Library containing implementation related to the research paper "Multi-task pre-training of deep neural networks for digital pathology" (Mormont et al.).
It can be used to load our pre-trained models or to build a multi-task classification architecture.
For an example, check the file
examples/feature_extract.py
.
The library provides a build_model
function to build a model and initialize it with our
pre-trained weights. To load our weights, the parameter pretrained
should be set to mtdp
.
from mtdp import build_model
model = build_model(arch="densenet121", pretrained="mtdp")
Alternatively, pretrained
can be set to imagenet
to load ImageNet pre-trained weights from PyTorch.
We currently provide pre-trained weights for the following architectures:
densenet121
resnet50
See an example script performing feature extraction using one of our model in the examples
folder (file feature_extract.py
).
If you want to bypass the library and download the raw PyTorch model files, you can access them at the following URLs:
densenet121
: https://dox.uliege.be/index.php/s/G72InP4xmJvOrVp/downloadresnet50
: https://dox.uliege.be/index.php/s/kvABLtVuMxW8iJy/download
For an example, see the
examples/multi_task_train.py
file.
Several steps for building the architecture:
- define a
DatasetFolder
/ImageFolder
for each of your individual dataset, - instantiate a
MultiImageFolders
object with all your dataset objects, - instantiate a
MultiHead
PyTorch module by passing it theMultiImageFolders
from step 2. The module will use the information of the tasks in order to build the multi-task architecture.