Pre-trained/Fine tuning for

[ggropican]

Hello Monai Developers!

I am currently trying to train a stack of 2D images(329 images) using 'unet_training_array.py' from the tutorial provided, however the training loss still high. So I am planning to use a pre-trained model instead but couldn't find any tutorial from Monai. Do you have any advice or tutorial for using a pre-trained model?

My image size is [1056,1056]
Thank you very much in advance

[Nic-Ma]

Hi @yiheng-wang-nv ,

Could you please help share some comments here?

Thanks in advance.

[yiheng-wang-nv]

Hi @ggropican , monai does not provide pretrained weights for UNet. As for your task, could you please provide some more information, such as 1) what transforms do you used? 2) what are the training and validation losses for each epoch?
I may need more information to provide more detailed advices, thanks!

[ggropican]

Hello @yiheng-wang-nv , The raw images are size of [2088,1984] and there are 140 images. I am operating at batch size of 21, learning rate of 1e-4 and these are my current transforms:
train_imtrans = Compose(
[
LoadImage(image_only=True),
AddChannel(),
Resize(spatial_size=(1052,1000),mode='bilinear'),
SpatialPad((1056,1056),"symmetric"),
ScaleIntensity(0,1),
EnsureType(),
]
)
epoch1
training loss: 0.9980
epoch2
training loss: 0.9976
val_meandice: 0.3238
epoch3
training loss: 0.9974
epoch4
training loss: 0.9872
val_meandice: 0.2547
epoch5
training loss: 0.9804
epoch6
training loss: 0.9624
val_meandice: 0.1399
epoch7
training loss: 0.9383
epoch8
training loss: 0.9283
val_meandice: 0.0635
epoch9
training loss: 0.9206
epoch10
training loss: 0.9256
val_meandice: 0.0298
epoch11
training loss: 0.9202
epoch12
training loss: 0.9356
val_meandice: 0.01425

(I got hold of a pre-trained UNET model on a pth file, do you have any recommendations on how to load it?)
Thank you very much!