lyndonchan / hsn_v1 Goto Github PK

Hi! I tried to follow the example with [demo_01_segment_patches.ipynb] but I'm getting a problem:
on Run all batches section:

confscores = np.zeros((len(hsn.input_files_all), len(hsn.hn.class_names)))
...
...

RuntimeError: Attempted to aggregate unsupported object Tensor("sequential_1/activation_8/Sigmoid:0", shape=(16, 51), dtype=float32).

Any way to solve this error?

Best,
Duerna

Hello,

I want to retrain the model of Patch-level Classification CNN, so I searched in the ADP project for the database. And I found the ADP website.
I noticed that the database on ADP Website is TO BE released and under Ethics Board Review. May I ask when I can download the database? I've already activated my account on the website. But I can't download the database for the reasons stated before. Or, there is any other way to get the database?

With many thanks

Hello, lyndonchan!

Sorry for bothering you again. I've already got the ADP database from ADP website. However, the label in the database is not exactly the same as you used in HistoSegNet.

In hsn paper, you use 'the CNN pre-trained on predicting the 31 HTTs in the third level of the ADP database'. But in hsn code, the classifier has 51 classes. My understanding is that you use some augmentation method, e.g.Hierarchical Binary Relevance, as you mentioned in ADP paper.

Here is the problem, the ADP repo only provides a training database containing ‘ADP_EncodedLabels_Release1_Flat.csv', no 'ADP_EncodedLabels_Release1.csv'. As the htt class definition in ADP repo, I can't train model in <dataset_type == 'ADP-Release1'>, only can in <dataset_type == 'ADP-Release1-Flat'>. The 'Flat' label file has totally different classes from the dataset you used in HistoSegNet. So even I got the image, I can't reproduce your patch-level classification CNN training.

So I wonder if you or your team could provide the other label file, which can support 51 classes training?

Hello, I am also doing research on pathological image segmentation recently. I want to cite your paper, so I need to run the code of your paper, but I did not find the relevant training files. Take the liberty to ask, can you provide me with the training files?

I noticed your paper (very nice work!) mention about HistoSegNet can generalize to other datasets without retraining.
If I am not misunderstanding, I can load my own image and get the segment information?

If so, can you point out how to input my images?

With many thanks,
Duerna