Caffe
Accelerator Server
important
When you use a machine learning environment, please change the current directory to /work_da area and submit the job to CA_001g queue.
The following commands are available.
For details, see the -help option.
caffe classification classify compute_image_mean convert_cifar_data convert_imageset convert_mnist_data convert_mnist_siamese_data detect device_query draw_net extract_features finetune_net net_speed_benchmark test_net train train_net upgrade_net_proto_binary upgrade_net_proto_text upgrade_solver_proto_text
The job script is as follows.
#!/bin/sh #PBS -l select=1[:ncpus=number of CPU][:ngpus=number of GPU] #PBS -q CA_001g #PBS -N jobname #PBS -v DOCKER_IMAGE=conda3/mlenv:tag cd $PBS_O_WORKDIR command option
(Example)
#!/bin/sh #PBS -l select=1 #PBS -q CA_001g #PBS -N Caffe #PBS -v DOCKER_IMAGE=conda3/mlenv:cuda10.1-007 cd $PBS_O_WORKDIR caffe train --solver=examples/cifar10/cifar10_quick_solver.prototxt
The command for interactive mode is as follows.
$ qsub -I -q CA_001g -v DOCKER_IMAGE=conda3/mlenv:tag
例）
$ qsub -I -q CA_001g -v DOCKER_IMAGE=conda3/mlenv:cuda10.1-007 qsub: waiting for job 26269.gpu1 to start qsub: job 26269.gpu1 ready bash-4.2$ cd $PBS_O_WORKDIR bash-4.2$ caffe train --solver=examples/cifar10/cifar10_quick_solver.prototxt