-
Notifications
You must be signed in to change notification settings - Fork 2
/
model.py
167 lines (140 loc) · 6.71 KB
/
model.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
import os
import numpy as np
import skimage.transform as transform
import matplotlib
matplotlib.use('agg')
import matplotlib.pyplot as plt
from tqdm import tqdm
from utils import quaternion_multiply, inv_log_quaternion, log_quaternion
from gen_map import GenMap
plt.style.use("seaborn")
class Model:
def __init__(self, batch_size=100, dim_state=6, dim_control=7, dim_observation=2,
image_size=(96, 96, 3), reconstruct_size=(64, 64, 3), reconstruction_accuracy=0.,
training=False, learning_rate=1e-3):
self.dim_state = dim_state
self.dim_control = dim_control
self.dim_observation = dim_observation
self.batch_size = batch_size
self.image_size_origin = tuple(image_size)
self.reconstruct_size_origin = tuple(reconstruct_size)
self.reconstruct_size = np.prod(reconstruct_size)
self.image_size = np.prod(image_size)
self.norm_xyz = self.norm_q = 1.
self.network = GenMap(arch_encoder_attribute=[self.dim_state, 512, self.dim_observation],
image_size=image_size,
dim_features=512,
num_conv_layer=4,
batch_size=batch_size,
rec_log_var=reconstruction_accuracy,
training=training,
reconstruct_size=self.reconstruct_size_origin,
learning_rate=learning_rate)
def set_norm(self, norm_xyz, norm_q):
self.norm_xyz = norm_xyz
self.norm_q = norm_q
def step(self, x, u, dt):
"""Apply quaternion for step"""
xyz = (x[:3, 0] * self.norm_xyz + u[:3, 0] * dt) / self.norm_xyz
qt = quaternion_multiply(u[3:, 0], inv_log_quaternion(x[3:, 0] * self.norm_q))
x1 = np.concatenate((xyz, log_quaternion(qt) / self.norm_q), axis=0).reshape(-1, 1)
return x1
def emit(self, state):
state = state.reshape(-1, self.dim_state)
observation = self.network.infer_by_x(state)
observation = observation.reshape(self.dim_observation, -1)
return observation
def generate(self, state):
state = state.reshape(-1, self.dim_state)
img = self.network.construct_y_from_x(state)
img = img.reshape(self.reconstruct_size_origin)
img = np.clip(img, 0, 1.)
return img
def reconstruct(self, img):
img = img.reshape((-1,) + self.image_size_origin)
rec = self.network.reconstruct(img)
rec = rec.reshape(self.reconstruct_size_origin)
rec = np.clip(rec, 0, 1.)
return rec
def wrap(self, img):
img = img.reshape((-1,) + self.image_size_origin) # for CNN
observation = self.network.infer_by_y(img)
cov_observation = self.network.cov_infer_by_y(img)
observation = observation.reshape(self.dim_observation, -1)
return observation, cov_observation
def train(self, data_loader, model_dir, save_every=5, epoch=100):
def update(loss_list, loss_list1, loss_list2, ep):
"""
Plot the intemediate training loss curve, save it to the model dir, and save the checkpoint.
:param loss_list: overall loss
:param loss_list1: reconstruction loss
:param loss_list2: KL-divergence loss
:param ep: current epoch index
:return: None
"""
f = plt.figure()
plt.plot(loss_list, label='overall')
plt.plot(loss_list1, label='reconstruction')
plt.plot(loss_list2, label='kl_div')
plt.title("training curve")
plt.xlabel("number of batch")
plt.ylabel("-ELBO")
p = int(0.05 * len(loss_list))
_max = np.max([loss_list[p], loss_list1[p], loss_list2[p]])
_min = np.min([loss_list, loss_list1, loss_list2])
_range = _max - _min
print("Error: {}, Reconstruction: {}, KL-div: {}".format(np.mean(loss_list[-100:]),
np.mean(loss_list1[-100:]),
np.mean(loss_list2[-100:])))
plt.ylim([_min - 0.03 * _range, _max + _range])
plt.legend()
path = os.path.join(model_dir, "training_curve.png")
plt.savefig(path)
checkpoint_path = os.path.join(model_dir, "model_{}.ckpt".format(ep))
self.save(checkpoint_path)
plt.close(f)
def get_epoch(dataset):
train_inputs, train_targets = dataset.get_train_set(sample=False)
reconstructs = []
for im in train_inputs:
reconstructs.append(transform.resize(im, self.reconstruct_size_origin,
anti_aliasing=True, mode='constant').reshape(-1)) # * 255.)
reconstructs = np.array(reconstructs)
return train_inputs, train_targets, reconstructs
loss_all, loss_all1, loss_all2 = [], [], []
input_len = len(data_loader.times_train)
inputs, targets, reconstructs = get_epoch(data_loader)
print("Training the observer...")
with tqdm(total=(epoch * input_len)) as tbar:
for e in range(epoch):
loss_list, loss_list1, loss_list2 = [], [], []
num_iters = int(input_len / self.batch_size)
if num_iters * self.batch_size < input_len:
num_iters += 1
for i in range(num_iters):
idx = np.random.choice(input_len, size=self.batch_size)
batch_inputs = inputs[idx, ...]
batch_targets = targets[idx, ...]
batch_reconstructs = reconstructs[idx, ...]
loss, loss_1, loss_2 = self.network.batch_train(batch_targets, batch_inputs, batch_reconstructs)
loss_list.append(loss)
loss_list1.append(loss_1)
loss_list2.append(loss_2)
final_mse = np.mean(loss_list)
tbar.set_postfix(err="%.3f" % final_mse)
tbar.update(input_len)
loss_all.extend(loss_list)
loss_all1.extend(loss_list1)
loss_all2.extend(loss_list2)
if (e+1) % save_every == 0:
update(loss_all, loss_all1, loss_all2, e + 1)
# save the final model
checkpoint_path = os.path.join(model_dir, 'model.ckpt')
self.save(checkpoint_path)
return loss_all
def save(self, path):
self.network.save(path)
def restore(self, path):
self.network.restore(path)
def close(self):
self.network.destruct()