1
|
|
|
from abc import ABCMeta, abstractmethod |
2
|
|
|
|
3
|
|
|
import theano |
4
|
|
|
from theano import tensor |
5
|
|
|
from six import add_metaclass |
6
|
|
|
|
7
|
|
|
from blocks.bricks.base import application, Brick |
8
|
|
|
|
9
|
|
|
|
10
|
|
|
@add_metaclass(ABCMeta) |
11
|
|
|
class Cost(Brick): |
12
|
|
|
@abstractmethod |
13
|
|
|
@application |
14
|
|
|
def apply(self, *args, **kwargs): |
15
|
|
|
pass |
16
|
|
|
|
17
|
|
|
|
18
|
|
|
@add_metaclass(ABCMeta) |
19
|
|
|
class CostMatrix(Cost): |
20
|
|
|
"""Base class for costs which can be calculated element-wise. |
21
|
|
|
|
22
|
|
|
Assumes that the data has format (batch, features). |
23
|
|
|
|
24
|
|
|
""" |
25
|
|
|
@application(outputs=["cost"]) |
26
|
|
|
def apply(self, *args, **kwargs): |
27
|
|
|
return self.cost_matrix(*args, **kwargs).sum(axis=1).mean() |
28
|
|
|
|
29
|
|
|
@abstractmethod |
30
|
|
|
@application |
31
|
|
|
def cost_matrix(self, *args, **kwargs): |
32
|
|
|
pass |
33
|
|
|
|
34
|
|
|
|
35
|
|
|
class BinaryCrossEntropy(CostMatrix): |
36
|
|
|
@application |
37
|
|
|
def cost_matrix(self, y, y_hat): |
|
|
|
|
38
|
|
|
cost = tensor.nnet.binary_crossentropy(y_hat, y) |
39
|
|
|
return cost |
40
|
|
|
|
41
|
|
|
|
42
|
|
|
class AbsoluteError(CostMatrix): |
43
|
|
|
@application |
44
|
|
|
def cost_matrix(self, y, y_hat): |
|
|
|
|
45
|
|
|
cost = abs(y - y_hat) |
46
|
|
|
return cost |
47
|
|
|
|
48
|
|
|
|
49
|
|
|
class SquaredError(CostMatrix): |
50
|
|
|
@application |
51
|
|
|
def cost_matrix(self, y, y_hat): |
|
|
|
|
52
|
|
|
cost = tensor.sqr(y - y_hat) |
53
|
|
|
return cost |
54
|
|
|
|
55
|
|
|
|
56
|
|
|
class CategoricalCrossEntropy(Cost): |
57
|
|
|
@application(outputs=["cost"]) |
58
|
|
|
def apply(self, y, y_hat): |
|
|
|
|
59
|
|
|
cost = tensor.nnet.categorical_crossentropy(y_hat, y).mean() |
60
|
|
|
return cost |
61
|
|
|
|
62
|
|
|
|
63
|
|
|
class MisclassificationRate(Cost): |
64
|
|
|
"""Calculates the misclassification rate for a mini-batch. |
65
|
|
|
|
66
|
|
|
Parameters |
67
|
|
|
---------- |
68
|
|
|
top_k : int, optional |
69
|
|
|
If the ground truth class is within the `top_k` highest |
70
|
|
|
responses for a given example, the model is considered |
71
|
|
|
to have predicted correctly. Default: 1. |
72
|
|
|
|
73
|
|
|
Notes |
74
|
|
|
----- |
75
|
|
|
Ties for `top_k`-th place are broken pessimistically, i.e. |
76
|
|
|
in the (in practice, rare) case that there is a tie for `top_k`-th |
77
|
|
|
highest output for a given example, it is considered an incorrect |
78
|
|
|
prediction. |
79
|
|
|
|
80
|
|
|
""" |
81
|
|
|
def __init__(self, top_k=1): |
82
|
|
|
self.top_k = top_k |
83
|
|
|
super(MisclassificationRate, self).__init__() |
84
|
|
|
|
85
|
|
|
@application(outputs=["error_rate"]) |
86
|
|
|
def apply(self, y, y_hat): |
|
|
|
|
87
|
|
|
# Support checkpoints that predate self.top_k |
88
|
|
|
top_k = getattr(self, 'top_k', 1) |
89
|
|
|
if top_k == 1: |
90
|
|
|
mistakes = tensor.neq(y, y_hat.argmax(axis=1)) |
91
|
|
|
else: |
92
|
|
|
row_offsets = theano.tensor.arange(0, y_hat.flatten().shape[0], |
93
|
|
|
y_hat.shape[1]) |
94
|
|
|
truth_score = y_hat.flatten()[row_offsets + y] |
95
|
|
|
# We use greater than _or equals_ here so that the model |
96
|
|
|
# _must_ have its guess in the top k, and cannot extend |
97
|
|
|
# its effective "list of predictions" by tying lots of things |
98
|
|
|
# for k-th place. |
99
|
|
|
higher_scoring = tensor.ge(y_hat, truth_score.dimshuffle(0, 'x')) |
100
|
|
|
# Because we used greater-than-or-equal we have to correct for |
101
|
|
|
# counting the true label. |
102
|
|
|
num_higher = higher_scoring.sum(axis=1) - 1 |
103
|
|
|
mistakes = tensor.ge(num_higher, top_k) |
104
|
|
|
return mistakes.mean(dtype=theano.config.floatX) |
105
|
|
|
|