无法对接收多个输入的模型执行网格搜索 (Keras)
Unable to perform Grid Search for models receiving more than one input (Keras)
我创建了 final_model 来接收两个输入（长度为 8 的序列）。它们中的每一个都由两个不同的模型处理，model_A 和 model_B。然后合并两个输出，返回 model_C 的输入，最终 returns 整个模型的输出。
这是图形概览
这是代码：
model_A = models.Sequential()
model_A.add(layers.Dense(16, activation='relu',input_shape=(n,)))
model_A.add(layers.Dense(3))
model_B = models.Sequential()
model_B.add(layers.Dense(16, activation='relu',input_shape=(n,)))
model_B.add(layers.Dense(3))
concatenated = layers.concatenate([model_A.output, model_B.output])
model_C = layers.Dense(16, activation='relu')(concatenated)
out = layers.Dense(3, activation='softmax')(model_C)
final_model = models.Model([model_A.input, model_B.input], out)
当我适合我的模型时一切正常：
opt = keras.optimizers.Adam(learning_rate=0.001)
final_model.compile(optimizer=opt,loss='categorical_crossentropy',metrics ['accuracy'])
history = final_model.fit([X_train,x_train], y_train, epochs=500, batch_size=1000)
但是，我无法使用网格搜索来优化超参数。的确，用下面的代码
batch_size = [10, 20]
epochs = [10, 50]
param_grid = dict(batch_size=batch_size, epochs=epochs)
grid = GridSearchCV(estimator=final_model, param_grid=param_grid, n_jobs=-1,
cv=3,scoring="accuracy")
grid_result = grid.fit([X_train,x_train], y_train)
我收到这个错误：
Found input variables with inconsistent numbers of samples: [2, 40000]
注意X_train和x_train的形状都是[40000,8]。
有没有办法在多输入的情况下使用网格搜索？
- 这可能会回答您的问题。
另一方面，还有其他特殊用途的超参数搜索库，例如ray。
请注意，@Sean 链接的答案中提到的 tf.keras.wrappers.scikit_learn.KerasClassifier 现在已弃用。目前的做法是 using scikeras. The models are wrapped in using subclasses of BaseWrapper (more details here).
这是一个玩具示例，用于演示如何使用您的模型执行此操作：
import numpy as np
from tensorflow.keras import models, layers, Input
from tensorflow import keras
from sklearn.base import BaseEstimator
from sklearn.preprocessing import FunctionTransformer
from sklearn.metrics import accuracy_score
from sklearn.model_selection import GridSearchCV
from scikeras.wrappers import KerasClassifier, BaseWrapper
# Splits X and returns the two inputs.
def split_input(X):
# Get back 2x(40000, 8).
return [X[:, :int(np.shape(X)[1]/2)], X[:, int(np.shape(X)[1]/2):]]
# Wrapper estimator that will call split_input via the FunctionTransformer.
# This will split the input before feeding them to the model.
# Instead of BaseWrapper, you can also subclass scikeras.wrappers.KerasClassifier (see documentation for differences).
class MultiInputEstimator(BaseWrapper):
@property
def feature_encoder(self):
return FunctionTransformer(func=split_input)
# Score passed to the grid search.
@staticmethod
def scorer(y_true, y_pred, **kwargs):
return accuracy_score(np.argmax(y_true, axis=1), np.argmax(y_pred, axis=1))
# Should return the fully compiled model. Pass eventual parameters here.
def get_model(input_shape, n_dense_1):
model_A = models.Sequential()
model_A.add(layers.Dense(n_dense_1, activation='relu',input_shape=input_shape))
model_A.add(layers.Dense(3))
model_B = models.Sequential()
model_B.add(layers.Dense(n_dense_1, activation='relu',input_shape=input_shape))
model_B.add(layers.Dense(3))
concatenated = layers.concatenate([model_A.output, model_B.output])
model_C = layers.Dense(16, activation='relu')(concatenated)
out = layers.Dense(3, activation='softmax')(model_C)
final_model = models.Model([model_A.input, model_B.input], out)
opt = keras.optimizers.Adam(learning_rate=0.001)
final_model.compile(optimizer=opt,loss='categorical_crossentropy', metrics=['accuracy'])
return final_model
X1 = np.random.random((40000, 8))
X2 = np.random.random((40000, 8))
# Stick the two datasets together. `split_input` will take care of separating them.
X = np.hstack([X1, X2]) # (40000, 16) important: of course X.shape[0] should match y.shape[0].
y = np.zeros((40000, 3))
y[0:20000, 0] = 1
y[20000:30000, 1] = 1
y[30000:, 2] = 1
# get_model is called to return the fully compiled model which is wrapped in our MultiInputEstimator instance.
clf = MultiInputEstimator(model=get_model, model__input_shape=(int(X.shape[1]/2),), model__n_dense_1=16)
params = {'model__n_dense_1': [16, 32 ,128]}
grid = GridSearchCV(estimator=clf, param_grid=params, cv=5, verbose=True)
grid_res = grid.fit(X=X, y=y)
这里唯一的新元素是 MultiInputEstimator 包装器和 split_input 函数。这个想法是通过合并输入来欺骗网格搜索具有单个输入数据集。然后，我们在输入数据上使用 scikeras.wrappers.BaseWrapper 到 运行 和 FunctionTransformer，这将在将数据集提供给模型之前拆分数据集。
正在访问模型：
由于模型现在“包装”在 clf 中，我们可以在构建后通过 clf.model_ 访问它。当我们适应它时会发生这种情况，例如 clf.fit(X, y) 或初始化它 (clf.initialize(X, y))。如果我们传递 Keras Model 实例而不是像 get_model 这样的构建函数，那么模型将直接可用。绘制模型的示例：
from tensorflow import keras
clf.initialize(X, y)
keras.utils.plot_model(clf.model_, show_shapes=True)
试试这个解决方案，使用 keras.wrappers.scikit_learn 和 KerasClassifier，您只需要重塑数据：
import numpy as np
from tensorflow import keras
from tensorflow.keras import layers
import tensorflow as tf
models=tf.keras
num_classes = 10
input_shape = (28, 28, 1)
(x_train, y_train), (x_test, y_test) = keras.datasets.mnist.load_data()
x_train = x_train.astype("float32") / 255
x_test = x_test.astype("float32") / 255
x_train = np.expand_dims(x_train, -1).reshape(-1,28,28,)
x_test = np.expand_dims(x_test, -1).reshape(-1,28,28,)
X_train=x_train
y_train = keras.utils.to_categorical(y_train, num_classes)
y_test = keras.utils.to_categorical(y_test, num_classes)
def create_model():
model_A = models.Sequential()
model_A.add(layers.Dense(16, activation='relu',input_shape=(28,28,1)))
model_A.add(layers.Dense(10))
model_A.add(layers.Flatten())
model_B = models.Sequential()
model_B.add(layers.Dense(16, activation='relu',input_shape=(28,28,1)))
model_B.add(layers.Dense(10))
model_B.add(layers.Flatten())
concatenated = layers.concatenate([model_A.output, model_B.output])
model_C = layers.Dense(16, activation='relu')(concatenated)
out = layers.Dense(10, activation='softmax')(model_C)
final_model = models.Model([model_A.input, model_B.input], out)
final_model.compile(loss="categorical_crossentropy", optimizer=keras.optimizers.Adam(learning_rate=0.001), metrics=["accuracy"])
return final_model
#opt =
from keras.wrappers.scikit_learn import KerasClassifier
model = KerasClassifier(build_fn=create_model)
history = model.fit([X_train,x_train], y_train,epochs=500,batch_size=1000)
from sklearn.model_selection import GridSearchCV
batch_size = [300, 600]
epochs = [100, 200]
param_grid = dict(batch_size=batch_size, epochs=epochs)
grid = GridSearchCV(estimator=model, param_grid=param_grid, n_jobs=-1, cv=3,scoring="accuracy")
grid_result = grid.fit(np.array([X_train,x_train]).reshape(-1,28,28,2), y_train)
我创建了 final_model 来接收两个输入（长度为 8 的序列）。它们中的每一个都由两个不同的模型处理，model_A 和 model_B。然后合并两个输出，返回 model_C 的输入，最终 returns 整个模型的输出。
这是图形概览
这是代码：
model_A = models.Sequential()
model_A.add(layers.Dense(16, activation='relu',input_shape=(n,)))
model_A.add(layers.Dense(3))
model_B = models.Sequential()
model_B.add(layers.Dense(16, activation='relu',input_shape=(n,)))
model_B.add(layers.Dense(3))
concatenated = layers.concatenate([model_A.output, model_B.output])
model_C = layers.Dense(16, activation='relu')(concatenated)
out = layers.Dense(3, activation='softmax')(model_C)
final_model = models.Model([model_A.input, model_B.input], out)
当我适合我的模型时一切正常：
opt = keras.optimizers.Adam(learning_rate=0.001)
final_model.compile(optimizer=opt,loss='categorical_crossentropy',metrics ['accuracy'])
history = final_model.fit([X_train,x_train], y_train, epochs=500, batch_size=1000)
但是，我无法使用网格搜索来优化超参数。的确，用下面的代码
batch_size = [10, 20]
epochs = [10, 50]
param_grid = dict(batch_size=batch_size, epochs=epochs)
grid = GridSearchCV(estimator=final_model, param_grid=param_grid, n_jobs=-1,
cv=3,scoring="accuracy")
grid_result = grid.fit([X_train,x_train], y_train)
我收到这个错误：
Found input variables with inconsistent numbers of samples: [2, 40000]
注意X_train和x_train的形状都是[40000,8]。
有没有办法在多输入的情况下使用网格搜索？
另一方面，还有其他特殊用途的超参数搜索库，例如ray。
请注意，@Sean 链接的答案中提到的 tf.keras.wrappers.scikit_learn.KerasClassifier 现在已弃用。目前的做法是 using scikeras. The models are wrapped in using subclasses of BaseWrapper (more details here).
这是一个玩具示例，用于演示如何使用您的模型执行此操作：
import numpy as np
from tensorflow.keras import models, layers, Input
from tensorflow import keras
from sklearn.base import BaseEstimator
from sklearn.preprocessing import FunctionTransformer
from sklearn.metrics import accuracy_score
from sklearn.model_selection import GridSearchCV
from scikeras.wrappers import KerasClassifier, BaseWrapper
# Splits X and returns the two inputs.
def split_input(X):
# Get back 2x(40000, 8).
return [X[:, :int(np.shape(X)[1]/2)], X[:, int(np.shape(X)[1]/2):]]
# Wrapper estimator that will call split_input via the FunctionTransformer.
# This will split the input before feeding them to the model.
# Instead of BaseWrapper, you can also subclass scikeras.wrappers.KerasClassifier (see documentation for differences).
class MultiInputEstimator(BaseWrapper):
@property
def feature_encoder(self):
return FunctionTransformer(func=split_input)
# Score passed to the grid search.
@staticmethod
def scorer(y_true, y_pred, **kwargs):
return accuracy_score(np.argmax(y_true, axis=1), np.argmax(y_pred, axis=1))
# Should return the fully compiled model. Pass eventual parameters here.
def get_model(input_shape, n_dense_1):
model_A = models.Sequential()
model_A.add(layers.Dense(n_dense_1, activation='relu',input_shape=input_shape))
model_A.add(layers.Dense(3))
model_B = models.Sequential()
model_B.add(layers.Dense(n_dense_1, activation='relu',input_shape=input_shape))
model_B.add(layers.Dense(3))
concatenated = layers.concatenate([model_A.output, model_B.output])
model_C = layers.Dense(16, activation='relu')(concatenated)
out = layers.Dense(3, activation='softmax')(model_C)
final_model = models.Model([model_A.input, model_B.input], out)
opt = keras.optimizers.Adam(learning_rate=0.001)
final_model.compile(optimizer=opt,loss='categorical_crossentropy', metrics=['accuracy'])
return final_model
X1 = np.random.random((40000, 8))
X2 = np.random.random((40000, 8))
# Stick the two datasets together. `split_input` will take care of separating them.
X = np.hstack([X1, X2]) # (40000, 16) important: of course X.shape[0] should match y.shape[0].
y = np.zeros((40000, 3))
y[0:20000, 0] = 1
y[20000:30000, 1] = 1
y[30000:, 2] = 1
# get_model is called to return the fully compiled model which is wrapped in our MultiInputEstimator instance.
clf = MultiInputEstimator(model=get_model, model__input_shape=(int(X.shape[1]/2),), model__n_dense_1=16)
params = {'model__n_dense_1': [16, 32 ,128]}
grid = GridSearchCV(estimator=clf, param_grid=params, cv=5, verbose=True)
grid_res = grid.fit(X=X, y=y)
这里唯一的新元素是 MultiInputEstimator 包装器和 split_input 函数。这个想法是通过合并输入来欺骗网格搜索具有单个输入数据集。然后，我们在输入数据上使用 scikeras.wrappers.BaseWrapper 到 运行 和 FunctionTransformer，这将在将数据集提供给模型之前拆分数据集。
正在访问模型：
由于模型现在“包装”在 clf 中，我们可以在构建后通过 clf.model_ 访问它。当我们适应它时会发生这种情况，例如 clf.fit(X, y) 或初始化它 (clf.initialize(X, y))。如果我们传递 Keras Model 实例而不是像 get_model 这样的构建函数，那么模型将直接可用。绘制模型的示例：
from tensorflow import keras
clf.initialize(X, y)
keras.utils.plot_model(clf.model_, show_shapes=True)
试试这个解决方案，使用 keras.wrappers.scikit_learn 和 KerasClassifier，您只需要重塑数据：
import numpy as np
from tensorflow import keras
from tensorflow.keras import layers
import tensorflow as tf
models=tf.keras
num_classes = 10
input_shape = (28, 28, 1)
(x_train, y_train), (x_test, y_test) = keras.datasets.mnist.load_data()
x_train = x_train.astype("float32") / 255
x_test = x_test.astype("float32") / 255
x_train = np.expand_dims(x_train, -1).reshape(-1,28,28,)
x_test = np.expand_dims(x_test, -1).reshape(-1,28,28,)
X_train=x_train
y_train = keras.utils.to_categorical(y_train, num_classes)
y_test = keras.utils.to_categorical(y_test, num_classes)
def create_model():
model_A = models.Sequential()
model_A.add(layers.Dense(16, activation='relu',input_shape=(28,28,1)))
model_A.add(layers.Dense(10))
model_A.add(layers.Flatten())
model_B = models.Sequential()
model_B.add(layers.Dense(16, activation='relu',input_shape=(28,28,1)))
model_B.add(layers.Dense(10))
model_B.add(layers.Flatten())
concatenated = layers.concatenate([model_A.output, model_B.output])
model_C = layers.Dense(16, activation='relu')(concatenated)
out = layers.Dense(10, activation='softmax')(model_C)
final_model = models.Model([model_A.input, model_B.input], out)
final_model.compile(loss="categorical_crossentropy", optimizer=keras.optimizers.Adam(learning_rate=0.001), metrics=["accuracy"])
return final_model
#opt =
from keras.wrappers.scikit_learn import KerasClassifier
model = KerasClassifier(build_fn=create_model)
history = model.fit([X_train,x_train], y_train,epochs=500,batch_size=1000)
from sklearn.model_selection import GridSearchCV
batch_size = [300, 600]
epochs = [100, 200]
param_grid = dict(batch_size=batch_size, epochs=epochs)
grid = GridSearchCV(estimator=model, param_grid=param_grid, n_jobs=-1, cv=3,scoring="accuracy")
grid_result = grid.fit(np.array([X_train,x_train]).reshape(-1,28,28,2), y_train)