-
Notifications
You must be signed in to change notification settings - Fork 0
/
figure2_hypergradients_fluctuation.py
193 lines (156 loc) · 9.51 KB
/
figure2_hypergradients_fluctuation.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
"""
Here we measure hypergradients for several runs when perturbing
the training data and weight initialization. This must be done on toy
datasets where reverse-mode differentiation is tractable. This corresponds
to figure 2 in the paper.
"""
import torch.optim as optim
import pickle
import os
import warnings
import sys
import shutil
import torch
import torch.nn.functional as F
import torch.optim as optimw
from utils.helpers import *
from utils.datasets import *
from models.selector import *
class HyperGradFluctuation(object):
def __init__(self, args):
self.args = args
self.hypergrads_all = torch.zeros((self.args.n_runs, self.args.T))
self.cross_entropy = nn.CrossEntropyLoss()
self.init_lr_schedule()
## Loaders
self.infinite_train_loader, self.val_loader, _ = get_loaders(datasets_path=self.args.datasets_path,
dataset=self.args.dataset,
train_batch_size=self.args.train_batch_size,
val_batch_size=self.args.n_val_images,
val_source='test',
workers=self.args.workers,
train_infinite=True,
val_infinite=False)
for x,y in self.val_loader: self.X_val, self.Y_val = x.to(device=self.args.device), y.to(device=self.args.device)
## Set up experiment folder
self.experiment_path = os.path.join(self.args.log_directory_path, self.args.experiment_name)
if os.path.isfile(os.path.join(self.experiment_path, 'hypergrads.pth.tar')):
if args.use_gpu: raise FileExistsError(f'Experiment already ran and exists at {self.experiment_path}. \nStopping now')
else:
if os.path.exists(self.experiment_path):
shutil.rmtree(self.experiment_path)
os.makedirs(self.experiment_path)
## Save and Print Args
copy_file(os.path.realpath(__file__), self.experiment_path) # save this python file in logs folder
print('\n---------')
with open(os.path.join(self.experiment_path, 'args.txt'), 'w+') as f:
for k, v in self.args.__dict__.items():
print(k, v)
f.write("{} \t {}\n".format(k, v))
print('---------\n')
def init_lr_schedule(self):
if self.args.inner_lr_cosine_anneal:
dummy_opt = optim.SGD([torch.ones([1], requires_grad=True)], lr=self.args.inner_lr_init)
dummy_scheduler = optim.lr_scheduler.CosineAnnealingLR(dummy_opt, T_max=self.args.T)
lrs = []
for i in range(self.args.T):
lrs.append(dummy_scheduler.get_last_lr()[0])
dummy_opt.step()
dummy_scheduler.step()
self.inner_lrs = torch.tensor(lrs, requires_grad=True, device=self.args.device)
else:
self.inner_lrs = torch.full((self.args.T,), self.args.inner_lr_init, requires_grad=True, device=self.args.device)
def inner_and_outer_loop(self):
for self.inner_step_idx, (x_train, y_train) in enumerate(self.infinite_train_loader):
x_train, y_train = x_train.to(self.args.device, self.args.dtype), y_train.to(self.args.device)
train_logits = self.classifier.forward_with_param(x_train, self.weights)
train_loss = self.cross_entropy(train_logits, y_train)
grads = torch.autograd.grad(train_loss, self.weights, create_graph=True)[0]
if self.args.clamp_inner_grads: grads.clamp_(-self.args.clamp_inner_grads_range, self.args.clamp_inner_grads_range)
self.velocity = self.args.inner_momentum * self.velocity + (grads + self.args.inner_weight_decay * self.weights)
self.weights = self.weights - self.inner_lrs[self.inner_step_idx] * self.velocity
if self.args.greedy:
self.compute_hypergradients() #only populates .grad of one item in self.inner_lrs
self.weights.detach_().requires_grad_()
self.velocity.detach_().requires_grad_()
if self.inner_step_idx+1 == self.args.T: break
if not self.args.greedy: self.compute_hypergradients() #populates .grad of all items in self.inner_lrs
def compute_hypergradients(self):
val_logits = self.classifier.forward_with_param(self.X_val, self.weights)
val_loss = self.cross_entropy(val_logits, self.Y_val)
val_loss.backward()
def run(self):
for self.run_idx in range(self.args.n_runs):
self.classifier = select_model(True, self.args.dataset, self.args.architecture,
self.args.init_type, self.args.init_param,
self.args.device).to(self.args.device)
self.weights = self.classifier.get_param()
self.velocity = torch.zeros(self.weights.numel(), device=self.args.device)
self.inner_and_outer_loop()
self.hypergrads_all[self.run_idx] = self.inner_lrs.grad.detach()
self.inner_lrs.grad.data.zero_()
self.save_final()
def save_final(self):
torch.save({'args': self.args,
'hypergrads_all': self.hypergrads_all},
os.path.join(self.experiment_path, 'hypergrads.pth.tar'))
print(f"Saved hypergrads to {os.path.join(self.experiment_path, 'hypergrads.pth.tar')}")
# ________________________________________________________________________________
# ________________________________________________________________________________
# ________________________________________________________________________________
def make_experiment_name(args):
experiment_name = f'Hg_{args.dataset}_{args.init_type}_T{args.T}_tbs{args.train_batch_size}_mom{args.inner_momentum}_wd{args.inner_weight_decay}_ilr{args.inner_lr_init}'
if args.inner_lr_cosine_anneal: experiment_name += f'cosine'
if args.greedy: experiment_name += f'_GREEDY'
if args.dtype == torch.float64: experiment_name += '_FL64'
experiment_name += f'_S{args.seed}'
return experiment_name
def main(args):
set_torch_seeds(args.seed)
torch.backends.cudnn.enabled = True
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = True
t0 = time.time()
hypervariance_learner = HyperGradFluctuation(args)
hypervariance_learner.run()
total_time = time.time() - t0
with open(os.path.join(args.log_directory_path, args.experiment_name, 'TOTAL_TIME_' + format_time(total_time)), 'w+') as f:
f.write("NA")
if __name__ == "__main__":
import argparse
print('Running...')
parser = argparse.ArgumentParser(description='Welcome to GreedyGrad')
## Main
parser.add_argument('--T', type=int, default=250, help='number of batches for the task and to learn a schedule over')
parser.add_argument('--n_runs', type=int, default=100, help='how many times to compute hypergrads, with different train-val-split each time')
parser.add_argument('--dataset', type=str, default='SVHN')
parser.add_argument('--greedy', type=str2bool, default=False)
parser.add_argument('--architecture', type=str, default='LeNet')
parser.add_argument('--init_type', type=str, default='xavier', choices=['normal', 'xavier', 'kaiming', 'orthogonal', 'zero', 'default'], help='network initialization scheme')
parser.add_argument('--init_param', type=float, default=1, help='network initialization param: gain, std, etc.')
parser.add_argument('--n_val_images', type=int, default=2000, help='ignored unless val_source=train') #20% of 60k=12000
## Inner Loop
parser.add_argument('--inner_lr_init', type=float, default=0.01, help='Used to initialize inner learning rate(s).')
parser.add_argument('--inner_lr_cosine_anneal', type=str2bool, default=True, help='Initial schedule is cosine annealing')
parser.add_argument('--inner_momentum', type=float, default=0.9, help='SGD inner momentum')
parser.add_argument('--inner_weight_decay', type=float, default=0.0, help='SGD + ADAM inner weight decay')
parser.add_argument('--train_batch_size', type=int, default=128)
parser.add_argument('--clamp_inner_grads', type=str2bool, default=True)
parser.add_argument('--clamp_inner_grads_range', type=float, default=1, help='clamp inner grads for each batch to +/- that')
## Misc
parser.add_argument('--datasets_path', type=str, default="~/Datasets/Pytorch/")
parser.add_argument('--log_directory_path', type=str, default="./logs/")
parser.add_argument('--dtype', type=str, default='float32', choices=['float32', 'float64'])
parser.add_argument('--seed', type=int, default=0)
parser.add_argument('--workers', type=int, default=0)
parser.add_argument('--use_gpu', type=str2bool, default=True)
args = parser.parse_args()
args.dataset_path = os.path.join(args.datasets_path, args.dataset)
args.use_gpu = args.use_gpu and torch.cuda.is_available()
args.device = torch.device('cuda') if args.use_gpu else torch.device('cpu')
if args.dtype == 'float64':
torch.set_default_tensor_type(torch.DoubleTensor) # changes weights and tensors but not loaders
args.dtype = torch.float64 if args.dtype == 'float64' else torch.float32
print('\nRunning on device: {}'.format(args.device))
args.experiment_name = make_experiment_name(args)
main(args)