-
Notifications
You must be signed in to change notification settings - Fork 1
/
net.py
328 lines (284 loc) · 12 KB
/
net.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
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
import os
import torch
import torch.nn as nn
import numpy as np
from numpy import pi, ndarray
import matplotlib.pyplot as plt
from matplotlib import figure
from matplotlib.collections import LineCollection
from torch.autograd.functional import jacobian
from matplotlib.patches import Ellipse
from itertools import zip_longest
from sklearn.datasets import make_circles, make_moons
class RiemannianMetric:
def __init__(self, matrix=None, dim=2):
self.dims = dim
if matrix is None:
self.matrix = np.eye(self.dims)
else:
self.matrix = matrix
def transform_tensor_entry(self, jacobi: ndarray, index: tuple) -> int:
su = 0
i1, i2 = index
for all in range(self.dims):
for bll in range(self.dims):
su += jacobi[all][i1] * jacobi[bll][i2] * self.matrix[all][bll]
return su
def transform_coordinates(self, jacobi: ndarray) -> object:
su = 0 * self.matrix
for al in range(self.dims):
for bl in range(self.dims):
indices = (al, bl)
su[al][bl] = self.transform_tensor_entry(jacobi, indices)
self.matrix = su
return self
class CustomLoss(nn.BCEWithLogitsLoss):
def forward(self, input: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
return torch.functional.F.binary_cross_entropy_with_logits(input, target, self.weight,
pos_weight=self.pos_weight, reduction=self.reduction)
class Layer(nn.Module):
def __init__(self, act_func, identity_conn=False):
super(Layer, self).__init__()
self.identity_connection = identity_conn
if act_func is None:
self.act_func = lambda x: x
else:
self.act_func = act_func
self.linear_map = nn.Linear(2, 2)
self.linear_map2 = nn.Linear(2, 2)
def forward(self, x):
if self.identity_connection:
return x + self.act_func(self.linear_map(x))
else:
return self.act_func(self.linear_map(x))
def print_ln(ll, obj='grid'):
print(f'Plotting {obj} for Layer {ll}')
class Net(nn.Module):
def __init__(self, layers: int, identity_conn: bool, c2: bool, init_gain: float, activation):
super(Net, self).__init__()
self.num_layers = layers
self.init_gain = init_gain
self.activations = []
self.jacobians = []
self.linear_out = nn.Linear(2, 1)
self.layers = nn.ModuleList([Layer(activation, identity_conn) for _ in range(layers)])
if c2:
self.forward_func = self.forward_c2
else:
self.forward_func = self.forward_c1
@staticmethod
def get_cmap(n, name='hsv'):
"""
@param n: number of required colors
@param name: a standard mpl colormap name
@return: list of n distinct RGB colors
"""
return plt.cm.get_cmap(name, n)
@staticmethod
def plot_grid(x, y, ax=None, **kwargs):
ax = ax or plt.gca()
segs1 = np.stack((x, y), axis=2)
segs2 = segs1.transpose(1, 0, 2)
ax.add_collection(LineCollection(segs1, **kwargs, zorder=1))
ax.add_collection(LineCollection(segs2, **kwargs, zorder=1))
ax.autoscale()
@staticmethod
def plot_liner_classifier(plot: figure, w: ndarray, b: ndarray, xmin: float, xmax: float) -> None:
"""
@param xmin:
@param xmax:
@param plot: matplotlib figure object
@param w: weights of linear classifier
@param b: bias of linear classifier
"""
def f(x):
return (-w[0, 0] * x - b[0]) / w[0, 1]
plot.plot([xmin, xmax], [f(xmin), f(xmax)], 'k')
def init_weights(self):
def init(m):
if type(m) == nn.Linear:
torch.nn.init.xavier_normal_(m.weight, gain=self.init_gain)
m.bias.data.fill_(0)
self.apply(init)
def init_forward(self):
self.activations = []
self.jacobians = []
def save_forward(self, coordinates, coord_change_func):
new_coords = coord_change_func(coordinates)
self.jacobians.append(jacobian(coord_change_func, coordinates))
self.activations.append(new_coords)
def forward_c1(self, x, save_activations=False):
self.init_forward()
for i, layer in enumerate(self.layers):
a = layer(x)
if save_activations:
self.save_forward(x, layer)
x = a
return self.linear_out(x)
def forward_c2(self, x, save_activations=False):
self.init_forward()
prev_layer = x
for i, layer in enumerate(self.layers):
a = layer(x)
if i != 0:
# y = 2*x(l) + f(x(l);l) - x(l-1)
a += x - prev_layer
if save_activations:
# not sure about this yet
self.save_forward(x, layer)
prev_layer = x
x = a
return self.linear_out(x)
def forward(self, x: torch.Tensor, save_activations: bool = False) -> torch.Tensor:
return self.forward_func(x, save_activations)
def plot_points(self, plots: figure, cls_a: ndarray, cls_b: ndarray, plot_classifier: bool = False) -> None:
# forward pass data points
iter_plots = iter(plots)
a = torch.from_numpy(cls_a).float()
b = torch.from_numpy(cls_b).float()
self.forward(a, save_activations=True)
act_cls1 = self.activations
self.forward(b, save_activations=True)
act_cls2 = self.activations
plot = next(iter_plots)
plot.scatter(cls_a[:, 0], cls_a[:, 1])
plot.scatter(cls_b[:, 0], cls_b[:, 1])
for ln, layer_activations in enumerate(zip(act_cls1, act_cls2), start=1):
print_ln(ln, 'points')
cls1, cls2 = map(lambda t: t.detach().numpy(), layer_activations)
plot = next(iter_plots)
plot.scatter(cls1[:, 0], cls1[:, 1])
plot.scatter(cls2[:, 0], cls2[:, 1])
if ln == self.num_layers:
cls12 = np.append(cls1, cls2)
xmin = np.min(cls12)
xmax = np.max(cls12)
params = self.parameters()
w = next(params).detach().numpy()
b = next(params).detach().numpy()
if plot_classifier:
Net.plot_liner_classifier(plot, w, b, xmin, xmax)
def plot_grids(self, plots: figure, xmin, xmax, grid_dim_x: int, grid_dim_y: int, plot_tensors: bool) -> None:
# calculate grid points
iter_plots = iter(plots)
grid_size = grid_dim_x * grid_dim_y
grid_x, grid_y = np.meshgrid(np.linspace(xmin, xmax, grid_dim_x), np.linspace(xmin, xmax, grid_dim_y))
grid_numpy_array = np.array([grid_x.reshape(grid_size), grid_y.reshape(grid_size)]).T
grid_tensor = torch.from_numpy(grid_numpy_array).float()
# forward pass grid points
self.forward(grid_tensor, save_activations=True)
plot = next(iter_plots)
print_ln(0)
Net.plot_grid(grid_x, grid_y, ax=plot, color="lightgrey")
for e, grid in enumerate(self.activations, start=1):
plot = next(iter_plots)
print_ln(e)
grid = grid.detach().numpy()
xx = grid.T.reshape(2, grid_dim_x, grid_dim_y)[0]
yy = grid.T.reshape(2, grid_dim_x, grid_dim_y)[1]
Net.plot_grid(xx, yy, ax=plot, color="lightgrey")
if plot_tensors:
self.plot_tensors(plots, grid_numpy_array, grid_tensor)
def plot_tensors(self, plots: figure, grid_numpy_array: ndarray, grid_tensor: torch.Tensor) -> None:
# every point gets it's own color for the metric tensor plot
cmap = Net.get_cmap(len(grid_numpy_array))
print("Plotting tensor glyphs...")
for e, grid_point in enumerate(grid_numpy_array):
point = grid_tensor[e]
self.forward(point, save_activations=True)
g = RiemannianMetric()
g_numpy = g.matrix
iter_jacobi = reversed(self.jacobians)
for en, layer in enumerate(zip_longest(reversed(self.activations), reversed(plots),
fillvalue=torch.from_numpy(grid_point))):
point, plot = layer
x, y = point.detach().numpy()
if en != 0:
jacobi = next(iter_jacobi)
g_numpy = g.transform_coordinates(jacobi.detach().numpy()).matrix
eig_vals, eig_vecs = np.linalg.eig(g_numpy)
eig_vals = np.sqrt(eig_vals)
indices = np.argsort(eig_vals)
angle = np.arccos(eig_vecs[indices[1]][0] / np.linalg.norm(eig_vecs[indices[1]]))
width, height = eig_vals[indices[1]], eig_vals[indices[0]]
plot.add_artist(Ellipse((x, y), width, height, angle * 360 / (2 * pi),
zorder=3, facecolor=cmap(e), edgecolor='k', lw=0.5))
def plot_geometry(self, a_numpy, b_numpy, grid_dim_x=15, grid_dim_y=15, plot_rows=1, plot_cols=11,
plot_grids=False, plot_points=True, plot_tensors=False, plot_classifier=False, save_img='test'):
"""
@param a_numpy:
@param b_numpy:
@param grid_dim_x:
@param grid_dim_y:
@param plot_rows:
@param plot_cols:
@param plot_grids:
@param plot_points:
@param plot_tensors:
@param plot_classifier:
@return:
"""
# prepare plots
fig, plots = plt.subplots(plot_rows, plot_cols, figsize=[5 * plot_cols, 5 * plot_rows])
if plot_rows > 1:
plots = plots.flatten()
xmin = min(np.min(a_numpy), np.min(b_numpy))
xmax = max(np.max(a_numpy), np.max(b_numpy))
if plot_grids:
self.plot_grids(plots, xmin, xmax, grid_dim_x, grid_dim_y, plot_tensors)
if plot_points:
self.plot_points(plots, a_numpy, b_numpy, plot_classifier)
for e, plot in enumerate(plots):
plot.set_title(f'Layer {e}')
if not os.path.exists('img_out'):
os.mkdir('img_out')
i = 0
while os.path.exists(f'img_out/{save_img}-{i}.png'):
i += 1
plt.savefig(f'img_out/{save_img}-{i}')
plt.show()
def spiral2d(data_points: int, n_noise: float, spiral_length_rad: float) -> tuple:
"""
@note idea from https://gist.github.com/45deg
@param data_points:
@param n_noise:
@param spiral_length_rad:
@return:
"""
dims = 2
points_per_cls = int(data_points/2)
theta = np.sqrt(np.random.rand(points_per_cls))*spiral_length_rad
r_a = 2*theta + pi
data_a = np.array([np.cos(theta)*r_a, np.sin(theta)*r_a]).T
x_a = data_a + n_noise * np.random.randn(points_per_cls, dims)
r_b = -2*theta - pi
data_b = np.array([np.cos(theta)*r_b, np.sin(theta)*r_b]).T
x_b = data_b + n_noise * np.random.randn(points_per_cls, dims)
res_a = np.append(x_a, np.zeros((points_per_cls, 1)), axis=1)
res_b = np.append(x_b, np.ones((points_per_cls, 1)), axis=1)
res = np.append(res_a, res_b, axis=0)
np.random.shuffle(res)
return res, x_a, x_b
def accuracy(output: torch.Tensor, target: torch.Tensor) -> float:
correct = 0
incorrect = 0
total = 0
n = output.shape[0]
output = output.view(n).detach().numpy()
target = target.view(n).detach().numpy()
for z in zip(output, target):
o, t = z
correct_positive = o > 0 and t == 1
correct_negative = o < 0 and t == 0
if correct_positive or correct_negative:
correct += 1
else:
incorrect += 1
total += 1
return correct / total
def data(n_samples: int, n_noise: float, n_shuffle: bool, data_type='circles'):
if data_type == 'moons':
return make_moons(n_samples=n_samples, noise=n_noise, shuffle=n_shuffle)
if data_type != 'circles':
print("Warning: Only 'spiral', 'moons' or 'circles' are admissible data manifolds. Default will be used.")
return make_circles(n_samples=n_samples, noise=n_noise, shuffle=n_shuffle)