Paper title: Better utilization of materials' compositions for predicting their properties: Materials Composition Visualization Network.
There are two open access datasets in the paper: NIMS and BMG-RMG-CRA can be obtained through the paper or by contacting the corresponding author of the paper, so the repository does not put all the experimental data, but only 10 samples of the data collated by the authors of this paper, and the dataset of the second experiment SRIM needs to be obtained by contacting the corresponding author of this paper.
original_data.csv: Original data csv file.
Image.csv : Generated by DataTransform.py file and sorted by ImgTemplate.xlsx template.
original_images.csv: Generated via DataTransform.py file and not sorted.
DataTransform.py: You can transform original data to image data by ImgTemplate.xlsx file.
ImgTemplate.xlsx: Image Template files. It must be .xlsx files or you can change code in DataTransform.py files. You can change image template, but must be 7*8 table and names should be in paper's appendix.
The network structure in paper. You can change original channel by yourself. If you want to change image channel, you should pay attention to the input images' size is 24*21.
If your dataset is much bigger than the paper's dataset, I suggest that you can try more complex network structure. If you can't get good result by MCVN, you should try other network structure in original channel and fusion channel.
My data loader will return a dictionary. The key is target name, the value is Data.TensorDataset and torch.Tensor, you can find more information in files.
It's my own trainer with early stopping method, you can build your own trainner or change the trainer with your thought.
Main_Example.py is a example using MCVN, but the parameters are not the ones used in the paper, such as the learning rate in the paper is obtained by grid search.