Skip to content

Commit

Permalink
Merge pull request caicloud#4 from Ymx6567/master
Browse files Browse the repository at this point in the history 
Update directory structure
  • Loading branch information
@perhapszzy
perhapszzy authored Feb 14, 2017
2 parents 2d2a9a0 + f3632d4 commit c316ed5
Show file tree
Hide file tree
Showing 21 changed files with 222 additions and 99 deletions.
115 changes: 88 additions & 27 deletions Deep_Learning_with_TensorFlow/0.12.0/Chapter04/1. 自定义损失函数.ipynb
View file
Original file line number Diff line number Diff line change
Expand Up @@ -45,14 +45,14 @@
"cell_type": "code",
"execution_count": 3,
"metadata": {
"collapsed": true
"collapsed": false
},
"outputs": [],
"source": [
"# 定义损失函数使得预测少了的损失大，于是模型应该偏向多的方向预测。\n",
"loss_less = 10\n",
"loss_more = 1\n",
"loss = tf.reduce_sum(tf.select(tf.greater(y, y_), (y - y_) * loss_more, (y_ - y) * loss_less))\n",
"loss = tf.reduce_sum(tf.where(tf.greater(y, y_), (y - y_) * loss_more, (y_ - y) * loss_less))\n",
"train_step = tf.train.AdamOptimizer(0.001).minimize(loss)"
]
},
Expand All @@ -73,7 +73,7 @@
"source": [
"rdm = RandomState(1)\n",
"X = rdm.rand(128,2)\n",
"Y = [[x1+x2+rdm.rand()] for (x1, x2) in X]"
"Y = [[x1+x2+rdm.rand()/10.0-0.05] for (x1, x2) in X]"
]
},
{
Expand All @@ -99,24 +99,24 @@
" [ 1.4855988 ]] \n",
"\n",
"After 1000 training step(s), w1 is: \n",
"[[ 0.10801763]\n",
" [ 2.33051562]] \n",
"[[ 0.01247112]\n",
" [ 2.1385448 ]] \n",
"\n",
"After 2000 training step(s), w1 is: \n",
"[[ 0.69744456]\n",
" [ 2.66114545]] \n",
"[[ 0.45567414]\n",
" [ 2.17060661]] \n",
"\n",
"After 3000 training step(s), w1 is: \n",
"[[ 1.01532638]\n",
" [ 2.6881988 ]] \n",
"[[ 0.69968724]\n",
" [ 1.8465308 ]] \n",
"\n",
"After 4000 training step(s), w1 is: \n",
"[[ 1.29930842]\n",
" [ 2.56087017]] \n",
"[[ 0.89886665]\n",
" [ 1.29736018]] \n",
"\n",
"Final w1 is: \n",
"[[ 1.50898325]\n",
" [ 2.31391478]]\n"
"[[ 1.01934695]\n",
" [ 1.04280889]]\n"
]
}
],
Expand Down Expand Up @@ -154,35 +154,97 @@
"output_type": "stream",
"text": [
"After 0 training step(s), w1 is: \n",
"[[-0.81031823]\n",
" [ 1.4855988 ]] \n",
"[[-0.81231821]\n",
" [ 1.48359871]] \n",
"\n",
"After 1000 training step(s), w1 is: \n",
"[[ 0.06754676]\n",
" [ 1.68602526]] \n",
"[[ 0.18643527]\n",
" [ 1.07393336]] \n",
"\n",
"After 2000 training step(s), w1 is: \n",
"[[ 0.62151301]\n",
" [ 1.49415982]] \n",
"[[ 0.95444274]\n",
" [ 0.98088616]] \n",
"\n",
"After 3000 training step(s), w1 is: \n",
"[[ 0.85797542]\n",
" [ 1.32460475]] \n",
"[[ 0.95574027]\n",
" [ 0.9806633 ]] \n",
"\n",
"After 4000 training step(s), w1 is: \n",
"[[ 1.01017809]\n",
" [ 1.2008121 ]] \n",
"[[ 0.95466018]\n",
" [ 0.98135227]] \n",
"\n",
"Final w1 is: \n",
"[[ 1.08652937]\n",
" [ 1.13757098]]\n"
"[[ 0.95525807]\n",
" [ 0.9813394 ]]\n"
]
}
],
"source": [
"loss_less = 1\n",
"loss_more = 10\n",
"loss = tf.reduce_sum(tf.select(tf.greater(y, y_), (y - y_) * loss_more, (y_ - y) * loss_less))\n",
"loss = tf.reduce_sum(tf.where(tf.greater(y, y_), (y - y_) * loss_more, (y_ - y) * loss_less))\n",
"train_step = tf.train.AdamOptimizer(0.001).minimize(loss)\n",
"\n",
"with tf.Session() as sess:\n",
" init_op = tf.global_variables_initializer()\n",
" sess.run(init_op)\n",
" STEPS = 5000\n",
" for i in range(STEPS):\n",
" start = (i*batch_size) % 128\n",
" end = (i*batch_size) % 128 + batch_size\n",
" sess.run(train_step, feed_dict={x: X[start:end], y_: Y[start:end]})\n",
" if i % 1000 == 0:\n",
" print(\"After %d training step(s), w1 is: \" % (i))\n",
" print sess.run(w1), \"\\n\"\n",
" print \"Final w1 is: \\n\", sess.run(w1)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"#### 6. 定义损失函数为MSE。"
]
},
{
"cell_type": "code",
"execution_count": 7,
"metadata": {
"collapsed": false
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"After 0 training step(s), w1 is: \n",
"[[-0.81031823]\n",
" [ 1.4855988 ]] \n",
"\n",
"After 1000 training step(s), w1 is: \n",
"[[-0.13337609]\n",
" [ 1.81309223]] \n",
"\n",
"After 2000 training step(s), w1 is: \n",
"[[ 0.32190299]\n",
" [ 1.52463484]] \n",
"\n",
"After 3000 training step(s), w1 is: \n",
"[[ 0.67850214]\n",
" [ 1.25297272]] \n",
"\n",
"After 4000 training step(s), w1 is: \n",
"[[ 0.89473999]\n",
" [ 1.08598232]] \n",
"\n",
"Final w1 is: \n",
"[[ 0.97437561]\n",
" [ 1.0243336 ]]\n"
]
}
],
"source": [
"loss = tf.losses.mean_squared_error(y, y_)\n",
"train_step = tf.train.AdamOptimizer(0.001).minimize(loss)\n",
"\n",
"with tf.Session() as sess:\n",
Expand Down Expand Up @@ -222,4 +284,3 @@
"nbformat": 4,
"nbformat_minor": 1
}

28 changes: 14 additions & 14 deletions Deep_Learning_with_TensorFlow/0.12.0/Chapter10/2. 多GPU并行.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -16,15 +16,15 @@
N_GPU = 4

# 定义日志和模型输出的路径。
MODEL_SAVE_PATH = "/path/to/logs_and_models/"
MODEL_SAVE_PATH = "logs_and_models/"
MODEL_NAME = "model.ckpt"
DATA_PATH = "/path/to/data.tfrecords"
DATA_PATH = "output.tfrecords"

# 定义输入队列得到训练数据，具体细节可以参考第七章。
def get_input():
filename_queue = tf.train.string_input_producer([DATA_PATH])
reader = tf.TFRecordReader()
_, serialized_example = reader.read(filename_queue)
_, serialized_example = reader.read(filename_queue)

# 定义数据解析格式。
features = tf.parse_single_example(
Expand All @@ -44,17 +44,17 @@ def get_input():
# 定义输入队列并返回。
min_after_dequeue = 10000
capacity = min_after_dequeue + 3 * BATCH_SIZE
return tf.train.shuffle_batch(
return tf.train.shuffle_batch(
[retyped_image, label],
batch_size=BATCH_SIZE,
capacity=capacity,
min_after_dequeue=min_after_dequeue)

# 定义损失函数。
def get_loss(x, y_, regularizer, scope):
y = mnist_inference.inference(x, regularizer)
cross_entropy = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(y, y_))
regularization_loss = tf.add_n(tf.get_collection('losses', scope))
y = mnist_inference.inference(x, regularizer)
cross_entropy = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(y, y_))
regularization_loss = tf.add_n(tf.get_collection('losses', scope))
loss = cross_entropy + regularization_loss
return loss

Expand Down Expand Up @@ -82,7 +82,7 @@ def average_gradients(tower_grads):
# 主训练过程。
def main(argv=None):
# 将简单的运算放在CPU上，只有神经网络的训练过程放在GPU上。
with tf.Graph().as_default(), tf.device('/cpu:0'):
with tf.Graph().as_default(), tf.device('/cpu:0'):

# 定义基本的训练过程
x, y_ = get_input()
Expand Down Expand Up @@ -110,29 +110,29 @@ def main(argv=None):
grads = average_gradients(tower_grads)
for grad, var in grads:
if grad is not None:
tf.histogram_summary('gradients_on_average/%s' % var.op.name, grad)
tf.summary.histogram('gradients_on_average/%s' % var.op.name, grad)

# 使用平均梯度更新参数。
apply_gradient_op = opt.apply_gradients(grads, global_step=global_step)
for var in tf.trainable_variables():
tf.histogram_summary(var.op.name, var)
tf.summary.histogram(var.op.name, var)

# 计算变量的滑动平均值。
variable_averages = tf.train.ExponentialMovingAverage(MOVING_AVERAGE_DECAY, global_step)
variables_averages_op = variable_averages.apply(tf.trainable_variables())
# 每一轮迭代需要更新变量的取值并更新变量的滑动平均值。
train_op = tf.group(apply_gradient_op, variables_averages_op)

saver = tf.train.Saver(tf.all_variables())
summary_op = tf.merge_all_summaries()
init = tf.initialize_all_variables()
saver = tf.train.Saver(tf.global_variables())
summary_op = tf.summary.merge_all()
init = tf.global_variables_initializer()
with tf.Session(config=tf.ConfigProto(
allow_soft_placement=True, log_device_placement=True)) as sess:
# 初始化所有变量并启动队列。
init.run()
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
summary_writer = tf.train.SummaryWriter(MODEL_SAVE_PATH, sess.graph)
summary_writer = tf.summary.FileWriter(MODEL_SAVE_PATH, sess.graph)

for step in range(TRAINING_STEPS):
# 执行神经网络训练操作，并记录训练操作的运行时间。
Expand Down
Loading

0 comments on commit c316ed5

Please sign in to comment.