GPU utilization become zero after long term training about detectron

[gaopeng-eugene]

python2 tools/train_net.py
--cfg configs/getting_started/tutorial_8gpu_e2e_faster_rcnn_R-50-FPN.yaml
OUTPUT_DIR /tmp/detectron-output

problem the GPU utilization become zero, however CPU is using a lot resource.

[gaopeng-eugene]

terminate called after throwing an instance of 'caffe2::EnforceNotMet'
what(): [enforce fail at context_gpu.h:230] error == cudaSuccess. 4 vs 0. Error at: /home/pgao/caffe2/caffe2/caffe2/core/context_gpu.h:230: unspecified launch failure Error from operator:
input: "gpu_0/res5_0_branch2c_w_grad" output: "gpu_2/res5_0_branch2c_w_grad" name: "" type: "Copy" device_option { device_type: 1 cuda_gpu_id: 2 }
terminate called recursively
*** Aborted at 1516782983 (unix time) try "date -d @1516782983" if you are using GNU date ***
terminate called recursively
terminate called recursively
terminate called recursively
terminate called recursively
terminate called recursively
terminate called recursively
terminate called recursively
terminate called recursively
terminate called recursively
terminate called recursively
terminate called recursively
terminate called recursively
terminate called recursively
terminate called recursively
terminate called recursively
terminate called recursively
terminate called recursively
terminate called recursively
terminate called recursively
terminate called recursively
terminate called recursively

[terrychenism]

same problem...

[rbgirshick]

@gaopeng-eugene, @terrychenism: can you try switching to the NCCL implementation of AllReduce to see if that resolves the problem? Instructions for building Caffe2 with NCCL support and enabling NCCL in Detectron can be found in #32.

[terrychenism]

fix it by using UCCL