在 parameters.py 中，定义了各类参数。
1 # training data directory 2 TRAINING_DATA_DIR = './data/' 3 4 # checkpoint directory 5 CHECKPOINT_DIR = './training_checkpoints/' 6 7 # training details 8 BATCH_SIZE = 16 9 BUFFER_SIZE = 128 10 EPOCHS = 15
在 numpy_dataset.py 中，创建了 5000 组训练数据集，模拟 y = x^3 + 1，并二进制格式写入文件。
1 from parameters import TRAINING_DATA_DIR 2 3 import numpy as np 4 import matplotlib.pyplot as plt 5 import os 6 7 8 # create training data 9 X = np.linspace(-1, 1, 5000) 10 np.random.shuffle(X) 11 y = X ** 3 + 1 + np.random.normal(0, 0.01, (5000,)) 12 13 # plot training data 14 plt.scatter(X, y) 15 plt.show() 16 17 # save data 18 if not os.path.exists(TRAINING_DATA_DIR): 19 os.makedirs(TRAINING_DATA_DIR) 20 21 X.tofile(os.path.join(TRAINING_DATA_DIR + 'training_data_X.bin')) 22 y.tofile(os.path.join(TRAINING_DATA_DIR + 'training_data_y.bin'))
在 subclassed_model.py 中，通过对 tf.keras.models.Model 进行子类化，设计了两个自定义模型。
1 import tensorflow as tf 2 tf.enable_eager_execution() 3 4 5 # model definition 6 class Encoder(tf.keras.models.Model): 7 def __init__(self): 8 super(Encoder, self).__init__() 9 self.fc1 = tf.keras.layers.Dense(units=16, activation='relu') 10 self.fc2 = tf.keras.layers.Dense(units=8, activation='relu') 11 12 def call(self, inputs): 13 r = self.fc1(inputs) 14 return self.fc2(r) 15 16 17 class Decoder(tf.keras.models.Model): 18 def __init__(self): 19 super(Decoder, self).__init__() 20 self.fc = tf.keras.layers.Dense(units=1, activation=None) 21 22 def call(self, inputs): 23 return self.fc(inputs)
在 loss_function.py 中，定义了损失函数。
1 import tensorflow as tf 2 tf.enable_eager_execution() 3 4 5 def loss(real, pred): 6 return tf.losses.mean_squared_error(labels=real, predictions=pred)
在 training.py 中，使用在 numpy_dataset.py 中创建的数据集训练模型，之后使用 model.save_weights() 保存 Keras Subclassed Model 模型，并创建验证集验证模型。
1 from parameters import TRAINING_DATA_DIR, CHECKPOINT_DIR, BATCH_SIZE, BUFFER_SIZE, EPOCHS 2 from subclassed_model import * 3 from loss_function import loss 4 5 import os 6 import numpy as np 7 import matplotlib.pyplot as plt 8 9 10 # load training data 11 training_X = np.fromfile(os.path.join(TRAINING_DATA_DIR, 'training_data_X.bin'), dtype=np.float64) 12 training_y = np.fromfile(os.path.join(TRAINING_DATA_DIR, 'training_data_y.bin'), dtype=np.float64) 13 14 # plot training data 15 plt.scatter(training_X, training_y) 16 plt.show() 17 18 # training dataset 19 training_dataset = tf.data.Dataset.from_tensor_slices((training_X, training_y)).batch(BATCH_SIZE).shuffle(BUFFER_SIZE) 20 21 # model instance 22 encoder = Encoder() 23 decoder = Decoder() 24 25 # optimizer 26 optimizer = tf.train.AdamOptimizer() 27 28 # checkpoint 29 checkpoint_prefix_encoder = os.path.join(CHECKPOINT_DIR, 'encoder/', 'ckpt') 30 checkpoint_prefix_decoder = os.path.join(CHECKPOINT_DIR, 'decoder/', 'ckpt') 31 32 if not os.path.exists(os.path.dirname(checkpoint_prefix_encoder)): 33 os.makedirs(os.path.dirname(checkpoint_prefix_encoder)) 34 if not os.path.exists(os.path.dirname(checkpoint_prefix_decoder)): 35 os.makedirs(os.path.dirname(checkpoint_prefix_decoder)) 36 37 # training step 38 for epoch in range(EPOCHS): 39 epoch_loss = 0 40 41 for (batch, (tx, ty)) in enumerate(training_dataset): 42 x = tf.cast(tx, tf.float32) 43 y = tf.cast(ty, tf.float32) 44 x = tf.expand_dims(x, axis=1) # tf.Tensor([...], shape=(BATCH_SIZE, 1), dtype=float32) 45 y = tf.expand_dims(y, axis=1) # tf.Tensor([...], shape=(BATCH_SIZE, 1), dtype=float32) 46 47 with tf.GradientTape() as tape: 48 y_ = encoder(x) # tf.Tensor([...], shape=(BATCH_SIZE, 8), dtype=float32) 49 prediction = decoder(y_) # tf.Tensor([...], shape=(BATCH_SIZE, 1), dtype=float32) 50 batch_loss = loss(real=y, pred=prediction) 51 52 variables = encoder.variables + decoder.variables 53 grads = tape.gradient(batch_loss, variables) 54 optimizer.apply_gradients(zip(grads, variables), global_step=tf.train.get_or_create_global_step()) 55 56 epoch_loss += batch_loss 57 58 if (batch + 1) % 100 == 0: 59 print('Epoch {} Batch {} Loss {:.4f}'.format(epoch + 1, 60 batch + 1, 61 batch_loss.numpy())) 62 63 print('Epoch {} Loss {:.4f}'.format(epoch + 1, 64 epoch_loss / len(training_X))) 65 66 if (epoch + 1) % 5 == 0: 67 encoder.save_weights(checkpoint_prefix_encoder) 68 decoder.save_weights(checkpoint_prefix_decoder) 69 70 # create evaluation data 71 X = np.linspace(-1, 1, 3000) 72 np.random.shuffle(X) 73 74 evaluation_X = tf.data.Dataset.from_tensor_slices(X).batch(BATCH_SIZE) 75 ey = [] 76 77 for (batch, ex) in enumerate(evaluation_X): 78 x = tf.cast(ex, tf.float32) 79 x = tf.expand_dims(x, axis=1) 80 prediction = decoder(encoder(x)) 81 for i in range(len(prediction.numpy())): 82 ey.append(prediction.numpy()[i]) 83 84 plt.scatter(X, ey) 85 plt.show() 86 87 # evaluate 88 eval_x = [[0.5]] 89 tensor_x = tf.convert_to_tensor(eval_x) 90 print(decoder(encoder(tensor_x)))
验证集评价结果如下图所示。
使用测试样例 eval_x 进行测试，测试结果如下。
tf.Tensor([[1.122567]], shape=(1, 1), dtype=float32)
在 evaluate.py 中，使用 model.load_weights() 恢复 Keras Subclassed Model 模型，并在验证集上进行验证，验证结果如下图所示。
1 from parameters import CHECKPOINT_DIR, BATCH_SIZE 2 from subclassed_model import * 3 4 import os 5 import numpy as np 6 import matplotlib.pyplot as plt 7 8 9 # load model 10 enc = Encoder() 11 dec = Decoder() 12 13 enc.load_weights(tf.train.latest_checkpoint(os.path.join(CHECKPOINT_DIR, 'encoder/'))) 14 dec.load_weights(tf.train.latest_checkpoint(os.path.join(CHECKPOINT_DIR, 'decoder/'))) 15 16 # create evaluation data 17 X = np.linspace(-1, 1, 3000) 18 np.random.shuffle(X) 19 20 evaluation_X = tf.data.Dataset.from_tensor_slices(X).batch(BATCH_SIZE) 21 ey = [] 22 23 for (batch, ex) in enumerate(evaluation_X): 24 x = tf.cast(ex, tf.float32) 25 x = tf.expand_dims(x, axis=1) 26 prediction = dec(enc(x)) 27 for i in range(len(prediction.numpy())): 28 ey.append(prediction.numpy()[i]) 29 30 plt.scatter(X, ey) 31 plt.show() 32 33 # evaluate 34 eval_x = [[0.5]] 35 tensor_x = tf.convert_to_tensor(eval_x) 36 print(dec(enc(tensor_x))) 37 38 # model summary 39 enc.summary() 40 dec.summary()
使用测试样例 eval_x 进行测试，测试结果如下。
tf.Tensor([[1.122567]], shape=(1, 1), dtype=float32)
恢复模型的测试结果，与训练后模型的测试结果一致，且无需 build 模型。
版权声明：本文为博主原创文章，欢迎转载，转载请注明作者及原文出处！
