The mask-rcnn structure graph is the same as faster-rcnn?

[lilichu]

hi!
I output the structure graph of mask-rcnn, but I find it is the same as the structure graph of faster-rcnn.

The net such as the following:

INFO net.py: 250: _[mask]_roi_feat_shuffled   : (13, 256, 14, 14)    => _[mask]_roi_feat            : (13, 256, 14, 14)    ------- (op: BatchPermutation)
INFO net.py: 250: mask_rois_idx_restore_int32 : (13,)                => _[mask]_roi_feat            : (13, 256, 14, 14)    ------|
INFO net.py: 250: _[mask]_roi_feat            : (13, 256, 14, 14)    => _[mask]_fcn1                : (13, 256, 14, 14)    ------- (op: Conv)
INFO net.py: 250: _[mask]_fcn1                : (13, 256, 14, 14)    => _[mask]_fcn1                : (13, 256, 14, 14)    ------- (op: Relu)
INFO net.py: 250: _[mask]_fcn1                : (13, 256, 14, 14)    => _[mask]_fcn2                : (13, 256, 14, 14)    ------- (op: Conv)
INFO net.py: 250: _[mask]_fcn2                : (13, 256, 14, 14)    => _[mask]_fcn2                : (13, 256, 14, 14)    ------- (op: Relu)
INFO net.py: 250: _[mask]_fcn2                : (13, 256, 14, 14)    => _[mask]_fcn3                : (13, 256, 14, 14)    ------- (op: Conv)
INFO net.py: 250: _[mask]_fcn3                : (13, 256, 14, 14)    => _[mask]_fcn3                : (13, 256, 14, 14)    ------- (op: Relu)
INFO net.py: 250: _[mask]_fcn3                : (13, 256, 14, 14)    => _[mask]_fcn4                : (13, 256, 14, 14)    ------- (op: Conv)
INFO net.py: 250: _[mask]_fcn4                : (13, 256, 14, 14)    => _[mask]_fcn4                : (13, 256, 14, 14)    ------- (op: Relu)
INFO net.py: 250: _[mask]_fcn4                : (13, 256, 14, 14)    => conv5_mask                  : (13, 256, 28, 28)    ------- (op: ConvTranspose)
INFO net.py: 250: conv5_mask                  : (13, 256, 28, 28)    => conv5_mask                  : (13, 256, 28, 28)    ------- (op: Relu)
INFO net.py: 250: conv5_mask                  : (13, 256, 28, 28)    => mask_fcn_logits             : (13, 13, 28, 28)     ------- (op: Conv)
INFO net.py: 250: mask_fcn_logits             : (13, 13, 28, 28)     => loss_mask                   : ()                   ------- (op: SigmoidCrossEntropyLoss)
INFO net.py: 250: masks_int32                 : (13, 10192)          => loss_mask                   : ()                   ------|

After roi_feat, it should be fcn1, fcn2, fcn3, fcn4... But why does the output of structure graph not contain these?
The structure graph of mask-rcnn is such as:

the graph is the same as faster-rcnn. Thanks!

[gadcam]

@lilichu As far as I understand it, the structure is coded in two separate nets :

• model.net which should really look like a Faster R-CNN
• model.mask_net which implements the mask specific parts
• OR model.keypoint_net in the case of a keypoint net

I think you should try to print the model.mask_net part.

See

Detectron/detectron/modeling/model_builder.py

Lines 280 to 286 in e5bb3a8

	# Inference uses a cascade of box predictions, then mask predictions.
	# This requires separate nets for box and mask prediction.
	# So we extract the mask prediction net, store it as its own network,
	# then restore model.net to be the bbox-only network
	model.mask_net, blob_mask = c2_utils.SuffixNet(
	'mask_net', model.net, len(bbox_net.op), blob_mask
	)

You can also check the function run_model_pb in #449 to get a quick idea of how the full prediction looks like in practice.
https://github.com/facebookresearch/Detectron/pull/449/files#diff-077e2ae71af52724d94c3522525d2e7fR524