python-3.x - Keras 模型似乎没有加载重量？
问题描述
我可以让模型有效地训练。但是，我现在在加载模型并再次在新数据上测试它时遇到了问题。
通过在保存新数据并对其进行预测之前将新数据加载到模型中，我能够验证我的模型本身在新数据上的表现是否良好。（参见代码示例）
我确实尝试遵循 Keras 关于如何正确加载和保存模型的文档（这似乎很容易），但它似乎对我来说并没有正常工作。任何帮助深表感谢。
下面是我用于训练和测试模型以及保存和加载模型的代码。
def Build_Model_CNN_Text(word_index, embeddings_index, nclasses,
MAX_SEQUENCE_LENGTH=500, EMBEDDING_DIM=50,
dropout=0.5):
"""
Function to build a Convolutional Neural Network (CNN). Using the
rectified linear unit (ReLU) activation function, along with the softmax
activation function at the end and sparse_categorical_crossentropy loss with
the adam optimizer. The original of this code came from Swayam Mittal's
Medium.com aritcle "Deep Learning Techniques for Text Classification."
https://medium.com/datadriveninvestor/deep-learning-techniques-for-text-
classification-9392ca9492c7
Parameters
----------
word_index : dict
The word index that was created from the tokenizer fucntion.
embeddings_index : dict
The embedded index of the word.
nclasses : int
The number of classes that are provided to the model.
MAX_SEQUENCE_LENGTH : int
Max length of each sequence, by default 500.
EMBEDDING_DIM : int
Dimension for word embedding, by default 50.
dropout : float, optional
Used to help prevent overfitting of the model, by default 0.5.
Returns
-------
Model
A untrained CNN model that is ready to fit data.
"""
model = Sequential()
embedding_matrix = np.random.random((len(word_index) + 1, EMBEDDING_DIM))
for word, i in word_index.items():
embedding_vector = embeddings_index.get(word)
if embedding_vector is not None:
# words not found in embedding index will be all-zeros.
if len(embedding_matrix[i]) != len(embedding_vector):
print("could not broadcast input array from shape",
str(len(embedding_matrix[i])),
"into shape", str(len(embedding_vector)),
" Please make sure your"
" EMBEDDING_DIM is equal to embedding_vector file,GloVe,")
exit(1)
embedding_matrix[i] = embedding_vector
embedding_layer = Embedding(len(word_index) + 1,
EMBEDDING_DIM,
weights=[embedding_matrix],
input_length=MAX_SEQUENCE_LENGTH,
trainable=True)
# applying a more complex convolutional approach
convs = []
filter_sizes = []
layer = 5
print("Filter ", layer)
for fl in range(0, layer):
filter_sizes.append((fl + 2))
node = 128
sequence_input = Input(shape=(MAX_SEQUENCE_LENGTH,), dtype='int32')
embedded_sequences = embedding_layer(sequence_input)
for fsz in filter_sizes:
l_conv = Conv1D(node, kernel_size=fsz, activation='relu')(
embedded_sequences)
l_pool = MaxPooling1D(5)(l_conv)
convs.append(l_pool)
l_merge = Concatenate(axis=1)(convs)
l_cov1 = Conv1D(node, 5, activation='relu')(l_merge)
l_cov1 = Dropout(dropout)(l_cov1)
l_pool1 = MaxPooling1D(5)(l_cov1)
l_cov2 = Conv1D(node, 5, activation='relu')(l_pool1)
l_cov2 = Dropout(dropout)(l_cov2)
l_pool2 = MaxPooling1D(30)(l_cov2)
l_flat = Flatten()(l_pool2)
l_dense = Dense(1024, activation='relu')(l_flat)
l_dense = Dropout(dropout)(l_dense)
l_dense = Dense(512, activation='relu')(l_dense)
l_dense = Dropout(dropout)(l_dense)
preds = Dense(nclasses, activation='softmax')(l_dense)
model = Model(sequence_input, preds)
model.compile(loss='sparse_categorical_crossentropy',
optimizer='adam',
metrics=['acc'])
return model
def use_cnn_model():
"""
Function used to train the CNN model. It starts by gathering the dataset
by using the import_dataset function from the make_dataset package. The
determine_if_spam function is then called on the dataframe and loaded into
the 'data' variable. The data is then split and tokenized and the CNN model
is built.
The model is then fit with the data and number of epochs, batch_size,
and verbosity level, returning a trained model. The trained model is then
used to predict on the test data and the output matix is printed to the
terminal. The model is then saved in an .h5 format.
"""
dirty_data = import_dataset()
data = determine_if_spam(dirty_data)
data = data.dropna(subset=['body'])
split_data = split_dataset(data)
X_train = split_data[0]
X_test = split_data[1]
y_train = split_data[2]
y_test = split_data[3]
X_train_Glove, X_test_Glove, word_index, embeddings_index = loadData_Tokenizer(
X_train, X_test)
model_CNN = Build_Model_CNN_Text(word_index, embeddings_index, 20)
model_CNN.summary()
model_CNN.fit(X_train_Glove, y_train,
validation_data=(X_test_Glove, y_test),
epochs=10,
batch_size=128,
verbose=2)
model_CNN.save('saved_model.h5')
predicted = model_CNN.predict(X_test_Glove)
predicted = np.argmax(predicted, axis=1)
print(metrics.classification_report(y_test, predicted))
new_dirty_data = import_dataset()
new_data = determine_if_spam(new_dirty_data)
new_data = data.dropna(subset=['body'])
new_split_data = split_dataset(new_data)
new_X_train = new_split_data[0]
new_X_test = new_split_data[1]
new_y_train = new_split_data[2]
new_y_test = new_split_data[3]
new_X_train_Glove, new_X_test_Glove, new_word_index, \
new_embeddings = loadData_Tokenizer(new_X_train, new_X_test)
new_predicted = model_CNN.predict(new_X_test_Glove)
new_predicted = np.argmax(new_predicted, axis=1)
print("New data prediction: \n)
print(metrics.classification_report(new_y_test, new_predicted))
def loaded_model():
"""
Function to load the trained model to use on the new data.
Returns
-------
Model
The trained model that is saved in the saved_model.h5 file.
"""
new_model = tf.keras.models.load_model('saved_model.h5')
return new_model
def main():
"""
Main function that is used to call all other functions.
"""
model = loaded_model()
dirty_data = import_dataset()
data = determine_if_spam(dirty_data)
data = data.dropna(subset=['body'])
split_data = split_dataset(data)
X_train = split_data[0]
X_test = split_data[1]
y_train = split_data[2]
y_test = split_data[3]
X_train_Glove, X_test_Glove, word_index, \
embeddings_index = loadData_Tokenizer(X_train, X_test)
loss, acc = model.evaluate(X_test_Glove, y_test)
print('Restored model, accuracy: {:5.2f}%'.format(100 * acc))
predicted = model.predict(X_test_Glove)
predicted = np.argmax(predicted, axis=1)
print(metrics.classification_report(y_test, predicted))
print(metrics.confusion_matrix(y_test, predicted))
print("Accuracy Score: ", metrics.accuracy_score(y_test,predicted))
这是调用 use_cnn_model() 函数的结果
Trained Data
precision recall f1-score support
0.0 0.99 1.00 0.99 851
1.0 0.75 0.77 0.76 110
2.0 0.33 0.23 0.27 48
3.0 0.90 0.96 0.93 303
4.0 0.99 0.93 0.96 306
5.0 0.99 0.98 0.99 325
6.0 1.00 1.00 1.00 584
7.0 0.98 0.98 0.98 246
accuracy 0.96 2773
macro avg 0.87 0.86 0.86 2773
weighted avg 0.96 0.96 0.96 2773
New data prediction
precision recall f1-score support
0.0 0.99 1.00 0.99 832
1.0 0.78 0.72 0.75 115
2.0 0.37 0.33 0.35 49
3.0 0.87 0.96 0.91 283
4.0 0.98 0.92 0.95 337
5.0 0.98 0.97 0.98 333
6.0 1.00 1.00 1.00 572
7.0 0.98 0.97 0.97 252
accuracy 0.96 2773
macro avg 0.87 0.86 0.86 2773
weighted avg 0.96 0.96 0.96 2773
这是加载函数后我的 main() 函数的输出。
Restored model, accuracy: 34.11%
precision recall f1-score support
0.0 1.00 0.81 0.89 3628
1.0 0.01 0.01 0.01 652
2.0 0.01 0.06 0.02 205
3.0 0.00 0.00 0.00 1375
4.0 0.25 0.50 0.33 1387
5.0 0.21 0.37 0.27 1177
6.0 0.21 0.02 0.04 2190
7.0 0.04 0.04 0.04 1661
accuracy 0.34 12275
macro avg 0.22 0.23 0.20 12275
weighted avg 0.39 0.34 0.34 12275
[[2926 0 0 702 0 0 0 0]
[ 0 8 9 0 130 50 50 405]
[ 0 55 12 138 0 0 0 0]
[ 0 14 24 4 117 1109 10 97]
[ 0 185 264 92 687 14 13 132]
[ 0 1 1 2 154 431 113 475]
[ 0 79 87 8 1081 444 51 440]
[ 0 276 556 103 627 26 5 68]]
Accuracy Score: 0.3410997963340122
感谢您对此提供的任何帮助，我知道这可能不是最容易遵循的。希望，老实说，我只是错过了一些简单的事情。
解决方案
推荐阅读
- azure - 如何使用 azure devOps 在 azure app 服务上部署 nupkg 文件
- angular - 是否有任何替代方法可以在角度的所有元素上编写点击功能
- java - Picasso“目标不能为空”（使用 Firebase 数据库）
- python - 如何从多个模块 python 日志记录
- regex - Perl 正则表达式仅匹配 Diggernaut 中的第一次出现
- r - 线性回归模拟
- security - 使用 JWT 保护 SPA 的最佳实践
- javascript - 使用 ApiGateWay 创建会话和重定向
- c++ - 命名空间或类只有一个成员？
- c++ - 在派生类的基类构造函数中使用覆盖函数