forked from tomgoldstein/loss-landscape
-
Notifications
You must be signed in to change notification settings - Fork 0
/
projection.py
253 lines (203 loc) · 7.79 KB
/
projection.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
"""
Project a model or multiple models to a plane spaned by given directions.
"""
import numpy as np
import torch
import os
import copy
import h5py
import net_plotter
import model_loader
import h5_util
from sklearn.decomposition import PCA
def tensorlist_to_tensor(weights):
""" Concatnate a list of tensors into one tensor.
Args:
weights: a list of parameter tensors, e.g. net_plotter.get_weights(net).
Returns:
concatnated 1D tensor
"""
return torch.cat([w.view(w.numel()) if w.dim() > 1 else torch.FloatTensor(w) for w in weights])
def nplist_to_tensor(nplist):
""" Concatenate a list of numpy vectors into one tensor.
Args:
nplist: a list of numpy vectors, e.g., direction loaded from h5 file.
Returns:
concatnated 1D tensor
"""
v = []
for d in nplist:
w = torch.tensor(d*np.float64(1.0))
# Ignoreing the scalar values (w.dim() = 0).
if w.dim() > 1:
v.append(w.view(w.numel()))
elif w.dim() == 1:
v.append(w)
return torch.cat(v)
def npvec_to_tensorlist(direction, params):
""" Convert a numpy vector to a list of tensors with the same shape as "params".
Args:
direction: a list of numpy vectors, e.g., a direction loaded from h5 file.
base: a list of parameter tensors from net
Returns:
a list of tensors with the same shape as base
"""
if isinstance(params, list):
w2 = copy.deepcopy(params)
idx = 0
for w in w2:
w.copy_(torch.tensor(direction[idx:idx + w.numel()]).view(w.size()))
idx += w.numel()
assert(idx == len(direction))
return w2
else:
s2 = []
idx = 0
for (k, w) in params.items():
s2.append(torch.Tensor(direction[idx:idx + w.numel()]).view(w.size()))
idx += w.numel()
assert(idx == len(direction))
return s2
def cal_angle(vec1, vec2):
""" Calculate cosine similarities between two torch tensors or two ndarraies
Args:
vec1, vec2: two tensors or numpy ndarraies
"""
if isinstance(vec1, torch.Tensor) and isinstance(vec1, torch.Tensor):
return torch.dot(vec1, vec2)/(vec1.norm()*vec2.norm()).item()
elif isinstance(vec1, np.ndarray) and isinstance(vec2, np.ndarray):
return np.ndarray.dot(vec1, vec2)/(np.linalg.norm(vec1)*np.linalg.norm(vec2))
def project_1D(w, d):
""" Project vector w to vector d and get the length of the projection.
Args:
w: vectorized weights
d: vectorized direction
Returns:
the projection scalar
"""
assert len(w) == len(d), 'dimension does not match for w and '
scale = torch.dot(w, d)/d.norm()
return scale.item()
def project_2D(d, dx, dy, proj_method):
""" Project vector d to the plane spanned by dx and dy.
Args:
d: vectorized weights
dx: vectorized direction
dy: vectorized direction
proj_method: projection method
Returns:
x, y: the projection coordinates
"""
if proj_method == 'cos':
# when dx and dy are orthorgonal
x = project_1D(d, dx)
y = project_1D(d, dy)
elif proj_method == 'lstsq':
# solve the least squre problem: Ax = d
A = np.vstack([dx.numpy(), dy.numpy()]).T
[x, y] = np.linalg.lstsq(A, d.numpy())[0]
return x, y
def project_trajectory(dir_file, w, s, dataset, model_name, model_files,
dir_type='weights', proj_method='cos'):
"""
Project the optimization trajectory onto the given two directions.
Args:
dir_file: the h5 file that contains the directions
w: weights of the final model
s: states of the final model
model_name: the name of the model
model_files: the checkpoint files
dir_type: the type of the direction, weights or states
proj_method: cosine projection
Returns:
proj_file: the projection filename
"""
proj_file = dir_file + '_proj_' + proj_method + '.h5'
if os.path.exists(proj_file):
print('The projection file exists! No projection is performed unless %s is deleted' % proj_file)
return proj_file
# read directions and convert them to vectors
directions = net_plotter.load_directions(dir_file)
dx = nplist_to_tensor(directions[0])
dy = nplist_to_tensor(directions[1])
xcoord, ycoord = [], []
for model_file in model_files:
net2 = model_loader.load(dataset, model_name, model_file)
if dir_type == 'weights':
w2 = net_plotter.get_weights(net2)
d = net_plotter.get_diff_weights(w, w2)
elif dir_type == 'states':
s2 = net2.state_dict()
d = net_plotter.get_diff_states(s, s2)
d = tensorlist_to_tensor(d)
x, y = project_2D(d, dx, dy, proj_method)
print ("%s (%.4f, %.4f)" % (model_file, x, y))
xcoord.append(x)
ycoord.append(y)
f = h5py.File(proj_file, 'w')
f['proj_xcoord'] = np.array(xcoord)
f['proj_ycoord'] = np.array(ycoord)
f.close()
return proj_file
def setup_PCA_directions(args, model_files, w, s):
"""
Find PCA directions for the optimization path from the initial model
to the final trained model.
Returns:
dir_name: the h5 file that stores the directions.
"""
# Name the .h5 file that stores the PCA directions.
folder_name = args.model_folder + '/PCA_' + args.dir_type
if args.ignore:
folder_name += '_ignore=' + args.ignore
folder_name += '_save_epoch=' + str(args.save_epoch)
os.system('mkdir ' + folder_name)
dir_name = folder_name + '/directions.h5'
# skip if the direction file exists
if os.path.exists(dir_name):
f = h5py.File(dir_name, 'a')
if 'explained_variance_' in f.keys():
f.close()
return dir_name
# load models and prepare the optimization path matrix
matrix = []
for model_file in model_files:
print (model_file)
net2 = model_loader.load(args.dataset, args.model, model_file)
if args.dir_type == 'weights':
w2 = net_plotter.get_weights(net2)
d = net_plotter.get_diff_weights(w, w2)
elif args.dir_type == 'states':
s2 = net2.state_dict()
d = net_plotter.get_diff_states(s, s2)
if args.ignore == 'biasbn':
net_plotter.ignore_biasbn(d)
d = tensorlist_to_tensor(d)
matrix.append(d.numpy())
# Perform PCA on the optimization path matrix
print ("Perform PCA on the models")
pca = PCA(n_components=2)
pca.fit(np.array(matrix))
pc1 = np.array(pca.components_[0])
pc2 = np.array(pca.components_[1])
print("angle between pc1 and pc2: %f" % cal_angle(pc1, pc2))
print("pca.explained_variance_ratio_: %s" % str(pca.explained_variance_ratio_))
# convert vectorized directions to the same shape as models to save in h5 file.
if args.dir_type == 'weights':
xdirection = npvec_to_tensorlist(pc1, w)
ydirection = npvec_to_tensorlist(pc2, w)
elif args.dir_type == 'states':
xdirection = npvec_to_tensorlist(pc1, s)
ydirection = npvec_to_tensorlist(pc2, s)
if args.ignore == 'biasbn':
net_plotter.ignore_biasbn(xdirection)
net_plotter.ignore_biasbn(ydirection)
f = h5py.File(dir_name, 'w')
h5_util.write_list(f, 'xdirection', xdirection)
h5_util.write_list(f, 'ydirection', ydirection)
f['explained_variance_ratio_'] = pca.explained_variance_ratio_
f['singular_values_'] = pca.singular_values_
f['explained_variance_'] = pca.explained_variance_
f.close()
print ('PCA directions saved in: %s' % dir_name)
return dir_name