feiyuchen7 / m3net Goto Github PK

您好，我对您的文章很感兴趣，在复现时遇到了一些问题，希望得到您的指点！
我在复现IEMOCAP数据集时，ACC和F1只能达到70，论文中写有72，我用的也是RTX 3090，实验中参数和您设置一样，这可能是由于不同参数有不同的效果，因此我也在尝试中。

我对超图神经网络应用于对话情绪识别上很感兴趣，因此我删掉了Multi-frequency propagation模块，ACC和F1特别低，分别为61和60，论文中论文中消融实验结果为69，想问下是否我的操作有问题，以下是我的操作过程，我仅在model_hyper.py文件中修改了相关代码，我在输出时只保留了Multivariate propagation模块的输出值。想问下正确的过程是怎样的，是否要在其他文件上做调整，这个问题一直困惑着我，希望得到您的指点，谢谢！

您好！我在尝试复现您的论文

按照您给出的参数设置
python -u train.py --base-model 'GRU' --dropout 0.5 --lr 0.0001 --batch-size 16 --graph_type='hyper' --epochs=80 --graph_construct='direct' --multi_modal --mm_fusion_mthd='concat_DHT' --modals='avl' --Dataset='IEMOCAP' --norm BN --num_L=3 --num_K=4
和指定的torch-geometric 版本1.7.2

在IEMOCAP数据集上，我得出的结果只达到了70.73，我也尝试过调整一些参数，例如use_modal、use_residue等，use_modal设置成True的话可以略微提高结果的accuracy，但是还是达不到论文的72，然后use_residue这个参数调整对结果没有明显的改变。

不知道是否哪里有遗漏，非常需要您的指点！拜托🙏

Namespace(no_cuda=False, base_model='GRU', graph_model=True, nodal_attention=True, windowp=10, windowf=10, lr=0.0001, l2=3e-05, rec_dropout=0.1, dropout=0.5, batch_size=16, epochs=80, class_weight=True, active_listener=False, attention='general', tensorboard=True, graph_type='hyper', use_topic=False, alpha=0.2, multiheads=6, graph_construct='direct', use_gcn=False, use_residue=False, multi_modal=True, mm_fusion_mthd='concat_DHT', modals='avl', av_using_lstm=False, Deep_GCN_nlayers=4, Dataset='IEMOCAP', use_speaker=True, use_modal=True, norm='BN', testing=False, num_L=3, num_K=4)
Running on GPU
construct hyper
Graph NN with GRU as base model.
<dataloader.IEMOCAPDataset object at 0x000001C9ED8ADF00>

Great job on the project! 👏 I'm interested in using this model on a custom dataset. Could you please provide a feature extractor script to facilitate this process? especially for dense net

Thanks in advance! 😊

Hello, congratulations on completing and publishing your work! I'm interested in graph methods, so I want to reproduce this method. However, I encountered the following issues while trying to replicate it:

1.The feature dimensions in the provided file do not match those in the code.
2.Even after modifying the dimension settings in the code based on the feature file, I still can't run it successfully. The core components of the code, model_hyper.py and HypergraphConv.py, are throwing errors.

Could you please help me with these questions? I'm looking forward to your response.

Hello, congratulations on completing and publishing your work! I appreciate your excellent work on ERC task, so I want to reproduce this method. However, I couldn't reproduce the results of the IEMOCAP and MELD datasets mentioned in your paper.
Here is what I have done and the results I have got, I would appreciate if you could tell me what's wrong with my implementation.

1. download the feature files you offer and place it in proper position as set in dataloader.py:

2. run the command you give in readme:
   Here are the results of IEMOCA, which is much lower than the paper as it only achieves around 66 in acc and F1-score. And results of paper is about 72 in F1-score and acc.

Hello, I'm trying to use your code, but I can't replicate your results. I downloaded your code and the dataset and used the code line that you gave on GitHub, but the metrics are different. Also, I have a problem finding the Multi-frequency Graph construction in the code. I would be grateful if you could help me.

Thank you for your impressive work!
I try to reproduce the result using your training example, while it reports "ValueError: Encountered tensor with size 2607 in dimension 0, but expected size 917." in line 150 in HypergraphConv.py, which is "out = self.propagate(hyperedge_index, x=out, norm=D, alpha=alpha,
size=(num_edges, num_nodes))"
Then what should i do to fix it
Thanks!!