文章目錄

• • 1. CNN 卷積神經網絡
  • 2. 預訓練模型
  • 3. RNN 循環神經網絡

學習于：簡單粗暴 TensorFlow 2

1. CNN 卷積神經網絡

卷積神經網絡，卷積后尺寸計算

• tf.keras.layers.Conv2D， tf.keras.layers.MaxPool2D

# CNN 模型 class myCNN(tf.keras.Model):def __init__(self):super().__init__()self.conv1 = tf.keras.layers.Conv2D(filters=32,kernel_size=[5,5],padding='same',activation='relu')self.pool1 = tf.keras.layers.MaxPool2D(pool_size=[2,2],strides=2)self.conv2 = tf.keras.layers.Conv2D(filters=64,kernel_size=[5,5],padding='same',activation='relu')self.pool2 = tf.keras.layers.MaxPool2D(pool_size=[2,2],strides=2)self.flatten = tf.keras.layers.Reshape(target_shape=(7*7*64,))self.dense1 = tf.keras.layers.Dense(units=1024,activation='relu')self.dense2 = tf.keras.layers.Dense(units=10)def call(self, inputs): # [m, 28 , 28 , 1]# size = (n+2p-f)/s + 1，same 保持尺寸不變x = self.conv1(inputs) # [m, 28 , 28 , 32]# 池化時 p 常為 0x = self.pool1(x) # [m, 14 , 14 , 32]x = self.conv2(x) # [m, 14 , 14 , 64]x = self.pool2(x) # [m, 7 , 7 , 64]x = self.flatten(x) # [m, 7*7*64]x = self.dense1(x) # [m, 1024]x = self.dense2(x) # [m, 10]outputs = tf.nn.softmax(x) # [m, 10]return outputs

2. 預訓練模型

• mymodel = tf.keras.applications.MobileNetV2()，可以調用 VGG16 、 VGG19 、 ResNet 、 MobileNet 等內置模型，使用預訓練好的權重初始化網絡

import tensorflow as tf import tensorflow_datasets as tfdsnum_epoch = 2 batch_size = 16 learning_rate = 1e-3version = tf.__version__ gpu_ok = tf.config.list_physical_devices('GPU') print("tf version:", version, "\nuse GPU", gpu_ok)# 會自動加載數據集，從網絡下載 dataset = tfds.load("tf_flowers", split=tfds.Split.TRAIN, as_supervised=True) # 數據歸一化、打亂、分批 dataset = dataset.map(lambda img, label: (tf.image.resize(img, (224, 224)) / 255.0, label)).shuffle(1024).batch(batch_size)# 加載預訓練模型 model = tf.keras.applications.MobileNetV2(include_top=True, weights=None, classes=5) # weights 默認： "imagenet"， 也可以不用預訓練權值，include_top 有最后的FC層 optimizer = tf.keras.optimizers.Adam(learning_rate=learning_rate) for e in range(num_epoch):for images, labels in dataset:with tf.GradientTape() as tape:pred = model(images, training=True) # *****設置training模式*****loss = tf.keras.losses.sparse_categorical_crossentropy(y_true=labels, y_pred=pred)loss = tf.reduce_mean(loss)print("loss: {}".format(loss.numpy()))grads = tape.gradient(loss, model.trainable_variables)optimizer.apply_gradients(grads_and_vars=zip(grads, model.trainable_variables))

注：如果數據集下載緩慢，我傳到csdn了，免費下載

3. RNN 循環神經網絡

• 數據預處理，字符 與 idx 的相互轉換映射， 字符集
• 獲取 batch_size 個樣本、每個樣本的下一個字符（標簽）

import tensorflow as tf import numpy as npclass Dataloader():def __init__(self):path = tf.keras.utils.get_file('nietzsche.txt',origin='https://s3.amazonaws.com/text-datasets/nietzsche.txt')with open(path, encoding='utf-8') as f:self.raw_text = f.read().lower()self.chars = sorted(list(set(self.raw_text)))self.char_idx = dict((c, i) for i, c in enumerate(self.chars))self.idx_char = dict((i, c) for i, c in enumerate(self.chars))self.text = [self.char_idx[c] for c in self.raw_text] # 原始文本的 idsdef get_batch(self, seq_len, batch_size):seq = []next_char = []for i in range(batch_size):idx = np.random.randint(0, len(self.text) - seq_len)seq.append(self.text[idx: idx + seq_len])next_char.append(self.text[idx + seq_len])return np.array(seq), np.array(next_char) # [batch_size, seq_len] [batch_size]

• 建模，tf.keras.layers.LSTMCell

class myRNN(tf.keras.Model):def __init__(self, num_chars, batch_size, seq_len):super().__init__()self.num_chars = num_charsself.seq_len = seq_lenself.batch_size = batch_sizeself.cell = tf.keras.layers.LSTMCell(units=256)self.dense = tf.keras.layers.Dense(units=self.num_chars)def call(self, inputs, from_logits=False):inputs = tf.one_hot(inputs, depth=self.num_chars) # [batch_size, seq_len, num_chars]# 獲取初始狀態state = self.cell.get_initial_state(batch_size=self.batch_size, dtype=tf.float32) # [batch_size, 256]for t in range(self.seq_len):output, state = self.cell(inputs[:, t, :], state)logits = self.dense(output) # [batch_size, num_chars]if from_logits: # 控制是否softmax歸一化return logitselse:return tf.nn.softmax(logits)def predict(self, inputs, temperature=1.0): # temperature調節詞匯的豐富度，對概率進行調整batch_size, _ = tf.shape(inputs)logits = self(inputs, from_logits=True) # self 調用 __call()__, __call()__ 調用 call()prob = tf.nn.softmax(logits / temperature).numpy() # [batch_size 64, num_chars]return np.array([np.random.choice(self.num_chars, p=prob[i, :])for i in range(batch_size.numpy())]) # 每個樣本，下一個字符 按預測概率隨機采樣

• 訓練

num_batches = 1000 seq_len = 40 # 輸入序列長度 batch_size = 64 learning_rate = 1e-3data_loader = Dataloader() model = myRNN(num_chars=len(data_loader.chars),batch_size=batch_size,seq_len=seq_len)# 優化器 optimizer = tf.keras.optimizers.Adam(learning_rate=learning_rate)# 訓練迭代 for i in range(num_batches):# 獲取批量數據X, y = data_loader.get_batch(seq_len, batch_size)# 梯度記錄器with tf.GradientTape() as tape:# 前向傳播y_pred = model(X)# 計算損失loss = tf.keras.losses.sparse_categorical_crossentropy(y, y_pred)loss = tf.reduce_mean(loss)print("batch：{}, loss {}".format(i, loss.numpy()))# 計算梯度grads = tape.gradient(loss, model.variables)# 更新參數optimizer.apply_gradients(zip(grads, model.variables))

• 預測

# 拿出一個樣本 X_, _ = data_loader.get_batch(seq_len, 1) for diversity in [0.2, 0.5, 1.0, 1.2]: # 詞匯豐富度X = X_print('diversity {}'.format(diversity))for t in range(400): # 輸出長度 400 y_pred = model.predict(X, diversity) # 預測字符的 idprint(data_loader.idx_char[y_pred[0]], end='', flush=True) # 輸出預測的字符X = np.concatenate([X[:, 1:], np.expand_dims(y_pred, axis=1)], axis=-1) # 滑動1位，作為下一次的輸入print("\n")

輸出：

diversity 0.2the the sere the s and the s and the the sere the the s and in the sere the ches the the sere the the sore the the s and the the s and the serend the seres the the the serely the the the s all the the the s and the the sere the the the sere the ther the the sorece the ninge sore the the the s of sell the pint the s the the the the the the the s of the serere the the s and the sere the s the the tdiversity 0.5 ere---and the ne ous bored bo s the the ande sereng to then hithe the the he sesthard on the non there the mores sor the the thit fus the ches sored the seresit and the the ntithe s at all sent for the fas theng the d end the ind che the the serangen the for ole the soll dund, and chered and the pereropher of the resiged the the s lore not the the s as the s the dethere hor the s mone se soull diversity 1.0 tyrive oop art rrame nd michicosentiun, luind the trourd tho t ts.cseseyreve oud s mhendgcomrools bored ere s oll ow ons, here blprlen, pforzede ntor, in this mis je,iof tore. hon bf cerign then thect nene hfurlilin of fallll devety irtes" the whiy ins puncaliridut drerales alder, as inen waveructache semaltou no aven it yuranty ahd oar in -whe s urofeg to the serecying sicoradt-i0nior anetheragl diversity 1.2 y, dpr thoucg,", ind soncrea5sfporcul os_; fac alin th thel. (owel, thenniv poteer" hithichp-hispin2, ho thas d-lher wrekek---fe l seh rabf ssit afolyicud i iedy, d chendle-hand-a ne lef urovut, phetif po'n. wskin ef; phtors eve mdd ali all an icig tedt g main aisec cowstixgeof adt vinnd thas phinte lllivita ou is toup tualy as isscppomeofea2y ie?y ounscded!wheor ome sllat , hhe"se, ouondibis

豐富度 大一些的時候，預測出來的字符多樣一些

總結

以上是生活随笔為你收集整理的TensorFlow 2.0 - CNN / 预训练 / RNN的全部內容，希望文章能夠幫你解決所遇到的問題。

如果覺得生活随笔網站內容還不錯，歡迎將生活随笔推薦給好友。