《Hands-On Machine Learning with Scikit-Learn and TensorFlow》代码

# Copyright 2016 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License Version 2.0 （the “License“）;
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing software
# distributed under the License is distributed on an “AS IS“ BASIS
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
“““Contains a model definition for AlexNet.

This work was first described in:
  ImageNet Classification with Deep Convolutional Neural Networks
  Alex Krizhevsky Ilya Sutskever and Geoffrey E. Hinton

and later refined in:
  One weird trick for parallelizing convolutional neural networks
  Alex Krizhevsky 2014

Here we provide the implementation proposed in “One weird trick“ and not
“ImageNet Classification“ as per the paper the LRN layers have been removed.

Usage:
  with slim.arg_scope（alexnet.alexnet_v2_arg_scope（））:
    outputs end_points = alexnet.alexnet_v2（inputs）

@@alexnet_v2
“““

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import tensorflow as tf

slim = tf.contrib.slim
trunc_normal = lambda stddev: tf.truncated_normal_initializer（0.0 stddev）


def alexnet_v2_arg_scope（weight_decay=0.0005）:
  with slim.arg_scope（[slim.conv2d slim.fully_connected]
                      activation_fn=tf.nn.relu
                      biases_initializer=tf.constant_initializer（0.1）
                      weights_regularizer=slim.l2_regularizer（weight_decay））:
    with slim.arg_scope（[slim.conv2d] padding=‘SAME‘）:
      with slim.arg_scope（[slim.max_pool2d] padding=‘VALID‘） as arg_sc:
        return arg_sc


def alexnet_v2（inputs
               num_classes=1000
               is_training=True
               dropout_keep_prob=0.5
               spatial_squeeze=True
               scope=‘alexnet_v2‘）:
  “““AlexNet version 2.

  Described in: http://arxiv.org/pdf/1404.5997v2.pdf
  Parameters from:
  github.com/akrizhevsky/cuda-convnet2/blob/master/layers/
  layers-imagenet-1gpu.cfg

  Note: All the fully_connected layers have been transformed to conv2d layers.
        To use in classification mode resize input to 224x224. To use in fully
        convolutional mode set spatial_squeeze to false.
        The LRN layers have been removed and change the initializers from
        random_normal_initializer to xavier_initializer.

  Args:
    inputs: a tensor of size [batch_size height width channels].
    num_classes: number of predicted classes.
    is_training: whether or not the model is being trained.
    dropout_keep_prob: the probability that activations are kept in the dropout
      layers during traini

 属性            大小     日期    时间   名称
----------- ---------  ---------- -----  ----
     目录           0  2016-11-25 08:34  handson-ml-master\
     文件         244  2016-11-25 08:34  handson-ml-master\.binder_start
     文件         112  2016-11-25 08:34  handson-ml-master\.gitignore
     文件      274510  2016-11-25 08:34  handson-ml-master\01_the_machine_learning_landscape.ipynb
     文件     2118822  2016-11-25 08:34  handson-ml-master\02_end_to_end_machine_learning_project.ipynb
     文件      960196  2016-11-25 08:34  handson-ml-master\03_classification.ipynb
     文件      743852  2016-11-25 08:34  handson-ml-master\04_training_linear_models.ipynb
     文件      867639  2016-11-25 08:34  handson-ml-master\05_support_vector_machines.ipynb
     文件      201263  2016-11-25 08:34  handson-ml-master\06_decision_trees.ipynb
     文件      567343  2016-11-25 08:34  handson-ml-master\07_ensemble_learning_and_random_forests.ipynb
     文件     1560349  2016-11-25 08:34  handson-ml-master\08_dimensionality_reduction.ipynb
     文件       83498  2016-11-25 08:34  handson-ml-master\09_up_and_running_with_tensorflow.ipynb
     文件      326697  2016-11-25 08:34  handson-ml-master\10_introduction_to_artificial_neural_networks.ipynb
     文件      275788  2016-11-25 08:34  handson-ml-master\11_deep_learning.ipynb
     文件       15732  2016-11-25 08:34  handson-ml-master\12_distributed_tensorflow.ipynb
     文件     1507818  2016-11-25 08:34  handson-ml-master\13_convolutional_neural_networks.ipynb
     文件      897957  2016-11-25 08:34  handson-ml-master\14_recurrent_neural_networks.ipynb
     文件      328928  2016-11-25 08:34  handson-ml-master\15_autoencoders.ipynb
     文件      757620  2016-11-25 08:34  handson-ml-master\16_reinforcement_learning.ipynb
     文件        1927  2016-11-25 08:34  handson-ml-master\Dockerfile
     文件       10175  2016-11-25 08:34  handson-ml-master\LICENSE
     文件        3639  2016-11-25 08:34  handson-ml-master\README.md
     目录           0  2016-11-25 08:34  handson-ml-master\datasets\
     目录           0  2016-11-25 08:34  handson-ml-master\datasets\housing\
     文件        3679  2016-11-25 08:34  handson-ml-master\datasets\housing\README.md
     文件     1423529  2016-11-25 08:34  handson-ml-master\datasets\housing\housing.csv
     文件      409488  2016-11-25 08:34  handson-ml-master\datasets\housing\housing.tgz
     目录           0  2016-11-25 08:34  handson-ml-master\datasets\inception\
     文件       31674  2016-11-25 08:34  handson-ml-master\datasets\inception\imagenet_class_names.txt
     目录           0  2016-11-25 08:34  handson-ml-master\datasets\lifesat\
     文件        4311  2016-11-25 08:34  handson-ml-master\datasets\lifesat\README.md
............此处省略70个文件信息