Training on custom dataset and Inference details

[abhigoku10]

@2014gaokao Hi thanks for opensourcing your work appreciate the effort , i have few queries

1. What are the steps to train on the custom dataset , hwo the annotations should be
2. Has glove.pkl and adj.pkl to be generated based on the dataset if so for custom dataset it will also change ??
3. What is the inference time taken for one image ? on the system you have run
4. can we change the resnet to other architecture like darknet shuffnet efficient net Thanks in advance

[2014gaokao]

针对自定义的数据集，要根据标签重新产生邻接矩阵adj.pkl和词向量glove.pkl，可以参考"Multi-Label Image Recognition with Graph Convolutional Networks"这篇论文；推理时间我没关注过，不知道；网络结构肯定可以随自己改的啊。

[thunanguyen]

Can you please write how to train on a custom dataset in English so that non-Chinese users can understand please. Thank you

[2014gaokao]

Sorry for my English. The most important files are glove.pkl and adj.pkl. Glove.pkl contains the word embedding vectors of labels. Adj.pkl contains an adjacency matrix representing the structure of a graph P, each node Pij = Mij / Ni, Ni represents the total number of occurrences of each label, Mij represents the number of occurrences of label pairs in the training data set. The input of the model consists of image features, glove.pkl and adj.pkl. Plesase refer to "Multi-Label Image Recognition with Graph Convolutional Networks" if you want more details.

[thunanguyen]

Thanks, I understand the algorithm now but for the implementation, what code do I have to change if I want to use a new dataset? Or do I just need to follow the format of RAPv2?

[2014gaokao]

I have no idea, I am done with deep learning, you know life is not easy.