ --res_only_norm --img_1x1_dropout 0.1 --spm_pre_norm \ --gpo_1x1 --gpo_rnn \ --weight_decay 1e-4 --grad_clip 1 --lr_warmup -1 --unif_residual \ --workers 4 --dropout 0.1 --caption_drop_prob 0.2 --butd_drop_prob 0.2
But the performance are unsatisfactory, the results are as follows.
so what's my problem in this trainning script, i will be very grateful if you can point out my mistake, thank you!