forked from alherit/cr-dqn
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathinstructions.txt
76 lines (46 loc) · 2.97 KB
/
instructions.txt
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
1) Install dqn_zoo:
git clone https://github.com/deepmind/dqn_zoo.git
cd dqn_zoo/
git checkout d798e3c6fdefe619708323c0b5a3f18f912a28c1
and follow dqn_zoo instructions to install its python dependencies (listed in docker-requirement.txt).
2) add the dqn_zoo directory to PYTHONPATH environment variable
3) The experiments were run as follow
######## SYNTHETIC EXPERIMENTS
### SET n_jobs accordingly to the number of cpus available
### DISABLE GPU
export CUDA_VISIBLE_DEVICES=
## FC CRAMER
nohup python synthetic.py --net fc --layers 2 --nodes_fc 45 --lr 0.001 --epochs 1000 --trials 100 --bs 32 --loss cramer --njobs 30 &> fc_cramer.out &
## FC QR0
nohup python synthetic.py --net fc --layers 2 --nodes_fc 45 --lr 0.001 --epochs 1000 --trials 100 --bs 32 --loss qr_loss --huber 0. --njobs 30 &> fc_qr0.out &
## FC QR0.1
nohup python synthetic.py --net fc --layers 2 --nodes_fc 45 --lr 0.001 --epochs 1000 --trials 100 --bs 32 --loss qr_loss --huber 0.1 --njobs 30 &> fc_qr0.1.out &
## FC QR1
nohup python synthetic.py --net fc --layers 2 --nodes_fc 45 --lr 0.001 --epochs 1000 --trials 100 --bs 32 --loss qr_loss --huber 1. --njobs 30 &> fc_qr1.out &
## FC W1
nohup python synthetic.py --net fc --layers 2 --nodes_fc 45 --lr 0.001 --epochs 1000 --trials 100 --bs 32 --loss wasserstein1 --njobs 30 &> fc_w1.out &
## NC CRAMER
nohup python synthetic.py --net nc --layers 2 --nodes_nc 32 --lr 0.001 --epochs 1000 --trials 100 --bs 32 --loss cramer --njobs 30 &> nc_cramer.out &
## NC QR0
nohup python synthetic.py --net nc --layers 2 --nodes_nc 32 --lr 0.001 --epochs 1000 --trials 100 --bs 32 --loss qr_loss --huber 0. --njobs 30 &> nc_qr0.out &
## NC SOFTPLUS CRAMER
nohup python synthetic.py --net nc --softplus --layers 2 --nodes_nc 32 --lr 0.001 --epochs 1000 --trials 100 --bs 32 --loss cramer --njobs 30 &> nc_softplus_cramer.out &
######## ATARI EXPERIMENTS
Better performance is achieved by assigning the process to a single cpu using cpu affinity (taskset).
Choose a proper cpu number to replace X and set output filenames accordingly
Pseudo-code for launching:
For GAME in <GAMES>
For SEED in 1,2,3
##NC_QR_DQN_1
nohup taskset -c X python run_atari.py --environment_name=<GAME> --seed=<SEED> --nc --nocramer --huber_param=1. --use_gym --results_csv_path=res.csv &> out.txt &
##NC_QR_DQN_0
nohup taskset -c X python run_atari.py --environment_name=<GAME> --seed=<SEED> --nc --nocramer --huber_param=0. --use_gym --results_csv_path=res.csv --scale_grad &> out.txt &
##NC_CR_DQN
nohup taskset -c X python run_atari.py --environment_name=<GAME> --seed=<SEED> --nc --use_gym --results_csv_path=res.csv &> out.txt &
##QR_DQN_0
nohup taskset -c X python run_atari.py --environment_name=<GAME> --seed=<SEED> --huber_param=0. --use_gym --results_csv_path=res.csv &> out.txt &
##CR_DQN
nohup taskset -c X python run_atari.py --environment_name=<GAME> --seed=<SEED> --use_gym --results_csv_path=res.csv &> out.txt &
#### PLOT
Works with Pandas 1.3.1
python plot_atari_curves.py