-
Notifications
You must be signed in to change notification settings - Fork 925
/
cnf.py
286 lines (224 loc) · 9.71 KB
/
cnf.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
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
import os
import argparse
import glob
from PIL import Image
import numpy as np
import matplotlib
matplotlib.use('agg')
import matplotlib.pyplot as plt
from sklearn.datasets import make_circles
import torch
import torch.nn as nn
import torch.optim as optim
parser = argparse.ArgumentParser()
parser.add_argument('--adjoint', action='store_true')
parser.add_argument('--viz', action='store_true')
parser.add_argument('--niters', type=int, default=1000)
parser.add_argument('--lr', type=float, default=1e-3)
parser.add_argument('--num_samples', type=int, default=512)
parser.add_argument('--width', type=int, default=64)
parser.add_argument('--hidden_dim', type=int, default=32)
parser.add_argument('--gpu', type=int, default=0)
parser.add_argument('--train_dir', type=str, default=None)
parser.add_argument('--results_dir', type=str, default="./results")
args = parser.parse_args()
if args.adjoint:
from torchdiffeq import odeint_adjoint as odeint
else:
from torchdiffeq import odeint
class CNF(nn.Module):
"""Adapted from the NumPy implementation at:
https://gist.github.com/rtqichen/91924063aa4cc95e7ef30b3a5491cc52
"""
def __init__(self, in_out_dim, hidden_dim, width):
super().__init__()
self.in_out_dim = in_out_dim
self.hidden_dim = hidden_dim
self.width = width
self.hyper_net = HyperNetwork(in_out_dim, hidden_dim, width)
def forward(self, t, states):
z = states[0]
logp_z = states[1]
batchsize = z.shape[0]
with torch.set_grad_enabled(True):
z.requires_grad_(True)
W, B, U = self.hyper_net(t)
Z = torch.unsqueeze(z, 0).repeat(self.width, 1, 1)
h = torch.tanh(torch.matmul(Z, W) + B)
dz_dt = torch.matmul(h, U).mean(0)
dlogp_z_dt = -trace_df_dz(dz_dt, z).view(batchsize, 1)
return (dz_dt, dlogp_z_dt)
def trace_df_dz(f, z):
"""Calculates the trace of the Jacobian df/dz.
Stolen from: https://github.com/rtqichen/ffjord/blob/master/lib/layers/odefunc.py#L13
"""
sum_diag = 0.
for i in range(z.shape[1]):
sum_diag += torch.autograd.grad(f[:, i].sum(), z, create_graph=True)[0].contiguous()[:, i].contiguous()
return sum_diag.contiguous()
class HyperNetwork(nn.Module):
"""Hyper-network allowing f(z(t), t) to change with time.
Adapted from the NumPy implementation at:
https://gist.github.com/rtqichen/91924063aa4cc95e7ef30b3a5491cc52
"""
def __init__(self, in_out_dim, hidden_dim, width):
super().__init__()
blocksize = width * in_out_dim
self.fc1 = nn.Linear(1, hidden_dim)
self.fc2 = nn.Linear(hidden_dim, hidden_dim)
self.fc3 = nn.Linear(hidden_dim, 3 * blocksize + width)
self.in_out_dim = in_out_dim
self.hidden_dim = hidden_dim
self.width = width
self.blocksize = blocksize
def forward(self, t):
# predict params
params = t.reshape(1, 1)
params = torch.tanh(self.fc1(params))
params = torch.tanh(self.fc2(params))
params = self.fc3(params)
# restructure
params = params.reshape(-1)
W = params[:self.blocksize].reshape(self.width, self.in_out_dim, 1)
U = params[self.blocksize:2 * self.blocksize].reshape(self.width, 1, self.in_out_dim)
G = params[2 * self.blocksize:3 * self.blocksize].reshape(self.width, 1, self.in_out_dim)
U = U * torch.sigmoid(G)
B = params[3 * self.blocksize:].reshape(self.width, 1, 1)
return [W, B, U]
class RunningAverageMeter(object):
"""Computes and stores the average and current value"""
def __init__(self, momentum=0.99):
self.momentum = momentum
self.reset()
def reset(self):
self.val = None
self.avg = 0
def update(self, val):
if self.val is None:
self.avg = val
else:
self.avg = self.avg * self.momentum + val * (1 - self.momentum)
self.val = val
def get_batch(num_samples):
points, _ = make_circles(n_samples=num_samples, noise=0.06, factor=0.5)
x = torch.tensor(points).type(torch.float32).to(device)
logp_diff_t1 = torch.zeros(num_samples, 1).type(torch.float32).to(device)
return(x, logp_diff_t1)
if __name__ == '__main__':
t0 = 0
t1 = 10
device = torch.device('cuda:' + str(args.gpu)
if torch.cuda.is_available() else 'cpu')
# model
func = CNF(in_out_dim=2, hidden_dim=args.hidden_dim, width=args.width).to(device)
optimizer = optim.Adam(func.parameters(), lr=args.lr)
p_z0 = torch.distributions.MultivariateNormal(
loc=torch.tensor([0.0, 0.0]).to(device),
covariance_matrix=torch.tensor([[0.1, 0.0], [0.0, 0.1]]).to(device)
)
loss_meter = RunningAverageMeter()
if args.train_dir is not None:
if not os.path.exists(args.train_dir):
os.makedirs(args.train_dir)
ckpt_path = os.path.join(args.train_dir, 'ckpt.pth')
if os.path.exists(ckpt_path):
checkpoint = torch.load(ckpt_path)
func.load_state_dict(checkpoint['func_state_dict'])
optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
print('Loaded ckpt from {}'.format(ckpt_path))
try:
for itr in range(1, args.niters + 1):
optimizer.zero_grad()
x, logp_diff_t1 = get_batch(args.num_samples)
z_t, logp_diff_t = odeint(
func,
(x, logp_diff_t1),
torch.tensor([t1, t0]).type(torch.float32).to(device),
atol=1e-5,
rtol=1e-5,
method='dopri5',
)
z_t0, logp_diff_t0 = z_t[-1], logp_diff_t[-1]
logp_x = p_z0.log_prob(z_t0).to(device) - logp_diff_t0.view(-1)
loss = -logp_x.mean(0)
loss.backward()
optimizer.step()
loss_meter.update(loss.item())
print('Iter: {}, running avg loss: {:.4f}'.format(itr, loss_meter.avg))
except KeyboardInterrupt:
if args.train_dir is not None:
ckpt_path = os.path.join(args.train_dir, 'ckpt.pth')
torch.save({
'func_state_dict': func.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
}, ckpt_path)
print('Stored ckpt at {}'.format(ckpt_path))
print('Training complete after {} iters.'.format(itr))
if args.viz:
viz_samples = 30000
viz_timesteps = 41
target_sample, _ = get_batch(viz_samples)
if not os.path.exists(args.results_dir):
os.makedirs(args.results_dir)
with torch.no_grad():
# Generate evolution of samples
z_t0 = p_z0.sample([viz_samples]).to(device)
logp_diff_t0 = torch.zeros(viz_samples, 1).type(torch.float32).to(device)
z_t_samples, _ = odeint(
func,
(z_t0, logp_diff_t0),
torch.tensor(np.linspace(t0, t1, viz_timesteps)).to(device),
atol=1e-5,
rtol=1e-5,
method='dopri5',
)
# Generate evolution of density
x = np.linspace(-1.5, 1.5, 100)
y = np.linspace(-1.5, 1.5, 100)
points = np.vstack(np.meshgrid(x, y)).reshape([2, -1]).T
z_t1 = torch.tensor(points).type(torch.float32).to(device)
logp_diff_t1 = torch.zeros(z_t1.shape[0], 1).type(torch.float32).to(device)
z_t_density, logp_diff_t = odeint(
func,
(z_t1, logp_diff_t1),
torch.tensor(np.linspace(t1, t0, viz_timesteps)).to(device),
atol=1e-5,
rtol=1e-5,
method='dopri5',
)
# Create plots for each timestep
for (t, z_sample, z_density, logp_diff) in zip(
np.linspace(t0, t1, viz_timesteps),
z_t_samples, z_t_density, logp_diff_t
):
fig = plt.figure(figsize=(12, 4), dpi=200)
plt.tight_layout()
plt.axis('off')
plt.margins(0, 0)
fig.suptitle(f'{t:.2f}s')
ax1 = fig.add_subplot(1, 3, 1)
ax1.set_title('Target')
ax1.get_xaxis().set_ticks([])
ax1.get_yaxis().set_ticks([])
ax2 = fig.add_subplot(1, 3, 2)
ax2.set_title('Samples')
ax2.get_xaxis().set_ticks([])
ax2.get_yaxis().set_ticks([])
ax3 = fig.add_subplot(1, 3, 3)
ax3.set_title('Log Probability')
ax3.get_xaxis().set_ticks([])
ax3.get_yaxis().set_ticks([])
ax1.hist2d(*target_sample.detach().cpu().numpy().T, bins=300, density=True,
range=[[-1.5, 1.5], [-1.5, 1.5]])
ax2.hist2d(*z_sample.detach().cpu().numpy().T, bins=300, density=True,
range=[[-1.5, 1.5], [-1.5, 1.5]])
logp = p_z0.log_prob(z_density) - logp_diff.view(-1)
ax3.tricontourf(*z_t1.detach().cpu().numpy().T,
np.exp(logp.detach().cpu().numpy()), 200)
plt.savefig(os.path.join(args.results_dir, f"cnf-viz-{int(t*1000):05d}.jpg"),
pad_inches=0.2, bbox_inches='tight')
plt.close()
img, *imgs = [Image.open(f) for f in sorted(glob.glob(os.path.join(args.results_dir, f"cnf-viz-*.jpg")))]
img.save(fp=os.path.join(args.results_dir, "cnf-viz.gif"), format='GIF', append_images=imgs,
save_all=True, duration=250, loop=0)
print('Saved visualization animation at {}'.format(os.path.join(args.results_dir, "cnf-viz.gif")))