forked from bilocq/Trust-the-Critics
-
Notifications
You must be signed in to change notification settings - Fork 0
/
ttc.py
165 lines (130 loc) · 6.69 KB
/
ttc.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
"""
Cleaned up code for training Trust The Critics (TTC) scheme. Will train a list of critic networks that can subsequently be used to push the source
dataset to the target dataset.
IMPORTANT INPUTS
- source (defaults to noise): The name of the distribution or dataset to push towards the target.
- target (defaults to mnist): The name of a dataset.
- data (required): A directory where the necessary data is located.
- temp_dir (required): A directory where the trained critics will be saved.
- model: The choice of architectures for the critics.
- num_crit: The number of critic networks used to push the source to the target.
- theta: The step size parameter described in the paper
Note: the default source and target correspond to generating MNIST-like samples.
OUTPUTS
Running this script will save the following files in temp_dir:
- The trained critic networks, saved under temp_dir/model_dicts/critic0.pth, temp_dir/model_dicts/critic1, etc.
- A .pkl log file containing, among other things, the step sizes for each critic (as in eq. (14) of the paper). Saved under temp_dir/log.pkl
- A .txt file containing the configuration of the experiment, saved under temp_dir/train_config.txt.
"""
import os, sys
sys.path.append(os.getcwd())
sys.path.append(os.path.join(os.getcwd(),'TTC_utils'))
import argparse
import time
import log
import json
import random
import numpy as np
import torch
from torch import optim
import dataloader
import networks
from get_training_time import write_training_time
from critic_trainer import critic_trainer
#################
#Get command line args
#################
parser = argparse.ArgumentParser('Training code for TTC')
parser.add_argument('--source', type=str, default='noise', choices=['noise', 'untrained_gen', 'noisybsds500', 'photo', 'unit_sphere'])
parser.add_argument('--target', type=str, default='mnist', choices=['cifar10','mnist','fashion', 'celeba', 'bsds500', 'monet', 'celebaHQ', 'all_zero'])
parser.add_argument('--data', type=str, required=True, help = 'directory where data is located')
parser.add_argument('--temp_dir', type=str, required=True, help = 'temporary directory for saving')
parser.add_argument('--model', type=str, default='dcgan', choices=['dcgan', 'infogan', 'arConvNet', 'sndcgan','bsndcgan', 'norm_taker'])
parser.add_argument('--dim', type=int, default=64, help = 'int determining network dimensions')
parser.add_argument('--seed', type=int, default=-1, help = 'Set random seed for reproducibility')
parser.add_argument('--lamb', type=float, default=1000., help = 'parameter multiplying gradient penalty')
parser.add_argument('--theta', type=float, default=0.5, help = 'parameter determining step size as fraction of W1 distance')
parser.add_argument('--sigma', type=float, default=0.02, help = 'std of noise. Only relevant if doing denoising. Effective value 1/2 of this')
parser.add_argument('--critters', type=int, default=5, help = 'number of critic iters')
parser.add_argument('--bs', type=int, default=128, help = 'batch size')
parser.add_argument('--plus', action= 'store_true', help = 'take one sided penalty')
parser.add_argument('--num_workers', type=int, default = 0, help = 'number of data loader processes')
parser.add_argument('--num_crit', type=int, default = 5, help = 'number of critics to train')
parser.add_argument('--num_step', type=int, default=1, help = 'how many steps to use in gradient descent')
parser.add_argument('--beta_1', type=float, default=0.5, help = 'beta_1 for Adam')
parser.add_argument('--beta_2', type=float, default=0.999, help = 'beta_2 for Adam')
args = parser.parse_args()
use_cuda = torch.cuda.is_available()
print('Arguments:')
for p in vars(args).items():
print(' ',p[0]+': ',p[1])
print('\n')
#code to get deterministic behaviour
if args.seed != -1: #if non-default seed
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark=False #If true, optimizes convolution for hardware, but gives non-deterministic behaviour
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
np.random.seed(args.seed)
random.seed(args.seed)
else:
torch.backends.cudnn.benchmark=True
print('using benchmark')
#begin definitions ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
###################
# Initialize Dataset iterators
###################
target_loader = getattr(dataloader, args.target)(args, train=True)
args.num_chan = target_loader.in_channels
args.hpix = target_loader.hpix
args.wpix = target_loader.wpix
source_loader = getattr(dataloader, args.source)(args, train=True)
###################
#save args in config file
###################
config_file_name = os.path.join(args.temp_dir, 'train_config.txt')
with open(config_file_name, 'w') as f:
json.dump(args.__dict__, f, indent=2)
####################
#Initialize networks and optimizers
####################
critic_list = [None]*args.num_crit
steps = [1]*args.num_crit
for i in range(args.num_crit):
critic_list[i] = getattr(networks, args.model)(args.dim, args.num_chan, args.hpix, args.wpix)
if use_cuda:
for i in range(args.num_crit):
critic_list[i] = critic_list[i].cuda()
optimizer_list = [None]*args.num_crit
print('Adam parameters are {} and {}'.format(args.beta_1, args.beta_2))
for i in range(args.num_crit):
optimizer_list[i] = optim.Adam(critic_list[i].parameters(), lr=1e-4, betas=(args.beta_1, args.beta_2))
abs_start = time.time()
#main training loop
for iteration in range(args.num_crit):
############################
# (1) Train D network
###########################
#trains critic at critic_list[iteration], and reports current W1 distance
critic_list, Wasserstein_D = critic_trainer(critic_list, optimizer_list, iteration, steps, target_loader, source_loader, args)
###########################
# (2) Pick step size
###########################
steps[iteration] = args.theta*Wasserstein_D.detach()
###########################
# (3) freeze critic and save
###########################
for p in critic_list[iteration].parameters():
p.requires_grad = False # this critic is now fixed
if iteration< args.num_crit -1:
critic_list[iteration+1].load_state_dict(critic_list[iteration].state_dict())#initialize next critic at current critic
log.plot('steps', steps[iteration].cpu().data.numpy())
log.flush(args.temp_dir)
log.tick()
#Save critic model dicts
if (iteration % 10 == 9) or (iteration ==args.num_crit -1):
os.makedirs(os.path.join(args.temp_dir,'model_dicts'), exist_ok = True)
for j in range(iteration+1):
torch.save(critic_list[j].state_dict(), os.path.join(args.temp_dir,'model_dicts','critic{}.pth'.format(j)))
print(steps)
write_training_time(args)