-
Notifications
You must be signed in to change notification settings - Fork 1
/
1D_pipeline.py
229 lines (205 loc) · 7.65 KB
/
1D_pipeline.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
import os.path
from utils.custom_run import *
from simulate_data.unit_test_data import *
import seaborn as sns
import matplotlib.pyplot as plt
from GP_baseline.gp_baseline_exact import *
from GP_baseline.gp_baseline_vi import *
from sklearn.model_selection import train_test_split
import numpy as np
import imageio
import shutil
# plt.rcParams['text.usetex'] = True
# plt.rcParams['text.latex.preamble'] = [r'\usepackage{amsmath}']
sns.set()
print(sns.color_palette("tab10")
)
#TODO: upgrade m_Q wtf haha
nn_params = {
'layers_x': [10,10],
'cat_size_list': [],
'transformation': torch.tanh,
'output_dim': 1,
}
VI_params={
'q_kernel':'r_param_scaling',#'nn_kernel',#'r_param_scaling',#'r_param_simple'
'p_kernel':'rbf',
'm_p':0.0,
'reg':1e-2,
'r':50,
'APQ': True,
}
h_space={
'depth_x':[2],
'width_x':[50],
'bs':[1000],
'lr':[1e-2],
'm_P':[1.0],
'sigma':[1e-3],
'transformation':[torch.nn.Tanh()],
'm_factor':[1.],
'parametrize_Z': [False],
'use_all_m': [False],
'm_q_choice': ['mlp'],
'x_s':[200],
#You get negative variance WTF?!
}
#KRR issue is likely related to initialization!
training_params = {
'patience': 1000,
'device': 'cuda:0',
'epochs':1000,
'lr':1e-2,
'model_name':'GWI',
'savedir':'regression_test_3',
'seed':0,
'hyperits':1,
'init_its':250,
'dataset':'None',
'fold': 0,
'bs': 1000,
}
def generate_gif(filenames,dir,gif_name):
with imageio.get_writer(f'{dir}/{gif_name}.gif', mode='I') as writer:
for filename in filenames:
image = imageio.imread(filename)
writer.append_data(image)
def random_split(X,y):
X_tr, X_val, y_tr, y_val=train_test_split(X, y, test_size=0.1, random_state=42)
return X_tr, X_val, y_tr, y_val
def forecast_split(X,y,factor=0.9):
N,d = X.shape
tr_ind = int(round(factor*N))
X_tr=X[:tr_ind]
y_tr=y[:tr_ind]
X_val=X[tr_ind:]
y_val=y[tr_ind:]
return X_tr, X_val, y_tr, y_val
def remove_random_chunks(X,y,chunks_to_remove=5,total_chunks=20):
np.random.seed(99)
x_chunks = torch.chunk(X,chunks=total_chunks,dim=0)
y_chunks = torch.chunk(y,chunks=total_chunks,dim=0)
remove_these=np.random.choice([i for i in range(20)],chunks_to_remove)
x_tr_list = []
y_tr_list = []
x_val_list=[]
y_val_list = []
for i in range(total_chunks):
if i in remove_these:
x_val_list.append(x_chunks[i])
y_val_list.append(y_chunks[i])
else:
x_tr_list.append(x_chunks[i])
y_tr_list.append(y_chunks[i])
X_tr = torch.cat(x_tr_list,dim=0)
y_tr = torch.cat(y_tr_list,dim=0)
X_val = torch.cat(x_val_list,dim=0)
y_val = torch.cat(y_val_list,dim=0)
return X_tr, X_val, y_tr, y_val
# def partial_split
def plot_stuff_2(sigma,d_val,method,index,dir,epoch):
if not os.path.exists(dir):
os.makedirs(dir)
sns.heatmap(sigma)
fname=f'{dir}/{method}_{index}_{epoch}.png'
plt.title(f'Epoch: {epoch}, d: {d_val}')
plt.savefig(fname)
plt.close()
plt.clf()
return fname
def plot_stuff(equation,X_inducing,Y_inducing,X,X_tr,y_tr,X_val,y_val,y_hat,l,u,method,index,dir,epoch):
if not os.path.exists(dir):
os.makedirs(dir)
sns.scatterplot(X_tr.squeeze(),y_tr.squeeze(),alpha=0.35,color=(0.12156862745098039, 0.4666666666666667, 0.7058823529411765))
try:
sns.scatterplot(X_inducing.squeeze(),Y_inducing.squeeze(),color= (0.17254901960784313, 0.6274509803921569, 0.17254901960784313),alpha=1.0)
except Exception as e:
pass
sns.scatterplot(X_val.squeeze(),y_val.squeeze(),color=(0.8392156862745098, 0.15294117647058825, 0.1568627450980392),alpha=0.35)
ax=sns.lineplot(X.squeeze(),y_hat.squeeze())
ax.fill_between(X.squeeze(),l,u, color=(0.12156862745098039, 0.4666666666666667, 0.7058823529411765), alpha=.15)
fname=f'{dir}/{method}_{index}_{epoch}.png'
plt.title(fr'${equation}$',fontsize=15)
plt.xlabel(r"$x$",fontsize=18)
plt.ylabel(r"$y(x)$",fontsize=18)
plt.savefig(fname,bbox_inches = 'tight',
pad_inches = 0.05)
plt.close()
plt.clf()
return fname
def get_u_l(y_hat,y_hat_q):
l = (y_hat - 1.96 * y_hat_q)
u = (y_hat + 1.96 * y_hat_q)
return u,l
def sim_run(index,method):
p_z=False
dir_name = f'gwi_gif_{index}_param_z={p_z}'
dir_name_2 = f'heatmap_gwi_gif_{index}_param_z={p_z}'
if index==1:
X,y=sim_sin_curve()
equation = r'y(x)=\sin(x)+0.1x+\epsilon'
elif index==2:
X,y=sim_sin_curve_2()
equation = r'y(x)=\sin(x)+0.1x^2+\epsilon'
elif index==3:
X,y = sim_sin_curve_3(noise=0.25)
equation = r'y(x)=\sin(3\pi x)+0.3\cos(9\pi x)+0.5\sin(7\pi x)+ \epsilon'
X_tr, X_val, y_tr, y_val=remove_random_chunks(X,y,chunks_to_remove=8,total_chunks=20)
# X_tr, X_val, y_tr, y_val=forecast_split(X,y,factor=0.75)
# method = 'GWI'
print(y_tr.std().item())
VI_params['y_var'] = y_tr.std().item()
dirname = f'regression_test_1_False'
training_params['savedir'] = dirname
if os.path.exists(dirname):
shutil.rmtree(dirname)
if method=='GWI':
e=regression_object(X=X_tr.float(),Y=y_tr.float(),
hyper_param_space=h_space, VI_params=VI_params, train_params=training_params)
# mvp_experiment_object(X=X_tr, Y=y_tr, nn_params=nn_params, VI_params=VI_params, train_params=training_params)
e.run()
x_inducing=e.Z.cpu().numpy()
y_inducing=e.Y_Z.cpu().numpy()
y_hat,b=e.pred_mean_std(X.cuda())
l, u = get_u_l(y_hat.squeeze(), b.squeeze())
elif method=='GP':
e = gp_full_baseline(train_x=X_tr.squeeze(),train_y=y_tr.squeeze(),train_params=training_params)
e.to('cuda:0')
e.train_model()
y_hat,l,u = e.eval_model(X.squeeze().cuda())
y_hat,l,u = y_hat.cpu(),l.cpu(),u.cpu()
x_inducing=X.cpu().squeeze()
y_inducing=y.cpu().squeeze()
elif method=='SVGP':
e = gp_svi_baseline(train_x=X_tr,train_y=y_tr,train_params=training_params,VI_params=VI_params)
e.to('cuda:0')
e.train_model()
x_inducing=e.inducing_points.cpu().numpy()
y_inducing=e.inducing_points_y.cpu().numpy()
y_hat,l,u = e.eval_model(X.cuda())
y_hat,l,u = y_hat.cpu(),l.cpu(),u.cpu()
fname = plot_stuff(equation=equation,X_inducing=x_inducing, Y_inducing=y_inducing, X=X, X_tr=X_tr.cpu(), y_tr=y_tr.cpu(), X_val=X_val.cpu(),
y_val=y_val.cpu(), y_hat=y_hat.cpu(), l=l.cpu(), u=u.cpu(), method=method, index=index, dir=dir_name, epoch=1000)
# filenames=[]
# for i,(a,b) in enumerate(zip(e.preds,e.vars)):
# if i%10==0:
# l,u=get_u_l(a,b)
# fname=plot_stuff(X_inducing=x_inducing,Y_inducing=y_inducing,X=X,X_tr=X_tr,y_tr=y_tr,X_val=X_val,y_val=y_val,y_hat=a,l=l,u=u,method=method,index=index,dir=dir_name,epoch=i)
# filenames.append(fname)
# generate_gif(filenames,dir_name,f'line_plot_{index}')
# filenames=[]
# for i,(a,b) in enumerate(zip(e.d_vals,e.mat_list)):
# if i%10==0:
# fname = plot_stuff_2(b,a,method,index,dir_name_2,i)
# filenames.append(fname)
# generate_gif(filenames,dir_name_2,f'heatmap_{index}')
if __name__ == '__main__':
torch.random.manual_seed(np.random.randint(0,100000))
# for i in [1,2,3]:
# for i in [1]:
# for i in [1,2,3]:
for i in [3]:
sim_run(i,'GWI')
# sim_run(i,'GP')
# sim_run(i,'SVGP')
#FIGURE OUT SCALING ISSUE