train_val.prototxt and solver.prototxt for Resnet-152 layer

[ghost]

trafficstars

Hi All, Past few weeks i have been trying to Train ResNet-152 on BVLC caffe with Imagenet 2012 dataset. But my accuracy or learning remains flat all the time, can some one share the train_val.prototxt and solver.prototxt. i tried pynetbuilder but some how couldn't crack it. If someone is successful can some one share these prototxt files

[ghost]

solver.prototxt.txt train_val1_xavier_init_prototxt.txt train_val2_use_global_stats_flase_prototxt.txt

Following are my prototxt whcih i tried... where the accuracy is flat after few epoch of Training on 4 GPU

[acgtyrant]

You can try to change weight_filler from xavier to msra.

[ps48]

@tyrosathish I am facing the similar problem, I tried similar approach but the loss rate is just flat.

[mathfinder]

@tyrosathish can you share me your resnet-152.caffemodel?i can't parse the caffemodel which i download from the link.

Check failed: ReadProtoFromBinaryFile(param_file, param) Failed to parse NetParameter file: ResNet-152-model.caffemodel

[codegank]

Hi guys,

I am facing the same problem too, my loss is almost around 6.9 and can not decrease. Anybody have fixed this problem?

[HolmesShuan]

.prototxt I try to re-implement ResNet-18 on ILSVRC12, but my accuracy is about 86.66%@Top-5. Baseline is about 90%+ @Torch.

[thigi]

@HolmesShuan Why are you using xavier filler for the Fully-Connected layer and msra filler for all the convolutional layers. I once did not use a weight filler for the Fully-Connected layer and retrieved better results. Can you explain that?

[HolmesShuan]

@thigi I find the unofficial prototxt here. To be honest, I haven't studied it yet. Sorry. :(

[whitewalker007]

Hi! @thigi. I just saw those prototxt files there and got confused as the guy has used type:Data instead type:MemoryData in the TEST layer. Please enlighten me on 'if this change will or will not effect the performance of ResNet. Thanks in advance!

[shichaosuper]

met the same problem too

[zangkaiqiang]

same problem too.. anyone get the answer?