-
These models were trained for the kaggle compitition and won a BRONZE medal.
-
Given a fragment, this segmentation model detects the ink.
-
The output of the model is a binary mask where the 1 represents the presence of ink and 0 for no ink.
-
This above mask is then converted into the RLE(Run Length Encodings).
The following example has the letter "E" it's RLE encoding is: "11 7 20 1 23 1 26 1 29 1 32 1 35 1 38 1 44 1"
- The data is taken from the kaggle compitition.
-
The dataset is 3d x-ray scans of detached fragments of ancient papyrus scrolls.
-
The training data had 3 fragments.
-
It had the slices from the 3d x-ray surface volume. Each file contains a greyscale slice in the z-direction. Each fragment contains 65 slices. Combined this image stack gives us width * height * 65 number of voxels per fragment.
The 65 slices of the first fragment
-
The inklables were given which was a binary mask which showed 1 for the presense of the ink.
-
Further there was even the mask of the fragment which basically shows where the data is present in the fragment.
- Training_notebooks: Contains the notebooks used for training. Much more description about them is given in the training readme.
- Inference_notebooks: Contains the notebooks used for Inference. Much more description about them is given in the Inference readme.
- tiles_extracted: This is the notebook that reduces the dataset by merging the volume scans with the mask. This removes the unecessary tiles for training the model
-
Trained multiple U-Net models using the segmentation-models-pytorch.
-
Trained models with the backbone
- mit_b2 , mitb3 , mit_b4 , mit_b5
- VGG19
- resnet50 resnet34
- efficientnet_b3
- resnest
- regnety etc
-
Few of them can be found in the repo.
-
Trained a model using stratified K fold technique
-
During the training the fragments were broken into the sizes of 224 by 224 and kept a stride of 224//4
-
trained for 15 epochs by keeping one of the 3 fragments as the cross validation
-
The calculated metric was fbeta score keeping the beta value 0.5
-
Was getting the CV score of single model around 63
-
used the weighted loss of dice loss and the Tversky loss and BCE loss
fig: the first image is the actual mask, the second one is the predicted mask, third one is the predicted mask after applying a threshhold 0.4
- Did the ensembles of different models trained, tried average ensemble.
- Ensembling the models increased the fbeta score from 65 to 74.
- Tried weighted ensembles tooo
-
Performed TTA(Test Time Augmentations) during the final inference.
-
This included the rotation of the image during the inference.(did 90,180,270,360 degree roatations)
x = [x, *[tc.rot90(x, k=i, dims=(-2, -1)) for i in range(1, 4)]] x = tc.cat(x, dim=0) -
after predicting on these augmentations, again rotated the output then took the average of the outputs
x=[tc.rot90(x[i],k=-i,dims=(-2,-1)) for i in range(4)] x=tc.stack(x,dim=0)
- Out of the 65 channels, the center channels contained the information.
- So took the center 8 channels and made them into group of 3 because the model trained took only 3 channels in input
- The groups of 3 channels were (1,2,3) , (4,5,6) , (6,7,8)
- So this resulted in output of 3 masks which was later averaged to get the final mask