Question

我试图从Keras example about siamese network更改代码。但奇怪的是，无论损失减少，精度始终为0.5000。我现在的假设是我错误地修改了create_pair函数，我想尝试将类的数量改为4：

原件：

def create_pairs(x, digit_indices):
'''Positive and negative pair creation.
Alternates between positive and negative pairs.
'''
pairs = []
labels = []
n = min([len(digit_indices[d]) for d in range(10)]) - 1
for d in range(10):
    for i in range(n):
        z1, z2 = digit_indices[d][i], digit_indices[d][i + 1]
        pairs += [[x[z1], x[z2]]]
        inc = random.randrange(1, 10)
        dn = (d + inc) % 10
        z1, z2 = digit_indices[d][i], digit_indices[dn][i]
        pairs += [[x[z1], x[z2]]]
        labels += [1, 0]
return np.array(pairs), np.array(labels)

和第93-97行：

digit_indices = [np.where(y_train == i)[0] for i in range(10)]
tr_pairs, tr_y = create_pairs(x_train, digit_indices)

digit_indices = [np.where(y_test == i)[0] for i in range(10)]
te_pairs, te_y = create_pairs(x_test, digit_indices)

这是我的代码：

def create_pairs(x, digit_indices):
'''Positive and negative pair creation.
Alternates between positive and negative pairs.
'''
pairs = []
labels = []
n = min([len(digit_indices[d]) for d in range(4)]) - 1
for d in range(4):
    for i in range(n):
        z1, z2 = digit_indices[d][i], digit_indices[d][i + 1]
        pairs += [[x[z1], x[z2]]]
        inc = random.randrange(1, 4)
        dn = (d + inc) % 4
        z1, z2 = digit_indices[d][i], digit_indices[dn][i]
        pairs += [[x[z1], x[z2]]]
        labels += [1, 0]
return np.array(pairs), np.array(labels)

和第93-97行：

digit_indices = [np.where(y_train == i)[0] for i in range(4)]
tr_pairs, tr_y = create_pairs(x_train, digit_indices)

digit_indices = [np.where(y_test == i)[0] for i in range(4)]
te_pairs, te_y = create_pairs(x_test, digit_indices)

这是我的base_network（使用RNN的那个，而不是我在评论回复中谈过的转换网，两者都给出相同的结果，50％的准确度）：

def create_base_network(embedding_layer):
    seq = Sequential()
    seq.add(embedding_layer)
    seq.add(GRU(512, use_bias=True, dropout=0.5, recurrent_dropout=0.5, return_sequences=True))
    seq.add(GRU(512, use_bias=True, dropout=0.5, recurrent_dropout=0.5))
    seq.add(Dense(512, activation='relu'))
    seq.add(Dropout(0.1))
    seq.add(Dense(512, activation='relu'))
    return seq

嵌入层只是一个简单的手套矩阵。并且我还在合并后使用sigmoid激活函数添加另一个密集层。

有什么遗漏？或者那不是我应该如何改变它？提前致谢

Answer 1

连体代码错了，尚未修复。问题是在切换0和1时不对称的损失函数，但是keras代码假设它是。改变这一行

return K.mean(y_true * K.square(y_pred) + (1 - y_true) * K.square(K.maximum(margin - y_pred, 0)))

到

return K.mean((1 - y_true) * K.square(y_pred) + y_true * K.square(K.maximum(margin - y_pred, 0)))

和

labels += [1, 0]

到

labels += [0, 1]

如何修改Keras Siamese Network示例？

1 个答案: