error/ loss aways equal 87.3362. I change the lr but not work.

[mqchen1993]

I1128 08:59:24.073550 14972 solver.cpp:397] Test net output #0: Accuracy1 = 1
I1128 08:59:24.073571 14972 solver.cpp:397] Test net output #1: SoftmaxWithLoss1 = 87.3362 (* 1 = 87.3362 loss)
I1128 08:59:24.187857 14972 solver.cpp:218] Iteration 1170 (6.24905 iter/s, 62.4095s/390 iters), loss = 87.3363
I1128 08:59:24.187876 14972 solver.cpp:237] Train net output #0: SoftmaxWithLoss1 = 87.3365 (* 1 = 87.3365 loss)
I1128 08:59:24.187881 14972 sgd_solver.cpp:105] Iteration 1170, lr = 0.1
I1128 09:00:11.118723 14972 solver.cpp:330] Iteration 1560, Testing net (#0)
I1128 09:00:26.383985 14979 data_layer.cpp:73] Restarting data prefetching from start.
I1128 09:00:26.506765 14972 solver.cpp:397] Test net output #0: Accuracy1 = 1
I1128 09:00:26.506784 14972 solver.cpp:397] Test net output #1: SoftmaxWithLoss1 = 87.3362 (* 1 = 87.3362 loss)
I1128 09:00:26.621100 14972 solver.cpp:218] Iteration 1560 (6.24695 iter/s, 62.4305s/390 iters), loss = 87.3363
I1128 09:00:26.621127 14972 solver.cpp:237] Train net output #0: SoftmaxWithLoss1 = 87.3365 (* 1 = 87.3365 loss)
I1128 09:00:26.621134 14972 sgd_solver.cpp:105] Iteration 1560, lr = 0.1
I1128 09:01:13.576076 14972 solver.cpp:330] Iteration 1950, Testing net (#0)
I1128 09:01:33.771523 14979 data_layer.cpp:73] Restarting data prefetching from start.
I1128 09:01:34.000592 14972 solver.cpp:397] Test net output #0: Accuracy1 = 1
I1128 09:01:34.000615 14972 solver.cpp:397] Test net output #1: SoftmaxWithLoss1 = 87.3362 (* 1 = 87.3362 loss)
I1128 09:01:34.220914 14972 solver.cpp:218] Iteration 1950 (5.76949 iter/s, 67.597s/390 iters), loss = 87.3363
I1128 09:01:34.220949 14972 solver.cpp:237] Train net output #0: SoftmaxWithLoss1 = 87.3365 (* 1 = 87.3365 loss)
I1128 09:01:34.220955 14972 sgd_solver.cpp:105] Iteration 1950, lr = 0.1
I1128 09:03:16.823488 14972 solver.cpp:330] Iteration 2340, Testing net (#0)
I1128 09:03:48.873394 14979 data_layer.cpp:73] Restarting data prefetching from start.
I1128 09:03:49.115883 14972 solver.cpp:397] Test net output #0: Accuracy1 = 1
I1128 09:03:49.115906 14972 solver.cpp:397] Test net output #1: SoftmaxWithLoss1 = 87.3362 (* 1 = 87.3362 loss)
I1128 09:03:49.344960 14972 solver.cpp:218] Iteration 2340 (2.88635 iter/s, 135.119s/390 iters), loss = 87.3363

[razorx89]

Without further information about your training setup, I am unable to help you. What does your solver.prototxt look like? Which network did you use? On which dataset are you training? A constant high loss value normally indicates an issue with the optimizer settings (too high learning rate or weight decay).

[mqchen1993]

I trained on the cifar10. Attached is my prototxt file. Thanks! 2017-11-28 17:33 GMT+08:00 Sven Koitka <[email protected]>:

…

Without further information about your training setup, I am unable to help you. What does your solver.prototxt look like? Which network did you use? On which dataset are you training? A constant high loss value normally indicates an issue with the optimizer settings (too high learning rate or weight decay). — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#1 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AaydEHIf1FGjTVf30OrHnN0bhJwA038Fks5s69NogaJpZM4Qs7Sc> .

[razorx89]

No, it is not.

[mqchen1993]

Hello, why not this, this can not be used in cifar10. I downloaded from your homepage. Thanks! ​ 2017-11-28 17:55 GMT+08:00 Sven Koitka <[email protected]>:

…

No, it is not. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#1 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AaydEGJGcxtf6Ht0QJAgc0aPZiu5e0oGks5s69iWgaJpZM4Qs7Sc> .

[razorx89]

You did not attach any prototxt file above. Furthermore, I requested the solver.prototxt and not trainval.prototxt (solver.prototxt is not included in the examples).

[mqchen1993]

I am sorry, it may be my side of the transmission error. I'll pass it again. Thanks！ net: "examples/cifar10/wrn/cifar100_WRN-16-4_dropout_train_val.prototxt" test_iter: 100 test_interval: 1000 base_lr: 0.0001 #clip_gradients:30 display: 100 max_iter: 64000 lr_policy: "multistep" gamma: 0.1 momentum: 0.9 weight_decay: 0.0005 snapshot: 10000 snapshot_prefix: "examples/cifar10/wrn" solver_mode: GPU test_initialization: true stepvalue: 32000 stepvalue: 48000 2017-11-28 18:53 GMT+08:00 Sven Koitka <[email protected]>:

…

You did not attach any prototxt file above. Furthermore, I requested the solver.prototxt and not trainval.prototxt (solver.prototxt is not included in the examples). — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#1 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AaydEOkfG20f0-TUD-fmX_YNx-z1emlhks5s6-Y3gaJpZM4Qs7Sc> .

[razorx89]

Your learning rate seems very low. I had a look at my old experiments and here is one of my solver configs:

net: "cifar100_WRN-16-8_train_val.prototxt"
max_iter: 46875

test_iter: 100
test_interval: 391
test_initialization: false

display: 10
average_loss: 10

solver_mode: GPU
type: "Nesterov"
momentum: 0.9
base_lr: 0.1
weight_decay: 0.00025
lr_policy: "step"
stepsize: 15625
gamma: 0.1

snapshot: 3907
snapshot_prefix: "snapshot_cifar100_WRN-16-8"

Since you are specifying momentum you should also specify the solver type (defaults to SGD https://github.com/BVLC/caffe/blob/master/src/caffe/proto/caffe.proto#L212). Besides that it could be that your input data has errors. Please check on how you are building your lmdbs.

[mqchen1993]

I modified the solver file as you gave it. But still failed. Dataset is no problem, because I use other network training is possible. Thanks！ 2017-11-28 19:08 GMT+08:00 Sven Koitka <[email protected]>:

…

Your learning rate seems very low. I had a look at my old experiments and here is one of my solver configs: net: "cifar100_WRN-16-8_train_val.prototxt" max_iter: 46875 test_iter: 100 test_interval: 391 test_initialization: false display: 10 average_loss: 10 solver_mode: GPU type: "Nesterov" momentum: 0.9 base_lr: 0.1 weight_decay: 0.00025 lr_policy: "step" stepsize: 15625 gamma: 0.1 snapshot: 3907 snapshot_prefix: "snapshot_cifar100_WRN-16-8" Since you are specifying momentum you should also specify the solver type (defaults to SGD https://github.com/BVLC/caffe/ blob/master/src/caffe/proto/caffe.proto#L212). Besides that it could be that your input data has errors. Please check on how you are building your lmdbs. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#1 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AaydEE7lvDRgQV6M31Q94rXqxplwn0I-ks5s6-mSgaJpZM4Qs7Sc> .