def net_atten(input_data, input_pos, reuse=False):
with tf.variable_scope('net', reuse=reuse):
#得到input_data的嵌入
x = embed(input_data, cfg.vocab_size, cfg.vocab_dim)
#得到输入位置input_pos的嵌入
pos = embed(input_pos, 800, cfg.vocab_dim)
#将以上嵌入加在一起然后再做layernorm
x = x + pos
x = normalize(x, 'norm_' + get_uid())
#以下循环只要超过2次就会收敛的很慢
for _ in range(3):
#先预留一个short_cut做residue连接用
short_cut = x
#一个multi head attention 输出
x = mha_block2(x, 64, 8)
#矩阵乘到512,没有用激活函数
x = linear(x, 512, dummy_func)
#给x+short_cut这个residue连接做layernorm
x = normalize(x + short_cut, 'norm_' + get_uid())
#再加两个矩阵乘, 注意第一个有relu激活
ix = linear(x, 512, tf.nn.relu)
lx = linear(ix, 512, dummy_func)
#最后给lx+x这个residue连接做layernorm
x = normalize(lx + x, 'norm_' + get_uid())
#由于我拿声音特征做测试,所以输出为27维, 所以再做一个输出的矩阵乘
x = linear(x, 27, dummy_func)
output = tf.transpose(x)
return output
def __transform(self, input_data, masks):
"""Apply transformer encoder
"""
config = self.config
transformed_output = tf.identity(input_data)
if config.tf_used:
tf_keep_prob = tf.cond(self.is_train, lambda: config.tf_keep_prob,
lambda: 1.0)
tf_mh_keep_prob = tf.cond(self.is_train,
lambda: config.tf_mh_keep_prob,
lambda: 1.0)
tf_ffn_keep_prob = tf.cond(self.is_train,
lambda: config.tf_ffn_keep_prob,
lambda: 1.0)
# last dimension must be equal to model_dim because we use a residual connection.
model_dim = transformed_output.get_shape().as_list()[-1]
# sinusoidal positional signal
signal = positional_encoding(self.sentence_lengths,
self.sentence_length,
model_dim,
zero_pad=False,
scale=False,
scope='positional-encoding',
reuse=None)
transformed_output += signal
# block
for i in range(config.tf_num_layers):
x = transformed_output
# layer norm
x_norm = normalize(x, scope='layer-norm-sa-%s' % i, reuse=None)
# multi-head attention
y = self.__self_attention(x_norm,
masks,
model_dim=model_dim,
keep_prob=tf_mh_keep_prob,
scope='self-attention-%s' % i)
# residual and dropout
x = tf.nn.dropout(x_norm + y, keep_prob=tf_keep_prob)
# layer norm
x_norm = normalize(x,
scope='layer-norm-ffn-%s' % i,
reuse=None)
# position-wise feed forward net
y = self.__feedforward(x_norm,
masks,
model_dim=model_dim,
kernel_size=config.tf_ffn_kernel_size,
keep_prob=tf_ffn_keep_prob,
scope='feed-forward-%s' % i)
# residual and dropout
x = tf.nn.dropout(x_norm + y, keep_prob=tf_keep_prob)
transformed_output = x
# final layer norm
transformed_output = normalize(transformed_output,
scope='layer-norm',
reuse=None)
return transformed_output
def intersect(self, rayO, rayD):
N = -1.0 * ops.normalize(ops.einsum_cross(self.c - self.a, self.b - self.a))
d = np.dot(self.normal, self.a)
t_num = d - (np.dot(self.normal, rayO))
t_dem = np.dot(self.normal, rayD)
t = t_num / t_dem
P = rayO + t * rayD
d = self.intersect_plane(rayO, rayD, P, self.normal)
if d == np.inf:
return np.inf
# check edge 1
e1 = self.b - self.a
vp1 = P - self.a
c1 = ops.einsum_cross(e1, vp1)
if np.dot(N, c1) < 0:
return np.inf
# check edge 2
e2 = self.c - self.b
vp2 = P - self.b
c2 = ops.einsum_cross(e2, vp2)
if np.dot(N, c2) < 0:
return np.inf
# check edge 3
e3 = self.a - self.c
vp3 = P - self.c
c3 = ops.einsum_cross(e3, vp3)
if np.dot(N, c3) < 0:
return np.inf
return d
def simple_save_result_from_test_lb_whole(batch_img_list, i, result, labels, stage, config):
"""
"""
parent_path = os.path.join(config.PARENT_PATH, stage)
if not os.path.isdir(parent_path):
os.mkdir(parent_path)
# save labels
labels_str = ''
for l in labels:
labels_str += (l + '\n')
with open(parent_path + '/label.txt', 'w') as f:
f.write(labels_str)
parent_path = os.path.join(parent_path, 'whole_output')
if not os.path.isdir(parent_path):
os.mkdir(parent_path)
# save the ori
for idx in range(i.shape[0]):
skimage.io.imsave(parent_path + '/ori_' + str(idx) + '.bmp', i[idx])
# save the result
with open(parent_path+'/result.txt', 'w') as f:
f.write(result)
for img in batch_img_list:
tmp_img = img['img']
tmp_label = img['label'].replace('/', '-')
skimage.io.imsave(parent_path+'/'+tmp_label+'.bmp', normalize(tmp_img))
def reflect_transmit_rays(og_rayO, og_rayD, col, reflection,
transmission_depth, max_transmission_depth):
traced = trace_ray(og_rayO, og_rayD)
if not traced:
return False
# reflection: create a new ray
obj, M, N, col_ray = traced
rf_rayO = M + N * 0.001
rf_rayD = ops.normalize(og_rayD - 2 * np.dot(og_rayD, N) * N)
col_rf = reflection * col_ray
transmission_depth += 1
opacity = obj.get_opacity()
transparency = 1.0 - opacity
if transmission_depth <= max_transmission_depth:
# transmission of ray through transparent object
tr_rayO = M - N * 0.001
tf_rayD = og_rayD
transmitted = reflect_transmit_rays(
tr_rayO,
tf_rayD,
transparency * col,
reflection,
transmission_depth,
args.transmission_depth,
)
if transmitted:
_, _, col_tr, _, _ = transmitted
col += transparency * col_tr
col += opacity * col_rf
reflection *= obj.get_reflection()
col = np.clip(col, 0.0, 1.0)
return rf_rayO, rf_rayD, col, reflection, transmission_depth
def trace_ray(rayO, rayD):
t = np.inf
for group in scene:
# print("[debug] Testing intersection with", group)
intersected = False
bounds = scene[group]["bounds"]
t_obj = bounds.intersect(rayO, rayD)
if t_obj < t:
# print("[debug] Intersected with", group)
objects = scene[group]["objects"]
for obj in objects:
# print("[debug] Testing intersection with", obj.type)
t_obj = obj.intersect(rayO, rayD)
if t_obj < t:
# print("[debug] Intersected with", obj.type)
t = t_obj
intersected_object = obj
intersected = True
break
if intersected:
break
if t == np.inf:
return None
obj = intersected_object
# print("[debug]", obj.type)
M = rayO + rayD * t
# get properties of object
N = obj.get_normal(M)
color = obj.get_color(M)
toL = ops.normalize(L - M)
toO = ops.normalize(O - M)
# compute color
col_ray = ambient_c
col_ray += obj.get_diffuse_c() * max(np.dot(N, toL), 0) * color
col_ray += (obj.get_specular_c() * max(np.dot(
N, ops.normalize(toL + toO)), 0)**obj.get_specular_k() * color_light)
return obj, M, N, col_ray
def shade_pixel(x, y, q_z, depth_max):
col = np.zeros(3)
Q = np.array([x, y, q_z])
D = ops.normalize(Q - O)
depth = 0
rayO, rayD = O, D
reflection = 1.0
transmission_depth = 0
while depth < depth_max:
traced = reflect_transmit_rays(rayO, rayD, col, reflection,
transmission_depth,
args.transmission_depth)
if traced:
rayO, rayD, col, reflection, transmission_depth = traced
else:
break
depth += 1
return col
def net(self, input, is_training = True, reuse = False, scope = 'Encoder'):
with tf.variable_scope(scope, 'Encoder', [input], reuse = reuse ) as net_scope:
with slim.arg_scope(model_arg_scope(is_training = is_training, ac_fn = _leaky_relu)):
if reuse:
net_scope.reuse_variables()
#b*32*32
net = slim.conv2d(input, self.nef, [5,5], stride = 2, normalizer_fn = None, scope = 'e_conv_1')
#b*16*16
net = slim.conv2d(net, self.nef*2, [5,5], stride = 2, scope = 'e_conv_2')
#b*8*8
net = slim.conv2d(net, self.nef*4, [5,5], stride = 2, scope = 'e_conv_3')
#b*4*4
net = slim.conv2d(net, self.nef*8, [5,5], stride = 2, scope = 'e_conv_4')
#b*1*1
net = slim.conv2d(net, self.nz, [4,4], stride = 1, padding = 'VALID', normalizer_fn = None, activation_fn = None, scope = 'e_conv_5')
if self.noise == 'sphere':
net = normalize(net)
return net
def getBatch(batch_paths):
batch_images_x = np.empty((len(batch_paths), 256, 256, 3), dtype=np.float32)
batch_images_y = np.empty((len(batch_paths), 256, 256, 3), dtype=np.float32)
for i,img_path in enumerate(batch_paths):
img_xy = cv2.imread(img_path)
img_xy = cv2.resize(img_xy, (512, 256)).astype(np.float32)
img_xy = ops.normalize(img_xy)
img_x = img_xy[:,:256,:]
img_y = img_xy[:,256:,:]
if args.direction == 'ytox':
img_x, img_y = img_y, img_x
batch_images_x[i, ...] = img_x
batch_images_y[i, ...] = img_y
return tf.convert_to_tensor(batch_images_x), tf.convert_to_tensor(batch_images_y)
def simple_save_result_from_test_lb(batch_img_list, result, labels, stage, config, output_layer):
"""Simple save the result.
"""
if '/' in output_layer:
raise ValueError('output_layer should not have /, but %s'%output_layer)
parent_path = os.path.join(config.PARENT_PATH, stage)
if not os.path.isdir(parent_path):
os.mkdir(parent_path)
# save labels
labels_str = ''
for l in labels:
labels_str += (l + '\n')
with open(parent_path + '/label.txt', 'w') as f:
f.write(labels_str)
with open(parent_path + '/result.txt', 'w') as f:
f.write(result)
parent_path = os.path.join(parent_path, output_layer)
if not os.path.isdir(parent_path):
os.mkdir(parent_path)
tmp_count = 0
for single_batch in batch_img_list:
tmp_count += 1
tmp_path = os.path.join(parent_path, str(tmp_count))
if not os.path.isdir(tmp_path):
os.mkdir(tmp_path)
ori_img = single_batch[1]
skimage.io.imsave(tmp_path + '/ori.bmp', ori_img)
tmp_count_inner = 0
for single_kernel_lb in single_batch[0]:
tmp_count_inner += 1
skimage.io.imsave(tmp_path + '/' + single_kernel_lb['label'].replace('/', '-') + str(tmp_count_inner) + '.bmp',
normalize(single_kernel_lb['img']))
def assertNormalizes(self, src, dst, ntype):
value = ops.normalize(src, type=ntype)
message = '%s != %s' % (type(value).__name__, type(dst).__name__)
self.assertTrue(isinstance(value, type(dst)), message)
self.assertEqual(value, dst)
'''
h_pool1 = nn_ops.conv2d_transpose(h_conv2 + b_conv2,W_conv2,
[class_size,14,14,conv1_size],[1,1,1,1])
'''
index = 0
dir = "l1f/"
for t in W_conv1_t.eval():
with open(dir + 'filter' + str(index) + '.png', "wb") as file:
t = tf.constant(t)
#t = tf.expand_dims(tf.constant(t), 0)
#t_n = tf.squeeze(nn_ops.conv2d_transpose(t, W_conv1, [1,5,5,1],[1,1,1,1]), [0])
t = tf.image.resize_images(t,
50,
50,
method=tf.image.ResizeMethod.BICUBIC)
t = tf.constant(ops.normalize(t.eval(), 0, 255))
file.write(tf.image.encode_png(t).eval())
index += 1
W_conv2_t = tf.transpose(W_conv2, perm=[3, 0, 1, 2])
index = 0
dir = "l2f/"
#for t in W_conv1_t.eval():
for t in W_conv2_t.eval():
with open(dir + 'filter' + str(index) + '.png', "wb") as file:
t = tf.expand_dims(tf.constant(t), 0)
t_n = tf.squeeze(
nn_ops.conv2d_transpose(t, W_conv1, [1, 5, 5, 1], [1, 1, 1, 1]),
[0])
t_n = tf.image.resize_images(t_n,
def compute_normal(self):
return ops.normalize(ops.einsum_cross(self.c - self.a, self.b - self.a))
def simple_gray_nor_imshow(img): plt.imshow(normalize(img), cmap ='gray')
def assertNormalizes(self, src, dst, ntype):
value = ops.normalize(src, type=ntype)
message = '%s != %s' % (type(value).__name__, type(dst).__name__)
self.assertTrue(isinstance(value, type(dst)), message)
self.assertEqual(value, dst)
def initialize(m_opts):
m_vars = {}
np.random.seed(m_opts['random_state'])
data = loadmat(m_opts['dataset'])
print "Dataset loaded: ", m_opts['dataset']
m_vars['Y_train'] = sparsify(data['Y_tr'])
m_vars['X_train'] = sparsify(data['X_tr'])
m_vars['Y_test'] = sparsify(data['Y_te'])
m_vars['X_test'] = sparsify(data['X_te'])
print "Training data -- Y:", m_vars['Y_train'].shape, " X:", m_vars[
'X_train'].shape
print "Testing data -- Y:", m_vars['Y_test'].shape, "X: ", m_vars[
'X_test'].shape
m_vars['n_users'], m_vars['n_labels'] = m_vars['Y_train'].shape
m_vars['n_features'] = m_vars['X_train'].shape[1]
if m_opts['label_normalize']:
normalize(m_vars['Y_train'], norm='l2', axis=1, copy=False)
if m_opts['no_feat_normalize'] == False:
normalize(m_vars['X_train'], norm='l2', axis=1, copy=False)
normalize(m_vars['X_test'], norm='l2', axis=1, copy=False)
# m_vars['U'] = m_opts['init_std']*np.random.randn(m_vars['n_users'], m_opts['n_components']).astype(floatX)
m_vars['U_batch'] = np.zeros(
(m_opts['batch_size'], m_opts['n_components'])).astype(floatX)
m_vars['V'] = m_opts['init_std'] * np.random.randn(
m_vars['n_labels'], m_opts['n_components']).astype(floatX)
m_vars['W'] = m_opts['init_w'] * np.random.randn(
m_opts['n_components'], m_vars['n_features']).astype(floatX)
# accumulator of sufficient statistics of label factors
m_vars['sigma_v'] = [None] * m_vars['n_labels']
for i in range(m_vars['n_labels']):
m_vars['sigma_v'][i] = m_opts['lam_v'] * ssp.eye(
m_opts['n_components'], format="csr")
m_vars['x_v'] = np.zeros((m_vars['n_labels'], m_opts['n_components']))
if not m_opts['use_grad']:
# accumulator of sufficient statistics of W matrix
m_vars['sigma_W'] = m_opts['lam_w'] * ssp.eye(
m_vars['n_features'], m_vars['n_features'], format="csr")
m_vars['x_W'] = np.zeros(
(m_vars['n_features'], m_opts['n_components']))
if m_opts['observance']:
m_vars['a'], m_vars['b'] = m_opts['init_mu_a'], m_opts['init_mu_b']
# Beta random initialization
m_vars['mu'] = np.random.beta(m_vars['a'],
m_vars['b'],
size=(m_vars['n_labels']))
# constant initialization
# m_vars['mu'] = m_opts['init_mu']*np.ones(m_vars['n_labels']).astype(floatX)
else:
m_vars['mu'] = np.ones(m_vars['n_labels']).astype(floatX)
return m_vars
def get_normal(self, coords):
return -1.0 * ops.normalize(coords - self.center)
train_step.run(feed_dict={x: batch[0], y_: batch[1], keep_prob: 0.5})
'''
def t_map(func, tensor):
flat = tf.reshape(tensor, [-1]).eval()
flat_mapped = list(map(func, flat))
return tf.reshape(tf.constant(flat_mapped), tensor.get_shape())
def normalize(tensor, lower, upper):
minVal = tf.reduce_min(tensor).eval()
maxVal = tf.reduce_max(tensor).eval()
translated = t_map(lambda x: x+minVal-lower, tensor)
scaled = t_map(lambda x: numpy.uint8(x*(upper-lower)/(maxVal-minVal)), translated)
#print(tf.reduce_min(scaled).eval(),tf.reduce_max(scaled).eval())
return scaled
'''
W_conv1_t = tf.transpose(W_conv1, perm=[3,0,1,2])
index = 0
dir = "filters/"
for t in W_conv1_t.eval():
with open(dir+'filter'+str(index)+'.png', "wb") as file:
t = tf.constant(t)
t_n = tf.image.resize_images(t,50,50, method=tf.image.ResizeMethod.BICUBIC)
t_n = tf.constant(ops.normalize(t_n.eval(), 0, 255))
#t_n = t_map(lambda x: numpy.uint8(x), t_n)
file.write(tf.image.encode_png(t_n).eval())
index += 1
#for t in W_conv1.eval():
def extract_vgg_face(self, inputs):
inputs = normalize((F.hardtanh(inputs) * 0.5 + 0.5) * 255,
[129.1863, 104.7624, 93.5940], [1.0, 1.0, 1.0])
return self.vgg_face(inputs)
