Open donglinjy opened 7 years ago
Looks like you are running cos in trainwithvalidation mode which produces interleaved output of test after every test interval defined in the solver proto txt . If you just want to train please set the test interval and test iteration to 0 in the solver file
On Tue, Dec 20, 2016 at 5:53 PM PigApple notifications@github.com wrote:
When we run caffe alone, we can get the test net output
I1220 08:31:50.558744 2105 solver.cpp:317] Iteration 1000, loss = 0.0884654
*I1220 08:31:50.558794 2105 solver.cpp:337] Iteration 1000, Testing net (#0)I1220 08:31:50.863924 2105 solver.cpp:391] Test loss: 0.0866651I1220 08:31:50.863971 2105 solver.cpp:404] Test net output #0: accuracy = 0.972I1220 08:31:50.863983 2105 solver.cpp:404] Test net output #1 https://github.com/yahoo/CaffeOnSpark/pull/1: loss = 0.0866651 ( 1 = 0.0866651 loss)**
But in CaffeonSpark, instead of the caffe testint net output, we will get some thing cos generated by itself,
*I1212 08:58:36.250149 14322 CaffeNet.cpp:36] Iteration 200, Testing net (#0)I1212 08:58:36.250187 14322 CaffeNet.cpp:55] Test net output #0: accuracy = 0.88I1212 08:58:36.250206 14322 CaffeNet.cpp:55] Test net output
1 https://github.com/yahoo/CaffeOnSpark/pull/1: loss = 0.365099 ( 1 =
0.365099 loss)**
I1212 08:58:36.384349 14322 solver.cpp:228] Iteration 200, loss = 0.161356
I1212 08:58:36.384402 14322 solver.cpp:244] Train net output #0: loss = 0.161356 (* 1 = 0.161356 loss)
I1212 08:58:36.384413 14322 sgd_solver.cpp:106] Iteration 200, lr = 0.00985258
Any one know if this is a designed behavior that in CaffeonSpark, the caffe itself test net output will not print?
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/yahoo/CaffeOnSpark/issues/211, or mute the thread https://github.com/notifications/unsubscribe-auth/ACCTVf1-RghqlmrbhLT1aOJB6vKOCVUJks5rKIaKgaJpZM4LSffs .
@mriduljain , thanks. But it is my intention to run with the test output and the issue is that when I run in that mode, I only get the CaffeonSpark test net output but the caffe original output is not in the log. Do you know how can get it?
We added a wrapper around some of the Caffe functions to fit in the Spark framework. The output you see from CoS should be equivalent to what you see from Caffe.
When we run caffe alone, we can get the test net output I1220 08:31:50.558744 2105 solver.cpp:317] Iteration 1000, loss = 0.0884654 *I1220 08:31:50.558794 2105 solver.cpp:337] Iteration 1000, Testing net (#0) I1220 08:31:50.863924 2105 solver.cpp:391] Test loss: 0.0866651 I1220 08:31:50.863971 2105 solver.cpp:404] Test net output #0: accuracy = 0.972 I1220 08:31:50.863983 2105 solver.cpp:404] Test net output #1: loss = 0.0866651 ( 1 = 0.0866651 loss)**
But in CaffeonSpark, instead of the caffe testint net output, we will get some thing cos generated by itself, *I1212 08:58:36.250149 14322 CaffeNet.cpp:36] Iteration 200, Testing net (#0) I1212 08:58:36.250187 14322 CaffeNet.cpp:55] Test net output #0: accuracy = 0.88 I1212 08:58:36.250206 14322 CaffeNet.cpp:55] Test net output #1: loss = 0.365099 ( 1 = 0.365099 loss)* I1212 08:58:36.384349 14322 solver.cpp:228] Iteration 200, loss = 0.161356 I1212 08:58:36.384402 14322 solver.cpp:244] Train net output #0: loss = 0.161356 ( 1 = 0.161356 loss) I1212 08:58:36.384413 14322 sgd_solver.cpp:106] Iteration 200, lr = 0.00985258
Any one know if this is a designed behavior that in CaffeonSpark, the caffe itself test net output will not print?