Ask for training code about crfasrnn

[bittnt]

@hjwdzh
I believe you should check how to train FCN network from this. For example, https://gist.github.com/shelhamer/80667189b218ad570e82

To train our network, just append the multistagemeanfield layer at the bottom of a train.prototxt:
'''
layer { type: "Deconvolution" name: 'upsample-8' bottom: 'score-fr' top: 'bigscore' param { lr_mult: 0 } convolution_param { bias_term: false num_output: 21 kernel_size: 16 stride: 8 } }

layer { type: "Split" name: 'splitting' bottom: 'bigscore' top: 'unary' top: 'Q0' }

layer { name: "inference1" type: "MultiStageMeanfield" bottom: "unary" bottom: "Q0" bottom: "datamf" top: "upscore" param { lr_mult: 10000 } param { lr_mult: 10000 } param { lr_mult: 1000 } multi_stage_meanfield_param { num_iterations: 5 compatibility_mode: POTTS threshold: 2 theta_alpha: 160 theta_beta: 3 theta_gamma: 3 spatial_filter_weight: 3 bilateral_filter_weight: 5 } }

layer { name: "loss" type: "SoftmaxWithLoss" bottom: "upscore" bottom: "label" top: "loss" loss_param { ignore_label: 255 normalize: false } include: { phase: TRAIN } }
'''

[martinkersner]

I have a question related to learning rate inside of MultiStageMeanfield. In TVG_CRFRNN_COCO_VOC.prototxt you specify learning rates with values 0.001, 0.001 and 0.01. However, here you posted much higher learning rates 10000, 10000 and 1000. Using higher values I am able to achieve better results, however, I am still wondering which one did you use for training with base learning rate 1e-13 which is mentioned in paper.

[bittnt]

Glad to hear that you get better performance.

Setting up learning rate higher for this last layer , e.g. 1000 or 10000
helps to get top performance, within our end to end framework and base lr
at 1e-13.

For other purpose like that you might want to fine tune the network a bit
more within crf-rnn, in which case you can set them to 0.001 or some small
number.

On Thu, 25 Feb 2016 at 05:57, Martin Keršner [email protected]
wrote:

I have a question related to learning rate inside of MultiStageMeanfield.
In TVG_CRFRNN_COCO_VOC.prototxt you specify learning rates with values
0.001, 0.001 and 0.01. However, here you posted much higher learning rates
10000, 10000 and 1000. Using higher values I am able to achieve better
results, however, I am still wondering which one did you use for training
with base learning rate 1e-13 which is mentioned in paper.

—
Reply to this email directly or view it on GitHub
#28 (comment).

[sadeepj]

Closing old issues with no recent activity.