TensorFlow 從入門到精通(二):MNIST 例程原始碼分析
阿新 • • 發佈:2019-01-01
按照上節步驟, TensorFlow 預設安裝在 /usr/lib/python/site-packages/tensorflow/ (也有可能是 /usr/local/lib……)下,檢視目錄結構:
# tree -d -L 3 /usr/lib/python2.7/site-packages/tensorflow/
/usr/lib/python2.7/site-packages/tensorflow/
├── contrib
│ ├── bayesflow
│ │ └── python
│ ├── cmake
│ ├── copy_graph
│ │ └── python
│ ├── crf
│ │ └── python
│ ├── cudnn_rnn
│ │ ├── ops
│ │ └── python
│ ├── distributions
│ │ └── python
│ ├── factorization
│ │ └── python
│ ├── ffmpeg
│ │ └── ops
│ ├── framework
│ │ └── python
│ ├── graph_editor
│ ├── grid_rnn
│ │ └── python
│ ├── layers
│ │ ├── ops
│ │ └── python
│ ├── learn
│ │ └── python
│ ├── linear_optimizer
│ │ ├── ops
│ │ └── python
│ ├── lookup
│ ├── losses
│ │ └── python
│ ├── metrics
│ │ ├── ops
│ │ └── python
│ ├── opt
│ │ └── python
│ ├── quantization
│ │ ├── kernels
│ │ ├── ops
│ │ └── python
│ ├── rnn
│ │ └── python
│ ├── session_bundle
│ ├── slim
│ │ └── python
│ ├── tensorboard
│ │ └── plugins
│ ├── tensor_forest
│ │ ├── client
│ │ ├── data
│ │ ├── hybrid
│ │ └── python
│ ├── testing
│ │ └── python
│ ├── training
│ │ └── python
│ └── util
├── core
│ ├── example
│ ├── framework
│ ├── lib
│ │ └── core
│ ├── protobuf
│ └── util
├── examples
│ └── tutorials
│ └── mnist
├── include
│ ├── Eigen
│ │ └── src
│ ├── external
│ │ └── eigen_archive
│ ├── google
│ │ └── protobuf
│ ├── tensorflow
│ │ └── core
│ ├── third_party
│ │ └── eigen3
│ └── unsupported
│ └── Eigen
├── models
│ ├── embedding
│ ├── image
│ │ ├── alexnet
│ │ ├── cifar10
│ │ ├── imagenet
│ │ └── mnist
│ └── rnn
│ ├── ptb
│ └── translate
├── python
│ ├── client
│ ├── debug
│ │ └── cli
│ ├── framework
│ ├── lib
│ │ ├── core
│ │ └── io
│ ├── ops
│ ├── platform
│ ├── saved_model
│ ├── summary
│ │ └── impl
│ ├── training
│ ├── user_ops
│ └── util
│ └── protobuf
├── tensorboard
│ ├── backend
│ ├── dist
│ ├── lib
│ │ ├── css
│ │ └── python
│ └── plugins
│ └── projector
└── tools
└── pip_package
119 directories
上節執行 MNIST 例程的命令為:
# python -m tensorflow.models.image.mnist.convolutional
對應檔案為 /usr/lib/python2.7/site-packages/tensorflow/models/image/mnist/convolutional.py
開啟例程原始碼:
# Copyright 2015 Google Inc. 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.
# ==============================================================================
"""Simple, end-to-end, LeNet-5-like convolutional MNIST model example.
This should achieve a test error of 0.7%. Please keep this model as simple and
linear as possible, it is meant as a tutorial for simple convolutional models.
Run with --self_test on the command line to execute a short self-test.
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import gzip
import os
import sys
import time
import numpy
from six.moves import urllib
from six.moves import xrange # pylint: disable=redefined-builtin
import tensorflow as tf
# 資料來源
SOURCE_URL = 'http://yann.lecun.com/exdb/mnist/'
# 工作目錄,存放下載的資料
WORK_DIRECTORY = 'data'
# MNIST 資料集特徵:
# 影象尺寸 28x28
IMAGE_SIZE = 28
# 黑白影象
NUM_CHANNELS = 1
# 畫素值0~255
PIXEL_DEPTH = 255
# 標籤分10個類別
NUM_LABELS = 10
# 驗證集共 5000 個樣本
VALIDATION_SIZE = 5000
# 隨機數種子,可設為 None 表示真的隨機
SEED = 66478
# 批處理大小為64
BATCH_SIZE = 64
# 資料全集一共過10遍網路
NUM_EPOCHS = 10
# 驗證集批處理大小也是64
EVAL_BATCH_SIZE = 64
# 驗證時間間隔,每訓練100個批處理,做一次評估
EVAL_FREQUENCY = 100
tf.app.flags.DEFINE_boolean("self_test", False, "True if running a self test.")
FLAGS = tf.app.flags.FLAGS
# 如果下載過了資料,就不再重複下載
def maybe_download(filename):
"""Download the data from Yann's website, unless it's already here."""
if not tf.gfile.Exists(WORK_DIRECTORY):
tf.gfile.MakeDirs(WORK_DIRECTORY)
filepath = os.path.join(WORK_DIRECTORY, filename)
if not tf.gfile.Exists(filepath):
filepath, _ = urllib.request.urlretrieve(SOURCE_URL + filename, filepath)
with tf.gfile.GFile(filepath) as f:
size = f.Size()
print('Successfully downloaded', filename, size, 'bytes.')
return filepath
# 抽取資料,變為 4維張量[影象索引,y, x, c]
# 去均值、做歸一化,範圍變到[-0.5, 0.5]
def extract_data(filename, num_images):
"""Extract the images into a 4D tensor [image index, y, x, channels].
Values are rescaled from [0, 255] down to [-0.5, 0.5].
"""
print('Extracting', filename)
with gzip.open(filename) as bytestream:
bytestream.read(16)
buf = bytestream.read(IMAGE_SIZE * IMAGE_SIZE * num_images)
data = numpy.frombuffer(buf, dtype=numpy.uint8).astype(numpy.float32)
data = (data - (PIXEL_DEPTH / 2.0)) / PIXEL_DEPTH
data = data.reshape(num_images, IMAGE_SIZE, IMAGE_SIZE, 1)
return data
# 抽取影象標籤
def extract_labels(filename, num_images):
"""Extract the labels into a vector of int64 label IDs."""
print('Extracting', filename)
with gzip.open(filename) as bytestream:
bytestream.read(8)
buf = bytestream.read(1 * num_images)
labels = numpy.frombuffer(buf, dtype=numpy.uint8).astype(numpy.int64)
return labels
# 假資料,用於功能自測
def fake_data(num_images):
"""Generate a fake dataset that matches the dimensions of MNIST."""
data = numpy.ndarray(
shape=(num_images, IMAGE_SIZE, IMAGE_SIZE, NUM_CHANNELS),
dtype=numpy.float32)
labels = numpy.zeros(shape=(num_images,), dtype=numpy.int64)
for image in xrange(num_images):
label = image % 2
data[image, :, :, 0] = label - 0.5
labels[image] = label
return data, labels
# 計算分類錯誤率
def error_rate(predictions, labels):
"""Return the error rate based on dense predictions and sparse labels."""
return 100.0 - (
100.0 *
numpy.sum(numpy.argmax(predictions, 1) == labels) /
predictions.shape[0])
# 主函式
def main(argv=None): # pylint: disable=unused-argument
if FLAGS.self_test:
print('Running self-test.')
train_data, train_labels = fake_data(256)
validation_data, validation_labels = fake_data(EVAL_BATCH_SIZE)
test_data, test_labels = fake_data(EVAL_BATCH_SIZE)
num_epochs = 1
else:
# 下載資料
train_data_filename = maybe_download('train-images-idx3-ubyte.gz')
train_labels_filename = maybe_download('train-labels-idx1-ubyte.gz')
test_data_filename = maybe_download('t10k-images-idx3-ubyte.gz')
test_labels_filename = maybe_download('t10k-labels-idx1-ubyte.gz')
# 載入資料到numpy
train_data = extract_data(train_data_filename, 60000)
train_labels = extract_labels(train_labels_filename, 60000)
test_data = extract_data(test_data_filename, 10000)
test_labels = extract_labels(test_labels_filename, 10000)
# 產生評測集
validation_data = train_data[:VALIDATION_SIZE, ...]
validation_labels = train_labels[:VALIDATION_SIZE]
train_data = train_data[VALIDATION_SIZE:, ...]
train_labels = train_labels[VALIDATION_SIZE:]
num_epochs = NUM_EPOCHS
train_size = train_labels.shape[0]
# 訓練樣本和標籤將從這裡送入網路。
# 每訓練迭代步,佔位符節點將被送入一個批處理資料
# 訓練資料節點
train_data_node = tf.placeholder(
tf.float32,
shape=(BATCH_SIZE, IMAGE_SIZE, IMAGE_SIZE, NUM_CHANNELS))
# 訓練標籤節點
train_labels_node = tf.placeholder(tf.int64, shape=(BATCH_SIZE,))
# 評測資料節點
eval_data = tf.placeholder(
tf.float32,
shape=(EVAL_BATCH_SIZE, IMAGE_SIZE, IMAGE_SIZE, NUM_CHANNELS))
# 下面這些變數是網路的可訓練權值
# conv1 權值維度為 32 x channels x 5 x 5, 32 為特徵圖數目
conv1_weights = tf.Variable(
tf.truncated_normal([5, 5, NUM_CHANNELS, 32], # 5x5 filter, depth 32.
stddev=0.1,
seed=SEED))
# conv1 偏置
conv1_biases = tf.Variable(tf.zeros([32]))
# conv2 權值維度為 64 x 32 x 5 x 5
conv2_weights = tf.Variable(
tf.truncated_normal([5, 5, 32, 64],
stddev=0.1,
seed=SEED))
conv2_biases = tf.Variable(tf.constant(0.1, shape=[64]))
# 全連線層 fc1 權值,神經元數目為512
fc1_weights = tf.Variable( # fully connected, depth 512.
tf.truncated_normal(
[IMAGE_SIZE // 4 * IMAGE_SIZE // 4 * 64, 512],
stddev=0.1,
seed=SEED))
fc1_biases = tf.Variable(tf.constant(0.1, shape=[512]))
# fc2 權值,維度與標籤類數目一致
fc2_weights = tf.Variable(
tf.truncated_normal([512, NUM_LABELS],
stddev=0.1,
seed=SEED))
fc2_biases = tf.Variable(tf.constant(0.1, shape=[NUM_LABELS]))
# 兩個網路:訓練網路和評測網路
# 它們共享權值
# 實現 LeNet-5 模型,該函式輸入為資料,輸出為fc2的響應
# 第二個引數區分訓練網路還是評測網路
def model(data, train=False):
"""The Model definition."""
# 二維卷積,使用“不變形”補零(即輸出特徵圖與輸入尺寸一致)。
conv = tf.nn.conv2d(data,
conv1_weights,
strides=[1, 1, 1, 1],
padding='SAME')
# 加偏置、過啟用函式一塊完成
relu = tf.nn.relu(tf.nn.bias_add(conv, conv1_biases))
# 最大值下采樣
pool = tf.nn.max_pool(relu,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME')
# 第二個卷積層
conv = tf.nn.conv2d(pool,
conv2_weights,
strides=[1, 1, 1, 1],
padding='SAME')
relu = tf.nn.relu(tf.nn.bias_add(conv, conv2_biases))
pool = tf.nn.max_pool(relu,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME')
# 特徵圖變形為2維矩陣,便於送入全連線層
pool_shape = pool.get_shape().as_list()
reshape = tf.reshape(
pool,
[pool_shape[0], pool_shape[1] * pool_shape[2] * pool_shape[3]])
# 全連線層,注意“+”運算自動廣播偏置
hidden = tf.nn.relu(tf.matmul(reshape, fc1_weights) + fc1_biases)
# 訓練階段,增加 50% dropout;而評測階段無需該操作
if train:
hidden = tf.nn.dropout(hidden, 0.5, seed=SEED)
return tf.matmul(hidden, fc2_weights) + fc2_biases
# Training computation: logits + cross-entropy loss.
# 訓練階段計算: 對數+交叉熵 損失函式
logits = model(train_data_node, True)
loss = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(
logits, train_labels_node))
# 全連線層引數進行 L2 正則化
regularizers = (tf.nn.l2_loss(fc1_weights) + tf.nn.l2_loss(fc1_biases) +
tf.nn.l2_loss(fc2_weights) + tf.nn.l2_loss(fc2_biases))
# 將正則項加入損失函式
loss += 5e-4 * regularizers
# 優化器: 設定一個變數,每個批處理遞增,控制學習速率衰減
batch = tf.Variable(0)
# 指數衰減
learning_rate = tf.train.exponential_decay(
0.01, # 基本學習速率
batch * BATCH_SIZE, # 當前批處理在資料全集中的位置
train_size, # Decay step.
0.95, # Decay rate.
staircase=True)
# Use simple momentum for the optimization.
optimizer = tf.train.MomentumOptimizer(learning_rate,
0.9).minimize(loss,
global_step=batch)
# 用softmax 計算訓練批處理的預測概率
train_prediction = tf.nn.softmax(logits)
# 用 softmax 計算評測批處理的預測概率
eval_prediction = tf.nn.softmax(model(eval_data))
# Small utility function to evaluate a dataset by feeding batches of data to
# {eval_data} and pulling the results from {eval_predictions}.
# Saves memory and enables this to run on smaller GPUs.
def eval_in_batches(data, sess):
"""Get all predictions for a dataset by running it in small batches."""
size = data.shape[0]
if size < EVAL_BATCH_SIZE:
raise ValueError("batch size for evals larger than dataset: %d" % size)
predictions = numpy.ndarray(shape=(size, NUM_LABELS), dtype=numpy.float32)
for begin in xrange(0, size, EVAL_BATCH_SIZE):
end = begin + EVAL_BATCH_SIZE
if end <= size:
predictions[begin:end, :] = sess.run(
eval_prediction,
feed_dict={eval_data: data[begin:end, ...]})
else:
batch_predictions = sess.run(
eval_prediction,
feed_dict={eval_data: data[-EVAL_BATCH_SIZE:, ...]})
predictions[begin:, :] = batch_predictions[begin - size:, :]
return predictions
# Create a local session to run the training.
start_time = time.time()
with tf.Session() as sess:
# Run all the initializers to prepare the trainable parameters.
tf.initialize_all_variables().run()
print('Initialized!')
# Loop through training steps.
for step in xrange(int(num_epochs * train_size) // BATCH_SIZE):
# Compute the offset of the current minibatch in the data.
# Note that we could use better randomization across epochs.
offset = (step * BATCH_SIZE) % (train_size - BATCH_SIZE)
batch_data = train_data[offset:(offset + BATCH_SIZE), ...]
batch_labels = train_labels[offset:(offset + BATCH_SIZE)]
# This dictionary maps the batch data (as a numpy array) to the
# node in the graph it should be fed to.
feed_dict = {train_data_node: batch_data,
train_labels_node: batch_labels}
# Run the graph and fetch some of the nodes.
_, l, lr, predictions = sess.run(
[optimizer, loss, learning_rate, train_prediction],
feed_dict=feed_dict)
if step % EVAL_FREQUENCY == 0:
elapsed_time = time.time() - start_time
start_time = time.time()
print('Step %d (epoch %.2f), %.1f ms' %
(step, float(step) * BATCH_SIZE / train_size,
1000 * elapsed_time / EVAL_FREQUENCY))
print('Minibatch loss: %.3f, learning rate: %.6f' % (l, lr))
print('Minibatch error: %.1f%%' % error_rate(predictions, batch_labels))
print('Validation error: %.1f%%' % error_rate(
eval_in_batches(validation_data, sess), validation_labels))
sys.stdout.flush()
# Finally print the result!
test_error = error_rate(eval_in_batches(test_data, sess), test_labels)
print('Test error: %.1f%%' % test_error)
if FLAGS.self_test:
print('test_error', test_error)
assert test_error == 0.0, 'expected 0.0 test_error, got %.2f' % (
test_error,)
# 程式入口點
if __name__ == '__main__':
tf.app.run()