Question

我想在TPU上运行Keras代码对文本进行分类。当我的模型想要评估Val_acc时，显示此错误：

batch_size must be divisible by strategy.num_towers (17 vs 8)

这是我的代码：

from __future__ import unicode_literals
import pandas as pd
import openpyxl 
from hazm import *
import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import LabelEncoder
from sklearn.feature_extraction.text import TfidfVectorizer
from keras.preprocessing.text import one_hot
from keras.preprocessing.sequence import pad_sequences
from tensorflow.keras import layers,models

from keras.preprocessing.text import Tokenizer
from keras.preprocessing import sequence
from keras import optimizers
from keras.callbacks import *
from keras.utils import to_categorical
from keras.callbacks import EarlyStopping
from sklearn.model_selection import StratifiedKFold
%matplotlib inline

我的数据：

X: texts
Y: labels

k折：

kfold = StratifiedKFold(n_splits=10, shuffle=True, random_state=7)

培训步骤：

epochs = 10
batch_size = 64
scores=[]
for train, test in kfold.split(X, Y):
     model = RNN2()

     tpu_model = tf.contrib.tpu.keras_to_tpu_model(
                 model,
                 strategy=tf.contrib.tpu.TPUDistributionStrategy(
                tf.contrib.cluster_resolver.TPUClusterResolver(tpu='grpc://' 
                + os.environ['COLAB_TPU_ADDR'])
                   )
                 )
     tpu_model.compile(
                optimizer='adam',
                loss='binary_crossentropy',
                metrics=['accuracy']
               )

     tpu_model.fit(X_sequences_matrix[train],Y[train],
                   batch_size=batch_size,
                   epochs=epochs,validation_split=0.15)

这些模型登录训练步骤：

Epoch 1/10
INFO:tensorflow:New input shapes; (re-)compiling: mode=train, [TensorSpec(shape=(4, 920), dtype=tf.float32, name='input_110'), TensorSpec(shape=(4, 1), dtype=tf.float32, name='activation_34_target0')]
INFO:tensorflow:Overriding default placeholder.
INFO:tensorflow:Remapping placeholder for input
INFO:tensorflow:Cloning Adam {'lr': 0.0010000000474974513, 'beta_1': 0.8999999761581421, 'beta_2': 0.9990000128746033, 'decay': 0.0, 'epsilon': 1e-07, 'amsgrad': False}
INFO:tensorflow:Get updates: Tensor("loss_3/mul:0", shape=(), dtype=float32)
INFO:tensorflow:Started compiling
INFO:tensorflow:Finished compiling. Time elapsed: 11.963297128677368 secs
INFO:tensorflow:Setting weights on TPU model.
3648/3697 [============================>.] - ETA: 1s - loss: 0.3725 - acc: 0.8728 
AssertionError: batch_size must be divisible by strategy.num_towers (17 vs 8)

Answer 1

问题是您有8个TPU内核和public class CustomAuthenticationManager implements AuthenticationManager{ @Autowired private UtenteDao utenteDao; @Override public Authentication authenticate(Authentication authentication) throws AuthenticationException { String username = (String)authentication.getPrincipal(); String password = (String)authentication.getCredentials(); Utente utente = utenteDao.login(username, password); AuthUserDetail principal = new AuthUserDetail(utente); return principal.refreshGrantAuthority(); } }17。如图here所示，您的batch_size必须与TPU内核数可整。

在keras交叉验证中使用Google Colab TPU时，batch_size必须被strategy.num_towers（17 vs 8）错误整除

1 个答案: