ddqn_train.submit_file

#---------------------------------------------
# Name your batch so it's easy to distinguish in the q.
JobBatchName = "PBN RL DDQN"

# ---------------------------------------------------
# Universe (vanilla, docker)
universe     = docker
docker_image = registry.eps.surrey.ac.uk/pbn-rl:ed941756

# -------------------------------------------------
# Event, out and error logs
log    = logs/$(exp).$(steps).$(seed).log
output = outs/$(exp).$(steps).$(seed).log
error  = logs/error/$(exp).$(steps).$(seed).log

# -----------------------------------
# File Transfer, Input, Output
# should_transfer_files = YES

environment = "mount=$ENV(PWD) WANDB_API_KEY=$ENV(WANDB_API_KEY) WANDB_RUN_GROUP=$(group)"
requirements = (CUDAGlobalMemoryMb > 3000) && \
               (HasStornext) && \
               (CUDACapability > 7.0)


# --------------------------------------
# Resources
request_CPUs   = 1
request_memory = 5G
request_GPUs   = 1
+GPUMem        = 4000

#This job can checkpoint
+CanCheckpoint = true

# -----------------------------------
# Queue commands. We can use variables and flags to launch our command with multiple options (as you would from the command line)
arguments = $(script) --seed $(seed) --time-steps $(steps) --env $(env) --exp-name $(exp) --resume-training --checkpoint-dir $ENV(PWD)/models --log-dir $ENV(PWD)/logs --learning-starts $(learningstarts) --hyperparams $(hyperparams) --eval-only

script = $ENV(PWD)/train_ddqn.py

env_n = 28
batch_size = 256
env = gym-PBN/Bittner-$(env_n)-v0
exp = ddqn-$(env_n)-b$(batch_size)
group = ddqn-$(env_n)

hyperparams = exploration_fraction=0.25,batch_size=$(batch_size)
steps = 150000
learningstarts = 30000

# Hours the job will take at maximum
+JobRunTime = 3

queue 1 seed in 42 727 420 69 1337