Bad Visual feedback results for Columns and Cells

[kmanojkkmr]

Hello @shahrukhqasim and Team,

I'm getting very bad results for Columns and Cells Prediction. But rows prediction is good. Could someone look into this and help me on this.

I have modified "is_sampling_balanced = 0" from 1 to overcome "indices does not index into param shape" issue. In the pdf's Blue rectangle indicates - Ground_Truth = 0 and Predicted = 0 Pink rectangle indicates - Ground_Truth = 1 and Predicted = 0 Orange rectangle indicates - The test cell which we are using for prediction of Cells/Columns/Rows

02916_cells.pdf 02916_cols.pdf

Thanks in advance.

[Sharathmk99]

Hi @kmanojkkmr What was the value of samples_per_vertex. If your working on building table structure model, we can also collaborate and work together. Interested, please email me.

[zaocan666]

any solution yet？

[kmanojkkmr]

After completing around 5500 iterations, validation predictions are good. But the testing predictions are very bad.

[oysz2016]

I got similar results from my experiments，The results of the rows are much better than the results of the cols is any solution yet？

[zaocan666]

I got similar results from my experiments，The results of the rows are much better than the results of the cols is any solution yet？

problem solved by training more iterations. I trained for 30K iterations and the results are good

[kbrajwani]

hey have you tried to inference on new images?

[ivaylojelev]

Try setting loss_alpha, loss_beta and loss_gamma all to 1. The total loss that your model seeks to minimize is a weighted average of the column, row and cell loss, and these parameters are the weights. If you are using the default config, it is possible loss_alpha and loss_gamma are set to 0, so the column and cell loss are not actually minimizing.

If the problem persists, you can set loss_alpha to an even higher value.