2

我已将我的 keras-tf-model 包装到 Sklearn 管道中，该管道还进行了一些预处理和后处理。我想序列化这个模型并通过 MLflow 捕获它的依赖关系。

我试过mlflow.keras.save_model()了，好像不太合适。（它不是一个“纯”的 keras 模型并且没有save()属性）

我也试过mlflow.sklearn.save_model()and mlflow.pyfunc.save_model()，这都导致我出现同样的错误：

NotImplementedError: numpy() is only available when eager execution is enabled.

（这个错误似乎源于 python 和 tensorflow 之间的冲突，也许吧？）

我想知道，是否已经/通常可以用 mlflow 序列化这些“混合”模型？

请在下面找到一个最小的示例

# In[1]:


from mlflow.sklearn import save_model
import mlflow.sklearn
from sklearn.datasets import load_iris
from sklearn import tree

from tensorflow.keras.models import Sequential

import numpy as np

from sklearn.datasets import load_iris
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import OneHotEncoder

from keras.models import Sequential
from keras.layers import Dense
from keras.optimizers import Adam


# ### Save Keras Model

# In[2]:


iris_data = load_iris() 

x = iris_data.data
y_ = iris_data.target.reshape(-1, 1)

# One Hot encode the class labels
encoder = OneHotEncoder(sparse=False)
y = encoder.fit_transform(y_)

# Split the data for training and testing
train_x, test_x, train_y, test_y = train_test_split(x, y, test_size=0.20)

# Build the model
model = Sequential()

model.add(Dense(10, input_shape=(4,), activation='relu', name='fc1'))
model.add(Dense(10, activation='relu', name='fc2'))
model.add(Dense(3, activation='softmax', name='output'))

optimizer = Adam(lr=0.001)
model.compile(optimizer, loss='categorical_crossentropy', metrics=['accuracy'])

# Train the model
model.fit(train_x, train_y, verbose=2, batch_size=5, epochs=20)


# In[3]:


import mlflow.keras

mlflow.keras.save_model(model, "modelstorage/model40")


# ### Save Minimal SKlearn-Pipeline (with Keras)

# In[4]:


from category_encoders.target_encoder import TargetEncoder
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import StandardScaler
from keras.wrappers.scikit_learn import KerasClassifier


# In[5]:


def define_model():
    """
    Create fully connected network with given parameters.
    """
    keras_model = Sequential()

    keras_model.add(Dense(10, input_shape=(4,), activation='relu', name='fc1'))
    keras_model.add(Dense(10, activation='relu', name='fc2'))
    keras_model.add(Dense(3, activation='softmax', name='output'))

    optimizer = Adam(lr=0.001)
    keras_model.compile(optimizer, loss='categorical_crossentropy', metrics=['accuracy'])
    return model


# In[6]:


# target_encoder = TargetEncoder() 
scaler = StandardScaler()
keras_model = KerasClassifier(define_model, batch_size=5, epochs=20)


# In[7]:


pipeline = Pipeline([
#     ('encoding', target_encoder),
    ('scaling', scaler),
    ('modeling', keras_model)
])


# In[8]:


pipeline.fit(train_x, train_y)


# In[9]:


mlflow.keras.save_model(pipeline, "modelstorage/model42")   #not working


# In[10]:


import mlflow.sklearn

mlflow.sklearn.save_model(pipeline, "modelstorage/model43")

Output from modelstorage/model43/conda.yaml:

======================
channels:
- defaults
dependencies:
- python=3.6.7
- scikit-learn=0.21.2
- pip:
  - mlflow
  - cloudpickle==1.2.1
name: mlflow-env
======================

Doesn't seem to capture Tensorflow.
4

1 回答 1

3

您可以在保存模型时添加额外的依赖项，例如，如果您的管道中有 keras 步骤，您可以添加 keras 和 tensorflow：

  conda_env = mlflow.sklearn.get_default_conda_env()
  conda_env["dependencies"] = ['keras==2.2.4', 'tensorflow==1.14.0'] + conda_env["dependencies"]
  mlflow.sklearn.log_model(pipeline, "modelstorage/model43", conda_env = conda_env)
于 2019-11-01T17:18:24.763 回答