I can not change loss type to SquaredL2Distance get errors.

[oujieww]

I want to use
dist=model.net.SquaredL2Distance([dt_output, 'masks_int32'],'dist')
loss_mask=model.net.AveragedLoss( dist, 'loss_mask')
loss_gradients = blob_utils.get_loss_gradients(model, [loss_mask])
model.AddLosses('loss_mask')

i think it is ok to change another loss ,but i get some error which i have no idea of those.
Traceback (most recent call last):
File "../tools/train_net.py", line 291, in
main()
File "../tools/train_net.py", line 129, in main
checkpoints = train_model()
File "../tools/train_net.py", line 138, in train_model
model, start_iter, checkpoints, output_dir = create_model()
File "../tools/train_net.py", line 216, in create_model
model = model_builder.create(cfg.MODEL.TYPE, train=True,gpu_id=1)
File "/mnt/disk1/oujie/fasr-rcnn-mpii/3/Detectron-master/lib/modeling/model_builder.py", line 124, in create
return get_func(model_type_func)(model)
File "/mnt/disk1/oujie/fasr-rcnn-mpii/3/Detectron-master/lib/modeling/model_builder.py", line 89, in generalized_rcnn
freeze_conv_body=cfg.TRAIN.FREEZE_CONV_BODY
File "/mnt/disk1/oujie/fasr-rcnn-mpii/3/Detectron-master/lib/modeling/model_builder.py", line 229, in build_generic_detection_model
optim.build_data_parallel_model(model, _single_gpu_build_func)
File "/mnt/disk1/oujie/fasr-rcnn-mpii/3/Detectron-master/lib/modeling/optimizer.py", line 42, in build_data_parallel_model
model.AddGradientOperators(all_loss_gradients)
File "/usr/local/lib/python2.7/dist-packages/caffe2/python/model_helper.py", line 350, in AddGradientOperators
self.grad_map = self.net.AddGradientOperators(*args, **kwargs)
File "/usr/local/lib/python2.7/dist-packages/caffe2/python/core.py", line 1846, in AddGradientOperators
self._net.op[skip:], ys)
File "/usr/local/lib/python2.7/dist-packages/caffe2/python/core.py", line 1113, in GetBackwardPass
return ir.GetBackwardPass(ys)
File "/usr/local/lib/python2.7/dist-packages/caffe2/python/core.py", line 984, in GetBackwardPass
forward_op_idx, all_input_to_grad)
File "/usr/local/lib/python2.7/dist-packages/caffe2/python/core.py", line 934, in GenerateGradientsForForwardOp
forward_op, g_output)
File "/usr/local/lib/python2.7/dist-packages/caffe2/python/core.py", line 1086, in GetGradientForOp
format(op.type, e, str(op))
Exception: Exception when creating gradient for [Python]:[enforce fail at operator_gradient.h:166] g_output.at(i).IsDense(). Gradient of output gpu_1/rois is not provided! .

and i get some information from "ef GetGradientForOp(cls, op, g_output): in core.py"
[None, 'gpu_1/loss_kps_grad']
[u'gpu_1/kps_score_reshaped_grad', None]
[u'gpu_1/kps_score_grad']
[u'gpu_1/kps_score_lowres_grad']
[u'gpu_1/conv_fcn8_grad']
[u'gpu_1/conv_fcn8_grad']
[u'gpu_1/conv_fcn7_grad']
[u'gpu_1/conv_fcn7_grad']
[u'gpu_1/conv_fcn6_grad']
[u'gpu_1/conv_fcn6_grad']
[u'gpu_1/conv_fcn5_grad']
[u'gpu_1/conv_fcn5_grad']
[u'gpu_1/conv_fcn4_grad']
[u'gpu_1/conv_fcn4_grad']
[u'gpu_1/conv_fcn3_grad']
[u'gpu_1/conv_fcn3_grad']
[u'gpu_1/conv_fcn2_grad']
[u'gpu_1/conv_fcn2_grad']
[u'gpu_1/conv_fcn1_grad']
[u'gpu_1/conv_fcn1_grad']
[u'gpu_1/_[pose]roi_feat_grad']
[u'gpu_1/[pose]roi_feat_shuffled_grad', None]
[u'gpu_1/[pose]roi_feat_fpn5_grad']
[u'gpu_1/[pose]roi_feat_fpn4_grad']
[u'gpu_1/[pose]roi_feat_fpn3_grad']
[u'gpu_1/[pose]roi_feat_fpn2_grad']
['gpu_1/loss_mask_grad']
[u'gpu_1/dist_grad']
[u'gpu_1/dt_output_grad']
[u'gpu_1/mask_fcn_logits_grad']
[u'gpu_1/conv5_mask_grad']
[u'gpu_1/conv5_mask_grad']
[u'gpu_1/[mask]fcn4_grad']
[u'gpu_1/[mask]fcn4_grad']
[u'gpu_1/[mask]fcn3_grad']
[u'gpu_1/[mask]fcn3_grad']
[u'gpu_1/[mask]fcn2_grad']
[u'gpu_1/[mask]fcn2_grad']
[u'gpu_1/[mask]fcn1_grad']
[u'gpu_1/[mask]fcn1_grad']
[u'gpu_1/[mask]roi_feat_grad']
[u'gpu_1/[mask]roi_feat_shuffled_grad', None]
[u'gpu_1/[mask]roi_feat_fpn5_grad']
[u'gpu_1/[mask]roi_feat_fpn4_grad']
[u'gpu_1/[mask]roi_feat_fpn3_grad']
[u'gpu_1/[mask]_roi_feat_fpn2_grad']
['gpu_1/loss_bbox_grad']
[None, 'gpu_1/loss_cls_grad']
[u'gpu_1/bbox_pred_grad']
[u'gpu_1/cls_score_grad']
[u'gpu_1/fc7_grad']
[u'gpu_1/fc7_grad']
[u'gpu_1/fc6_grad']
[u'gpu_1/fc6_grad']
[u'gpu_1/roi_feat_grad']
[u'gpu_1/roi_feat_shuffled_grad', None]
[u'gpu_1/roi_feat_fpn5_grad']
[u'gpu_1/roi_feat_fpn4_grad']
[u'gpu_1/roi_feat_fpn3_grad']
[u'gpu_1/roi_feat_fpn2_grad']
['gpu_1/loss_rpn_bbox_fpn6_grad']
['gpu_1/loss_rpn_cls_fpn6_grad']
['gpu_1/loss_rpn_bbox_fpn5_grad']
['gpu_1/loss_rpn_cls_fpn5_grad']
['gpu_1/loss_rpn_bbox_fpn4_grad']
['gpu_1/loss_rpn_cls_fpn4_grad']
['gpu_1/loss_rpn_bbox_fpn3_grad']
['gpu_1/loss_rpn_cls_fpn3_grad']
['gpu_1/loss_rpn_bbox_fpn2_grad']
['gpu_1/loss_rpn_cls_fpn2_grad']
[None, None, None, None, None, None, None, u'gpu_1/masks_int32_grad', u'gpu_1/masks_int32_weights_grad', None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None]

anyone help me ? thanks

[xieshuqin]

The problem is that the GenerateProposalsOp and the CollectAndDistributeProposalsOp aren't expecting any gradients income.

For example, looking at the CrossEntropyLoss code, it only returns the gradients for X but not Y. Thus the above Op won't cause you problems. However, when you are using SquaredL2Distance, it will return both the gradients for X and Y, then you get the errors.

To fixed this, maybe you can try something like this,

blob_masks_int32 = core.ScopedBlobReference('masks_int32')
blob_masks_int32 = model.net.StopGradients(blob_masks_int32, blob_masks_int32)
dist=model.net.SquaredL2Distance([dt_output, blob_masks_int32], 'dist')
loss_mask=model.net.AveragedLoss( dist, 'loss_mask')
loss_gradients = blob_utils.get_loss_gradients(model, [loss_mask])

Or, you may want to write a new caffe2 op similar to SquaredL2Distance but only returns the gradient for X. That should do too.

[rbgirshick]

Yes, using StopGradient is the correct solution.