forked from cgpotts/cs224u
-
Notifications
You must be signed in to change notification settings - Fork 0
/
np_glove.py
200 lines (160 loc) · 5.96 KB
/
np_glove.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
import itertools
import numpy as np
import pandas as pd
import random
import sys
import utils
__author__ = "Christopher Potts"
__version__ = "CS224u, Stanford, Spring 2022"
class GloVe:
def __init__(self,
n=100,
xmax=100,
alpha=0.75,
max_iter=100,
eta=0.05,
tol=1e-5,
display_progress=True):
"""
Basic GloVe. This is mainly here as a reference implementation.
We recommend using `torch_glove.py` instead.
Parameters
----------
df : pd.DataFrame or np.array
This must be a square matrix.
n : int (default: 100)
The dimensionality of the output vectors.
xmax : int (default: 100)
Words with frequency greater than this are given weight 1.0.
Words with frequency under this are given weight (c/xmax)**alpha
where c is their count in mat (see the paper, eq. (9)).
alpha : float (default: 0.75)
Exponent in the weighting function (see the paper, eq. (9)).
max_iter : int (default: 100)
Number of training epochs.
eta : float (default: 0.05)
Controls the rate of SGD weight updates.
tol : float (default: 1e-4)
Stopping criterion for the loss.
display_progress : bool (default: True)
Whether to print iteration number and current error to stdout.
"""
self.n = n
self.xmax = xmax
self.alpha = alpha
self.max_iter = max_iter
self.eta = eta
self.tol = tol
self.display_progress = display_progress
def fit(self, df):
"""
Learn the GloVe matrix.
Parameters
----------
df : pd.DataFrame or np.array, shape `(n_vocab, n_vocab)`
This should be a matrix of (possibly scaled) co-occcurrence
counts.
Returns
-------
pd.DataFrame or np.array, shape `(n_vocab, self.n)`
The type will be the same as the user's `df`. If it's a
`pd.DataFrame`, the index will be the same as `df.index`.
"""
X = self.convert_input_to_array(df)
m = X.shape[0]
# Parameters:
W = utils.randmatrix(m, self.n) # Word weights.
C = utils.randmatrix(m, self.n) # Context weights.
B = utils.randmatrix(2, m) # Word and context biases.
# Precomputable GloVe values:
X_log = utils.log_of_array_ignoring_zeros(X)
X_weights = (np.minimum(X, self.xmax) / self.xmax)**self.alpha # eq. (9)
# Learning:
indices = list(range(m))
for iteration in range(self.max_iter):
epoch_error = 0.0
random.shuffle(indices)
for i, j in itertools.product(indices, indices):
if X[i, j] > 0.0:
weight = X_weights[i,j]
# Cost is J' based on eq. (8) in the paper:
diff = W[i].dot(C[j]) + B[0, i] + B[1, j] - X_log[i, j]
fdiff = diff * weight
# Gradients:
wgrad = fdiff * C[j]
cgrad = fdiff * W[i]
wbgrad = fdiff
wcgrad = fdiff
# Updates:
W[i] -= self.eta * wgrad
C[j] -= self.eta * cgrad
B[0, i] -= self.eta * wbgrad
B[1, j] -= self.eta * wcgrad
# One-half squared error term:
epoch_error += 0.5 * weight * (diff**2)
epoch_error /= m
if epoch_error <= self.tol:
utils.progress_bar(
"Converged on iteration {} with error {}".format(
iteration, epoch_error, self.display_progress))
break
utils.progress_bar(
"Finished epoch {} of {}; error is {}".format(
iteration, self.max_iter, epoch_error, self.display_progress))
# Return the sum of the word and context matrices, per the advice
# in section 4.2:
G = W + C
self.embedding = self.convert_output(G, df)
return self.embedding
def score(self, X):
"""
The goal of GloVe is to learn vectors whose dot products are
proportional to the log co-occurrence probability. This score
method assesses that directly using the current `self.embedding`.
Parameters
----------
X : pd.DataFrame or np.array, shape `(self.n_words, self.n_vocab)`
The original count matrix.
Returns
-------
float
The Pearson correlation.
"""
X = self.convert_input_to_array(X)
G = self.convert_input_to_array(self.embedding)
mask = X > 0
M = G.dot(G.T)
X_log = utils.log_of_array_ignoring_zeros(X)
row_log_prob = np.log(X.sum(axis=1))
row_log_prob = np.outer(row_log_prob, np.ones(X.shape[1]))
prob = X_log - row_log_prob
return np.corrcoef(prob[mask].ravel(), M[mask].ravel())[0, 1]
def convert_input_to_array(self, X):
if isinstance(X, pd.DataFrame):
X = X.values
return X
@staticmethod
def convert_output(X_pred, X):
if isinstance(X, pd.DataFrame):
X_pred = pd.DataFrame(X_pred, index=X.index)
return X_pred
def simple_example():
utils.fix_random_seeds()
X = np.array([
[4., 4., 2., 0.],
[4., 61., 8., 18.],
[2., 8., 10., 0.],
[0., 18., 0., 5.]])
mod = GloVe(n=2, max_iter=1000)
print(mod)
G = mod.fit(X)
print("\nLearned vectors:")
print(G)
print("We expect the dot product of learned vectors "
"to be proportional to the log co-occurrence probs. "
"Let's see how close we came:")
corr = mod.score(X)
print("Pearson's R: {} ".format(corr))
return corr
if __name__ == '__main__':
simple_example()