def generate(self, z=None):
with tf.name_scope('Generate'):
if self.prior == 'gaussian':
if z is None:
z = np.random_normal((self.batch_size, self.latent_dim))
return self.sess.run(self.x_mean, feed_dict={self.z: z})
elif self.prior == 'gmm':
if z is None:
targets = (self.gmm_pi, self.gmm_mu, self.gmm_log_var)
pis, means, log_vars = self.sess.run(targets)
pis /= pis.sum()
cluster = np.random.choice(range(self.num_clusters), p=pis)
mean = means[cluster]
std = np.sqrt(np.exp(log_vars[cluster]))
eps = np.random_normal(std.shape)
z = mean + std * eps
return self.sess.run(self.x_mean, feed_dict={self.z: z})
def init_params():
w = nd.random_normal(scale=1, shape=(num_inputs, 1))
b = nd.zeros(shape=(1, ))
params = [w, b]
vs = []
sqrs = []
for param in params:
param.attach_grad()
# 把算法中基于指数加权移动平均的变量初始化为和参数形状相同的零张量。
vs.append(param.zeros_like())
sqrs.append(param.zeros_like())
return params, vs, sqrs
def __init__(self, M1 = 0, M2 = 0, f = tf.nn.tanh, use_bias = True, zeros = False, copy_from_ESwNN = False, W_E = 0, b_E = 0):
self.use_bias = use_bias
if not copy_from_ESwNN:
if zeros:
W = np.zeros((M1, M2)).astype(np.float32)
else:
W = np.random_normal(shape = (M1, M2))
if use_bias:
b = np.zeros(M2).astype(np.float32)
else:
W = np.array(W_E).astype(np.float32)
b = np.array(b_E).astype(np.float32)
print(W)
print("\n\n")
print(b)
print("\n\n\n\n")
self.W = tf.Variable(W)
self.b = tf.Variable(b)
self.params = [self.W, self.b]
self.f = f
def __getitem__(self, idx):
# 1. Random select a video
# 2. slice the video frames into batches and select one batch([l_bound, r_bound])
selected_video_num = np.random.random_integers(
0,
len(self.annotations) - 1)
num_frames = sum(
1 for line in open(self.annotations[selected_video_num]))
num_batch = num_frames / self.config['BATCH_SIZE']
l_bound = (idx % num_batch) * self.config['BATCH_SIZE']
r_bound = (idx % num_batch + 1) * self.config['BATCH_SIZE']
if r_bound - 1 + self.config['TIME_STEP'] > num_frames:
r_bound = num_frames - self.config['TIME_STEP'] + 1
l_bound = r_bound - self.config['BATCH_SIZE']
if l_bound < 0:
raise Exception(
"Number of frames in every video must be more than batch size ( > %d )"
% self.config['BATCH_SIZE'])
if (l_bound + self.config['BATCH_SIZE'] - 1 +
self.config['TIME_STEP'] - 1) > num_frames - 1:
l_bound = num_frames - 1 - (self.config['BATCH_SIZE'] - 1 +
self.config['TIME_STEP'] - 1)
r_bound = l_bound + self.config['BATCH_SIZE']
labels = np.loadtxt(self.annotations[selected_video_num],
delimiter=',')
##########################################################################
# Make sure labels are not NAN
##########################################################################
while isNAN(labels[l_bound:(r_bound - 1 + self.config['TIME_STEP']),
...]):
# print("\nGround truth is Nan, choose another batch")
l_bound = (np.random.random_integers(0, 1000) %
num_batch) * self.config['BATCH_SIZE']
r_bound = l_bound + self.config['BATCH_SIZE']
if r_bound - 1 + self.config['TIME_STEP'] > num_frames:
r_bound = num_frames - self.config['TIME_STEP'] + 1
l_bound = r_bound - self.config['BATCH_SIZE']
if l_bound < 0:
raise Exception(
"Number of frames in every video must be more than batch size ( > %d )"
% self.config['BATCH_SIZE'])
# print("l_bound:", l_bound)
# print("r_bound:", r_bound)
# print("#frame:", num_frames)
# print('#batch:', num_batch)
detections = np.load(self.detected_label_namelist[selected_video_num])
features = np.load(self.features_namelist[selected_video_num])
# TODO
''' x_center, y_center, w, h (normalized)
'''
labels[:, 0] = (labels[:, 0] + labels[:, 2] / 2.0) / 1280
labels[:, 1] = (labels[:, 1] + labels[:, 3] / 2.0) / 720
labels[:, 2] = labels[:, 2] / 1280
labels[:, 3] = labels[:, 3] / 720
# Make input data
if isinstance(self.config['INPUT_SIZE'], list):
x_batch = np.zeros(
(self.config['BATCH_SIZE'], self.config['TIME_STEP'],
self.config['INPUT_SIZE'][0], self.config['INPUT_SIZE'][1],
self.config['INPUT_SIZE'][2]))
bbox_batch = np.zeros(
(self.config['BATCH_SIZE'], self.config['TIME_STEP'], 4))
# detection_batch = np.zeros((self.config['BATCH_SIZE'], self.config['GRID_H'], self.config['GRID_W'], self.config['BOX'], 4+1+self.config['CLASS']+1))
# b_batch = np.zeros((self.config['BATCH_SIZE'], 1, 1, 1, 1, 4)) # list of self.config['TRUE_self.config['BOX']_BUFFER'] GT boxes
# feat_batch = np.zeros((self.config['BATCH_SIZE'], self.config['INPUT_SIZE'][0], self.config['INPUT_SIZE'][1], self.config['INPUT_SIZE'][2]))
y_batch = np.zeros(
(self.config['BATCH_SIZE'], self.config['TIME_STEP'], 4))
# y_batch = np.zeros((self.config['BATCH_SIZE'], self.config['GRID_H'], self.config['GRID_W'], self.config['BOX'], 4+1+self.config['CLASS']+1))
# bbox_score_batch = np.zeros((self.config['BATCH_SIZE'], self.config['TIME_STEP'], self.config['INPUT_SIZE'][0], self.config['INPUT_SIZE'][1], 1))
else:
x_batch = np.zeros(
(self.config['BATCH_SIZE'], self.config['TIME_STEP'],
self.config['INPUT_SIZE']))
y_batch = np.zeros(
(self.config['BATCH_SIZE'], self.config['TIME_STEP'], 4))
instance_count = 0
for i in range(l_bound, r_bound):
# Every instance in every batch contains #time_step images
for j in range(self.config['TIME_STEP']):
detection = detections[i + j, ...]
feature = features[i + j, ...]
label = labels[i + j, ...]
# print("Detection shape:", detection.shape)
# print("Feature shape:", feature.shape)
# print("Detection:", detection)
# inputs = detection
# inputs = np.concatenate((feature.flatten(), detection))
# inputs = feature
# print("Input shape:", inputs.shape)
x_batch[instance_count, j, ...] = feature
if isNAN(detection):
# When no detection results in some frame, use the initial box plus some Gaussian random normals as detection results
print("No detection results, use auxiliary bbox.")
detection = detections[0, ...]
for value in detection:
tmp = value + np.random_normal(0, 0.5)
while tmp < 0 or tmp > 1:
tmp = value + np.random_normal(0, 0.5)
value = tmp
bbox_batch[instance_count, j, :] = detection
y_batch[instance_count, j, :] = label
if j == self.config['TIME_STEP'] - 1:
# detection = detections[i+j, ...]
# if isNAN(detection):
# detection = detections[0, ...]
# for value in detection:
# tmp = value + np.random_normal(0, 0.5)
# while tmp < 0 or tmp > 1:
# tmp = value + np.random_normal(0, 0.5)
# value = tmp
# print(detection)
# bbox_batch[instance_count, :] = detection
# label = labels[i+j, ...]
if isNAN(label):
raise ValueError("Label is nan!")
for value in label:
if value < 0:
print(label)
raise ValueError("Label value should > 0")
# feat_batch[instance_count, ...] = feature
# y_batch[instance_count, :] = label
# find the anchor that best predicts this box
# print("Label:", label)
instance_count += 1
if isNAN(x_batch):
raise ValueError("X batch NAN!!!!!!!!!!!!!1")
# if isNAN(bbox_batch):
# raise ValueError("BBOX batch NAN!!!!!!!!!!!!!1")
# print ' new batch created', idx
# print(feat_batch.shape)
return [x_batch, bbox_batch], y_batch
def generate(self, hidden = None):
if hidden is None:
hidden = np.random_normal(size = self.weights["b1"])
return self.sess.run(self.reconstruct, feed_dict = {self.hidden: hidden})
n_epochs = 2
batch_size = 150
def shuffle_batch(X, y, batch_size):
rnd_idx = np.random.permutation(len(X))
n_batches = len(X) // batch_size
for batch_idx in np.array_split(rnd_idx, n_batches):
X_batch, y_batch = X[batch_idx], y[batch_idx]
yield X_batch, y_batch
def plot_image(image, shape=[28, 28]):
plt.imshow(image.reshape(shape), cmap="Greys", interpolation="nearest")
plt.axis("off")
with tf.Session() as sess:
init.run()
for epoch in range(n_epochs):
n_batches = len(X_train)
print(epoch)
for iteration in range(n_batches):
X_batch, y_batch = next(shuffle_batch(X_train, y_train,
batch_size))
sess.run(training_op, feed_dict={X: X_batch})
codings_rnd = np.random_normal(size=[n_digits, n_hidden3])
outputs_val = outputs.eval(feed_dict={hidden3: codings_rnd})
for iteration in range(n_digits):
plt.subplot(n_digits, 10, iteration + 1)
plot_image(outputs_val[iteration])
def generate_data(self, z_mu=None):
if z_mu is None:
z_mu = np.random_normal(size=self.network_arch["n_z"])
return self.sess.run(self.x_reconstruction_mean, feed_dict={self.z_mean: z_mu})
#激励函数解决不能用线性方程y = Wx解决的问题
# y = AF(Wx) 掰弯利器 激励函数必须是可微分的
import tensorflow as tf
import numpy as np
#in_size表示输入多少个单位代表行 out_size代表列
def add_layer(inputs, in_size, out_size, activation_function=None):
Weights = tf.Variable(tf.random_normal([in_size, out_size]) )
biases = tf.Variable(tf.zores[1, out_size] + 0.1)
Wx_plus_b = tf.matmul(inputs, Weights) + biases
if activation_function is None:
outputs = Wx_plus_b
else:
outputs = activation_function(Wx_plus_b)
return outputs
x_data = np.linspace(-1, 1, 300)[:, np.newaxis]
noise = np.random_normal(0, 0.05, x_data.shape)
y_data = np.square(x_data) - 0.5 + noise
xs = tf.placeholder([None, 1])
ys =
sqr_bias_corr = sqr / (1. - beta2**t)
div = lr * v_bias_corr / (nd.sqrt(sqr_bias_corr) + eps_stable)
param[:] = param - div
if __name__ == '__main__':
import random
import numpy as nd
random.seed(1)
# 生成数据集。
num_inputs = 2
num_examples = 1000
true_w = [2, -3.4]
true_b = 4.2
X = nd.random_normal(scale=1, shape=(num_examples, num_inputs))
y = true_w[0] * X[:, 0] + true_w[1] * X[:, 1] + true_b
y += .01 * nd.random_normal(scale=1, shape=y.shape)
dataset = gluon.data.ArrayDataset(X, y)
# 构造迭代器。
import random
def data_iter(batch_size):
idx = list(range(num_examples))
random.shuffle(idx)
for batch_i, i in enumerate(range(0, num_examples, batch_size)):
j = nd.array(idx[i:min(i + batch_size, num_examples)])
yield batch_i, X.take(j), y.take(j)
# 初始化模型参数。
def init_weights(shape):
return tf.Variable(np.random_normal(shape, stddev=0.01))
import tensorflow as tf
import numpy as np
x_zero = np.random.multivariate_normal(mean=(-1, -1),
cov=.1 * np.eye(2),
size=(50, ))
y_zero = np.zeros(50)
x_one = np.random.multivariate_normal(mean=(1, 1),
cov=.1 * np.eye(2),
size=(50, ))
y_one = np.ones(50)
x = np.concatenate([x_zero, x_one])
y = np.concatenate([y_zero, y_one])
x_input = tf.placeholder(dtype=float32, shape=(100, 2))
y_input = tf.placeholder(dtype=float32, shape=(100, ))
w = tf.Variable(np.random_normal(size=(2, 1)))
b = tf.Variable(tf.zeros((1, )))
