如何解决序列自动编码器的Keras准确性与手动计算不同
我将基于GRU的seq2seq模型与字节对编码一起用于拼写校正问题。 这是我的模型架构/代码:
encoder_inputs = Input(shape=(padding_length,),name="EncoderInput_1")
embedded_encoder_inputs = Embedding(num_encoder_tokens,latent_dim,mask_zero=True)(encoder_inputs)
encoder = GRU(latent_dim,return_state=True)
_,state_h = encoder(embedded_encoder_inputs)
encoder_states = state_h
decoder_inputs = Input(shape=(padding_length,name="DecoderInput_1")
embedded_decoder_inputs = Embedding(num_decoder_tokens,mask_zero=True)(decoder_inputs)
decoder_lstm = GRU(latent_dim,return_sequences=True,return_state=True)
x,_ = decoder_lstm(embedded_decoder_inputs,initial_state=encoder_states)
decoder_dense = TimeDistributed(Dense(num_decoder_tokens,activation='softmax'))
decoder_outputs = decoder_dense(x)
model = Model([encoder_inputs,decoder_inputs],decoder_outputs)
model.compile(optimizer='adam',loss=sparse_categorical_crossentropy,metrics=["accuracy"])
model.fit([ip_encoded,op_encoded],op_offset_encoded,epochs=epochs,verbose=1,batch_size=batch_size,validation_split=0.2,
我用约700K的训练示例对其进行了10个时期的训练,最终验证准确性约为97%。
我研究了样本的预测和准确性(我正在使用model.evaluate来获得单个示例的准确性),但这与手动计算有所不同。
我想提一下我对手动计算过程中使用的精度计算的理解:
- 对于解码器的所有输入令牌,它会预测下一个令牌。如果预测令牌==预期令牌,则正确预测的“计数”将增加1。我对所有解码器输入令牌应用相同的计数,最后通过将“计数”除以输入令牌总数来计算准确性。
示例:( ip:模型输入,op:模型输出,ex:预期)
(注意:我假设填充令牌不用于度量计算)
(注意::1,
用于查找单个示例准确性的代码:
model.evaluate([ip_encoded[idx:idx+1],op_encoded[idx:idx+1]],op_offset_encoded[idx:idx+1],batch_size=64)
- (所有标记均正确预测,手动和Keras精度相同)
ip: Do you get a pattern similar to t hat shown in Fig.
op: Do you get a pattern similar to that shown in Fig.
ex: Do you get a pattern similar to that shown in Fig.
ip: [1,1313,416,742,261,1953,1239,299,259,308,270,911,283,814,16,2,0]
op: [1313,336,2]
ex: [1313,0]
1/1 [==============================] - 0s 1ms/step - loss: 0.0031 - accuracy: 1.0000
[0.003069164464250207,1.0]
- (手动准确度:17/23 = 0.73,keras:0.95)
ip: The equal part that we get in is,therefore,larger than the equal part we get in.
op: The equal part that we get in is,larger than the equal part in which we get .
ex: The equal part that we get in is,larger than the equal part we get in .
ip: [1,316,752,496,370,294,14,1578,2115,621,264,0]
op: [316,223,406,1964,2]
ex: [316,0]
1/1 [==============================] - 0s 1ms/step - loss: 0.0289 - accuracy: 0.9565
[0.028948727995157242,0.95652174949646]
- (手册:10/13 = 0.77,Keras:0.92)
ip: Can you tell what advanxtage the fungus derives from this association?
op: Can you tell what advantage affect the sewage from this association?
ex: Can you tell what advantage the fungus derives from this association?
ip: [1,1321,2797,985,2923,282,1424,596,9590,19016,389,428,6330,33,0]
op: [1321,6089,2211,4739,2]
ex: [1321,0]
1/1 [==============================] - 0s 1ms/step - loss: 0.0121 - accuracy: 0.9286
[0.012137999758124352,0.9285714030265808]
- 如果社区可以帮助我了解Keras准确性指标的内部结构并验证我的手动指标计算过程,这将非常有帮助。
版权声明:本文内容由互联网用户自发贡献,该文观点与技术仅代表作者本人。本站仅提供信息存储空间服务,不拥有所有权,不承担相关法律责任。如发现本站有涉嫌侵权/违法违规的内容, 请发送邮件至 dio@foxmail.com 举报,一经查实,本站将立刻删除。