一，准备数据
imdb数据集的目标是根据电影评论的文本内容预测评论的情感标签。
训练集有20000条电影评论文本，测试集有5000条电影评论文本，其中正面评论和负面评论都各占一半。
文本数据预处理较为繁琐，包括中文切词（本示例不涉及），构建词典，编码转换，序列填充，构建数据管道等等。
在tensorflow中完成文本数据预处理的常用方案有两种，第一种是利用tf.keras.preprocessing中的Tokenizer词典构建工具和tf.keras.utils.Sequence构建文本数据生成器管道。
第二种是使用tf.data.Dataset搭配.keras.layers.experimental.preprocessing.TextVectorization预处理层。
第一种方法较为复杂，其使用范例可以参考以下文章。
https://zhuanlan.zhihu.com/p/67697840
第二种方法为TensorFlow原生方式，相对也更加简单一些。
我们此处介绍第二种方法。
首先看一下train.csv中的部分内容是什么：
"0 It really boggles my mind when someone comes across a movie like this and claims it to be one of the worst slasher films out there. This is by far not one of the worst out there" still not a good movie but not the worst nonetheless. Go see something like Death Nurse or Blood Lake and then come back to me and tell me if you think the Night Brings Charlie is the worst. The film has decent camera work and editing which is way more than I can say for many more extremely obscure slasher films.
The film doesn't deliver on the on-screen deaths there's one death where you see his pruning saw rip into a neck but all other deaths are hardly interesting. But the lack of on-screen graphic violence doesn't mean this isn't a slasher film just a bad one.
The film was obviously intended not to be taken too seriously. The film came in at the end of the second slasher cycle so it certainly was a reflection on traditional slasher elements done in a tongue in cheek way. For example after a kill Charlie goes to the town's 'welcome' sign and marks the population down one less. This is something that can only get a laugh.
If you're into slasher films definitely give this film a watch. It is slightly different than your usual slasher film with possibility of two killers but not by much. The comedy of the movie is pretty much telling the audience to relax and not take the movie so god darn serious. You may forget the movie you may remember it. I'll remember it because I love the name. "0 Mary Pickford becomes the chieftain of a Scottish clan after the death of her father" and then has a romance. As fellow commenter Snow Leopard said the film is rather episodic to begin. Some of it is amusing such as Pickford whipping her clansmen to church while some of it is just there. All in all the story is weak especially the recycled contrived romance plot-line and its climax. The transfer is so dark it's difficult to appreciate the scenery but even accounting for that this doesn't appear to be director Maurice Tourneur's best work. Pickford and Tourneur collaborated once more in the somewhat more accessible 'The Poor Little Rich Girl ' typecasting Pickford as a child character. "0 Well" at least my theater group did lol. So of course I remember watching Grease since I was a little girl while it was never my favorite musical or story it does still hold a little special place in my heart since it's still a lot of fun to watch. I heard horrible things about Grease 2 and that's why I decided to never watch it but my boyfriend said that it really wasn't all that bad and my friend agreed so I decided to give it a shot but I called them up and just laughed. First off the plot is totally stolen from the first one and it wasn't really clever not to mention they just used the same characters but with different names and actors. Tell me how did the Pink Ladies and T-Birds continue years on after the former gangs left? Not to mention the creator face motor cycle enemy gee what a striking resemblance to the guys in the first film as well as these T-Birds were just stupid and ridiculous.
Another year at Rydell and the music and dancing hasn't stopped. But when a new student who is Sandy's cousin comes into the scene he is love struck by a pink lady Stephanie. But she must stick to the code where only Pink Ladies must stick with the T-Birds so the new student decides to train as a T-Bird to win her heart. So he dresses up as a rebel motor cycle bandit who can ride well and defeat the evil bikers from easily kicking the T-Bird's butts. But will he tell Stephanie who he really is or will she find out on her own? Well find out for yourself.
Grease 2 is like a silly TV show of some sort that didn't work. The gang didn't click as well as the first Grease did not to mention Frenchy coming back was a bit silly and unbelievable because I thought that she graduated from Rydell but apparently she didn't. The songs were not really that catchy; I'm glad that Michelle was able to bounce back so fast but that's probably because she was the only one with talent in this silly little sequel I wouldn't really recommend this film other than if you are curious but I warned you this is just a pathetic attempt at more money from the famous musical.
2/10 "1 I must give How She Move a near-perfect rating because the content is truly great. As a previous reviewer commented" I have no idea how this film has found itself in IMDBs bottom 100 list! That's absolutely ridiculous! Other films--particular those that share the dance theme--can't hold a candle to this one in terms of its combination of top-notch believable acting and amazing dance routines.
From start to finish the underlying story (this is not just about winning a competition) is very easy to delve into and surprisingly realistic. None of the main characters in this are 2-dimensional by any means and by the end of the film it's very easy to feel emotionally invested in them. (And even if you're not the crying type you might get a little weepy-eyed before the credits roll.)
I definitely recommend this film to dance-lovers and even more so to those who can appreciate a poignant and well-acted storyline. How She Move isn't perfect of course (what film is?) but it's definitely a cut above movies that use pretty faces to hide a half-baked plot and/or characters who lack substance. The actors and settings in this film make for a very realistic ride that is equally enthralling thanks to the amazing talent of the dancers! "0 I must say" when I read the storyline on the back of the case It sounded really interesting but when I started to watch the movie seemed boring at first and even more at the end. Some scenes are way too long and the story has not been worked out properly. "0 i am 13 and i hated this film its the worst film on earth i totally wasted my time watching it and was disappointed with it cause on the cover and on the back the film it looks pretty good" but i was wrong its bad. but when i saw delta she was totally different and a bad actress and i really didn't know how old the 2 girls was trying to be i was so confused. the film was in some parts confusing and i didn't enjoy it at all but i watched all the film just to see if it was going to get better but it didn't it was boring dull and did i say BORING.and i don't think many other people liked it as well as me.boring boring boring "0 The acting may be okay" the more u watch this movie the more u wish you weren't this movie is so horrible that if I could get a hold of every copy I would burn them all and not look back this movie is terrible!! "0 I've seen some bad things in my time. A half dead cow trying to get out of waist high mud; a head on collision between two cars; a thousand plates smashing on a kitchen floor; human beings living like animals.
But never in my life have I seen anything as bad as The Cat in the Hat.
This film is worse than 911" worse than Hitler worse than Vllad the Impaler worse than people who put kittens in microwaves.
It is the most disturbing film of all time easy.
I used to think it was a joke some elaborate joke and that Mike Myers was maybe a high cocaine sniffing drug addled betting junkie who lost a bet or something.
I shudder
每一个单元格里面都是一句话。
然后是构造训练集和测试集：
import numpy as np import pandas as pd from matplotlib import pyplot as plt import tensorflow as tf from tensorflow.keras import models,layers,preprocessing,optimizers,losses,metrics from tensorflow.keras.layers.experimental.preprocessing import TextVectorization import re,string train_data_path = "./data/imdb/train.csv" test_data_path = "./data/imdb/test.csv" MAX_WORDS = 10000 # 仅考虑最高频的10000个词 MAX_LEN = 200 # 每个样本保留200个词的长度 BATCH_SIZE = 20 # 构建管道 def split_line(line): arr = tf.strings.split(line,"\t") label = tf.expand_dims(tf.cast(tf.strings.to_number(arr[0]),tf.int32),axis = 0) text = tf.expand_dims(arr[1],axis = 0) return (text,label) ds_train_raw = tf.data.TextLineDataset(filenames = [train_data_path]) \ .map(split_line,num_parallel_calls = tf.data.experimental.AUTOTUNE) \ .shuffle(buffer_size = 1000).batch(BATCH_SIZE) \ .prefetch(tf.data.experimental.AUTOTUNE) ds_test_raw = tf.data.TextLineDataset(filenames = [test_data_path]) \ .map(split_line,num_parallel_calls = tf.data.experimental.AUTOTUNE) \ .batch(BATCH_SIZE) \ .prefetch(tf.data.experimental.AUTOTUNE) # 构建词典 def clean_text(text): lowercase = tf.strings.lower(text) stripped_html = tf.strings.regex_replace(lowercase, '
', ' ') cleaned_punctuation = tf.strings.regex_replace(stripped_html, '[%s]' % re.escape(string.punctuation),'') return cleaned_punctuation vectorize_layer = TextVectorization( standardize=clean_text, split = 'whitespace', max_tokens=MAX_WORDS-1, #有一个留给占位符 output_mode='int', output_sequence_length=MAX_LEN) ds_text = ds_train_raw.map(lambda text,label: text) vectorize_layer.adapt(ds_text) print(vectorize_layer.get_vocabulary()[0:100]) # 单词编码 ds_train = ds_train_raw.map(lambda text,label:(vectorize_layer(text),label)) \ .prefetch(tf.data.experimental.AUTOTUNE) ds_test = ds_test_raw.map(lambda text,label:(vectorize_layer(text),label)) \ .prefetch(tf.data.experimental.AUTOTUNE)
[b'the', b'and', b'a', b'of', b'to', b'is', b'in', b'it', b'i', b'this', b'that', b'was', b'as', b'for', b'with', b'movie', b'but', b'film', b'on', b'not', b'you', b'his', b'are', b'have', b'be', b'he', b'one', b'its', b'at', b'all', b'by', b'an', b'they', b'from', b'who', b'so', b'like', b'her', b'just', b'or', b'about', b'has', b'if', b'out', b'some', b'there', b'what', b'good', b'more', b'when', b'very', b'she', b'even', b'my', b'no', b'would', b'up', b'time', b'only', b'which', b'story', b'really', b'their', b'were', b'had', b'see', b'can', b'me', b'than', b'we', b'much', b'well', b'get', b'been', b'will', b'into', b'people', b'also', b'other', b'do', b'bad', b'because', b'great', b'first', b'how', b'him', b'most', b'dont', b'made', b'then', b'them', b'films', b'movies', b'way', b'make', b'could', b'too', b'any', b'after', b'characters']
二，定义模型
使用Keras接口有以下3种方式构建模型：使用Sequential按层顺序构建模型，使用函数式API构建任意结构模型，继承Model基类构建自定义模型。
此处选择使用继承Model基类构建自定义模型。
# 演示自定义模型范例，实际上应该优先使用Sequential或者函数式API tf.keras.backend.clear_session() class CnnModel(models.Model): def __init__(self): super(CnnModel, self).__init__() def build(self,input_shape): self.embedding = layers.Embedding(MAX_WORDS,7,input_length=MAX_LEN) self.conv_1 = layers.Conv1D(16, kernel_size= 5,name = "conv_1",activation = "relu") self.pool = layers.MaxPool1D() self.conv_2 = layers.Conv1D(128, kernel_size=2,name = "conv_2",activation = "relu") self.flatten = layers.Flatten() self.dense = layers.Dense(1,activation = "sigmoid") super(CnnModel,self).build(input_shape) def call(self, x): x = self.embedding(x) x = self.conv_1(x) x = self.pool(x) x = self.conv_2(x) x = self.pool(x) x = self.flatten(x) x = self.dense(x) return(x) model = CnnModel() model.build(input_shape =(None,MAX_LEN)) model.summary()
模型结构：
三，训练模型
训练模型通常有3种方法，内置fit方法，内置train_on_batch方法，以及自定义训练循环。此处我们通过自定义训练循环训练模型。
# 打印时间分割线 @tf.function def printbar(): ts = tf.timestamp() today_ts = ts%(24*60*60) hour = tf.cast(today_ts//3600+8,tf.int32)%tf.constant(24) minite = tf.cast((today_ts%3600)//60,tf.int32) second = tf.cast(tf.floor(today_ts%60),tf.int32) def timeformat(m): if tf.strings.length(tf.strings.format("{}",m))==1: return(tf.strings.format("0{}",m)) else: return(tf.strings.format("{}",m)) timestring = tf.strings.join([timeformat(hour),timeformat(minite), timeformat(second)],separator = ":") tf.print("=========="*8,end = "") tf.print(timestring) optimizer = optimizers.Nadam() loss_func = losses.BinaryCrossentropy() train_loss = metrics.Mean(name='train_loss') train_metric = metrics.BinaryAccuracy(name='train_accuracy') valid_loss = metrics.Mean(name='valid_loss') valid_metric = metrics.BinaryAccuracy(name='valid_accuracy') @tf.function def train_step(model, features, labels): with tf.GradientTape() as tape: predictions = model(features,training = True) loss = loss_func(labels, predictions) gradients = tape.gradient(loss, model.trainable_variables) optimizer.apply_gradients(zip(gradients, model.trainable_variables)) train_loss.update_state(loss) train_metric.update_state(labels, predictions) @tf.function def valid_step(model, features, labels): predictions = model(features,training = False) batch_loss = loss_func(labels, predictions) valid_loss.update_state(batch_loss) valid_metric.update_state(labels, predictions) def train_model(model,ds_train,ds_valid,epochs): for epoch in tf.range(1,epochs+1): for features, labels in ds_train: train_step(model,features,labels) for features, labels in ds_valid: valid_step(model,features,labels) #此处logs模板需要根据metric具体情况修改 logs = 'Epoch={},Loss:{},Accuracy:{},Valid Loss:{},Valid Accuracy:{}' if epoch%1==0: printbar() tf.print(tf.strings.format(logs, (epoch,train_loss.result(),train_metric.result(),valid_loss.result(),valid_metric.result()))) tf.print("") train_loss.reset_states() valid_loss.reset_states() train_metric.reset_states() valid_metric.reset_states() train_model(model,ds_train,ds_test,epochs = 6)
训练结果：
================================================================================14:45:06 Epoch=1,Loss:0.474225521,Accuracy:0.7376,Valid Loss:0.336961836,Valid Accuracy:0.8526 ================================================================================14:45:12 Epoch=2,Loss:0.245222151,Accuracy:0.9035,Valid Loss:0.326947063,Valid Accuracy:0.8666 ================================================================================14:45:17 Epoch=3,Loss:0.165854618,Accuracy:0.93795,Valid Loss:0.365531504,Valid Accuracy:0.867 ================================================================================14:45:23 Epoch=4,Loss:0.104812928,Accuracy:0.96395,Valid Loss:0.448238105,Valid Accuracy:0.861 ================================================================================14:45:29 Epoch=5,Loss:0.0595887862,Accuracy:0.98125,Valid Loss:0.602612,Valid Accuracy:0.8624 ================================================================================14:45:35 Epoch=6,Loss:0.0318539739,Accuracy:0.9905,Valid Loss:0.762770712,Valid Accuracy:0.8598
四，评估模型
通过自定义训练循环训练的模型没有经过编译，无法直接使用model.evaluate(ds_valid)方法
def evaluate_model(model,ds_valid): for features, labels in ds_valid: valid_step(model,features,labels) logs = 'Valid Loss:{},Valid Accuracy:{}' tf.print(tf.strings.format(logs,(valid_loss.result(),valid_metric.result()))) valid_loss.reset_states() train_metric.reset_states() valid_metric.reset_states() evaluate_model(model,ds_test)
评估结果：
Valid Loss:0.762770712,Valid Accuracy:0.8598
五，使用模型
可以使用以下方法:
- model.predict(ds_test)
- model(x_test)
- model.call(x_test)
- model.predict_on_batch(x_test)
推荐优先使用model.predict(ds_test)方法，既可以对Dataset，也可以对Tensor使用。
model.predict(ds_test)for x_test,_ in ds_test.take(1): print(model(x_test)) #以下方法等价： #print(model.call(x_test)) #print(model.predict_on_batch(x_test))
评估结果：
tf.Tensor( [[9.9007505e-01] [9.9999797e-01] [9.9836570e-01] [2.6509229e-06] [4.7592866e-01] [3.7760619e-05] [8.0391978e-08] [1.6816575e-05] [9.9996006e-01] [9.9695146e-01] [1.0000000e+00] [9.9962234e-01] [1.9009445e-08] [9.7622436e-01] [4.4549329e-06] [2.8802201e-01] [1.0730105e-04] [3.8324962e-03] [2.2874507e-03] [9.9966860e-01]], shape=(20, 1), dtype=float32)
六，保存模型
推荐使用TensorFlow原生方式保存模型。
model.save('./data/tf_model_savedmodel', save_format="tf") print('export saved model.') model_loaded = tf.keras.models.load_model('./data/tf_model_savedmodel') model_loaded.predict(ds_test)
结果：
WARNING:tensorflow:From /usr/local/lib/python3.6/dist-packages/tensorflow/python/ops/resource_variable_ops.py:1817: calling BaseResourceVariable.__init__ (from tensorflow.python.ops.resource_variable_ops) with constraint is deprecated and will be removed in a future version. Instructions for updating: If using Keras pass *_constraint arguments to layers. INFO:tensorflow:Assets written to: ./data/tf_model_savedmodel/assets export saved model. WARNING:tensorflow:No training configuration found in save file, so the model was *not* compiled. Compile it manually. array([[0.99007505], [0.999998 ], [0.9983657 ], ..., [0.9962114 ], [0.94716024], [1. ]], dtype=float32)
参考：
开源电子书地址：https://lyhue1991.github.io/eat_tensorflow2_in_30_days/
GitHub 项目地址：https://github.com/lyhue1991/eat_tensorflow2_in_30_days