Custom pretrained resnet-50 model

[Ruslan-Lunev]

Hello.
I have trained a custom resnet-50 model using the Facebook implementation for Torch.
But the following error appears during the training of the Deep Mask model:

stack traceback:
	.../lua/5.2/nn/SpatialZeroPadding.lua:87: in function 'updateGradInput'
	.../lua/5.2/nn/Module.lua:31: in function <.../lua/5.2/nn/Module.lua:29>
	[C]: in function 'xpcall'
	.../lua/5.2/nn/Container.lua:63: in function 'rethrowErrors'
	.../lua/5.2/nn/Sequential.lua:88: in function <.../lua/5.2/nn/Sequential.lua:78>
	[C]: in function 'xpcall'
	.../lua/5.2/nn/Container.lua:63: in function 'rethrowErrors'
	.../lua/5.2/nn/Sequential.lua:88: in function <.../lua/5.2/nn/Sequential.lua:78>
	[C]: in function 'xpcall'
	.../lua/5.2/nn/Container.lua:63: in function 'rethrowErrors'
	.../lua/5.2/nn/Sequential.lua:88: in function 'backward'
	.../deepmask/TrainerDeepMask.lua:88: in function 'train'
	train.lua:117: in main chunk

resnet-50 settings

Data structure:

/data
    /label1
        img01.png
        img02.png
        ...
    /label2
    ...

Image resolution: 256x256
There are options I have tried to train the model:

- th main.lua -data data -depth 50 -nClasses 35
- th main.lua -data data -depth 50 -nClasses 35 -nThreads 4 -shareGradInput true

Deep Mask settings

Images contain multiple objects the resnet-50 model was trained on.
Image resolution: 1024x1024
Coco annotations file contains polygons in format: x1, y1, x2, y2, ...

1. Did anyone try to use a custom pretrained model or does know how to fix that?
2. If I use https://s3.amazonaws.com/deepmask/models/resnet-50.t7 I get messages like:

[train] | epoch 00006 | s/batch 0.11 | loss: 0.09705 	
[test]  | epoch 00006 | IoU: mean 000.00 median 000.00 [email protected] 000.00 [email protected] 000.00 | acc 095.26 | bestmodel x

Is that normal or there is something wrong with the custom data?