Reproduction Metrics Issue on MOSI and MOSEI

[nitritez]

Hello authors,

I encountered some issues when reproducing the results on MOSI/MOSEI. For MOSI, the reproduced metrics differ significantly from those reported in the paper, while for MOSEI there is a slight performance drop.

Both datasets were downloaded from MMSA, and I used the config files from the current version of the repository.

I noticed that the feature dimensions for MOSI may be inconsistent with those reported in the paper. Could this indicate a problem, or is it related to the current config settings?

Thank you for your help.

"need_data_aligned": true,
"dst_feature_dim_nheads": [64, 8],
"nlevels": 5,
"conv1d_kernel_size_l": 1,
"conv1d_kernel_size_a": 1,
"conv1d_kernel_size_v": 1,
"num_prototypes": 7,
"gmm_em_iters": 5,
"gmm_var_floor": 0.0001,
"lambda_ot": 0.1,
"ot_num_iters": 50,
"use_cross_modal_attention": true,
"pde_dim": 64,
"mmd_bandwidth": 1.0,
"attn_dropout": 0.1,
"attn_dropout_a": 0.0,
"attn_dropout_v": 0.0,
"relu_dropout": 0.1,
"embed_dropout": 0.1,
"res_dropout": 0.05,
"output_dropout": 0.0,
"text_dropout": 0.0,
"attn_mask": true,
"task_loss": "mse",
"dec_loss_weight": 1.0,
"alpha": 0.05,
"beta": 0.07,
"ct_weight": 0.003,
"ct_temperature": 0.1,
"use_sequence_mask": false,
"batch_size": 32,
"learning_rate": 0.00006,
"bert_learning_rate": null,
"weight_decay": 0.005,
"num_epochs": 18,
"early_stop_patience": 6,
"scheduler_patience": 5,
"patience": 5,
"clip": 0.6,
"selection_metric": "MOSI",
"scheduler_metric": "MOSI",
"factor": 0.1