我正在训练一个ResNet模型来分类汽车品牌。
我在每次训练周期保存了模型的权重。
作为测试，我在第三个周期停止了训练。
# checkpoint = ModelCheckpoint("best_model.hdf5", monitor='loss', verbose=1)checkpoint_path = "weights/cp-{epoch:04d}.ckpt"checkpoint_dir = os.path.dirname(checkpoint_path)cp_callback = tf.keras.callbacks.ModelCheckpoint( checkpoint_path, verbose=1, # Save weights, every epoch. save_freq='epoch')model.save_weights(checkpoint_path.format(epoch=0))history = model.fit_generator( training_set, validation_data = test_set, epochs = 50, steps_per_epoch = len(training_set), validation_steps = len(test_set), callbacks = [cp_callback])然而，当我加载这些权重时，我不确定是否从最后保存的周期恢复，因为它显示的仍然是周期1/50。以下是我用来加载最后保存的模型的代码。
from keras.models import Sequential, load_model# load the modelnew_model = load_model('./weights/cp-0003.ckpt')# fit the modelhistory = new_model.fit_generator( training_set, validation_data = test_set, epochs = 50, steps_per_epoch = len(training_set), validation_steps = len(test_set), callbacks = [cp_callback])看起来是这样的：图像显示运行保存的权重从周期1/50重新开始
请问有人可以帮忙吗？
回答：
你可以使用 fit_generator 的 initial_epoch 参数。默认情况下，它被设置为0，但你可以将其设置为任何正数：
from keras.models import Sequential, load_modelimport tensorflow as tfcheckpoint_path = "weights/cp-{epoch:04d}.ckpt"checkpoint_dir = os.path.dirname(checkpoint_path)cp_callback = tf.keras.callbacks.ModelCheckpoint( checkpoint_path, verbose=1, # Save weights, every epoch. save_freq='epoch')model.save_weights(checkpoint_path.format(epoch=0))history = model.fit_generator( training_set, validation_data=test_set, epochs=3, steps_per_epoch=len(training_set), validation_steps=len(test_set), callbacks = [cp_callback])new_model = load_model('./weights/cp-0003.ckpt')# fit the modelhistory = new_model.fit_generator( training_set, validation_data=test_set, epochs=50, steps_per_epoch=len(training_set), validation_steps=len(test_set), callbacks=[cp_callback], initial_epoch=3)这将使你的模型额外训练50 – 3 = 47个周期。
关于你的代码，如果你使用Tensorflow 2.X，以下是一些备注：
fit_generator自从fit支持生成器后已被废弃- 你应该将你的导入从
from keras....替换为from tensorflow.keras...
