-
Notifications
You must be signed in to change notification settings - Fork 0
/
mynet.py
280 lines (238 loc) · 10.5 KB
/
mynet.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
import numpy as np
def compute_accuracy(y_hat, y):
y_hat_classes = np.argmax(y_hat, axis=1)
y_classes = np.argmax(y, axis=1)
return np.sum(np.where(y_hat_classes == y_classes, 1, 0)) / y.shape[0]
def get_batched_indices(N, batch_size=100):
N_batches = N // batch_size
indices = np.arange(N)
np.random.shuffle(indices)
batches = indices[: N_batches * batch_size]
batches = batches.reshape(N_batches, batch_size)
batches = list(batches)
if N != N_batches * batch_size:
batches.append(list(indices[N_batches * batch_size:]))
return batches
class ReLu:
def __init__(self):
self.x = None
self.g = None
def forward(self, x, mode=None):
self.x = x
return np.maximum(np.zeros(self.x.shape), self.x)
def backward(self, g):
self.g = g * (self.x > 0).astype(np.float32)
return self.g
def print_params(self):
pass
def Softmax_Loss(x, y):
exp_x = np.exp(x)
y_hat = exp_x / np.sum(exp_x, axis=0)
loss = -np.log(y_hat[y == 1]).sum() / x.shape[1]
grad = (y_hat - y) / x.shape[1]
return y_hat, loss, grad
def Softmax(x):
exp_x = np.exp(x)
return exp_x / np.sum(exp_x, axis=0)
class BatchNorm:
def __init__(self, n):
self.x = None
self.x_c = None
self.disp = None
self.mean = None
self.gamma = np.ones(n).reshape(-1, 1)
self.beta = np.zeros(n).reshape(-1, 1)
self.g = None
self.g_out = None
self.avg_mean = 0
self.avg_disp = 0
self.count = 0.0
def forward(self, x, mode=None):
if mode == 'train':
self.x = x
N = x.shape[1]
self.mean = np.sum(x, axis=1).reshape(-1, 1) / N
self.x_c = x - self.mean
self.disp = np.sum(np.power(self.x_c, 2), axis=1).reshape(-1, 1) / N + 0.00000001
self.count += 1
self.avg_mean = (self.count - 1) / self.count * self.avg_mean + self.mean / self.count
self.avg_disp = (self.count - 1) / self.count * self.avg_disp + self.disp / self.count
return self.x_c * np.power(self.disp, -1. / 2) * self.gamma + self.beta
if mode == 'test':
self.x = x
return (self.x - self.avg_mean) * np.power(self.avg_disp, -1. / 2) * self.gamma + self.beta
def backward(self, grad):
N = grad.shape[1]
self.g = grad
self.g_out = np.power(self.disp, -1. / 2) * self.gamma * (N * grad -
np.sum(grad, axis=1).reshape(-1, 1) -
np.power(self.disp, -1) *
self.x_c * np.sum(grad * self.x_c, axis=1).reshape(-1,
1)) / N
return self.g_out
def print_params(self):
print('BatchNorm layer \ngamma: \n', self.gamma, '\nbeta: \n', self.beta)
class Linear:
def __init__(self, m, n, dropout=0):
self.W = np.random.randn(m, n)
self.b = np.random.uniform(low=0, high=2, size=m)
self.x = None
self.g = None
self.dropout = dropout
def forward(self, x, mode=None):
self.x = x
return np.matmul(self.W, self.x) + self.b.reshape(self.b.shape[0], 1)
def backward(self, g):
self.g = g
return np.matmul(self.W.T, self.g)
def print_params(self):
print('Linear layer \nW: \n', self.W, '\nb:\n', self.b)
class MyNet:
def __init__(self, layers, reg='l1', l=0.01):
self.layers = layers
self.linear_layers = [el for el in self.layers if type(el) == Linear]
self.batch_norm_layers = [el for el in self.layers if type(el) == BatchNorm]
self.pred = None
self.loss = None
self.grad = None
self.lr = 0.0003
self.reg = reg
self.l = l
self.m_w = []
self.m_b = []
self.m_gamma = []
self.m_beta = []
self.v_w = []
self.v_b = []
self.v_gamma = []
self.v_beta = []
self.t = 0
self.epochs = 0
self.loss_history = []
for layer in self.linear_layers:
self.m_w.append(np.zeros_like(layer.W))
self.v_w.append(np.zeros_like(layer.W))
self.m_b.append(np.zeros_like(layer.b))
self.v_b.append(np.zeros_like(layer.b))
for layer in self.batch_norm_layers:
self.m_gamma.append(np.zeros_like(layer.gamma))
self.v_gamma.append(np.zeros_like(layer.gamma))
self.m_beta.append(np.zeros_like(layer.beta))
self.v_beta.append(np.zeros_like(layer.beta))
def forward(self, x, y, mode=None):
if x.ndim == 1:
x = x.reshape(1, -1)
if y.ndim == 1:
y = y.reshape(1, -1)
h = x.T
for layer in self.layers:
h = layer.forward(h, mode)
self.pred, self.loss, self.grad = Softmax_Loss(h, y.T)
return self.pred, self.loss
def predict(self, x, mode=None):
if x.ndim == 1:
x = x.reshape(1, -1)
h = x.T
for layer in self.layers:
h = layer.forward(h, mode)
self.pred = Softmax(h)
return self.pred
def backward(self):
cur_grad = self.grad
for layer in reversed(self.layers):
cur_grad = layer.backward(cur_grad)
def print_net(self):
for layer in self.layers:
print(type(layer), layer.x, layer.g)
def make_adam_step(self, beta1, beta2, epsilon):
for i in range(len(self.linear_layers)):
layer = self.linear_layers[i]
if self.reg == 'none':
grad = layer.g @ layer.x.T
elif self.reg == 'l1':
grad = layer.g @ layer.x.T + self.l * np.where(layer.W > 0, 1, -1)
elif self.reg == 'l2':
grad = layer.g @ layer.x.T + self.l * 2 * layer.W
self.m_w[i] = beta1 * self.m_w[i] + (1 - beta1) * grad
self.v_w[i] = beta2 * self.v_w[i] + (1 - beta2) * np.power(grad, 2)
m_hat = self.m_w[i] / (1 - np.power(beta1, self.t))
v_hat = self.v_w[i] / (1 - np.power(beta2, self.t))
layer.W -= self.lr * m_hat / (np.sqrt(v_hat) + epsilon)
grad = layer.g.sum(axis=1) / layer.x.shape[1]
self.m_b[i] = beta1 * self.m_b[i] + (1 - beta1) * grad
self.v_b[i] = beta2 * self.v_b[i] + (1 - beta2) * np.power(grad, 2)
m_hat = self.m_b[i] / (1 - np.power(beta1, self.t))
v_hat = self.v_b[i] / (1 - np.power(beta2, self.t))
layer.b -= self.lr * m_hat / (np.sqrt(v_hat) + epsilon)
for i in range(len(self.batch_norm_layers)):
layer = self.batch_norm_layers[i]
grad = np.sum(layer.g * layer.x_c * np.power(layer.disp, -1. / 2), axis=1).reshape(-1, 1)
self.m_gamma[i] = beta1 * self.m_gamma[i] + (1 - beta1) * grad
self.v_gamma[i] = beta2 * self.v_gamma[i] + (1 - beta2) * np.power(grad, 2)
m_hat = self.m_gamma[i] / (1 - np.power(beta1, self.t))
v_hat = self.v_gamma[i] / (1 - np.power(beta2, self.t))
layer.gamma -= self.lr * m_hat / (np.sqrt(v_hat) + epsilon)
grad = np.sum(layer.g, axis=1).reshape(-1, 1)
self.m_beta[i] = beta1 * self.m_beta[i] + (1 - beta1) * grad
self.v_beta[i] = beta2 * self.v_beta[i] + (1 - beta2) * np.power(grad, 2)
self.m_hat = self.m_beta[i] / (1 - np.power(beta1, self.t))
self.v_hat = self.v_beta[i] / (1 - np.power(beta2, self.t))
layer.beta -= self.lr * m_hat / (np.sqrt(v_hat) + epsilon)
def fit(self, x, y, epochs=5, bs=100):
beta1 = 0.9
beta2 = 0.999
epsilon = 0.0000000001
for epoch in range(epochs):
batches = get_batched_indices(x.shape[0], batch_size=bs)
loss = 0
for batch in batches:
self.t += 1
loss += self.forward(x[batch], y[batch])[1]
self.backward()
self.make_adam_step(beta1, beta2, epsilon)
self.epochs += 1
self.loss_history.append(loss)
if self.epochs % 20 == 0 and floor(self.epochs / 20) < 10:
self.lr = self.lr * 0.5 ** floor(self.epochs / 20)
print('epoch = ', epoch, 'loss = ', loss / len(batches))
return self
def check_w1_grad(self, x, y):
grad = np.zeros_like(self.linear_layers[0].W)
for i in range(self.linear_layers[0].W.shape[0]):
for j in range(self.linear_layers[0].W.shape[1]):
_, f0 = net.forward(x, y)
self.linear_layers[0].W[i][j] += 0.00000001
_, f1 = net.forward(x, y)
grad[i][j] = (f1 - f0) / 0.00000001
self.linear_layers[0].W[i][j] -= 0.00000001
return grad
def check_w2_grad(self, x, y):
grad = np.zeros_like(self.linear_layers[1].W)
for i in range(self.linear_layers[1].W.shape[0]):
for j in range(self.linear_layers[1].W.shape[1]):
_, f0 = net.forward(x, y)
self.linear_layers[1].W[i][j] += 0.00000001
_, f1 = net.forward(x, y)
grad[i][j] = (f1 - f0) / 0.00000001
self.linear_layers[1].W[i][j] -= 0.00000001
return grad
def check_batch_grad(self, x, y):
layer = self.batch_norm_layers[0]
beta_grad = np.zeros_like(layer.beta)
for i in range(layer.beta.shape[0]):
_, f0 = net.forward(x, y)
layer.beta[i] += 0.00000001
_, f1 = net.forward(x, y)
beta_grad[i] = (f1 - f0) / 0.00000001
layer.beta[i] -= 0.00000001
gamma_grad = np.zeros_like(layer.gamma)
for i in range(layer.gamma.shape[0]):
_, f0 = net.forward(x, y)
layer.gamma[i] += 0.00000001
_, f1 = net.forward(x, y)
gamma_grad[i] = (f1 - f0) / 0.00000001
layer.gamma[i] -= 0.00000001
print('beta_grad = \n', beta_grad, '\ngamma_grad = \n', gamma_grad)
def print_params(self):
for layer in self.layers:
layer.print_params()