def test_Generators():
num_range = (1, 31)
data_size = 300 # 4
make_geomeric_objects(dataset_size=300,
image_size=(120, 120),
locations=dd.five_star_not_huge,
num_range=(1, 31),
guassian=True,
noise=True)
input_data = np.load(dd.five_star_not_huge) # 1
print(input_data.shape)
size = 120 #5
epoch_size = 30 # 表示一波含有的图片
dd.clear(dd.show_data) # !!
image = tf.placeholder(tf.float32, [epoch_size, size, size, 1],
name="image")
tf.summary.image('five_star_not_huge_20', image,
max_outputs=epoch_size) # 2
merged = tf.summary.merge_all()
writer = tf.summary.FileWriter(dd.show_data +
"/five_star_not_huge_20") # 3
with tf.Session() as sess:
for i in range(data_size // epoch_size):
data = input_data[i * epoch_size:(i + 1) * epoch_size,
0:size * size]
data = np.reshape(data, (epoch_size, size, size, 1))
summary = sess.run(merged, feed_dict={image: data})
writer.add_summary(summary, i)
writer.close()
def test_tensor():
num = 800
size = 60
data_dir = dd.ploygons
show_dir = dd.show_data + "/polygon/"
dd.clear(show_dir)
tensor = tf.placeholder(tf.float32, [num, size * size], name="t")
tf.summary.image("x_generated",
tf.reshape(tensor, [-1, size, size, 1]),
max_outputs=10)
merged = tf.summary.merge_all()
writer = tf.summary.FileWriter(show_dir)
a = np.load(data_dir)
a = a[:, 0:-1]
# mnist
# from tensorflow.examples.tutorials.mnist import input_data
# mnist = input_data.read_data_sets('/home/martin/X-Brain/Martins3.github.io/source/_posts/nonsense/ML/project/MNIST_data', one_hot=True)
# x_value, _ = mnist.train.next_batch(100)
# a = x_value
with tf.Session() as sess:
summary = sess.run(merged, feed_dict={tensor: a})
writer.add_summary(summary, 0)
writer.close()
def naive_conv_with_bn():
geometric_data_dir = dd.five_star_noise # 2
if (not os.path.isfile(geometric_data_dir)):
makeData.make_data(geometric_data_dir)
print("generate random data finished !")
n_output = 10 # 1
polygon_data = np.load(geometric_data_dir)
learning_rate = 0.0001
batch_size = 50
train_test_ratio = 0.8
dataset_size = polygon_data.shape[0]
image_size = 60
n_input = 60 * 60 # 图片的大小
dev = 0.1
bias_init = 0.1
training_iters = 8000
def make_one_hot(truth, labels):
base = np.zeros((truth.shape[0], labels), dtype=np.float32)
for i in range(truth.shape[0]):
index = int(truth[i][0])
base[i][index] = 1
return base
np.random.shuffle(polygon_data)
train_data = polygon_data[0:int(train_test_ratio * dataset_size)]
test_data = polygon_data[int(train_test_ratio * dataset_size):dataset_size]
dd.clear(dd.show_data2)
x = tf.placeholder(tf.float32, [None, n_input])
ground_truth = tf.placeholder(tf.float32,
[None, n_output]) # 意味着开始的时候就是需要转化为 one-hot
is_training = tf.placeholder(tf.bool)
keep_prob = tf.placeholder(tf.float32) #dropout (keep probability)
# Create some wrappers for simplicity
def conv2d(x, W, b):
# Conv2D wrapper, with bias and relu activation
x = tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')
x = tf.nn.bias_add(x, b)
return tf.nn.relu(x)
def maxpool2d(x, k=2):
# MaxPool2D wrapper
return tf.nn.max_pool(x,
ksize=[1, k, k, 1],
strides=[1, k, k, 1],
padding='SAME')
# Create model
def conv_net(x, weights, biases, dropout):
# Reshape input picture
x = tf.reshape(x, shape=[-1, 60, 60, 1])
conv1 = conv2d(x, weights['wc1'], biases['bc1'])
conv1 = tf.contrib.layers.batch_norm(conv1,
decay=0.9,
center=True,
scale=True,
updates_collections=None,
is_training=is_training,
reuse=None,
trainable=True,
scope="bn_conv1")
conv1 = maxpool2d(conv1)
conv2 = conv2d(conv1, weights['wc2'], biases['bc2'])
conv2 = tf.contrib.layers.batch_norm(conv2,
decay=0.9,
center=True,
scale=True,
updates_collections=None,
is_training=is_training,
reuse=None,
trainable=True,
scope="bn_conv2")
conv2 = maxpool2d(conv2)
fc1 = tf.reshape(conv2, [-1, weights['wd1'].get_shape().as_list()[0]])
fc1 = tf.add(tf.matmul(fc1, weights['wd1']), biases['bd1'])
bn = tf.contrib.layers.batch_norm(fc1,
decay=0.9,
center=True,
scale=True,
updates_collections=None,
is_training=is_training,
reuse=None,
trainable=True,
scope="bn1")
fc1 = tf.nn.relu(bn)
fc1 = tf.nn.dropout(fc1, dropout) # dorpout 不可以去除
fc2 = tf.add(tf.matmul(fc1, weights['wd2']), biases['bd2'])
bn2 = tf.contrib.layers.batch_norm(fc2,
decay=0.9,
center=True,
scale=True,
updates_collections=None,
is_training=is_training,
reuse=None,
trainable=True,
scope="bn2")
fc2 = tf.nn.relu(bn2)
fc2 = tf.nn.dropout(fc2, dropout) # dorpout 不可以去除
fc3 = tf.add(tf.matmul(fc2, weights['wd3']), biases['bd3'])
bn3 = tf.contrib.layers.batch_norm(fc3,
decay=0.9,
center=True,
scale=True,
updates_collections=None,
is_training=is_training,
reuse=None,
trainable=True,
scope="bn3")
fc3 = tf.nn.relu(bn3)
fc3 = tf.nn.dropout(fc3, dropout) # dorpout 不可以去除
# Output, class prediction
out = tf.nn.relu(tf.add(tf.matmul(fc3, weights['out']), biases['out']))
return out
# Store layers weight & bias
feature_num1 = 6
feature_num2 = 12
kernel_size = 5
full_hidden_1 = 800
full_hidden_2 = 300
full_hidden_3 = 100
weights = {
# 5x5 conv, 1 input, 32 outputs
'wc1':
tf.Variable(
tf.truncated_normal([kernel_size, kernel_size, 1, feature_num1],
stddev=dev)),
# 5x5 conv, 32 inputs, 64 outputs
'wc2':
tf.Variable(
tf.truncated_normal(
[kernel_size, kernel_size, feature_num1, feature_num2],
stddev=dev)),
# fully connected, 7*7*64 inputs, 1024 outputs
'wd1':
tf.Variable(
tf.truncated_normal([15 * 15 * feature_num2, full_hidden_1],
stddev=dev)),
# 第二层 全连接
'wd2':
tf.Variable(
tf.truncated_normal([full_hidden_1, full_hidden_2], stddev=dev)),
'wd3':
tf.Variable(
tf.truncated_normal([full_hidden_2, full_hidden_3], stddev=dev)),
# 1024 inputs, 10 outputs (class prediction)
'out':
tf.Variable(tf.truncated_normal([full_hidden_3, n_output], stddev=dev))
}
biases = {
'bc1': tf.Variable(tf.constant(bias_init, shape=[feature_num1])),
'bc2': tf.Variable(tf.constant(bias_init, shape=[feature_num2])),
'bd1': tf.Variable(tf.constant(bias_init, shape=[full_hidden_1])),
'bd2': tf.Variable(tf.constant(bias_init, shape=[full_hidden_2])),
'bd3': tf.Variable(tf.constant(bias_init, shape=[full_hidden_3])),
'out': tf.Variable(tf.constant(bias_init, shape=[n_output]))
}
# Construct model
# (-1, n_output的个数)
# Construct model
pred = conv_net(x, weights, biases, keep_prob)
# Define loss and optimizer
cost = tf.reduce_sum(
tf.nn.softmax_cross_entropy_with_logits(logits=pred,
labels=ground_truth))
optimizer = tf.train.AdamOptimizer(
learning_rate=learning_rate).minimize(cost)
# Evaluate model
correct_pred = tf.equal(tf.argmax(pred, 1), tf.argmax(ground_truth, 1))
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
tf.summary.scalar("accuracy", accuracy)
tf.summary.scalar("cost", cost)
# Initializing the variables
init = tf.global_variables_initializer()
optimizer = tf.train.AdamOptimizer(
learning_rate=learning_rate).minimize(cost)
# Initializing the variables
init = tf.global_variables_initializer()
# Launch the graph
with tf.Session() as sess:
merged = tf.summary.merge_all()
sess.run(init)
train_writer = tf.summary.FileWriter(dd.show_data2 + "/train",
sess.graph)
test_writer = tf.summary.FileWriter(dd.show_data2 + "/test")
for i in range(training_iters):
if (i % 50 == 0):
summary, res, acc = sess.run(
[merged, cost, accuracy],
feed_dict={
x:
test_data[:, 0:n_input],
ground_truth:
make_one_hot(test_data[:, n_input:n_input + 1],
n_output),
keep_prob:
1,
is_training:
False
})
train_writer.add_summary(summary, i)
print(acc)
if (i % 100 == 0):
s_train_sample = train_data[0:1000]
s_train_x = s_train_sample[:, 0:n_input]
s_train_y = make_one_hot(
s_train_sample[:, n_input:n_input + 1], n_output)
s_acc, summary = sess.run(
[accuracy, merged],
feed_dict={
x: s_train_x,
ground_truth: s_train_y,
keep_prob: 1,
is_training: False
})
test_writer.add_summary(summary, i)
print("{1} trianing: {0}".format(s_acc, i))
np.random.shuffle(train_data)
train_sample = train_data[0:batch_size]
train_x = train_sample[:, 0:n_input]
train_y = make_one_hot(train_sample[:, n_input:n_input + 1],
n_output)
sess.run(optimizer,
feed_dict={
x: train_x,
ground_truth: train_y,
keep_prob: 0.6,
is_training: True
})
train_writer.close()
def naive_ResNet():
# super parameter
learning_rate = 0.0001
batch_size = 50
train_test_ratio = 0.8
dev = 0.1
bias_init = 0.1
training_iters = 8000
kernel_size = 3
image_size = 60
n_input = 60 * 60 # 图片的大小
n_output = 10 # 1
# prepare the data
geometric_data_dir = dd.five_star_noise # 2
if (not os.path.isfile(geometric_data_dir)):
makeData.make_data(geometric_data_dir)
print("generate random data finished !")
polygon_data = np.load(geometric_data_dir)
dataset_size = polygon_data.shape[0]
np.random.shuffle(polygon_data)
train_data = polygon_data[0:int(train_test_ratio * dataset_size)]
test_data = polygon_data[int(train_test_ratio * dataset_size):dataset_size]
dd.clear(dd.show_data2)
def make_one_hot(truth, labels):
base = np.zeros((truth.shape[0], labels), dtype=np.float32)
for i in range(truth.shape[0]):
index = int(truth[i][0])
base[i][index] = 1
return base
# fundamental functions
def weight_variable(shape, name=None):
initial = tf.truncated_normal(shape, stddev=dev)
return tf.Variable(initial, name=name)
def bais_variable(shape, name=None):
"""
所有创建的bias_variable都是一维向量, 所以shape 必须是数字
"""
init = tf.constant(bias_init, shape=[shape])
return tf.Variable(init, name=name)
def dense_layer(ipt, shape, keep_prob=None, name=None, is_training=None):
assert keep_prob != None and name != None and is_training != None
w = weight_variable(shape)
b = bais_variable(shape[1])
o = tf.matmul(ipt, w) + b
conv_bn = tf.contrib.layers.batch_norm(o,
decay=0.9,
center=True,
scale=True,
updates_collections=None,
is_training=is_training,
reuse=None,
trainable=True,
scope=name)
r = tf.nn.relu(conv_bn)
return tf.nn.dropout(r, keep_prob)
def conv_layer(ipt, filter_shape, conv_name, is_training):
"""
1. 最后有没有必要添加的一个relu ?
2. 所有的stride为 1
注意: name 必须不相同
"""
assert conv_name != None
conv_filter = weight_variable(filter_shape)
conv = tf.nn.conv2d(ipt,
conv_filter,
strides=[1, 1, 1, 1],
padding="SAME")
conv_bn = tf.contrib.layers.batch_norm(conv,
decay=0.9,
center=True,
scale=True,
updates_collections=None,
is_training=is_training,
reuse=None,
trainable=True,
scope=conv_name)
return tf.nn.relu(conv_bn)
def res_block(ipt,
output_channel,
name=None,
down_sample=False,
projection=False,
is_training=None):
"""
由于使用了conv_layer,所以一定需要使用不同name
默认不是down sample 的
1. 首先使用 把 输入的数据down sample
2. 两个conv
3. 两个数据相加起来
"""
assert name != None
input_depth = ipt.get_shape().as_list()[3]
if down_sample:
filter_ = [1, 2, 2, 1]
# 如果使用了 down sample 之后, ipt 被重新赋值, 所以只要深度 down sampeling
ipt = tf.nn.max_pool(ipt,
ksize=filter_,
strides=filter_,
padding='SAME')
# 中间有没有可以修改的机会
conv1 = conv_layer(
ipt, [kernel_size, kernel_size, input_depth, output_channel],
conv_name=name + "conv1",
is_training=is_training)
conv2 = conv_layer(
conv1, [kernel_size, kernel_size, output_channel, output_channel],
conv_name=name + "conv2",
is_training=is_training)
if input_depth != output_channel:
if projection:
# Option B: Projection shortcut
input_layer = conv_layer(ipt,
[1, 1, input_depth, output_channel],
2, name + "convProj")
else:
# Option A: Zero-padding
input_layer = tf.pad(ipt, [[0, 0], [0, 0], [0, 0],
[0, output_channel - input_depth]])
else:
input_layer = ipt
res = conv2 + input_layer
return res
# net 手动搭建, 不需要的什么花里胡哨的东西
# 需要随着随着 iter step 变化 learning rate 的东西
channel = {"res1": 8, "res2": 16, "res3": 32, "res4": 64}
weights = {"dc0": 15 * 15 * channel["res3"], "dc1": 600, "dc2": 100}
def res_net(input_x,
keep_prob,
is_training,
weights=weights,
channel=channel):
"""
x: 训练的数据, 不包含标签
dropout: 添加给dense connected
is_training: bn
使用 res res_block 的 时候的, 只有input channel 和 out channel 的数目不相同的
时候才会有问题
首先使用6组进行简单的测试
1. 要不要使用 full connected 在中间
2. 为什么开始的时候需要添加 一个 conv
3. projection的影响是什么 ?
4. 查看对于奇数的 max_pool 的使用的结果是什么 ?
"""
unique_name = [
"res1", "res2", "res3", "res4", "res5", "res6", "conv1", "full1",
"full2", "full3"
]
# reshape
x = tf.reshape(input_x, shape=[-1, 60, 60, 1])
# conv
conv1 = conv_layer(x, [kernel_size, kernel_size, 1, channel["res1"]],
conv_name=unique_name[6],
is_training=is_training)
# residual
res1 = res_block(conv1,
channel["res1"],
name=unique_name[0],
is_training=is_training)
res2 = res_block(res1,
channel["res1"],
name=unique_name[1],
down_sample=True,
is_training=is_training)
# 针对于现有模型的测试代码, 不具有兼容性
assert res2.get_shape().as_list()[1:] == [30, 30, 8]
res3 = res_block(res2,
channel["res2"],
name=unique_name[2],
is_training=is_training)
res4 = res_block(res3,
channel["res2"],
name=unique_name[3],
is_training=is_training)
assert res4.get_shape().as_list()[1:] == [30, 30, 16]
res5 = res_block(res4,
channel["res3"],
name=unique_name[4],
is_training=is_training)
res6 = res_block(res5,
channel["res3"],
name=unique_name[5],
down_sample=True,
is_training=is_training)
assert res6.get_shape().as_list()[1:] == [15, 15, 32]
# 转换到达普通的full层
# full
dc0 = tf.reshape(res6, shape=[-1, weights["dc0"]])
dc1 = dense_layer(dc0, [weights["dc0"], weights["dc1"]],
keep_prob=keep_prob,
name=unique_name[7],
is_training=is_training)
dc2 = dense_layer(dc1, [weights["dc1"], weights["dc2"]],
keep_prob=keep_prob,
name=unique_name[8],
is_training=is_training)
dc3 = dense_layer(dc2, [weights["dc2"], n_output],
keep_prob=keep_prob,
name=unique_name[9],
is_training=is_training)
return dc3
# place holders
x = tf.placeholder(tf.float32, [None, n_input])
ground_truth = tf.placeholder(tf.float32,
[None, n_output]) # 意味着开始的时候就是需要转化为 one-hot
is_training = tf.placeholder(tf.bool)
keep_prob = tf.placeholder(tf.float32) #dropout (keep probability)
# finish the graph
pred = res_net(x, keep_prob, is_training)
cost = tf.reduce_sum(
tf.nn.softmax_cross_entropy_with_logits(logits=pred,
labels=ground_truth))
optimizer = tf.train.AdamOptimizer(
learning_rate=learning_rate).minimize(cost)
correct_pred = tf.equal(tf.argmax(pred, 1), tf.argmax(ground_truth, 1))
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
tf.summary.scalar("accuracy", accuracy)
tf.summary.scalar("cost", cost)
optimizer = tf.train.AdamOptimizer(
learning_rate=learning_rate).minimize(cost)
init = tf.global_variables_initializer()
# begin the training
with tf.Session() as sess:
merged = tf.summary.merge_all()
sess.run(init)
train_writer = tf.summary.FileWriter(dd.show_data2 + "/train",
sess.graph)
test_writer = tf.summary.FileWriter(dd.show_data2 + "/test")
for i in range(training_iters):
if (i % 50 == 0):
summary, res, acc = sess.run(
[merged, cost, accuracy],
feed_dict={
x:
test_data[:, 0:n_input],
ground_truth:
make_one_hot(test_data[:, n_input:n_input + 1],
n_output),
keep_prob:
1,
is_training:
False
})
train_writer.add_summary(summary, i)
print(acc)
if (i % 200 == 0):
s_train_sample = train_data[0:1000]
s_train_x = s_train_sample[:, 0:n_input]
s_train_y = make_one_hot(
s_train_sample[:, n_input:n_input + 1], n_output)
s_acc, summary = sess.run(
[accuracy, merged],
feed_dict={
x: s_train_x,
ground_truth: s_train_y,
keep_prob: 1,
is_training: False
})
test_writer.add_summary(summary, i)
print("{1} trianing: {0}".format(s_acc, i))
np.random.shuffle(train_data)
train_sample = train_data[0:batch_size]
train_x = train_sample[:, 0:n_input]
train_y = make_one_hot(train_sample[:, n_input:n_input + 1],
n_output)
sess.run(optimizer,
feed_dict={
x: train_x,
ground_truth: train_y,
keep_prob: 0.6,
is_training: True
})
train_writer.close()