-
Notifications
You must be signed in to change notification settings - Fork 10
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
error/ loss aways equal 87.3362. I change the lr but not work. #1
Comments
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). |
I trained on the cifar10.
Attached is my prototxt file.
Thanks!
2017-11-28 17:33 GMT+08:00 Sven Koitka <notifications@github.com>:
… 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>
.
|
No, it is not. |
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 <notifications@github.com>:
… 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>
.
|
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). |
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 <notifications@github.com>:
… 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>
.
|
Your learning rate seems very low. I had a look at my old experiments and here is one of my solver configs:
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. |
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 <notifications@github.com>:
… 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>
.
|
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
The text was updated successfully, but these errors were encountered: