Des*_*eux 5 python deep-learning lstm keras mini-batch
问题陈述
我在带有 tensorflow 后端（tf 版本 1.13.1）的 keras 中训练多对多 LSTM，用于标记文本序列，以使用预训练的 GloVe 嵌入来预测序列中每个元素的标记。我的训练方案涉及小批量随机梯度下降，每个小批量矩阵按列填充零以确保输入到网络的长度相等。
至关重要的是，由于任务和数据的性质对我的小批量进行自定义限制，我没有使用 keras 嵌入层。我的目标是为我的零填充单元实现屏蔽机制，以确保损失计算不会将这些单元虚假地视为真正的数据点。
方法
正如 keras文档中所解释的，keras 提供了三种设置掩蔽层的方法：
keras.layers.Embedding图层配置mask_zero
为True。keras.layers.Masking图层；因为我没有使用嵌入层来编码我的训练数据，所以我无法使用带有掩码嵌入层的选项 (1)。因此，我选择了 (2) 并在初始化我的模型后立即添加了一个遮罩层。然而，这种变化似乎没有产生影响。事实上，不仅我的模型的准确性没有提高，在预测阶段模型仍然产生零预测。为什么我的屏蔽层不屏蔽零填充的单元格？这是否与在我的密集层中指定 3 个类而不是 2 个（因此将 0 作为单独的类包括在内）的事实有关？
现有资源的限制
已经提出并回答了类似的问题，但我无法使用它们来解决我的问题。虽然这篇文章没有得到直接回应，但评论中提到的一个链接帖子侧重于如何预处理数据以分配掩码值，这在这里没有争议。然而，掩蔽层初始化与此处使用的相同。这篇文章提到了同样的问题 - 遮罩层对性能没有影响 - 答案以与我相同的方式定义遮罩层，但再次侧重于将特定值转换为遮罩值。最后，这篇文章中的答案提供了相同的层初始化，而无需进一步详细说明。
玩具数据生成
为了重现我的问题，我生成了一个包含两个类 (1,2) 的玩具 10 批次数据集。批是一个可变长度的序列，后填充零，最大长度为 20 个嵌入，每个嵌入向量由 5 个单元组成，因此input_shape=(20,5). 这两个类的嵌入值是从不同但部分重叠的截断正态分布生成的，从而为网络创建了一个可学习但并非微不足道的问题。我在下面包含了玩具数据，以便您可以重现该问题。
import pandas as pd
from keras.models import Sequential
from keras.layers import LSTM, Dense, TimeDistributed, Bidirectional, Dropout, Masking
from keras import optimizers
# *** model initialization ***
model = Sequential()
model.add(Masking(mask_value=0., input_shape=(20, 5))) # <- masking layer here
model.add(Bidirectional(LSTM(20, return_sequences=True), input_shape=(20, 5)))
model.add(Dropout(0.2))
model.add(TimeDistributed(Dense(3, activation='sigmoid')))
sgd = optimizers.SGD(lr=0.01, decay=1e-6, momentum=0.9, nesterov=True)
model.compile(loss='sparse_categorical_crossentropy', optimizer=sgd, metrics=['mse'])
# *** model training ***
for epoch in range(10):
for X,y in data_train:
X = X.reshape(1, 20, 5)
y = y.reshape(1, 20, 1)
history = model.fit(X, y, epochs=1, batch_size=20, verbose=0)
# *** model prediction ***
preds = pd.DataFrame(columns=['true', 'pred'])
for index, (X,y) in enumerate(data_test):
X = X.reshape(1, 20, 5)
y = y.reshape(1, 20, 1)
y_pred = model.predict_classes(X, verbose=0)
df = pd.DataFrame(columns=['true', 'pred'])
df['true'] = [y[0, i][0] for i in range(20)]
df['pred'] = [y_pred[0, i] for i in range(20)]
preds = preds.append(df, ignore_index=True)
# convert true labels to int & drop padded rows (where y_true=0)
preds['true'] = [int(label) for label in preds['true']]
preds = preds[preds['true']!=0]
这是带有掩码的模型的摘要：
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
masking_2 (Masking) (None, 20, 5) 0
_________________________________________________________________
bidirectional_4 (Bidirection (None, 20, 40) 4160
_________________________________________________________________
dropout_4 (Dropout) (None, 20, 40) 0
_________________________________________________________________
time_distributed_4 (TimeDist (None, 20, 3) 123
=================================================================
Total params: 4,283
Trainable params: 4,283
Non-trainable params: 0
我训练了一个模型，一个模型没有掩蔽层，并使用以下方法计算精度：
np.round(sum(preds['true']==preds['pred'])/len(preds)*100,1)
没有遮罩的模型准确率为 53.3%，有遮罩的模型准确率为 33.3%。更令人惊讶的是，在两个模型中，我一直将零作为预测标签。为什么屏蔽层无法忽略零填充单元格？
重现问题的数据：
data_train = list(zip(X_batches_train, y_batches_train))
data_test = list(zip(X_batches_test, y_batches_test))
X_batches_train
[array([[-1.00612917, 1.47313952, 2.68021318, 1.54875809, 0.98385996,
1.49465265, 0.60429106, 1.12396908, -0.24041602, 1.77266187,
0.1961381 , 1.28019637, 1.78803092, 2.05151245, 0.93606708,
0.51554755, 0. , 0. , 0. , 0. ],
[-0.97596563, 2.04536053, 0.88367922, 1.013342 , -0.16605355,
3.02994344, 2.04080806, -0.25153046, -0.5964068 , 2.9607247 ,
-0.49722121, 0.02734492, 2.16949987, 2.77367066, 0.15628842,
2.19823207, 0. , 0. , 0. , 0. ],
[ 0.31546283, 3.27420503, 3.23550769, -0.63724013, 0.89150128,
0.69774266, 2.76627308, -0.58408384, -0.45681779, 1.98843041,
-0.31850477, 0.83729882, 0.45471165, 3.61974147, -1.45610756,
1.35217453, 0. , 0. , 0. , 0. ],
[ 1.03329532, 1.97471646, 1.33949611, 1.22857243, -1.46890642,
1.74105506, 1.40969261, 0.52465603, -0.18895266, 2.81025597,
2.64901037, -0.83415186, 0.76956826, 1.48730868, -0.16190164,
2.24389007, 0. , 0. , 0. , 0. ],
[-1.0676654 , 3.08429323, 1.7601179 , 0.85448051, 1.15537064,
2.82487842, 0.27891413, 0.57842569, -0.62392063, 1.00343057,
1.15348843, -0.37650332, 3.37355345, 2.22285473, 0.43444434,
0.15743873, 0. , 0. , 0. , 0. ]]),
array([[ 1.05258873, -0.17897376, -0.99932932, -1.02854121, 0.85159208,
2.32349131, 1.96526709, -0.08398597, -0.69474809, 1.32820222,
1.19514151, 1.56814867, 0.86013263, 1.48342922, 0. ,
0. , 0. , 0. , 0. , 0. ],
[ 0.1920635 , -0.48702788, 1.24353985, -1.3864121 , 0.16713229,
3.10134683, 0.61658271, -0.63360643, 0.86000807, 2.74876157,
2.87604877, 0.16339724, 2.87595396, 3.2846962 , 0. ,
0. , 0. , 0. , 0. , 0. ],
[ 0.1380241 , -0.76783029, 0.18814436, -1.18165209, -0.02981728,
1.49908113, 0.61521007, -0.98191097, 0.31250199, 1.39015803,
3.16213211, -0.70891214, 3.83881766, 1.92683533, 0. ,
0. , 0. , 0. , 0. , 0. ],
[ 1.39080778, -0.59179216, 0.80348201, 0.64638205, -1.40144268,
1.49751413, 3.0092166 , 1.33099666, 1.43714841, 2.90734268,
3.09688943, 0.32934884, 1.14592787, 1.58152023, 0. ,
0. , 0. , 0. , 0. , 0. ],
[-0.77164353, 0.50293096, 0.0717377 , 0.14487556, -0.90246591,
2.32612179, 1.98628857, 1.29683166, -0.12399569, 2.60184685,
3.20136653, 0.44056647, 0.98283455, 1.79026663, 0. ,
0. , 0. , 0. , 0. , 0. ]]),
array([[-0.93359914, 2.31840281, 0.55691601, 1.90930758, -1.58260431,
-1.05801881, 3.28012523, 3.84105406, -1.2127093 , 0.00490079,
1.28149304, 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ],
[-1.03105486, 2.7703693 , 0.16751813, 1.12127987, -0.44070271,
-0.0789227 , 2.79008301, 1.11456745, 1.13982551, -1.10128658,
0.87430834, 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ],
[-0.69710668, 1.72702833, -2.62599502, 2.34730002, 0.77756661,
0.16415884, 3.30712178, 1.67331828, -0.44022431, 0.56837829,
1.1566811 , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ],
[-0.71845983, 1.79908544, 0.37385522, 1.3870915 , -1.48823234,
-1.487419 , 3.0879945 , 1.74617784, -0.91538815, -0.24244522,
0.81393954, 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ],
[-1.38501563, 3.73330047, -0.52494265, 2.37133716, -0.24546709,
-0.28360782, 2.89384717, 2.42891743, 0.40144022, -1.21850571,
2.00370751, 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ]]),
array([[ 1.27989188, 1.16254538, -0.06889142, 1.84133355, 1.3234908 ,
1.29611702, 2.0019294 , -0.03220116, 1.1085194 , 1.96495985,
1.68544302, 1.94503544, 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ],
[ 1.3004439 , 2.48768923, 0.59809607, 2.38155155, 2.78705889,
1.67018683, 0.21731778, -0.59277191, 2.87427207, 2.63950475,
2.39211459, 0.93083423, 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ],
[ 2.39239371, 0.30900383, -0.97307155, 1.98100711, 0.30613735,
1.12827171, 0.16987791, 0.31959096, 1.30366416, 1.45881023,
2.45668401, 0.5218711 , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ],
[ 0.0826574 , 2.05100254, 0.013161 , 2.95120798, 1.15730011,
0.75537024, 0.13708569, -0.44922143, 0.64834001, 2.50640862,
2.00349347, 3.35573624, 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ],
[ 0.47135124, 2.10258532, 0.70212032, 2.56063126, 1.62466971,
2.64026892, 0.21309489, -0.57752813, 2.21335957, 0.20453233,
0.03106993, 3.01167822, 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ]]),
array([[-0.42125521, 0.54016939, 1.63016057, 2.01555253, -0.10961255,
-0.42549555, 1.55793753, -0.0998756 , 0.36417335, 3.37126414,
1.62151191, 2.84084192, 0.10831384, 0.89293054, -0.08671363,
0.49340353, 0. , 0. , 0. , 0. ],
[-0.37615411, 2.00581062, 2.30426605, 2.02205839, 0.65871664,
1.34478836, -0.55379752, -1.42787727, 0.59732227, 0.84969282,
0.54345723, 0.95849568, -0.17131602, -0.70425277, -0.5337757 ,
1.78207229, 0. , 0. , 0. , 0. ],
[-0.13863276, 1.71490034, 2.02677925, 2.60608619, 0.26916522,
0.35928298, -1.26521844, -0.59859219, 1.19162219, 1.64565259,
1.16787165, 2.95245196, 0.48681084, 1.66621053, 0.918077 ,
-1.10583747, 0. , 0. , 0. , 0. ],
[ 0.87763797, 2.38740754, 2.9111822 , 2.21184069, 0.78091173,
-0.53270909, 0.40100338, -0.83375593, 0.9860009 , 2.43898437,
-0.64499989, 2.95092003, -1.52360727, 0.44640918, 0.78131922,
-0.24401283, 0. , 0. , 0. , 0. ],
[ 0.92615066, 3.45437746, 3.28808981, 2.87207404, -1.60027223,
-1.14164941, -1.63807699, 0.33084805, 2.92963629, 3.51170824,
-0.3286093 , 2.19108385, 0.97812366, -1.82565766, -0.34034678,
-2.0485913 , 0. , 0. , 0. , 0. ]]),
array([[ 1.96438618e+00, 1.88104784e-01, 1.61114494e+00,
6.99567690e-04, 2.55271963e+00, 0.00000000e+00,
0.00000000e+00, 0.00000000e+00, 0.00000000e+00,
0.00000000e+00, 0.00000000e+00, 0.00000000e+00,
0.00000000e+00, 0.00000000e+00, 0.00000000e+00,
0.00000000e+00, 0.00000000e+00, 0.00000000e+00,
0.00000000e+00, 0.00000000e+00],
[ 2.41578815e+00, -5.70625661e-01, 2.15545894e+00,
-1.80948908e+00, 1.62049331e+00, 0.00000000e+00,
0.00000000e+00, 0.00000000e+00, 0.00000000e+00,
0.00000000e+00, 0.00000000e+00, 0.00000000e+00,
0.00000000e+00, 0.00000000e+00, 0.00000000e+00,
0.00000000e+00, 0.00000000e+00, 0.00000000e+00,
0.00000000e+00, 0.00000000e+00],
[ 1.97017040e+00, -1.62556528e+00, 2.49469152e+00,
4.18785985e-02, 2.61875866e+00, 0.00000000e+00,
0.00000000e+00, 0.00000000e+00, 0.00000000e+00,
0.00000000e+00, 0.00000000e+00, 0.00000000e+00,
0.00000000e+00, 0.00000000e+00, 0.00000000e+00,
0.00000000e+00, 0.00000000e+00, 0.00000000e+00,
0.00000000e+00, 0.00000000e+00],
[ 3.14277819e+00, 3.01098398e-02, 7.40376369e-01,
1.76517344e+00, 2.68922918e+00, 0.00000000e+00,
0.00000000e+00, 0.00000000e+00, 0.00000000e+00,
0.00000000e+00, 0.00000000e+00, 0.00000000e+00,
0.00000000e+00, 0.00000000e+00, 0.00000000e+00,
0.00000000e+00, 0.00000000e+00, 0.00000000e+00,
0.00000000e+00, 0.00000000e+00],
[ 2.06250296e+00, 4.67605528e-01, 1.55927230e+00,
1.85788889e-01, 1.30359922e+00, 0.00000000e+00,
0.00000000e+00, 0.00000000e+00, 0.00000000e+00,
0.00000000e+00, 0.00000000e+00, 0.00000000e+00,
0.00000000e+00, 0.00000000e+00, 0.00000000e+00,
0.00000000e+00, 0.00000000e+00, 0.00000000e+00,
0.00000000e+00, 0.00000000e+00]]),
array([[ 1.22152427, 3.74926839, 0.64415552, 2.35268329, 1.98754653,
2.89384829, 0.44589817, 3.94228743, 2.72405657, 0.86222004,
0.68681903, 3.89952458, 1.43454512, 0. , 0. ,
0. , 0. , 0. , 0. , 0. ],
[-0.02203262, 0.95065123, 0.71669023, 0.02919391, 2.30714524,
1.91843002, 0.73611294, 1.20560482, 0.85206836, -0.74221506,
-0.72886308, 2.39872927, -0.95841402, 0. , 0. ,
0. , 0. , 0. , 0. , 0. ],
[ 0.55775319, 0.33773314, 0.79932151, 1.94966883, 3.2113281 ,
2.70768249, -0.69745554, 1.23208345, 1.66199957, 1.69894081,
0.13124461, 1.93256147, -0.17787952, 0. , 0. ,
0. , 0. , 0. , 0. , 0. ],
[ 0.45089205, 2.62430534, -1.9517961 , 2.24040577, 1.75642049,
1.94962325, 0.26796497, 2.28418304, 1.44944487, 0.28723885,
-0.81081633, 1.54840214, 0.82652939, 0. , 0. ,
0. , 0. , 0. , 0. , 0. ],
[ 1.27678173, 1.17204606, -0.24738322, 1.02761617, 1.81060444,
2.37830861, 0.55260134, 2.50046334, 1.04652821, 0.03467176,
-2.07336654, 1.2628897 , 0.61604732, 0. , 0. ,
0. , 0. , 0. , 0. , 0. ]]),
array([[ 3.86138405, 2.35068317, -1.90187438, 0.600788 , 0.18011722,
1.3469559 , -0.54708828, 1.83798823, -0.01957845, 2.88713217,
3.1724991 , 2.90802072, 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ],
[ 1.26785642, 0.51076756, 0.32070756, 2.33758816, 2.08146669,
-0.60796736, 0.93777509, 2.70474711, 0.44785738, 1.61720609,
1.52890594, 3.03072971, 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ],
[ 3.30219394, 3.1515445 , 1.16550716, 2.07489374, 0.66441859,
0.97529244, 0.35176367, 1.22593639, -1.80698271, 1.19936482,
3.34017172, 2.15960657, 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ],
[ 2.34839018, 2.24827352, -1.61070856, 2.81044265, -1.21423372,
0.24633846, -0.82196609, 2.28616568, 0.033922 , 2.7557593 ,
1.16178372, 3.66959512, 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ],
[ 1.32913219, 1.63231852, 0.58642744, 1.55873546, 0.86354741,
2.06654246, -0.44036504, 3.22723595, 1.33279468, 0.05975892,
2.48518999, 3.44690602, 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ]]),
array([[ 0.61424344, -1.03068819, -1.47929328, 2.91514641, 2.06867196,
1.90384921, -0.45835234, 1.22054782, 0.67931536, 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ],
[ 2.76480464, 1.12442631, -2.36004758, 2.91912726, 1.67891181,
3.76873596, -0.93874096, -0.32397781, -0.55732374, 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ],
[ 0.39953353, -1.26828104, 0.44482517, 2.85604975, 3.08891062,
2.60268725, -0.15785176, 1.58549879, -0.32948578, 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ],
[ 1.65156484, -1.56545168, -1.42771206, 2.74216475, 1.8758154 ,
3.51169147, 0.18353058, -0.14704149, 0.00442783, 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ],
[ 1.27736372, 0.37407608, -1.25713475, 0.53171176, 1.53714914,
0.21015523, -1.06850669, -0.09755327, -0.92373834, 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ]]),
array([[-1.39160433, 0.21014669, -0.89792475, 2.6702794 , 1.54610601,
0.84699037, 2.96726482, 1.84236946, 0.02211578, 0.32842575,
1.02718924, 1.78447936, -1.20056829, 2.26699318, -0.23156537,
2.50124959, 1.93372501, 0.10264369, -1.70813962, 0. ],
[ 0.38823591, -1.30348049, -0.31599117, 2.60044143, 2.32929389,
1.40348483, 3.25758736, 1.92210728, -0.34150988, -1.22336921,
2.3567069 , 1.75456835, 0.28295694, 0.68114898, -0.457843 ,
1.83372069, 2.10177851, -0.26664178, -0.26549595, 0. ],
[ 0.08540346, 0.71507504, 1.78164285, 3.04418137, 1.52975256,
3.55159169, 3.21396003, 3.22720346, 0.68147142, 0.12466013,
-0.4122895 , 1.97986653, 1.51671949, 2.06096825, -0.6765908 ,
2.00145086, 1.73723014, 0.50186043, -2.27525744, 0. ],
[ 0.00632717, 0.3050794 , -0.33167875, 1.48109172, 0.19653696,
1.97504239, 2.51595821, 1.74499313, -1.65198805, -1.04424953,
-0.23786945, 1.18639347, -0.03568057, 3.82541131, 2.84039446,
2.88325909, 1.79827675, -0.80230291, 0.08165052, 0. ],
[ 0.89980086, 0.34690991, -0.60806566, 1.69472308, 1.38043417,
0.97139487, 0.21977176, 1.01340944, -1.69946943, -0.01775586,
-0.35851919, 1.81115864, 1.15105661, 1.21410373, 1.50667558,
1.70155313, 3.1410754 , -0.54806167, -0.51879299, 0. ]])]
y_batches_train
[array([1., 2., 2., 1., 1., 2., 2., 1., 1., 2., 1., 1., 2., 2., 1.
第一个问题：重塑后的 X 数据不是您所期望的。如果你看一下重塑后的第一个样本，它是：
array([[[-1.00612917, 1.47313952, 2.68021318, 1.54875809,
0.98385996],
[ 1.49465265, 0.60429106, 1.12396908, -0.24041602,
1.77266187],
[ 0.1961381 , 1.28019637, 1.78803092, 2.05151245,
0.93606708],
[ 0.51554755, 0. , 0. , 0. ,
0. ],
[-0.97596563, 2.04536053, 0.88367922, 1.013342 ,
-0.16605355],
[ 3.02994344, 2.04080806, -0.25153046, -0.5964068 ,
2.9607247 ],
[-0.49722121, 0.02734492, 2.16949987, 2.77367066,
0.15628842],
[ 2.19823207, 0. , 0. , 0. ,
0. ],
[ 0.31546283, 3.27420503, 3.23550769, -0.63724013,
0.89150128],
[ 0.69774266, 2.76627308, -0.58408384, -0.45681779,
1.98843041],
[-0.31850477, 0.83729882, 0.45471165, 3.61974147,
-1.45610756],
[ 1.35217453, 0. , 0. , 0. ,
0. ],
[ 1.03329532, 1.97471646, 1.33949611, 1.22857243,
-1.46890642],
[ 1.74105506, 1.40969261, 0.52465603, -0.18895266,
2.81025597],
[ 2.64901037, -0.83415186, 0.76956826, 1.48730868,
-0.16190164],
[ 2.24389007, 0. , 0. , 0. ,
0. ],
[-1.0676654 , 3.08429323, 1.7601179 , 0.85448051,
1.15537064],
[ 2.82487842, 0.27891413, 0.57842569, -0.62392063,
1.00343057],
[ 1.15348843, -0.37650332, 3.37355345, 2.22285473,
0.43444434],
[ 0.15743873, 0. , 0. , 0. ,
0. ]]])
所以实际上没有时间步被屏蔽，因为Masking层只屏蔽所有特征都为0的时间步，所以上面的20个时间步没有被屏蔽，因为它们都不是完全0的。
对于 Masking 层，为了确保将 mask 成功传播到输出层，您可以执行以下操作：
for i, l in enumerate(model.layers):
print(f'layer {i}: {l}')
print(f'has input mask: {l.input_mask}')
print(f'has output mask: {l.output_mask}')
layer 0: <tensorflow.python.keras.layers.core.Masking object at 0x6417b7f60>
has input mask: None
has output mask: Tensor("masking/Identity_1:0", shape=(None, 20), dtype=bool)
layer 1: <tensorflow.python.keras.layers.wrappers.Bidirectional object at 0x641e25cf8>
has input mask: Tensor("masking/Identity_1:0", shape=(None, 20), dtype=bool)
has output mask: Tensor("masking/Identity_1:0", shape=(None, 20), dtype=bool)
layer 2: <tensorflow.python.keras.layers.core.Dropout object at 0x641814128>
has input mask: Tensor("masking/Identity_1:0", shape=(None, 20), dtype=bool)
has output mask: Tensor("masking/Identity_1:0", shape=(None, 20), dtype=bool)
layer 3: <tensorflow.python.keras.layers.wrappers.TimeDistributed object at 0x6433b6ba8>
has input mask: Tensor("masking/Identity_1:0", shape=(None, 20), dtype=bool)
has output mask: Tensor("time_distributed/Reshape_3:0", shape=(None, 20), dtype=bool)
所以你可以看到最后一层也有output_mask，这意味着掩码已成功传播。您似乎对Keras中Masking的工作原理有误解，它实际上会生成一个掩码，这是一个布尔数组，掩码的形状是（无，时间步长），因为在您的模型定义中，时间步长维度始终保持相同，因此掩码将传播到末尾而不发生任何变化。然后，当 Keras 计算损失时（当然还有计算梯度时），掩码值为 False 的时间步将被忽略。Masking 层不会改变输出值，当然你的模型仍然会预测类 0，它唯一做的就是生成一个布尔数组，指示应跳过哪个时间步并将其传递到末尾（如果所有层都接受面具）。
因此，您可以做的就是按如下所示更改模型定义的一行，并将 y_labels 移动 1，这意味着您当前的类：
0 -> 0（因为这些时间步的损失将被忽略，对模型的训练没有贡献，所以是0还是1并不重要）
1 -> 0
2 -> 1
# I would prefer softmax if doing classification
# here we only need to specify 2 classes
# and actually TimeDistributed can be thrown away (at least in recent Keras versions)
model.add(TimeDistributed(Dense(2, activation='softmax')))
您还可以在这里查看我的答案/sf/answers/4151970371/，以了解如何使用/不使用掩蔽来计算损失。
| 归档时间： |
|
| 查看次数： |
1530 次 |
| 最近记录： |