Google Colab keras Sequential 模型返回一个空的 history.history 对象“”
Posted
技术标签:
【中文标题】Google Colab keras Sequential 模型返回一个空的 history.history 对象“”【英文标题】：Google Colab keras Sequential model returns an empty history.history object ""Google Colab keras Sequential 模型返回一个空的 history.history 对象“” 【发布时间】：2020-10-28 13:44:03 【问题描述】：我在使用 google colab 和我训练的 Keras 模型时遇到了困难。我正在使用各种魔术技巧，例如 Tensorboard、HParams、回调等。
最初，请原谅我只发布代码和屏幕截图，而不是数据。由于发布的代码在我的本地机器上成功运行，而不是在 colab 上，我猜这不是由于数据错误，而是代码错误。
我的本地机器TF版本：2.1.0
Colab 的 TF 版本：2.2.0
下面是我的代码。
库
import tensorflow as tf
import tensorflow_addons as tfa
import tensorflow_docs as tfdocs #!pip install git+https://github.com/tensorflow/docs
import tensorflow_docs.plots as tfplots
import tensorflow_docs.modeling as tfmodel
from tensorflow.keras import layers, regularizers, models
from tensorflow.keras import models
from tensorflow.keras.callbacks import ModelCheckpoint, EarlyStopping
from tensorflow.keras.utils import model_to_dot, plot_model
from tensorflow.keras.models import load_model, model_from_json
%load_ext tensorboard
from tensorboard.plugins.hparams import api as hp
import keras.backend as K
from tensorflow import keras
回调
def callback(folder_path, saved_model_name, logdir, hparams):
# Initialize parameters
monitor_metric = 'val_loss'
minimum_delta = 0.1
patience_limit = 1
verbose_value = 1
mode_value = 'min'
weights_fname = os.path.join(os.getcwd(), '0/1.h5'.format(folder_path, saved_model_name))
print(weights_fname)
# Initialize callbacks
callbacks = [
EarlyStopping(monitor=monitor_metric,
min_delta=minimum_delta,
patience=patience_limit,
verbose=verbose_value,
mode=mode_value,
restore_best_weights=True),
ModelCheckpoint(filepath=weights_fname,
monitor=monitor_metric,
verbose=verbose_value,
save_best_only=True,
save_weights_only=True),
tf.keras.callbacks.TensorBoard(logdir),
hp.KerasCallback(logdir, hparams)
]
return callbacks
超参数
HP_HIDDEN_UNITS = hp.HParam('batch_size', hp.Discrete([32, 64, 128]))
HP_EMBEDDING_DIM = hp.HParam('embedding_dim', hp.Discrete([100, 200, 300]))
HP_LEARNING_RATE = hp.HParam('learning_rate', hp.Discrete([0.001, 0.01, 0.1])) # Adam default: 0.001, SGD default: 0.01, RMSprop default: 0.001
HP_DECAY_STEPS_MULTIPLIER = hp.HParam('decay_steps_multiplier', hp.Discrete([100, 1000]))
METRIC_ACCURACY = 'accuracy'
创建和拟合模型函数
def create_fit_keras_model(hparams,
version_data_control,
optimizer_name,
validation_method,
callbacks,
folder_path,
optimizer_version = None):
sentenceLength_actors = X_train_seq_actors.shape[1]
vocab_size_frequent_words_actors = len(actors_tokenizer.word_index)
sentenceLength_plot = X_train_seq_plot.shape[1]
vocab_size_frequent_words_plot = len(plot_tokenizer.word_index)
sentenceLength_features = X_train_seq_features.shape[1]
vocab_size_frequent_words_features = len(features_tokenizer.word_index)
sentenceLength_reviews = X_train_seq_reviews.shape[1]
vocab_size_frequent_words_reviews = len(reviews_tokenizer.word_index)
model = keras.Sequential(name='MultyInput_Keras_Classification_model_0dim_1batchsize_2lr_3decaymultiplier_4'.format(hparams[HP_EMBEDDING_DIM], hparams[HP_HIDDEN_UNITS],
hparams[HP_LEARNING_RATE], hparams[HP_DECAY_STEPS_MULTIPLIER],
version_data_control))
actors = keras.Input(shape=(sentenceLength_actors,), name='actors_input')
plot = keras.Input(shape=(sentenceLength_plot,), name='plot_input')
features = keras.Input(shape=(sentenceLength_features,), name='features_input')
reviews = keras.Input(shape=(sentenceLength_reviews,), name='reviews_input')
emb1 = layers.Embedding(input_dim = vocab_size_frequent_words_actors + 2,
output_dim = hparams[HP_EMBEDDING_DIM],
embeddings_initializer = 'uniform',
mask_zero = True,
input_length = sentenceLength_actors,
name="actors_embedding_layer")(actors)
encoded_layer1 = layers.GlobalMaxPooling1D(name="globalmaxpooling_actors_layer")(emb1)
emb2 = layers.Embedding(input_dim = vocab_size_frequent_words_plot + 2,
output_dim = hparams[HP_EMBEDDING_DIM],
embeddings_initializer = 'uniform',
mask_zero = True,
input_length = sentenceLength_plot,
name="plot_embedding_layer")(plot)
encoded_layer2 = layers.GlobalMaxPooling1D(name="globalmaxpooling_plot_summary_Layer")(emb2)
emb3 = layers.Embedding(input_dim = vocab_size_frequent_words_features + 2,
output_dim = hparams[HP_EMBEDDING_DIM],
embeddings_initializer = 'uniform',
mask_zero = True,
input_length = sentenceLength_features,
name="features_embedding_layer")(features)
encoded_layer3 = layers.GlobalMaxPooling1D(name="globalmaxpooling_movie_features_layer")(emb3)
emb4 = layers.Embedding(input_dim = vocab_size_frequent_words_reviews + 2,
output_dim = hparams[HP_EMBEDDING_DIM],
embeddings_initializer = 'uniform',
mask_zero = True,
input_length = sentenceLength_reviews,
name="reviews_embedding_layer")(reviews)
encoded_layer4 = layers.GlobalMaxPooling1D(name="globalmaxpooling_user_reviews_layer")(emb4)
merged = layers.concatenate([encoded_layer1, encoded_layer2, encoded_layer3, encoded_layer4], axis=-1)
dense_layer_1 = layers.Dense(hparams[HP_HIDDEN_UNITS],
kernel_regularizer=regularizers.l2(neural_network_parameters['l2_regularization']),
activation=neural_network_parameters['dense_activation'],
name="1st_dense_hidden_layer_concatenated_inputs")(merged)
layers.Dropout(neural_network_parameters['dropout_rate'])(dense_layer_1)
output_layer = layers.Dense(y_train.shape[1],
activation=neural_network_parameters['output_activation'],
name='output_layer')(dense_layer_1)
model = keras.Model(inputs=[actors, plot, features, reviews], outputs=output_layer, name='MultyInput_Keras_Classification_model_0dim_1batchsize_2lr_3decaymultiplier_4'.format(hparams[HP_EMBEDDING_DIM],
hparams[HP_HIDDEN_UNITS],
hparams[HP_LEARNING_RATE],
hparams[HP_DECAY_STEPS_MULTIPLIER],
version_data_control))
print(model.summary())
if optimizer_name=="adam" and optimizer_version is None:
optimizer = optimizer_adam_v2(hparams[HP_LEARNING_RATE], hparams[HP_DECAY_STEPS_MULTIPLIER], X_train_seq_actors.shape[0], optimizer_parameters['validation_split_ratio'], hparams[HP_HIDDEN_UNITS])
elif optimizer_name=="sgd" and optimizer_version is None:
optimizer = optimizer_sgd_v1(hparams[HP_LEARNING_RATE])
elif optimizer_name=="rmsprop" and optimizer_version is None:
optimizer = optimizer_rmsprop_v1(hparams[HP_LEARNING_RATE])
model.compile(optimizer=optimizer,
loss=neural_network_parameters['model_loss'],
metrics=[neural_network_parameters['model_metric']])
plot_model(model, to_file=os.path.join(os.getcwd(), 'model_one/network_structure_multy_input_keras_model_0.png'.format(version_data_control)))
start_time = time.time()
if validation_method=="validation_split":
model.fit([X_train_seq_actors, X_train_seq_plot, X_train_seq_features, X_train_seq_reviews],
y_train,
steps_per_epoch=int(np.ceil((X_train_seq_actors.shape[0]*optimizer_parameters['validation_split_ratio'])//hparams[HP_HIDDEN_UNITS])),
epochs=fit_parameters["epoch"],
verbose=fit_parameters["verbose_fit"],
batch_size=hparams[HP_HIDDEN_UNITS],
validation_split=fit_parameters['validation_data_ratio'],
callbacks=callbacks)
elif validation_method=="validation_data":
model.fit([X_train_seq_actors, X_train_seq_plot, X_train_seq_features, X_train_seq_reviews],
y_train,
steps_per_epoch=int(np.ceil((X_train_seq_actors.shape[0]*optimizer_parameters['validation_split_ratio'])//hparams[HP_HIDDEN_UNITS])),
epochs=fit_parameters["epoch"],
verbose=fit_parameters["verbose_fit"],
batch_size=hparams[HP_HIDDEN_UNITS],
validation_data=([X_test_seq_actors, X_test_seq_plot, X_test_seq_features, X_test_seq_reviews], y_test),
callbacks=callbacks)
#save the model
save_model(model, folder_path, "multi_input_keras_model_0dim_1batchsize_2lr_3decaymultiplier_4".format(str(hparams[HP_EMBEDDING_DIM]), str(hparams[HP_HIDDEN_UNITS]), str(hparams[HP_LEARNING_RATE]), str(hparams[HP_DECAY_STEPS_MULTIPLIER]), version_data_control))
elapsed_time = time.time() - start_time
print("\nTraining time of the multi-input keras model has finished. Duration secs".format(format_timespan(elapsed_time)))
_, accuracy = model.evaluate([X_test_seq_actors, X_test_seq_plot, X_test_seq_features, X_test_seq_reviews], y_test, batch_size=hparams[HP_HIDDEN_UNITS], verbose=2)
return accuracy, model
运行一切 - 函数 -> 基本上这是我调用来训练我的模型的函数。
def run(run_dir, hparams, version_data_control, optimizer_name, validation_method, callbacks, folder_path):
with tf.summary.create_file_writer(run_dir).as_default():
hp.hparams(hparams) # record the values used in this trial
accuracy, model = create_fit_keras_model(hparams, version_data_control, optimizer_name, validation_method, callbacks, folder_path)
print(model.history.history)
tf.summary.scalar(METRIC_ACCURACY, accuracy, step=1)
return model
训练过程 -> 在我的本地机器上成功运行，但在 Google Colab 中却没有。
session_num = 1
for batch_size in HP_HIDDEN_UNITS.domain.values:
for embedding_dim in HP_EMBEDDING_DIM.domain.values:
for learning_rate in HP_LEARNING_RATE.domain.values:
for decay_steps_multiplier in HP_DECAY_STEPS_MULTIPLIER.domain.values:
hparams =
HP_HIDDEN_UNITS: batch_size,
HP_EMBEDDING_DIM: embedding_dim,
HP_LEARNING_RATE: learning_rate,
HP_DECAY_STEPS_MULTIPLIER: decay_steps_multiplier
run_name = "run-id 0/1".format(session_num, (len(HP_HIDDEN_UNITS.domain.values)*len(HP_EMBEDDING_DIM.domain.values)*len(HP_LEARNING_RATE.domain.values)*len(HP_DECAY_STEPS_MULTIPLIER.domain.values)))
print('--- Starting trial: %s/n' % run_name)
print(h.name: hparams[h] for h in hparams, "/n")
model_history=run('./logs/hparam_tuning/' + run_name, hparams, version_data_control, "adam", "validation_split",
callback("model_one/adam_v2_07072020",
"multi_input_keras_model_0dim_1batchsize_2lr_3decaymultiplier_4".format(str(embedding_dim), str(batch_size), str(learning_rate), str(decay_steps_multiplier), version_data_control),
"./logs/hparam_tuning/"+datetime.now().strftime("%Y%m%d-%H%M%S"),
hparams),
"model_one/adam_v2_07072020")
hist = pd.DataFrame(model_history.history.history)
hist['epoch'] = model_history.history.epoch
根据下面的屏幕截图，一切运行良好。虽然从 run() 函数返回的历史模型是完全空的。此外，即使日志位于 logdir 目录中，Tensorboard 上也不会打印任何内容。另外，我想补充一点，上面发布的代码在我的本地机器上完美运行，包括 Tensorboard 和回调，以及模型历史。但由于缺乏资源，我想在 colab 上运行 100 个 epoch，而不是像本例（演示）中那样运行 1 个。
我得到的错误
我的文件
提前感谢您的任何 cmets 和帮助。如果您可能想查看包含数据的完整代码，我可以与您分享我的 colab 笔记本。只需在 cmets 中发布您的 Gmail 帐户，我会与您分享！
【问题讨论】：
【参考方案1】：我通过将 model.fit() 方法分配给不同的对象解决了这个问题，如下所示：
fitted_model = model.fit([X_train_seq_actors, X_train_seq_plot, X_train_seq_features, X_train_seq_reviews],
y_train,
steps_per_epoch=int(np.ceil((X_train_seq_actors.shape[0]*optimizer_parameters['validation_split_ratio'])//hparams[HP_HIDDEN_UNITS])),
epochs=fit_parameters["epoch"],
verbose=fit_parameters["verbose_fit"],
batch_size=hparams[HP_HIDDEN_UNITS],
validation_split=fit_parameters['validation_data_ratio'],
callbacks=callbacks)
然后通过返回fitted_model 对象我可以成功运行这个hist = pd.DataFrame(model_history.history)
输出：
+----+----------+------------+------------+----------------+---------+
| | loss | accuracy | val_loss | val_accuracy | epoch |
|----+----------+------------+------------+----------------+---------|
| 1 | 0.295619 | 0.452375 | 0.186601 | 0.64396 | 1 |
+----+----------+------------+------------+----------------+---------+
希望这也能帮助其他用户。如果在同一问题上给出了类似的答案，请随意将其分配为 [重复] 问题。
【讨论】：
这是一个非常奇怪的。今天当我打电话给model.history.history返回空字典时，我突然遇到了类似的问题。所以我需要像你建议的那样分配单独的变量，它可以工作。
@kitokid 很高兴我帮助了你:)以上是关于Google Colab keras Sequential 模型返回一个空的 history.history 对象“”的主要内容，如果未能解决你的问题，请参考以下文章
使用 Keras 在 Google colab 上运行 3D CNN 的问题
Google Colab keras Sequential 模型返回一个空的 history.history 对象“”
使用 TPU 运行时在 Google Colab 上训练 Keras 模型时出错
Google Colab 上的 Keras YoloV3，AttributeError: module 'keras.backend' has no attribute 'control_flow_o