@yupbank Doesn't sound like a deadlock. You are training LogisticRegression 1000 times on a (100000, 200) input dataset. To do that on a (1000, 200) dataset takes already ~1 min on a 4 core cpu. Not sure what's the exact scaling of LogistcRegression with n_samples when using lbfgs, but this would probably take hours to compute.

MultioutputClassifier.fit never ends with base classifier supports n_jobs

How long have you waited?

~40 minutes.

Using how many CPU? This could take ≳ 1.3 CPU-hours in my estimation...

Also using n_jobs > 1 actually slow things down here,

Using a 100 output vector (instead of 1000 as in your example) on a 4 core CPU,

• MultiOutputClassifier(base_old, n_jobs=1):

  $ time python /tmp/test.py
  
  real    0m42.301s
  user    2m42.168s
  sys     0m2.268s
  

• MultiOutputClassifier(base_old, n_jobs=4):

  $ time python /tmp/test.py
  
  real    2m49.912s
  user    9m4.964s
  sys     2m8.304s
  

so using n_jobs=4 slows the computations almost exactly 4 times here. Probably for the same reason as #8216 due joblib.Parallel pickling / memmaping overhead...

interesting.. well.. i have made this finish within ~3mins with 1000 output vector with updated loss_function and gradient

It's hard to tell whether there is an actual problem here, or whether it is just that your snippet take a long time to run. I am going to close this one, @yupbank feel free to reopen if you have some new information to add to this issue.