-
Notifications
You must be signed in to change notification settings - Fork 5
/
crl_arithlang.py
218 lines (199 loc) · 9.55 KB
/
crl_arithlang.py
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
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
import argparse
import numpy as np
import os
import torch
import torch.nn.functional as F
import cmdp_trainer as ct
from dataloader.multilingual_dataset import ArithmeticLanguageWordEncoding
from learners.multilingual import CRL_MultitaskSequenceAgent
from log import Logger, mkdirp
from networks.encoders import Identity
from sample_episode.cmdp_sample_episode import sample_episode, sample_data
parser = argparse.ArgumentParser(description='Learned Functions')
################################################################################
# Update Every
parser.add_argument('--computation_update', type=int, default=256,
help='number of episodes before updating computation (default: 256)')
parser.add_argument('--computation_update_offset', type=int, default=0,
help='offset for updating computation (default: 0)')
parser.add_argument('--policy_update', type=int, default=1024,
help='number of episodes before updating policy (default: 1024)')
parser.add_argument('--policy_update_offset', type=int, default=0,
help='offset for updating policy (default: 0)')
# LR
parser.add_argument('--step_penalty', type=float, default=1e-2,
help='step_penalty (default: 1e-2)')
parser.add_argument('--plr', type=float, default=5e-4,
help='learning rate (default: 5e-4)')
parser.add_argument('--clr', type=float, default=1e-3,
help='learning rate (default: 1e-3)')
parser.add_argument('--lr-mult-min', type=float, default=1e-4,
help='minimum learning rate multiplier')
parser.add_argument('--anneal-policy-lr', action='store_true',
help='linearly anneal learning rate for policy')
parser.add_argument('--anneal-comp-lr', action='store_true',
help='linearly anneal learning rate for computation')
################################################################################
# Data
parser.add_argument('--bsize', type=int, default=1,
help='input batch size for training (default: 1)')
parser.add_argument('--env', type=str, default='arithlang',
help='arithlang')
parser.add_argument('--maxterms', nargs='+', type=int, default=[5,5,10],
help='number of arithmetic terms: train | val/test | extrap_val/extrap_test')
parser.add_argument('--numrange', nargs='+', type=int, default=[0,10],
help='')
parser.add_argument('--ops', type=str, default='+*-',
help='types of operations for math')
parser.add_argument('--samplefrom', type=int, default=1e4,
help='max number of problems to sample from (1e4)')
parser.add_argument('--episodecap', type=int, default=1e3,
help='max number of problems to record (1e3)')
parser.add_argument('--nlang', type=int, default=5,
help='number of languages (5)')
parser.add_argument('--pretrainmode', type=str, default='ed',
help='ed | ring')
################################################################################
# Model
parser.add_argument('--indim', type=int, default=784,
help='hidden dimension of model (default: 784)')
parser.add_argument('--hdimp', type=int, default=128,
help='hidden dimension of policy (default: 128)')
parser.add_argument('--hdimf', type=int, default=128,
help='hidden dimension of function (default: 128)')
parser.add_argument('--outdim', type=int, default=10,
help='hidden dimension of model (default: 10)')
parser.add_argument('--nactions', type=int, default=12,
help='number of functions (default: 12)')
parser.add_argument('--nreducers', type=int, default=3,
help='number of functions (default: 4)')
parser.add_argument('--ntranslators', type=int, default=8,
help='number of functions (default: 8)')
parser.add_argument('--nsteps', type=int, default=4,
help='number of steps of computation (default: 4)')
parser.add_argument('--model', type=str, default='crl_seq',
help='crl_seq')
parser.add_argument('--policylayers', type=int, default=1,
help='number of layers of the policy')
parser.add_argument('--functionlayers', type=int, default=1,
help='number of layers of the function')
################################################################################
# Algorithm
parser.add_argument('--ppo_optim_epochs', type=int, default=5,
help='hidden dimension of model (default: 5)')
parser.add_argument('--ppo_value_iters', type=int, default=1,
help='hidden dimension of model (default: 1)')
parser.add_argument('--ppo_minibatch_size', type=int, default=256,
help='hidden dimension of model (default: 256)')
parser.add_argument('--ppo_anneal_epochs', action='store_true',
help='anneal number of ppo epochs from ppo_optim_epochs to 1')
parser.add_argument('--ppo_clip', type=float, default=0.1,
help='hidden dimension of model (default: 0.1)')
parser.add_argument('--entropy_coeff', type=float, default=1e-4,
help='hidden dimension of model (default: 0.01)')
parser.add_argument('--gamma', type=float, default=0.99,
help='discount factor (default: 0.99)')
################################################################################
# Experimental Config
parser.add_argument('--max_episodes', type=int, default=1e7,
help='Maximum number of training episodes')
parser.add_argument('--seed', type=int, default=1,
help='random seed (default: 1)')
parser.add_argument('--no-cuda', action='store_true', default=False,
help='disables CUDA training')
parser.add_argument('--log_interval', type=int, default=1e3, metavar='N',
help='interval between training status logs (default: 1000)')
parser.add_argument('--plot', action='store_true', default=True,
help='plot')
parser.add_argument('--plot_every', type=int, default=10000,
help='plot every (default: 1e5=4)')
parser.add_argument('--track_selected', action='store_true',
help='track the function selection')
parser.add_argument('--printf', action='store_true',
help='print to file')
parser.add_argument('--save_every', type=int, default=10000,
help='number of episodes before saving')
parser.add_argument('--val_every', type=int, default=10000,
help='number of episodes before validation')
parser.add_argument('--numval', type=int, default=100,
help='number of validation examples')
parser.add_argument('--curr_every', type=int, default=1e5,
help='number of episodes before increasing the dataset')
parser.add_argument('--outputdir', type=str, default='runs/arith_verify/crl',
help='outputdir')
parser.add_argument('--eval', action='store_true',
help='eval mode')
parser.add_argument('--ckpt_every', action='store_true',
help='save ckpt every 1e5 iterations')
parser.add_argument('--resume', type=str, default='', metavar='R',
help='path of saved model')
parser.add_argument('--debug', action='store_true',
help='debug')
args = parser.parse_args()
args.cuda = not args.no_cuda and torch.cuda.is_available()
seeder = torch.cuda.manual_seed if args.cuda else torch.manual_seed
np.random.seed(args.seed)
seeder(args.seed)
if args.outputdir != '':
if not os.path.exists(args.outputdir):
os.mkdir(args.outputdir)
def process_args(args):
if args.debug:
args.max_episodes = 40
args.val_every = 20
args.save_every = 20
args.numval = 10
args.log_interval = 5
args.computation_update = 10
args.policy_update = 10
args.curr_every = 4
args.outputdir = 'runs/arith_verify/crl/debug'
return args
def build_expname(args):
expname = 'env-{}'.format(args.env)
expname += '_agent-{}'.format(args.model)
if args.debug: expname+= '_debug'
return expname
def main():
envbuilder = ArithmeticLanguageWordEncoding
root = 'data'
env = envbuilder(
max_terms=args.maxterms,
num_range=args.numrange,
ops=args.ops,
samplefrom=args.samplefrom,
episodecap=args.episodecap,
root=root,
curr=True,
nlang=args.nlang
)
args.indim = env.vocabsize
args.outdim = env.vocabsize
encoder = Identity
decoder = lambda: Identity(
predictor=lambda x: torch.max(x.data,-1)[1])
num_actions = args.nactions
expname = build_expname(args)
main_method = ct.eval if args.eval else ct.train
args.memory = False
logdir = expname
logger = Logger(expname=logdir, logdir=os.path.join(args.outputdir, logdir), params=args)
agent = CRL_MultitaskSequenceAgent(
indim=args.indim,
langsize=env.langsize,
zsize=env.zsize,
hdimp=args.hdimp,
hdimf=args.hdimf,
outdim=args.indim,
num_steps=args.nsteps,
num_actions=num_actions,
layersp=args.policylayers,
layersf=args.functionlayers,
encoder=encoder,
decoder=decoder,
args=args,
relax=True)
main_method(sample_data, lambda a, e, i, t, m: sample_episode(a, e, i, t, 'crl_seq', m), agent, logger, env, args)
if __name__ == "__main__":
args = process_args(args)
main()