how to reproduce ResNet-152 + Bi-GRU on COCO experiment

[Asakinevergup]

Hello! I'm reproducing your experiments. I followed your train_eval_coco.sh script and obtained the same results. However, when I attempted to reproduce the experiment with ResNet-152 + Bi-GRU, the results were not as expected. Could you please tell me how you configured it? Thank you very much!
PS: These are the parameters I used during training.
python3 train.py
--data_name coco --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-5 --txt_lr_scale 10 --img_pie_lr_scale 10 --txt_pie_lr_scale 10
--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 sine --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
--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

[NJNUwhb]

Hello! Have you tackled the problem?

[Asakinevergup]

Unfortunately, no but u can try more times, good luck!