-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathPI.py
executable file
·158 lines (131 loc) · 5.35 KB
/
PI.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
# modified from https://github.com/YujiaBao/Predict-then-Interpolate/blob/main/src/main.py
import argparse
import numpy as np
import torch
from torchvision import datasets
from torch import nn, optim, autograd
import torch.nn.functional as F
import copy
import os
from mydatasets import coloredmnist
from models import MLP, TopMLP
from utils import pretty_print, correct_pred,GeneralizedCELoss, EMA, mean_weight, mean_nll, mean_mse, mean_accuracy
def validation(model, envs, test_envs, lossf):
with torch.no_grad():
for env in envs + test_envs:
logits = model(env['images'])
env['nll'] = lossf(logits, env['labels'])
env['acc'] = mean_accuracy(logits, env['labels'])
test_worst_loss = torch.stack([env['nll'] for env in test_envs]).max()
test_worst_acc = torch.stack([env['acc'] for env in test_envs]).min()
train_loss = torch.stack([env['nll'] for env in envs]).mean()
train_acc = torch.stack([env['acc'] for env in envs]).mean()
return train_loss.detach().cpu().numpy(), train_acc.detach().cpu().numpy(), \
test_worst_loss.detach().cpu().numpy(),test_worst_acc.detach().cpu().numpy()
parser = argparse.ArgumentParser(description='Colored MNIST & CowCamel')
parser.add_argument('--verbose', type=bool, default=True)
parser.add_argument('--n_restarts', type=int, default=10)
parser.add_argument('--dataset', type=str, default='coloredmnist025')
parser.add_argument('--hidden_dim', type=int, default=390)
parser.add_argument('--n_top_layers', type=int, default=2)
parser.add_argument('--l2_regularizer_weight', type=float,default=0.0011)
parser.add_argument('--lr', type=float, default=0.0005 )
parser.add_argument('--steps1', type=int, default=51)
parser.add_argument('--steps3', type=int, default=701)
parser.add_argument('--lossf', type=str, default='nll')
parser.add_argument('--save_dir', type=str, default='.')
flags = parser.parse_args()
if flags.dataset == 'coloredmnist025':
envs, test_envs = coloredmnist(0.25, 0.1, 0.2, int_target = False)
elif flags.dataset == 'coloredmnist01':
envs, test_envs = coloredmnist(0.1, 0.2, 0.25, int_target = False)
logs = []
for step in range(flags.n_restarts):
## load datasets
num_envs = len(envs)
## init models
input_dim = 14*14*2
n_targets = 1
models = []
def get_topmlp_func():
return TopMLP(hidden_dim = flags.hidden_dim, n_top_layers=flags.n_top_layers, n_targets=n_targets).cuda()
for i in range(num_envs):
mlp = MLP(hidden_dim = flags.hidden_dim, input_dim=input_dim).cuda()
topmlp = get_topmlp_func()
model = torch.nn.Sequential(mlp, topmlp)
models.append(model)
## Stage1: train models for each env. earlystopping on a 10% validation dataset (controled by --step1).
for i in range(num_envs):
print(i)
x, y = envs[i]['images'], envs[i]['labels']
idx = np.arange(len(x))
np.random.shuffle(idx)
val_x, val_y = x[idx[:int(len(idx)*0.1)]], y[idx[:int(len(idx)*0.1)]]
x, y = x[idx[int(len(idx)*0.1):]], y[idx[int(len(idx)*0.1):]]
model = models[i]
optimizer = optim.Adam(model.parameters(), lr=flags.lr)
lossf = mean_nll if flags.lossf == 'nll' else mean_mse
for step in range(flags.steps1):
logits = model(x)
loss = lossf(logits, y)
weight_norm = 0
for w in model.parameters():
weight_norm += w.norm().pow(2)
loss += flags.l2_regularizer_weight * weight_norm
optimizer.zero_grad()
loss.backward()
optimizer.step()
if step % 50 == 0:
with torch.no_grad():
tr_acc = mean_accuracy(model(x),y)
val_acc = mean_accuracy(model(val_x),val_y)
#$print(tr_acc, val_acc)
pretty_print(np.int32(step),tr_acc.detach().cpu().numpy(), val_acc.detach().cpu().numpy())
# train_loss, train_acc, test_worst_loss, test_worst_acc = \
# validation(model, [envs[i]], test_envs, lossf)
# log = [np.int32(step), train_loss, train_acc,test_worst_loss, test_worst_acc]
# if flags.verbose:
# pretty_print(*log)
## Stage2: cross group splitting
created_envs = []
for i in range(num_envs):
x, y = envs[i]['images'], envs[i]['labels']
model = models[i]
model.eval()
pred_y = model(x)
correct, uncorrect = correct_pred(pred_y,y)
created_envs.append( {'images':x[correct], 'labels':y[correct]})
created_envs.append( {'images':x[uncorrect], 'labels':y[uncorrect]})
## Stage3: DRO training
### init model
mlp = MLP(hidden_dim = flags.hidden_dim, input_dim=input_dim).cuda()
topmlp = get_topmlp_func()
model = torch.nn.Sequential(mlp, topmlp)
optimizer = optim.Adam(model.parameters(), lr=flags.lr)
lossf = mean_nll if flags.lossf == 'nll' else mean_mse
### dro training
for step in range(flags.steps3):
losses = []
for env in created_envs:
x,y = env['images'], env['labels']
losses.append(lossf(model(x),y))
#print(losses)
losses = torch.stack(losses)
loss = losses[torch.argmax(losses)]
weight_norm = 0
for w in model.parameters():
weight_norm += w.norm().pow(2)
loss += flags.l2_regularizer_weight * weight_norm
optimizer.zero_grad()
loss.backward()
optimizer.step()
if step % 5 == 0:
train_loss, train_acc, test_worst_loss, test_worst_acc = \
validation(model, envs, test_envs, lossf)
log = [np.int32(step), train_loss, train_acc,test_worst_loss, test_worst_acc]
logs.append(log)
if flags.verbose:
pretty_print(*log)
if not os.path.exists(flags.save_dir):
os.mkdir(flags.save_dir)
np.save(os.path.join(flags.save_dir, '%s_%s_PI_%d.npy' % (flags.dataset,flags.lossf, flags.steps1)), logs)