def __init(self,
name,
kernel_shape,
bias_shape,
padding,
stride=1,
activation_func=None):
NetworkLayer.__init__(self, name)
with tf.variable_scope(name):
self.weights = tf.get_variable(
'w',
kernel_shape,
initializer=layers.xavier_initializer_conv2d(uniform=False))
self.bias = tf.get_variable(
'b', bias_shape, initializer=tf.random_normal_initializer)
assert self.weights.name == name + '/w:0' and self.bias.name == name + '/b:0'
tf.add_to_collection(self.weights.name, self.weights)
tf.add_to_collection(self.bias.name, self.bias)
self.__activation_func = activation_func
self.__padding = padding
self.__stride = stride
def encoder(self, x):
shapes = []
n_hidden = [1] + self.n_hidden
input = x
for i, k_size in enumerate(self.kernel_size):
w = tf.get_variable(
'enc_w{}'.format(i),
shape=[k_size, k_size, n_hidden[i], n_hidden[i + 1]],
initializer=layers.xavier_initializer_conv2d(),
regularizer=self.reg)
b = tf.get_variable('enc_b{}'.format(i),
shape=[n_hidden[i + 1]],
initializer=tf.zeros_initializer())
shapes.append(input.get_shape().as_list())
enc_i = tf.nn.conv2d(input,
w,
strides=[1, 2, 2, 1],
padding='SAME')
enc_i = tf.nn.bias_add(enc_i, b)
enc_i = tf.nn.relu(enc_i)
input = enc_i
return input, shapes
def conv2d(input, filter_shape, strides=2, padding='SAME', name='conv'):
with tf.variable_scope(name):
w = tf.get_variable(name='w',
shape=filter_shape,
dtype=tf.float32,
initializer=xavier_initializer_conv2d())
tf.summary.histogram('w_' + name, w)
b = tf.get_variable(name='b',
shape=[filter_shape[-1]],
dtype=tf.float32,
initializer=tf.zeros_initializer())
tf.summary.histogram('b_' + name, b)
conv = tf.nn.conv2d(input=input,
filter=w,
strides=[1, strides, strides, 1],
padding=padding,
use_cudnn_on_gpu=True)
return lrelu(conv + b, 0.02)
def __init__(self, label_size, network_architecture, activation=tf.nn.relu, max_grad_norm=1,
learning_rate=0.001, batch_size=100, save_path=None, load_model=None):
"""
:param dict network_architecture: dictionary with following elements
n_input: shape of input
n_z: dimensionality of latent space
:param activation: activation function (tensor flow function)
:param float learning_rate:
:param int batch_size:
"""
self.network_architecture = network_architecture
self.max_grad_norm = max_grad_norm
self.activation = activation
self.learning_rate = learning_rate
self.batch_size = batch_size
self.label_size = label_size
# Initializer
if "relu" in self.activation.__name__:
self.ini_c, self.ini = variance_scaling_initializer(), variance_scaling_initializer()
else:
self.ini_c, self.ini = xavier_initializer_conv2d(), xavier_initializer()
# Create network
self._create_network()
# Summary
tf.summary.scalar("loss", self.loss)
# Launch the session
self.sess = tf.Session(config=tf.ConfigProto(log_device_placement=False))
# Summary writer for tensor board
self.summary = tf.summary.merge_all()
if save_path:
self.writer = tf.summary.FileWriter(save_path, self.sess.graph)
# Load model
if load_model:
tf.reset_default_graph()
self.saver.restore(self.sess, load_model)
def gen_nlayer_conv_decoder(X, decoder_para):
'''
generate a n-layer convolutional encoder
@encoder_para: dict, parameters of method, including:
@@kernal_size: list, shape=[n:4], [[height, width, output_channel, input_channel], ... ];
notice that elements in kernal_size has the same order with that in encoder. That is to say
input_channel of encoder is the output_channel of decoder.
@@strides: list, shape=[n:4], [[batch_stride, height_stride, width_stride, channel_stride], ...]
@@output_size: list, shape=[4:], [batch, height, width, channel]
'''
kernal_size = decoder_para['kernal_size']
strides = decoder_para['strides']
output_size = decoder_para['output_size']
last_layer = X
n_layers = len(kernal_size)
for l_idx in xrange(n_layers):
knl_w = tf.get_variable('dec_w_%d'%l_idx, shape=kernal_size[n_layers - l_idx - 1],\
initializer=layers.xavier_initializer_conv2d())
knl_b = tf.Variable(tf.zeros([kernal_size[n_layers - l_idx - 1][-2]]))
stride = strides[n_layers - l_idx - 1]
#calculate the output shape
last_ly_shape = last_layer.get_shape().as_list()
out_height = last_ly_shape[1] * stride[1] if l_idx!=n_layers-1 else output_size[1]
out_width = last_ly_shape[2] * stride[2] if l_idx!=n_layers-1 else output_size[2]
out_channel = kernal_size[n_layers - l_idx - 1][-2] if l_idx!=n_layers-1 else output_size[3]
#generate last layer
last_layer = tf.nn.bias_add(tf.nn.conv2d_transpose(last_layer, knl_w, \
tf.stack([output_size[0], out_height, out_width, out_channel]), strides=stride, padding='SAME'),\
knl_b)
last_layer = tf.nn.relu(last_layer)
return last_layer
def gen_conv_encoder(encoder_para):
'''
generate a one layer convolutional encoder
@encoder_para: dict, parameters of method, including:
@@img_size: np.ndarray, shape=[4:], the size of input image
@@kernal_size: np.ndarray, shape=[4:], [width, height, input_channel, output_channel]
@@strides: np.ndarray, shape=[4:], [batch_stride, width_stride, height_stride, channel_stride]
'''
img_size = encoder_para['img_size']
kernal_size = encoder_para['kernal_size']
strides = encoder_para['strides']
# generate X and weight
X = tf.placeholder(tf.float32,
img_size) #batch_size, img_width, img_height, channel
enc_knl = tf.get_variable("enc_knl", shape=kernal_size, \
initializer=layers.xavier_initializer_conv2d())
enc_b = tf.Variable(tf.zeros([kernal_size[-1]], dtype=tf.float32))
#generate conv layer
conv_layer = tf.nn.conv2d(X, enc_knl, strides, padding='SAME')
conv_layer = tf.nn.relu(tf.nn.bias_add(conv_layer, enc_b))
return conv_layer, X
def conv2d(input_node,
channels_n,
level,
is_training,
stride=1,
activation=True):
input_shape = input_node.get_shape().as_list()
assert len(input_shape) == 4, 'Tensor with rank 4 is expected.'
in_channels_n = input_shape[3]
with tf.variable_scope(f'conv_{level}'):
Wconv1 = tf.get_variable(f"W_{level}",
shape=[3, 3, in_channels_n, channels_n],
initializer=xavier_initializer_conv2d())
bconv1 = tf.get_variable(f"b_{level}",
shape=[channels_n],
initializer=tf.zeros_initializer)
conv = tf.nn.conv2d(input_node,
Wconv1,
strides=[1, stride, stride, 1],
padding='SAME') + bconv1
batch_norm = tf.layers.batch_normalization(conv,
training=is_training)
output = tf.nn.relu(batch_norm) if activation else batch_norm
return output
def conv2d(self,
inputs,
filters=16,
kernel_size=3,
strides=1,
padding='VALID',
activation=None,
kernel_initializer=layers.xavier_initializer_conv2d(),
bias_initializer=tf.constant_initializer(0.2),
kernel_regularizer=layers.l2_regularizer(2e-4),
use_bias=True,
name="conv"):
return tf.layers.conv2d(inputs,
filters=filters,
kernel_size=kernel_size,
strides=strides,
padding=padding,
data_format="channels_first",
activation=activation,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer,
bias_initializer=bias_initializer,
use_bias=use_bias,
name=name)
def __init_model__(self, input_shape):
self.states = tf.placeholder(tf.float32, [None] + list(input_shape),
name='states')
self.actions = tf.placeholder(tf.float32, [None, 1], name='actions')
self.weights = tf.placeholder(tf.float32, [None, 1], name='weights')
self.learning_rate = tf.placeholder(tf.float32, name='learning_rate')
self.beta = tf.placeholder(tf.float32, name='beta')
conv1 = tf.layers.conv2d(
inputs=self.states,
filters=64,
kernel_size=8,
strides=4,
data_format='channels_first',
activation=tf.nn.tanh,
kernel_initializer=xavier_initializer_conv2d())
h1 = tf.layers.dense(inputs=tf.layers.flatten(conv1),
units=128,
activation=tf.nn.tanh,
kernel_initializer=xavier_initializer())
self.prob_1 = tf.layers.dense(inputs=h1,
units=1,
activation=tf.sigmoid,
kernel_initializer=xavier_initializer())
self.prob_1 = self.prob_1 * 0.9998 + 0.0001
prob_0 = 1 - self.prob_1
entropy = self.prob_1 * tf.log(1 / self.prob_1) + prob_0 * tf.log(
1 / prob_0)
regularizer = tf.reduce_mean(entropy)
loss = tf.losses.log_loss(labels=self.actions,
predictions=self.prob_1,
weights=self.weights)
optimizer = tf.train.AdamOptimizer(learning_rate=self.learning_rate)
self.train_op = optimizer.minimize(loss - self.beta * regularizer)
def construct_conv_weights(self):
weights = {}
dtype = tf.float32
conv_initializer = contrib_layers.xavier_initializer_conv2d(dtype=dtype)
fc_initializer = contrib_layers.xavier_initializer(dtype=dtype)
k = 3
weights['conv1'] = tf.get_variable('conv1', [k, k, self.channels, self.dim_hidden], initializer=conv_initializer, dtype=dtype)
weights['b1'] = tf.Variable(tf.zeros([self.dim_hidden]))
weights['conv2'] = tf.get_variable('conv2', [k, k, self.dim_hidden, self.dim_hidden], initializer=conv_initializer, dtype=dtype)
weights['b2'] = tf.Variable(tf.zeros([self.dim_hidden]))
weights['conv3'] = tf.get_variable('conv3', [k, k, self.dim_hidden, self.dim_hidden], initializer=conv_initializer, dtype=dtype)
weights['b3'] = tf.Variable(tf.zeros([self.dim_hidden]))
weights['conv4'] = tf.get_variable('conv4', [k, k, self.dim_hidden, self.dim_hidden], initializer=conv_initializer, dtype=dtype)
weights['b4'] = tf.Variable(tf.zeros([self.dim_hidden]))
if FLAGS.datasource == 'miniimagenet' or FLAGS.datasource == 'dclaw':
# assumes max pooling
weights['w5'] = tf.get_variable('w5', [self.dim_hidden*5*5, self.dim_output], initializer=fc_initializer)
weights['b5'] = tf.Variable(tf.zeros([self.dim_output]), name='b5')
else:
weights['w5'] = tf.Variable(tf.random_normal([self.dim_hidden, self.dim_output]), name='w5')
weights['b5'] = tf.Variable(tf.zeros([self.dim_output]), name='b5')
return weights
def encoder(self, imgs, activation, is_training, batch_size, img_shape,
channels):
with tf.variable_scope('encoder',
reuse=tf.AUTO_REUSE,
initializer=xavier_initializer_conv2d(),
regularizer=l2_regularizer(0.01)):
# stn >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>
n_fc = 6
initial = np.array([[1., 0, 0], [0, 1., 0]])
initial = initial.astype('float32').flatten()
W_fc1 = tf.Variable(
tf.zeros(shape=[img_shape * img_shape * channels, n_fc]),
name='W_fc1',
validate_shape=False)
b_fc1 = tf.Variable(initial_value=initial, name='b_fc1')
h_fc1 = tf.matmul(
tf.zeros([batch_size, img_shape * img_shape * channels]),
W_fc1) + b_fc1
h_trans = transformer(imgs, h_fc1)
# stn <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<
e_conv = self.coord_conv(h_trans,
48,
3,
padding='same',
activation=None)
e_conv = tf.layers.batch_normalization(e_conv,
training=is_training,
fused=True)
e_conv = activation(e_conv)
e_conv = tf.layers.max_pooling2d(e_conv, 2, 2)
e_conv = self.coord_conv(e_conv,
92,
3,
padding='same',
activation=None)
e_conv = tf.layers.batch_normalization(e_conv,
training=is_training,
fused=True)
e_conv = activation(e_conv)
e_conv = tf.layers.max_pooling2d(e_conv, 2, 2)
e_conv = self.coord_conv(e_conv,
256,
3,
padding='same',
activation=None)
e_conv = tf.layers.batch_normalization(e_conv,
training=is_training,
fused=True)
e_conv = activation(e_conv)
e_conv = tf.layers.max_pooling2d(e_conv, 2, 2)
e_conv = self.coord_conv(e_conv,
256,
3,
padding='same',
activation=None)
e_conv = tf.layers.batch_normalization(e_conv,
training=is_training,
fused=True)
e_conv = activation(e_conv)
e_conv = tf.layers.max_pooling2d(e_conv, 2, 2)
e_conv = self.coord_conv(e_conv,
256,
3,
padding='same',
activation=None)
e_conv = tf.layers.batch_normalization(e_conv,
training=is_training,
fused=True)
e_conv = activation(e_conv)
e_conv = tf.layers.max_pooling2d(e_conv, 2, 2)
e_conv = self.coord_conv(e_conv,
256,
3,
padding='same',
activation=None)
e_conv = tf.layers.batch_normalization(e_conv,
training=is_training,
fused=True)
e_conv = activation(e_conv)
e_conv = tf.layers.max_pooling2d(e_conv, 2, 2)
lv = tf.layers.flatten(e_conv)
return lv
def __init__(self, params, pretrained_model_path=""):
self.params = params
self.pretrained_model_path = pretrained_model_path
dicts = helper.load_dictionaries(self.params.dicts_file)
self.word2id = dicts["word2id"]
self.id2word = dicts["id2word"]
self.char2id = dicts["char2id"]
self.id2char = dicts["id2char"]
self.tag2id = dicts["tag2id"]
self.id2tag = dicts["id2tag"]
self.pretrained_emb = np.zeros(shape=(len(self.word2id),
self.params.word_dim))
if self.pretrained_model_path == "" and self.params.train != "" and self.params.pretrained_emb != "":
self.pretrained_emb = helper.load_word_emb(
self.word2id, self.pretrained_emb, self.params.pretrained_emb)
# build model
self.tf_word_ids = tf.placeholder(dtype=tf.int32,
shape=[None, None],
name="word_ids")
self.tf_sentence_lengths = tf.placeholder(dtype=tf.int32,
shape=[None],
name="sentence_lengths")
self.tf_labels = tf.placeholder(dtype=tf.int32,
shape=[None, None],
name="labels")
self.tf_dropout = tf.placeholder(dtype=tf.float32,
shape=[],
name="drop_out")
self.tf_learning_rate = tf.placeholder(dtype=tf.float32,
shape=[],
name="learning_rate")
self.tf_char_ids = tf.placeholder(dtype=tf.int32,
shape=[None, None, None],
name="char_ids")
self.tf_word_lengths = tf.placeholder(dtype=tf.int32,
shape=[None, None],
name="word_lengths")
self.tf_raw_word = tf.placeholder(dtype=tf.string,
shape=[None, None],
name="raw_word")
with tf.variable_scope("word_embedding"):
tf_word_embeddings = tf.Variable(self.pretrained_emb,
dtype=tf.float32,
trainable=True,
name="word_embedding")
self.input = tf.nn.embedding_lookup(tf_word_embeddings,
self.tf_word_ids,
name="embedded_words")
with tf.variable_scope("char_cnn"):
tf_char_embeddings = tf.get_variable(
name="char_embeddings",
dtype=tf.float32,
shape=[len(self.char2id), self.params.char_dim],
trainable=True,
initializer=xavier_initializer())
embedded_cnn_chars = tf.nn.embedding_lookup(
tf_char_embeddings,
self.tf_char_ids,
name="embedded_cnn_chars")
conv1 = tf.layers.conv2d(
inputs=embedded_cnn_chars,
filters=self.params.nb_filters_1,
kernel_size=(1, 3),
strides=(1, 1),
padding="same",
name="conv1",
kernel_initializer=xavier_initializer_conv2d())
conv2 = tf.layers.conv2d(
inputs=conv1,
filters=self.params.nb_filters_2,
kernel_size=(1, 3),
strides=(1, 1),
padding="same",
name="conv2",
kernel_initializer=xavier_initializer_conv2d())
char_cnn = tf.reduce_max(conv2, axis=2)
self.input = tf.concat([self.input, char_cnn], axis=-1)
with tf.variable_scope("elmo_emb"):
elmo = hub.Module("/elmo2", trainable=False)
embeddings = \
elmo(inputs={"tokens": self.tf_raw_word, "sequence_len": self.tf_sentence_lengths}, signature="tokens",
as_dict=True)["elmo"] # num_sent, max_sent_len, 1024
elmo_emb = tf.layers.dense(inputs=embeddings,
units=self.params.elmo_dim,
activation=None)
self.input = tf.concat([self.input, elmo_emb], axis=-1)
self.input = tf.nn.dropout(self.input, self.tf_dropout)
with tf.variable_scope("bi_lstm_words"):
cell_fw = tf.contrib.rnn.LSTMCell(self.params.word_hidden_size)
cell_bw = tf.contrib.rnn.LSTMCell(self.params.word_hidden_size)
(output_fw, output_bw), _ = tf.nn.bidirectional_dynamic_rnn(
cell_fw,
cell_bw,
self.input,
sequence_length=self.tf_sentence_lengths,
dtype=tf.float32)
self.output = tf.concat([output_fw, output_bw], axis=-1)
ntime_steps = tf.shape(self.output)[1]
self.output = tf.reshape(self.output,
[-1, 2 * params.word_hidden_size])
layer1 = tf.nn.dropout(
tf.layers.dense(inputs=self.output,
units=params.word_hidden_size,
activation=None,
kernel_initializer=xavier_initializer()),
self.tf_dropout)
pred = tf.layers.dense(inputs=layer1,
units=len(self.tag2id),
activation=None,
kernel_initializer=xavier_initializer())
self.logits = tf.reshape(pred, [-1, ntime_steps, len(self.tag2id)])
# compute loss value using crf
log_likelihood, self.transition_params = tf.contrib.crf.crf_log_likelihood(
self.logits, self.tf_labels, self.tf_sentence_lengths)
with tf.variable_scope("loss_and_opt"):
self.tf_loss = tf.reduce_mean(-log_likelihood)
optimizer = tf.train.AdamOptimizer(
learning_rate=self.tf_learning_rate)
self.tf_train_op = optimizer.minimize(self.tf_loss)
def __init__(self, params=None):
self.params = params
self.keep_prob = params["keep_prob"]
self.nb_epochs = params["nb_epochs"]
# load word and char dictionaries
dicts = pickle.load(open(params["dicts_file"], "rb"))
self.w2i = dicts["w2i"]
self.i2w = dicts["i2w"]
self.c2i = dicts["c2i"]
self.i2c = dicts["i2c"]
self.word_dim = params["word_dim"]
self.word_vocab_size = len(self.w2i)
self.char_dim = params["char_dim"]
self.char_vocab_size = len(self.c2i)
print("Sizes of word and char dictionaries: {}, {}".format(
self.word_vocab_size, self.char_vocab_size))
self.word_emb = np.zeros(shape=(self.word_vocab_size, self.word_dim))
# load word embedding
if "word_emb" in params:
print("pre-trained word embedding {} is being loaded ...".format(
params["word_emb"]))
self.load_word_emb(params["word_emb"])
tf.reset_default_graph()
# [sent, word]
self.tf_word_ids = tf.placeholder(dtype=tf.int32,
shape=[None, None],
name="word_ids")
# real length sents
self.tf_sentence_lengths = tf.placeholder(dtype=tf.int32,
shape=[None],
name="sentence_lengths")
# [sent, word, char]
self.tf_char_ids = tf.placeholder(dtype=tf.int32,
shape=[None, None, None],
name="char_ids")
# binary matrix representing the relationship between sents: [sent, sent]
self.tf_target_matrix = tf.placeholder(dtype=tf.int32,
shape=[None],
name="target_matrix")
# keep_prob
self.tf_keep_prob = tf.placeholder(dtype=tf.float32,
shape=[],
name="keep_prob")
# learning rate
self.tf_learning_rate = tf.placeholder(dtype=tf.float32,
shape=[],
name="learning_rate")
# load word embedding
with tf.variable_scope("word_embedding"):
tf_word_embeddings = tf.Variable(self.word_emb,
dtype=tf.float32,
trainable=True,
name="word_embedding")
embedded_words = tf.nn.embedding_lookup(tf_word_embeddings,
self.tf_word_ids,
name="embedded_words")
self.input = embedded_words # sent, word, word_dim
# CNN network to capture character-level features
with tf.variable_scope("char_cnn"):
tf_char_embeddings = tf.get_variable(
name="char_embeddings",
dtype=tf.float32,
shape=[self.char_vocab_size, self.char_dim],
trainable=True,
initializer=xavier_initializer())
conv = tf.nn.embedding_lookup(tf_char_embeddings,
self.tf_char_ids,
name="embedded_cnn_chars")
for i, (ks, fil) in enumerate(self.params["conv"]):
conv = tf.layers.conv2d(
inputs=conv, # sent, word, char, feature
filters=fil,
kernel_size=(1, ks),
strides=(1, 1),
padding="same",
name="conv_{}".format(i),
kernel_initializer=xavier_initializer_conv2d())
self.char_cnn = tf.reduce_max(conv,
axis=2) # sent, word, cnn_feature
self.input = tf.nn.dropout(
tf.concat([self.input, self.char_cnn], axis=-1),
self.tf_keep_prob) # [sents, words, word_dim + cnn_features]
# Bi-LSTM to generate final input representation in combination with both left and right contexts
with tf.variable_scope("bi_lstm_words"):
cell_fw = tf.contrib.rnn.LSTMCell(self.params["word_lstm_units"])
cell_bw = tf.contrib.rnn.LSTMCell(self.params["word_lstm_units"])
(output_fw, output_bw), _ = tf.nn.bidirectional_dynamic_rnn(
cell_fw,
cell_bw,
self.input,
sequence_length=self.tf_sentence_lengths,
dtype=tf.float32)
bilstm_output = tf.concat([output_fw, output_bw],
axis=-1) # [sents, words, 2*lstm_units]
mask = tf.where(
condition=tf.equal(self.tf_word_ids, self.w2i["<PAD>"]),
x=-1e10 * tf.ones_like(self.tf_word_ids, dtype=tf.float32),
y=tf.zeros_like(self.tf_word_ids, dtype=tf.float32))
mask = tf.tile(tf.expand_dims(mask, -1),
(1, 1, 2 * self.params["word_lstm_units"]))
bilstm_output = bilstm_output + mask
bilstm_output = tf.reduce_max(bilstm_output,
axis=1) # [sents, 2*word_lstm_units]
with tf.variable_scope("att"):
c = bilstm_output[:-1, :]
q = bilstm_output[-1:, :]
de = c.get_shape()[-1]
w = tf.get_variable(dtype=tf.float32,
shape=[1, de],
trainable=True,
name="w")
w1 = tf.get_variable(dtype=tf.float32,
shape=[de, de],
trainable=True,
name="w1")
w2 = tf.get_variable(dtype=tf.float32,
shape=[de, de],
trainable=True,
name="w2")
b1 = tf.get_variable(dtype=tf.float32,
shape=[de, 1],
trainable=True,
name="b1")
b = tf.get_variable(dtype=tf.float32,
shape=[1],
trainable=True,
name="b")
w1cT = tf.matmul(w1, tf.transpose(c))
w2q = tf.matmul(w2, tf.transpose(q))
s = tf.add(tf.add(w1cT, w2q), b1)
f = tf.squeeze(tf.matmul(w, tf.tanh(s)) + b, axis=0)
e = tf.nn.softmax(f)
h = c * e[:, None]
h = tf.add(h, tf.squeeze(q, axis=0))
with tf.variable_scope("loss_and_opt"):
self.logits = tf.nn.dropout(
tf.layers.dense(inputs=h, units=2, activation=None),
self.tf_keep_prob)
self.loss = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits_v2(
labels=tf.one_hot(self.tf_target_matrix, 2),
logits=self.logits,
name="loss_function"))
self.pred = tf.argmax(self.logits, axis=-1)
eq = tf.cast(
tf.equal(tf.cast(self.pred, tf.int32), self.tf_target_matrix),
tf.float32)
self.acc = tf.reduce_mean(eq)
self.opt = tf.train.AdamOptimizer(
learning_rate=self.tf_learning_rate).minimize(self.loss)
def _initialize_weights(self):
all_weights = dict()
all_weights['enc_w0'] = tf.get_variable(
"enc_w0",
shape=[
self.kernel_size[0], self.kernel_size[0], 1, self.n_hidden[0]
],
initializer=layers.xavier_initializer_conv2d(),
regularizer=self.reg)
all_weights['enc_b0'] = tf.Variable(
tf.zeros([self.n_hidden[0]], dtype=tf.float32))
all_weights['enc_w1'] = tf.get_variable(
"enc_w1",
shape=[
self.kernel_size[1], self.kernel_size[1], self.n_hidden[0],
self.n_hidden[1]
],
initializer=layers.xavier_initializer_conv2d(),
regularizer=self.reg)
all_weights['enc_b1'] = tf.Variable(
tf.zeros([self.n_hidden[1]], dtype=tf.float32))
all_weights['enc_w2'] = tf.get_variable(
"enc_w2",
shape=[
self.kernel_size[2], self.kernel_size[2], self.n_hidden[1],
self.n_hidden[2]
],
initializer=layers.xavier_initializer_conv2d(),
regularizer=self.reg)
all_weights['enc_b2'] = tf.Variable(
tf.zeros([self.n_hidden[2]], dtype=tf.float32))
all_weights['dec_w0'] = tf.get_variable(
"dec_w0",
shape=[
self.kernel_size[2], self.kernel_size[2], self.n_hidden[1],
self.n_hidden[2]
],
initializer=layers.xavier_initializer_conv2d(),
regularizer=self.reg)
all_weights['dec_b0'] = tf.Variable(
tf.zeros([self.n_hidden[1]], dtype=tf.float32))
all_weights['dec_w1'] = tf.get_variable(
"dec_w1",
shape=[
self.kernel_size[1], self.kernel_size[1], self.n_hidden[0],
self.n_hidden[1]
],
initializer=layers.xavier_initializer_conv2d(),
regularizer=self.reg)
all_weights['dec_b1'] = tf.Variable(
tf.zeros([self.n_hidden[0]], dtype=tf.float32))
all_weights['dec_w2'] = tf.get_variable(
"dec_w2",
shape=[
self.kernel_size[0], self.kernel_size[0], 1, self.n_hidden[0]
],
initializer=layers.xavier_initializer_conv2d(),
regularizer=self.reg)
all_weights['dec_b2'] = tf.Variable(tf.zeros([1], dtype=tf.float32))
return all_weights
def conv(X, in_ch, out_ch, name):
with tf.variable_scope(name) as scope:
W_conv = tf.get_variable(name='weights', shape=[3, 3, in_ch, out_ch], initializer=xavier_initializer_conv2d())
h_bn = tf.layers.batch_normalization(tf.nn.conv2d(X, W_conv, strides=[1, 1, 1, 1], padding='SAME'), training =trainphase)
h_conv = my_relu(h_bn)
return h_conv
def _initialize_weights(self):
all_weights = dict()
all_weights['enc_w0'] = tf.get_variable(
"enc_w0",
shape=[
self.kernel_size[0], self.kernel_size[0], 1, self.n_hidden[0]
],
initializer=layers.xavier_initializer_conv2d(),
regularizer=self.reg)
all_weights['enc_b0'] = tf.Variable(
tf.zeros([self.n_hidden[0]], dtype=tf.float32))
all_weights['enc_w1'] = tf.get_variable(
"enc_w1",
shape=[
self.kernel_size[1], self.kernel_size[1], self.n_hidden[0],
self.n_hidden[1]
],
initializer=layers.xavier_initializer_conv2d(),
regularizer=self.reg)
all_weights['enc_b1'] = tf.Variable(
tf.zeros([self.n_hidden[1]], dtype=tf.float32))
all_weights['enc_w2'] = tf.get_variable(
"enc_w2",
shape=[
self.kernel_size[2], self.kernel_size[2], self.n_hidden[1],
self.n_hidden[2]
],
initializer=layers.xavier_initializer_conv2d(),
regularizer=self.reg)
all_weights['enc_b2'] = tf.Variable(
tf.zeros([self.n_hidden[2]], dtype=tf.float32))
all_weights['Coef'] = tf.Variable(
1.0e-4 * tf.ones([self.batch_size, self.batch_size], tf.float32),
name='Coef')
all_weights['dec_w0'] = tf.get_variable(
"dec_w0",
shape=[
self.kernel_size[2], self.kernel_size[2], self.n_hidden[1],
self.n_hidden[2]
],
initializer=layers.xavier_initializer_conv2d(),
regularizer=self.reg)
all_weights['dec_b0'] = tf.Variable(
tf.zeros([self.n_hidden[1]], dtype=tf.float32))
all_weights['dec_w1'] = tf.get_variable(
"dec_w1",
shape=[
self.kernel_size[1], self.kernel_size[1], self.n_hidden[0],
self.n_hidden[1]
],
initializer=layers.xavier_initializer_conv2d(),
regularizer=self.reg)
all_weights['dec_b1'] = tf.Variable(
tf.zeros([self.n_hidden[0]], dtype=tf.float32))
all_weights['dec_w2'] = tf.get_variable(
"dec_w2",
shape=[
self.kernel_size[0], self.kernel_size[0], 1, self.n_hidden[0]
],
initializer=layers.xavier_initializer_conv2d(),
regularizer=self.reg)
all_weights['dec_b2'] = tf.Variable(tf.zeros([1], dtype=tf.float32))
all_weights['oenc_w0'] = tf.get_variable(
"oenc_w0",
shape=[
self.kernel_size[0], self.kernel_size[0], self.n_hidden[0], 1
],
initializer=layers.xavier_initializer_conv2d(),
regularizer=self.reg)
all_weights['oenc_b0'] = tf.Variable(tf.zeros([1], dtype=tf.float32))
all_weights['oenc_w1'] = tf.get_variable(
"oenc_w1",
shape=[
self.kernel_size[1], self.kernel_size[1], self.n_hidden[2],
self.n_hidden[0]
],
initializer=layers.xavier_initializer_conv2d(),
regularizer=self.reg)
all_weights['oenc_b1'] = tf.Variable(
tf.zeros([self.n_hidden[2]], dtype=tf.float32))
all_weights['oenc_w2'] = tf.get_variable(
"oenc_w2",
shape=[
self.kernel_size[2], self.kernel_size[2], self.n_hidden[2],
self.n_hidden[1]
],
initializer=layers.xavier_initializer_conv2d(),
regularizer=self.reg)
all_weights['oenc_b2'] = tf.Variable(
tf.zeros([self.n_hidden[2]], dtype=tf.float32))
all_weights['oCoef'] = tf.Variable(
1.0e-4 * tf.ones([self.batch_size, self.batch_size], tf.float32),
name='Coef')
all_weights['odec_w0'] = tf.get_variable(
"odec_w0",
shape=[
self.kernel_size[2], self.kernel_size[2], self.n_hidden[1],
self.n_hidden[2]
],
initializer=layers.xavier_initializer_conv2d(),
regularizer=self.reg)
all_weights['odec_b0'] = tf.Variable(
tf.zeros([self.n_hidden[2]], dtype=tf.float32))
all_weights['odec_w1'] = tf.get_variable(
"odec_w1",
shape=[
self.kernel_size[1], self.kernel_size[1], self.n_hidden[0],
self.n_hidden[2]
],
initializer=layers.xavier_initializer_conv2d(),
regularizer=self.reg)
all_weights['odec_b1'] = tf.Variable(
tf.zeros([self.n_hidden[2]], dtype=tf.float32))
all_weights['odec_w2'] = tf.get_variable(
"odec_w2",
shape=[self.kernel_size[0], self.kernel_size[0], 3, 3],
initializer=layers.xavier_initializer_conv2d(),
regularizer=self.reg)
all_weights['odec_b2'] = tf.Variable(tf.zeros([3], dtype=tf.float32))
all_weights['odec_w3'] = tf.get_variable(
"odec_w3",
shape=[self.kernel_size[0], self.kernel_size[0], 3, 3],
initializer=layers.xavier_initializer_conv2d(),
regularizer=self.reg)
all_weights['odec_b3'] = tf.Variable(tf.zeros([3], dtype=tf.float32))
return all_weights
def __conv_relu(self, input, kernel_shape, bias_shape, name):
weights = tf.get_variable("W"+name, kernel_shape, initializer=xavier_initializer_conv2d())
biases = tf.get_variable("b"+name, bias_shape, initializer=tf.constant_initializer(0.0))
conv = tf.nn.conv2d(input, weights, strides=[1, 1, 1, 1], padding='SAME', name=name+"_conv")
return tf.nn.relu(conv + biases)
def siamese_net(input, reuse=False, is_training=True): # without batch norm
with tf.name_scope("model"):
with tf.variable_scope("conv1_1") as scope:
# ,normalizer_fn=slim.batch_norm,normalizer_params={'is_training':is_training}
# tf.truncated_normal_initializer(stddev=0.001)
conv1_1 = layers.conv2d(
input,
48, [3, 3],
activation_fn=tf.nn.relu,
weights_initializer=layers.xavier_initializer_conv2d(),
weights_regularizer=slim.l2_regularizer(0.0001),
scope=scope,
reuse=reuse)
with tf.variable_scope("conv1_2") as scope:
conv1_2 = layers.conv2d(
conv1_1,
48, [3, 3],
activation_fn=tf.nn.relu,
weights_initializer=layers.xavier_initializer_conv2d(),
weights_regularizer=slim.l2_regularizer(0.0001),
scope=scope,
reuse=reuse)
with tf.variable_scope("conv2_1") as scope:
conv2_1 = layers.avg_pool2d(conv1_2, [2, 2])
with tf.variable_scope("conv2_2") as scope:
conv2_2 = layers.conv2d(
conv2_1,
96, [3, 3],
activation_fn=tf.nn.relu,
weights_initializer=layers.xavier_initializer_conv2d(),
weights_regularizer=slim.l2_regularizer(0.0001),
scope=scope,
reuse=reuse)
with tf.variable_scope("conv2_3") as scope:
conv2_3 = layers.conv2d(
conv2_2,
96, [3, 3],
activation_fn=tf.nn.relu,
weights_initializer=layers.xavier_initializer_conv2d(),
weights_regularizer=slim.l2_regularizer(0.0001),
scope=scope,
reuse=reuse)
with tf.variable_scope("conv3_1") as scope:
conv3_1 = layers.avg_pool2d(conv2_3, [2, 2])
with tf.variable_scope("conv3_2") as scope:
conv3_2 = layers.conv2d(
conv3_1,
192, [3, 3],
activation_fn=tf.nn.relu,
weights_initializer=layers.xavier_initializer_conv2d(),
weights_regularizer=slim.l2_regularizer(0.0001),
scope=scope,
reuse=reuse)
with tf.variable_scope("conv3_3") as scope:
conv3_3 = layers.conv2d(
conv3_2,
192, [3, 3],
activation_fn=tf.nn.relu,
weights_initializer=layers.xavier_initializer_conv2d(),
weights_regularizer=slim.l2_regularizer(0.0001),
scope=scope,
reuse=reuse)
with tf.variable_scope("conv4_1") as scope:
conv4_1 = layers.avg_pool2d(conv3_3, [2, 2])
with tf.variable_scope("conv4_2") as scope:
conv4_2 = layers.conv2d(
conv4_1,
384, [3, 3],
activation_fn=tf.nn.relu,
weights_initializer=layers.xavier_initializer_conv2d(),
weights_regularizer=slim.l2_regularizer(0.0001),
scope=scope,
reuse=reuse)
with tf.variable_scope("conv4_3") as scope:
conv4_3 = layers.conv2d(
conv4_2,
384, [3, 3],
activation_fn=tf.nn.relu,
weights_initializer=layers.xavier_initializer_conv2d(),
weights_regularizer=slim.l2_regularizer(0.0001),
scope=scope,
reuse=reuse)
with tf.variable_scope("conv43") as scope:
conv43 = layers.conv2d_transpose(
conv4_3,
192, [2, 2],
stride=2,
activation_fn=None,
weights_initializer=layers.xavier_initializer_conv2d(),
weights_regularizer=slim.l2_regularizer(0.0001),
scope=scope,
reuse=reuse)
with tf.variable_scope("conv3_4") as scope:
conv3_4 = tf.concat([conv43, conv3_3], 3)
with tf.variable_scope("conv3_5") as scope:
conv3_5 = layers.conv2d(
conv3_4,
192, [3, 3],
activation_fn=tf.nn.relu,
weights_initializer=layers.xavier_initializer_conv2d(),
weights_regularizer=slim.l2_regularizer(0.0001),
scope=scope,
reuse=reuse)
with tf.variable_scope("conv3_6") as scope:
conv3_6 = layers.conv2d(
conv3_5,
192, [3, 3],
activation_fn=tf.nn.relu,
weights_initializer=layers.xavier_initializer_conv2d(),
weights_regularizer=slim.l2_regularizer(0.0001),
scope=scope,
reuse=reuse)
with tf.variable_scope("conv32") as scope:
conv32 = layers.conv2d_transpose(
conv3_6,
64, [2, 2],
stride=2,
activation_fn=None,
weights_initializer=layers.xavier_initializer_conv2d(),
weights_regularizer=slim.l2_regularizer(0.0001),
scope=scope,
reuse=reuse)
with tf.variable_scope("conv2_4") as scope:
conv2_4 = tf.concat([conv32, conv2_3], 3)
with tf.variable_scope("conv2_5") as scope:
conv2_5 = layers.conv2d(
conv2_4,
96, [3, 3],
activation_fn=tf.nn.relu,
weights_initializer=layers.xavier_initializer_conv2d(),
weights_regularizer=slim.l2_regularizer(0.0001),
scope=scope,
reuse=reuse)
with tf.variable_scope("conv2_6") as scope:
conv2_6 = layers.conv2d(
conv2_5,
96, [3, 3],
activation_fn=tf.nn.relu,
weights_initializer=layers.xavier_initializer_conv2d(),
weights_regularizer=slim.l2_regularizer(0.0001),
scope=scope,
reuse=reuse)
with tf.variable_scope("conv21") as scope:
conv21 = layers.conv2d_transpose(
conv2_6,
48, [2, 2],
stride=2,
activation_fn=None,
weights_initializer=layers.xavier_initializer_conv2d(),
weights_regularizer=slim.l2_regularizer(0.0001),
scope=scope,
reuse=reuse)
with tf.variable_scope("conv1_3") as scope:
conv1_3 = tf.concat([conv21, conv1_2], 3)
with tf.variable_scope("conv1_4") as scope:
conv1_4 = layers.conv2d(
conv1_3,
48, [3, 3],
activation_fn=tf.nn.relu,
weights_initializer=layers.xavier_initializer_conv2d(),
weights_regularizer=slim.l2_regularizer(0.0001),
scope=scope,
reuse=reuse)
with tf.variable_scope("conv1_5") as scope:
conv1_5 = layers.conv2d(
conv1_4,
24, [3, 3],
activation_fn=tf.nn.relu,
weights_initializer=layers.xavier_initializer_conv2d(),
weights_regularizer=slim.l2_regularizer(0.0001),
scope=scope,
reuse=reuse)
with tf.variable_scope("conv1_6") as scope:
conv1_6 = layers.conv2d(
conv1_5,
1, [3, 3],
activation_fn=None,
weights_initializer=layers.xavier_initializer_conv2d(),
weights_regularizer=slim.l2_regularizer(0.0001),
scope=scope,
reuse=reuse)
return conv1_6
def mru_conv_block_v3(inp, ht, filter_depth, sn,
stride, dilate=1,
activation_fn=tf.nn.relu,
normalizer_fn=None,
normalizer_params=None,
weights_initializer=ly.xavier_initializer_conv2d(),
biases_initializer_mask=tf.constant_initializer(
value=0.5),
biases_initializer_h=tf.constant_initializer(value=-1),
data_format='NCHW',
weight_decay_rate=1e-8,
norm_mask=False,
norm_input=True,
deconv=False):
def norm_activ(tensor_in):
if normalizer_fn is not None:
_normalizer_params = normalizer_params or {}
tensor_normed = normalizer_fn(tensor_in, **_normalizer_params)
else:
tensor_normed = tf.identity(tensor_in)
if activation_fn is not None:
tensor_normed = activation_fn(tensor_normed)
return tensor_normed
channel_index = 1 if data_format == 'NCHW' else 3
reduce_dim = [2, 3] if data_format == 'NCHW' else [1, 2]
hidden_depth = ht.get_shape().as_list()[channel_index]
regularizer = ly.l2_regularizer(
weight_decay_rate) if weight_decay_rate > 0 else None
weights_initializer_mask = weights_initializer
biases_initializer = tf.zeros_initializer()
if norm_mask:
mask_normalizer_fn = normalizer_fn
mask_normalizer_params = normalizer_params
else:
mask_normalizer_fn = None
mask_normalizer_params = None
if deconv:
if stride == 2:
ht = upsample(ht, data_format=data_format)
elif stride != 1:
raise NotImplementedError
ht_orig = tf.identity(ht)
# Normalize hidden state
with tf.variable_scope('norm_activation_in') as sc:
if norm_input:
full_inp = tf.concat([norm_activ(ht), inp], axis=channel_index)
else:
full_inp = tf.concat([ht, inp], axis=channel_index)
# update gate
rg = conv2d2(full_inp, hidden_depth, 3, sn=sn, stride=1, rate=dilate,
data_format=data_format, activation_fn=lrelu,
normalizer_fn=mask_normalizer_fn, normalizer_params=mask_normalizer_params,
weights_regularizer=regularizer,
weights_initializer=weights_initializer_mask,
biases_initializer=biases_initializer_mask,
scope='update_gate')
rg = (rg - tf.reduce_min(rg, axis=reduce_dim, keep_dims=True)) / (
tf.reduce_max(rg, axis=reduce_dim, keep_dims=True) - tf.reduce_min(rg, axis=reduce_dim, keep_dims=True))
# Input Image conv
img_new = conv2d2(inp, hidden_depth, 3, sn=sn, stride=1, rate=dilate,
data_format=data_format, activation_fn=None,
normalizer_fn=None, normalizer_params=None,
biases_initializer=biases_initializer,
weights_regularizer=regularizer,
weights_initializer=weights_initializer)
ht_plus = ht + rg * img_new
with tf.variable_scope('norm_activation_merge_1') as sc:
ht_new_in = norm_activ(ht_plus)
# new hidden state
h_new = conv2d2(ht_new_in, filter_depth, 3, sn=sn, stride=1, rate=dilate,
data_format=data_format, activation_fn=activation_fn,
normalizer_fn=normalizer_fn, normalizer_params=normalizer_params,
biases_initializer=biases_initializer,
weights_regularizer=regularizer,
weights_initializer=weights_initializer)
h_new = conv2d2(h_new, filter_depth, 3, sn=sn, stride=1, rate=dilate,
data_format=data_format, activation_fn=None,
normalizer_fn=None, normalizer_params=None,
biases_initializer=biases_initializer,
weights_regularizer=regularizer,
weights_initializer=weights_initializer)
# new hidden state out
# linear project for filter depth
if ht.get_shape().as_list()[channel_index] != filter_depth:
ht_orig = conv2d2(ht_orig, filter_depth, 1, sn=sn, stride=1,
data_format=data_format, activation_fn=None,
normalizer_fn=None, normalizer_params=None,
biases_initializer=biases_initializer,
weights_regularizer=regularizer,
weights_initializer=weights_initializer)
ht_new = ht_orig + h_new
if not deconv:
if stride == 2:
ht_new = mean_pool(ht_new, data_format=data_format)
elif stride != 1:
raise NotImplementedError
return ht_new
def model(x, n_classes, keep_prob):
mu = 0
sigma = 0.1
depth = {
'D_1': 16,
'D_2': 32,
'D_3': 512,
'D_4': 8192,
'D_5': 256,
}
weights = {
# 'W_conv1':tf.Variable(tf.truncated_normal(shape=[5, 5, 1, depth["D_1"]], mean = mu, stddev = sigma, name = 'weight1')),
# 'W_conv2':tf.Variable(tf.truncated_normal(shape=[5, 5, depth["D_1"], depth["D_2"]], mean = mu, stddev = sigma, name = 'weight2')),
# 'W_conv3':tf.Variable(tf.truncated_normal(shape=[5, 5, depth["D_2"], depth["D_3"]], mean = mu, stddev = sigma, name = 'weight3')),
# 'W_fc1': tf.Variable(tf.truncated_normal(shape=[8192, depth["D_4"]], mean = mu, stddev = sigma, name = 'weight4')),
# 'W_fc2': tf.Variable(tf.truncated_normal(shape=[depth["D_4"], depth["D_5"]], mean = mu, stddev = sigma, name = 'weight5')),
# 'W_out': tf.Variable(tf.truncated_normal(shape=[depth["D_5"], n_classes], mean = mu, stddev = sigma, name = 'weight_out')),
'W_conv1':
tf.Variable(xavier_initializer_conv2d()([5, 5, 1, depth["D_1"]]),
name='weight1'),
'W_conv2':
tf.Variable(
xavier_initializer_conv2d()([5, 5, depth["D_1"], depth["D_2"]]),
name='weight2'),
'W_conv3':
tf.Variable(
xavier_initializer_conv2d()([5, 5, depth["D_2"], depth["D_3"]]),
name='weight3'),
'W_fc1':
tf.Variable(xavier_initializer_conv2d()([8192, depth["D_4"]]),
name='weight4'),
'W_fc2':
tf.Variable(xavier_initializer_conv2d()([depth["D_4"], depth["D_5"]]),
name='weight5'),
'W_out':
tf.Variable(xavier_initializer_conv2d()([depth["D_5"], n_classes]),
name='weight_out'),
}
biases = {
'B_1': tf.Variable(tf.zeros(depth['D_1']), name='bias_1'),
'B_2': tf.Variable(tf.zeros(depth['D_2']), name='bias_2'),
'B_3': tf.Variable(tf.zeros(depth['D_3']), name='bias_3'),
'B_4': tf.Variable(tf.zeros(depth['D_4']), name='bias_4'),
'B_5': tf.Variable(tf.zeros(depth['D_5']), name='bias_5'),
'B_out': tf.Variable(tf.zeros(n_classes), name='bias_out')
}
# print(x)
# Layer 1: Convolutional. Input = 32x32x1. Output = 32x32x1.
conv1 = tf.nn.conv2d(
x, weights['W_conv1'], strides=[1, 1, 1, 1
], padding='SAME') + biases['B_1']
conv1 = tf.nn.relu(conv1) # Activation.
# Pooling. Input = 32x32x1. Output = 14x14x16.
conv1 = tf.nn.max_pool(conv1,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME')
# Layer 2: Convolutional. Output = 10x10x32.
conv2 = tf.nn.conv2d(
conv1, weights['W_conv2'], strides=[1, 1, 1, 1
], padding='SAME') + biases['B_2']
conv2 = tf.nn.relu(conv2) # Activation.
# Pooling. Input = 10x10x32. Output = 5x5x32.
conv2 = tf.nn.max_pool(conv2,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME')
# Layer 2: Convolutional. Output = 10x10x32.
conv3 = tf.nn.conv2d(
conv2, weights['W_conv3'], strides=[1, 1, 1, 1
], padding='SAME') + biases['B_3']
conv3 = tf.nn.relu(conv3) # Activation.
# Pooling. Input = 10x10x32. Output = 5x5x32.
conv3 = tf.nn.max_pool(conv3,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME')
# Flatten. Input = 5x5x32. Output = 800.
fc0 = flatten(conv3)
# print("fc0", fc0, "weights:", weights["W_fc1"])
# Layer 3: Fully Connected. Input = 800. Output = 120.
fc1 = tf.matmul(fc0, weights['W_fc1']) + biases['B_4']
fc1 = tf.nn.relu(fc1) # Activation.
# Layer 4: Fully Connected. Input = 120. Output = 84.
fc2 = tf.matmul(fc1, weights['W_fc2']) + biases['B_5']
fc2 = tf.nn.relu(fc2) # Activation.
# Layer 5: Fully Connected. Input = 84. Output = 10.
logits = tf.matmul(fc2, weights['W_out']) + biases['B_out']
return logits
def apply_logit(self):
with tf.variable_scope(self.modelname + "conv_1") as scope:
self.W_conv1 = tf.get_variable(
name='weights',
shape=[1, 5, 1, 5],
initializer=xavier_initializer_conv2d())
self.h_bn1 = tf.layers.batch_normalization(
tf.nn.conv2d(self.X,
self.W_conv1,
strides=[1, 1, 1, 1],
padding='VALID'),
training=self.trainphase)
self.h_conv1 = tf.nn.leaky_relu(self.h_bn1)
with tf.variable_scope(self.modelname + "conv_2") as scope:
self.W_conv2 = tf.get_variable(
name='weights',
shape=[2, 5, 1, 5],
initializer=xavier_initializer_conv2d())
self.h_bn2 = tf.layers.batch_normalization(
tf.nn.conv2d(self.X,
self.W_conv2,
strides=[1, 1, 1, 1],
padding='VALID'),
training=self.trainphase)
self.h_conv2 = tf.nn.leaky_relu(self.h_bn2)
with tf.variable_scope(self.modelname + "conv_3") as scope:
self.W_conv3 = tf.get_variable(
name='weights',
shape=[3, 5, 1, 5],
initializer=xavier_initializer_conv2d())
self.h_bn3 = tf.layers.batch_normalization(
tf.nn.conv2d(self.X,
self.W_conv3,
strides=[1, 1, 1, 1],
padding='VALID'),
training=self.trainphase)
self.h_conv3 = tf.nn.leaky_relu(self.h_bn3)
with tf.variable_scope(self.modelname + "conv_4") as scope:
self.W_conv4 = tf.get_variable(
name='weights',
shape=[4, 5, 1, 5],
initializer=xavier_initializer_conv2d())
self.h_bn4 = tf.layers.batch_normalization(
tf.nn.conv2d(self.X,
self.W_conv4,
strides=[1, 1, 1, 1],
padding='VALID'),
training=self.trainphase)
self.h_conv4 = tf.nn.leaky_relu(self.h_bn4)
self.pooled1 = tf.reshape(self.h_conv1,
shape=[-1, int(self.duration), 5])
self.pooled1 = tf.reduce_mean(self.pooled1, reduction_indices=2)
self.pooled1 = tf.squeeze(self.pooled1)
self.pooled2 = tf.reshape(self.h_conv2,
shape=[-1, int(self.duration - 1), 5])
self.pooled2 = tf.reduce_mean(self.pooled2, reduction_indices=2)
self.pooled2 = tf.squeeze(self.pooled2)
self.pooled3 = tf.reshape(self.h_conv3,
shape=[-1, int(self.duration - 2), 5])
self.pooled3 = tf.reduce_mean(self.pooled3, reduction_indices=2)
self.pooled3 = tf.squeeze(self.pooled3)
'''
self.pooled4 = tf.reshape(self.h_conv4, shape=[-1, int(self.duration - 3), 5])
self.pooled4 = tf.reduce_mean(self.pooled4, reduction_indices=2)
self.pooled4 = tf.squeeze(self.pooled4)
'''
'''
self.pooled1_flatten = tf.reshape(self.h_conv1,shape=[-1,int(self.duration)*5])
self.pooled2_flatten = tf.reshape(self.h_conv2,shape=[-1,int(self.duration-1)*5])
self.pooled3_flatten = tf.reshape(self.h_conv3, shape=[-1, int(self.duration-2)*5])
self.pooled4_flatten = tf.reshape(self.h_conv4, shape=[-1, int(self.duration-3)*5])
'''
#self.pooled1_flatten = tf.reshape(self.pooled1,shape=[-1,self.duration])
self.pooled2_flatten = tf.reshape(self.pooled2,
shape=[-1, self.duration - 1])
self.pooled3_flatten = tf.reshape(self.pooled3,
shape=[-1, self.duration - 2])
#self.pooled4_flatten = tf.reshape(self.pooled4, shape=[-1, self.duration - 3])
#self.flatten = tf.concat([self.pooled1_flatten,self.pooled2_flatten,self.pooled3_flatten,self.pooled4_flatten],1)
self.flatten = tf.concat([self.pooled2_flatten, self.pooled3_flatten],
1)
num_unit = 300
H = self.linear(self.flatten, self.flatten.shape[1], num_unit, 'L1')
H = self.linear(H, num_unit, num_unit, "L2")
#H = self.linear(H,num_unit,num_unit,"L3")
#H = self.linear(H, num_unit, num_unit, "L4")
#H = self.linear(H, num_unit, num_unit, "L5")
#H = self.linear(H, num_unit, num_unit, "L6")
#H = self.linear(H,num_unit,num_unit,"L7")
#H = self.linear(H,num_unit,num_unit,"L8")
#H = self.linear(H,num_unit,num_unit,"L9")
#H = self.linear(H,num_unit,num_unit,"L10")
#H = self.linear(H,100,100,"L7")
#H = self.linear(H,100,100,"L8")
#H = self.linear(H,100,100,"L9")
self.output_features = self.linear(H, num_unit, 128, "Llast")
self.output = self.linear(self.output_features,
128,
1,
"output",
active_f=None)
return self.output
def __init__(self,
n_tags: int,
word_vocab,
word_dim: int,
word_emb_path: str,
word_emb_name: str = None,
char_vocab_size: int = None,
pos_vocab_size: int = None,
chunk_vocab_size: int = None,
char_dim: int = None,
elmo_dim: int = None,
pos_dim: int = None,
chunk_dim: int = None,
cap_dim: int = None,
cap_vocab_size: int = 5,
lstm_hidden_size: int = 256,
dropout_keep_prob: float = 0.5,
**kwargs) -> None:
assert n_tags != 0, 'Number of classes equal 0! It seems that vocabularies is not loaded.' \
' Check that all vocabulary files are downloaded!'
if 'learning_rate_drop_div' not in kwargs:
kwargs['learning_rate_drop_div'] = 10.0
if 'learning_rate_drop_patience' not in kwargs:
kwargs['learning_rate_drop_patience'] = 5.0
if 'clip_norm' not in kwargs:
kwargs['clip_norm'] = 5.0
super().__init__(**kwargs)
word2id = word_vocab.t2i
self._dropout_ph = tf.placeholder_with_default(dropout_keep_prob,
shape=[],
name='dropout')
self.training_ph = tf.placeholder_with_default(False,
shape=[],
name='is_training')
self._y_ph = tf.placeholder(tf.int32, [None, None], name='y_ph')
self._xs_ph_list = []
self._input_features = []
# use for word contextualized bi-lstm, elmo
self.real_sent_lengths_ph = tf.placeholder(tf.int32, [None],
name="real_sent_lengths")
self._xs_ph_list.append(self.real_sent_lengths_ph)
# Word emb
with tf.variable_scope("word_emb"):
word_ids_ph = tf.placeholder(tf.int32, [None, None],
name="word_ids")
self._xs_ph_list.append(word_ids_ph)
word_embeddings = self.load_pretrained_word_emb(
word_emb_path, word_emb_name, word_dim, word2id)
word_lookup_table = tf.Variable(word_embeddings,
dtype=tf.float32,
trainable=True,
name="word_embeddings")
word_emb = tf.nn.embedding_lookup(word_lookup_table,
word_ids_ph,
name="embedded_word")
self._input_features.append(word_emb)
# POS feature
if pos_dim is not None:
with tf.variable_scope("pos_emb"):
pos_ph = tf.placeholder(tf.int32, [None, None], name="pos_ids")
self._xs_ph_list.append(pos_ph)
tf_pos_embeddings = tf.get_variable(
name="pos_embeddings",
dtype=tf.float32,
shape=[pos_vocab_size, pos_dim],
trainable=True,
initializer=xavier_initializer())
embedded_pos = tf.nn.embedding_lookup(tf_pos_embeddings,
pos_ph,
name="embedded_pos")
self._input_features.append(embedded_pos)
# Chunk feature
if chunk_dim is not None:
with tf.variable_scope("chunk_emb"):
chunk_ph = tf.placeholder(tf.int32, [None, None],
name="chunk_ids")
self._xs_ph_list.append(chunk_ph)
tf_chunk_embeddings = tf.get_variable(
name="chunk_embeddings",
dtype=tf.float32,
shape=[chunk_vocab_size, chunk_dim],
trainable=True,
initializer=xavier_initializer())
embedded_chunk = tf.nn.embedding_lookup(tf_chunk_embeddings,
chunk_ph,
name="embedded_chunk")
self._input_features.append(embedded_chunk)
# Capitalization feature
if cap_dim is not None:
with tf.variable_scope("cap_emb"):
cap_ph = tf.placeholder(tf.int32, [None, None], name="cap_ids")
self._xs_ph_list.append(cap_ph)
tf_cap_embeddings = tf.get_variable(
name="cap_embeddings",
dtype=tf.float32,
shape=[cap_vocab_size, cap_dim],
trainable=True,
initializer=xavier_initializer())
embedded_cap = tf.nn.embedding_lookup(tf_cap_embeddings,
cap_ph,
name="embedded_cap")
self._input_features.append(embedded_cap)
# Character feature
if char_dim is not None:
with tf.variable_scope("char_emb"):
char_ids_ph = tf.placeholder(tf.int32, [None, None, None],
name="char_ids")
self._xs_ph_list.append(char_ids_ph)
tf_char_embeddings = tf.get_variable(
name="char_embeddings",
dtype=tf.float32,
shape=[char_vocab_size, char_dim],
trainable=True,
initializer=xavier_initializer())
embedded_cnn_chars = tf.nn.embedding_lookup(
tf_char_embeddings, char_ids_ph, name="embedded_cnn_chars")
conv1 = tf.layers.conv2d(
inputs=embedded_cnn_chars,
filters=128,
kernel_size=(1, 3),
strides=(1, 1),
padding="same",
name="conv1",
kernel_initializer=xavier_initializer_conv2d())
conv2 = tf.layers.conv2d(
inputs=conv1,
filters=128,
kernel_size=(1, 3),
strides=(1, 1),
padding="same",
name="conv2",
kernel_initializer=xavier_initializer_conv2d())
char_cnn = tf.reduce_max(conv2, axis=2)
self._input_features.append(char_cnn)
# ELMo
if elmo_dim is not None:
with tf.variable_scope("elmo_emb"):
padded_x_tokens_ph = tf.placeholder(tf.string, [None, None],
name="padded_x_tokens")
self._xs_ph_list.append(padded_x_tokens_ph)
elmo = hub.Module("https://tfhub.dev/google/elmo/2",
trainable=True)
emb = elmo(inputs={
"tokens": padded_x_tokens_ph,
"sequence_len": self.real_sent_lengths_ph
},
signature="tokens",
as_dict=True)["elmo"]
elmo_emb = tf.layers.dense(emb, elmo_dim, activation=None)
self._input_features.append(elmo_emb)
features = tf.nn.dropout(tf.concat(self._input_features, axis=2),
self._dropout_ph)
with tf.variable_scope("bi_lstm_words"):
cell_fw = tf.contrib.rnn.LSTMCell(lstm_hidden_size)
cell_bw = tf.contrib.rnn.LSTMCell(lstm_hidden_size)
(output_fw, output_bw), _ = tf.nn.bidirectional_dynamic_rnn(
cell_fw,
cell_bw,
features,
sequence_length=self.real_sent_lengths_ph,
dtype=tf.float32)
self.output = tf.concat([output_fw, output_bw], axis=-1)
ntime_steps = tf.shape(self.output)[1]
self.output = tf.reshape(self.output, [-1, 2 * lstm_hidden_size])
layer1 = tf.nn.dropout(
tf.layers.dense(inputs=self.output,
units=lstm_hidden_size,
activation=None,
kernel_initializer=xavier_initializer()),
self._dropout_ph)
pred = tf.layers.dense(inputs=layer1,
units=n_tags,
activation=None,
kernel_initializer=xavier_initializer())
self.logits = tf.reshape(pred, [-1, ntime_steps, n_tags])
log_likelihood, self.transition_params = tf.contrib.crf.crf_log_likelihood(
self.logits, self._y_ph, self.real_sent_lengths_ph)
# loss and opt
with tf.variable_scope("loss_and_opt"):
self.loss = tf.reduce_mean(-log_likelihood)
self.train_op = self.get_train_op(self.loss)
self.sess = tf.Session()
self.sess.run(tf.global_variables_initializer())
self.load()
def redResidua(input):
out = tf.reshape(input, [-1, 128, 128, 1])
out = conv2d(out,
64,
3,
padding='same',
activation=tf.nn.relu,
kernel_initializer=xavier_initializer_conv2d())
out = batch_normalization(out)
out = conv2d(out,
64,
3,
padding='same',
activation=tf.nn.relu,
kernel_initializer=xavier_initializer_conv2d())
out = batch_normalization(out)
out = max_pool2d(out, 3, stride=2)
max1 = out
out = conv2d(out,
64,
3,
padding='same',
activation=tf.nn.relu,
kernel_initializer=xavier_initializer_conv2d())
out = batch_normalization(out)
out = conv2d(out,
64,
3,
padding='same',
activation=tf.nn.relu,
kernel_initializer=xavier_initializer_conv2d())
out = batch_normalization(out)
out = out + max1
r1 = out
out = conv2d(out,
64,
3,
padding='same',
activation=tf.nn.relu,
kernel_initializer=xavier_initializer_conv2d())
out = batch_normalization(out)
out = conv2d(out,
64,
3,
padding='same',
activation=tf.nn.relu,
kernel_initializer=xavier_initializer_conv2d())
out = batch_normalization(out)
out = out + r1
out = conv2d(out,
128,
3,
padding='same',
activation=tf.nn.relu,
kernel_initializer=xavier_initializer_conv2d())
out = batch_normalization(out)
out = max_pool2d(out, 3, stride=2)
max2 = out
out = conv2d(out,
128,
3,
padding='same',
activation=tf.nn.relu,
kernel_initializer=xavier_initializer_conv2d())
out = batch_normalization(out)
out = conv2d(out,
128,
3,
padding='same',
activation=tf.nn.relu,
kernel_initializer=xavier_initializer_conv2d())
out = batch_normalization(out)
out = out + max2
r3 = out
out = conv2d(out,
128,
3,
padding='same',
activation=tf.nn.relu,
kernel_initializer=xavier_initializer_conv2d())
out = batch_normalization(out)
out = conv2d(out,
128,
3,
padding='same',
activation=tf.nn.relu,
kernel_initializer=xavier_initializer_conv2d())
out = batch_normalization(out)
out = out + r3
out = conv2d(out,
256,
3,
padding='same',
activation=tf.nn.relu,
kernel_initializer=xavier_initializer_conv2d())
out = batch_normalization(out)
out = max_pool2d(out, 3, stride=2)
max3 = out
out = conv2d(out,
256,
3,
padding='same',
activation=tf.nn.relu,
kernel_initializer=xavier_initializer_conv2d())
out = batch_normalization(out)
out = conv2d(out,
256,
3,
padding='same',
activation=tf.nn.relu,
kernel_initializer=xavier_initializer_conv2d())
out = batch_normalization(out)
out = out + max3
r5 = out
out = conv2d(out,
256,
3,
padding='same',
activation=tf.nn.relu,
kernel_initializer=xavier_initializer_conv2d())
out = batch_normalization(out)
out = conv2d(out,
256,
3,
padding='same',
activation=tf.nn.relu,
kernel_initializer=xavier_initializer_conv2d())
out = batch_normalization(out)
out = out + r5
out = conv2d(out,
256,
3,
padding='same',
activation=tf.nn.relu,
kernel_initializer=xavier_initializer_conv2d())
out = batch_normalization(out)
out = max_pool2d(out, 3, stride=2)
out = tf.contrib.layers.flatten(out)
out = fully_connected(out,
1024,
activation_fn=tf.nn.relu,
weights_initializer=xavier_initializer())
out = fully_connected(out,
100,
activation_fn=None,
weights_initializer=xavier_initializer())
return out
def inference(self, img):
self.padding = 'SAME'
# initializer = tf.truncated_normal_initializer(stddev=0.01)
self.initializer = layers.xavier_initializer_conv2d()
self.initializer_b = layers.xavier_initializer()
self.regularizer = slim.l2_regularizer(0.0005)
self.activation = slim.layers.nn.leaky_relu
self.activation = slim.layers.nn.relu
print(self.n_filters_first_conv)
with tf.variable_scope("first-conv"):
stack = slim.conv2d(img,
num_outputs=self.n_filters_first_conv,
kernel_size=[3, 3],
weights_initializer=self.initializer,
activation_fn=self.activation,
padding=self.padding,
weights_regularizer=self.regularizer,
biases_initializer=self.initializer_b)
print(stack.get_shape())
n_filters = self.n_filters_first_conv
pool_heads = []
for i in range(self.n_pool + 1):
print("Dense Block ", i + 1)
with tf.variable_scope("block" + str(i + 1)):
for j in range(self.n_layers_per_block[i]):
with tf.variable_scope("layer" + str(j + 1)):
#with tf.variable_scope("1x1conv"):
#conv=self.conv(stack,self.growth_rate*4,kernel_size=1)
with tf.variable_scope("3x3conv"):
conv = self.conv(stack,
self.growth_rate,
kernel_size=3)
stack = tf.concat([stack, conv], axis=-1)
n_filters += self.growth_rate
print(stack.get_shape(), " n=", n_filters)
pool_heads.append(stack)
if self.n_pool == i:
continue
with tf.variable_scope("Downsample" + str(i + 1)):
stack = self.downSample(stack, int(n_filters * 0.5))
print(n_filters / 2)
stack = pool_heads[len(pool_heads) - 1]
print("Turning up:", stack.get_shape())
for i in range(self.n_pool):
print("Dense Block ", i + self.n_pool + 1)
with tf.variable_scope("Upsample" + str(i + 1)):
concat = pool_heads[len(pool_heads) - i - 2]
n_filters_keep = self.growth_rate * self.n_layers_per_block[
self.n_pool + i]
print(n_filters)
stack = self.upsample(stack, concat,
int(n_filters_keep * self.compression))
with tf.variable_scope("unpool-block" + str(i + 1)):
for j in range(self.n_layers_per_block[self.n_pool + i + 1]):
with tf.variable_scope("layer" + str(j + 1)):
#with tf.variable_scope("1x1conv" + str(j + 1)):
#conv=self.conv(stack,self.growth_rate*4,kernel_size=1)
with tf.variable_scope("3x3conv" + str(j + 1)):
conv = self.conv(stack, self.growth_rate)
stack = tf.concat([stack, conv], axis=-1)
print(stack.get_shape())
with tf.variable_scope("last-conv"):
batchnorm = tf.layers.batch_normalization(stack)
logits = slim.conv2d(inputs=batchnorm,
num_outputs=self.n_classes + 1,
kernel_size=1,
padding='SAME',
weights_initializer=self.initializer,
biases_initializer=self.initializer_b,
activation_fn=None)
print("logits=", logits.get_shape())
with tf.variable_scope("softmax"):
pred = tf.argmax(tf.nn.softmax(logits), axis=-1)
print(pred.get_shape())
p = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)
#print(p)
for i, a in enumerate(p):
if (i % 8) == 3:
#print(i)
tf.summary.histogram(
a.name,
tf.Graph.get_tensor_by_name(tf.get_default_graph(),
a.name))
pred = tf.expand_dims(pred, axis=-1)
print("pred==", pred.get_shape())
tf.summary.image(
"pred",
tf.multiply(tf.constant(255, dtype=tf.uint8),
tf.cast(pred, tf.uint8)))
# tf.summary.histogram("bias", kernel)
return logits, pred
def _build_model(self, inputs):
self.inputs = inputs
if self.data_format == 'NCHW':
channel_axis = 1
_inputs = tf.cast(tf.transpose(inputs, [0, 3, 1, 2]), tf.float32)
else:
channel_axis = 3
_inputs = tf.cast(inputs, tf.float32)
self.L = []
with arg_scope([layers.avg_pool2d], \
padding='VALID', data_format=self.data_format):
with tf.variable_scope('SRM_preprocess'):
W_SRM = tf.get_variable('W', initializer=SRM_Kernels, \
dtype=tf.float32, \
regularizer=None)
b = tf.get_variable('b', shape=[30], dtype=tf.float32, \
initializer=tf.constant_initializer(0.))
self.L.append(tf.nn.bias_add( \
tf.nn.conv2d(_inputs, \
W_SRM, [1,1,1,1], 'VALID', \
data_format=self.data_format), b, \
data_format=self.data_format, name='Layer1'))
self.L.append(tf.clip_by_value(self.L[-1], \
-self.tlu_threshold, self.tlu_threshold, \
name='TLU'))
with tf.variable_scope('ConvNetwork'):
with arg_scope([my_layers.conv2d], num_outputs=30, \
kernel_size=3, stride=1, padding='VALID', \
data_format=self.data_format, \
activation_fn=tf.nn.relu, \
weights_initializer=layers.xavier_initializer_conv2d(), \
weights_regularizer=layers.l2_regularizer(5e-4), \
biases_initializer=tf.constant_initializer(0.2), \
biases_regularizer=None), arg_scope([layers.batch_norm], \
decay=0.9, center=True, scale=True, \
updates_collections=None, is_training=self.is_training, \
fused=True, data_format=self.data_format):
if self.with_bn:
self.L.append(layers.batch_norm(self.L[-1], \
scope='Norm1'))
self.L.append(my_layers.conv2d(self.L[-1], \
scope='Layer2'))
if self.with_bn:
self.L.append(layers.batch_norm(self.L[-1], \
scope='Norm2'))
self.L.append(my_layers.conv2d(self.L[-1], \
scope='Layer3'))
if self.with_bn:
self.L.append(layers.batch_norm(self.L[-1], \
scope='Norm3'))
self.L.append(my_layers.conv2d(self.L[-1], \
scope='Layer4'))
if self.with_bn:
self.L.append(layers.batch_norm(self.L[-1], \
scope='Norm4'))
self.L.append(layers.avg_pool2d(self.L[-1], \
kernel_size=[2,2], scope='Stride1'))
with arg_scope([my_layers.conv2d], kernel_size=5, \
num_outputs=32):
self.L.append(my_layers.conv2d(self.L[-1], \
scope='Layer5'))
if self.with_bn:
self.L.append(layers.batch_norm(self.L[-1], \
scope='Norm5'))
self.L.append(layers.avg_pool2d(self.L[-1], \
kernel_size=[3,3], \
scope='Stride2'))
self.L.append(my_layers.conv2d(self.L[-1], \
scope='Layer6'))
if self.with_bn:
self.L.append(layers.batch_norm(self.L[-1], \
scope='Norm6'))
self.L.append(layers.avg_pool2d(self.L[-1], \
kernel_size=[3,3], \
scope='Stride3'))
self.L.append(my_layers.conv2d(self.L[-1], \
scope='Layer7'))
if self.with_bn:
self.L.append(layers.batch_norm(self.L[-1], \
scope='Norm7'))
self.L.append(layers.avg_pool2d(self.L[-1], \
kernel_size=[3,3], \
scope='Stride4'))
self.L.append(my_layers.conv2d(self.L[-1], \
num_outputs=16, \
scope='Layer8'))
if self.with_bn:
self.L.append(layers.batch_norm(self.L[-1], \
scope='Norm8'))
self.L.append(my_layers.conv2d(self.L[-1], \
num_outputs=16, stride=3, \
scope='Layer9'))
if self.with_bn:
self.L.append(layers.batch_norm(self.L[-1], \
scope='Norm9'))
self.L.append(layers.flatten(self.L[-1]))
self.L.append(layers.fully_connected(self.L[-1], num_outputs=2, \
activation_fn=None, normalizer_fn=None, \
weights_initializer=tf.random_normal_initializer(mean=0., stddev=0.01), \
biases_initializer=tf.constant_initializer(0.), scope='ip'))
self.outputs = self.L[-1]
return self.outputs
def _build_critic_network(self):
""" Common shared critic. Take the observation(state) from all agencies
First observation state_inputs[0] must be the observation of 'self' agent.
Variable:
num_assist (int) : number of assisting state
critic_state_input (placeholder) : state placeholder for critic evaluation
assist_states (list placeholder) : state placeholder for assists critic evaluation
assist_mask (boolean placeholder) : mask of assist inputs (Block if flowpath is blocked)
"""
num_assist = self.num_agent-1
self.critic_state_input = tf.placeholder(shape=self.in_size, dtype=tf.float32, name='cr_state_hold')
self.assist_states = [tf.placeholder(shape=self.in_size, dtype=tf.float32, name='assist_states')
for _ in range(num_assist)]
self.mask = tf.placeholder(shape=[None, self.num_agent], dtype=tf.float32, name='mask')
scope = 'critic'
critic_evaluations = []
with tf.variable_scope(scope):
for input_tensor in [self.critic_state_input]+self.assist_states:
net = layers.conv2d(input_tensor,
32,
[3,3],
activation_fn=tf.nn.relu,
weights_initializer=layers.xavier_initializer_conv2d(),
biases_initializer=tf.zeros_initializer(),
padding='VALID',
scope='conv1',
reuse=tf.AUTO_REUSE)
net = layers.max_pool2d(net, [2,2])
net = layers.conv2d(net,
64,
[2,2],
activation_fn=tf.nn.relu,
weights_initializer=layers.xavier_initializer_conv2d(),
biases_initializer=tf.zeros_initializer(),
padding='VALID',
scope='conv2',
reuse=tf.AUTO_REUSE)
net = layers.flatten(net)
net = layers.fully_connected(net, 128, scope='dense1', reuse=tf.AUTO_REUSE)
net = layers.fully_connected(net, 1, activation_fn=None, scope='dense2', reuse=tf.AUTO_REUSE)
#if input_tensor in self.assist_states:
# net = tf.stop_gradient(net)
critic_evaluations.append(tf.reshape(net,[-1,1]))
with tf.name_scope('Concat'):
net = tf.concat(critic_evaluations, 1) # [None, 4]
reweight = tf.get_variable(name='critic_reweight',
shape=[self.num_agent],
dtype=tf.float32,
initializer=tf.constant_initializer(value=1.0/self.num_agent)
)
shift = tf.get_variable(name='critic_shift',
shape=[1],
dtype=tf.float32,
initializer=tf.zeros_initializer
)
net = tf.multiply(net, reweight) + shift
net = tf.reduce_sum(tf.multiply(net, self.mask), axis=1)
vars_list = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope=self.scope+'/'+scope)
return net, vars_list
def _create_weight(self, name, shape):
init = layers.xavier_initializer_conv2d(dtype=tf.float32)
return tf.Variable(init(shape=shape), name=name)
def _initialize_weights(self):
all_weights = dict()
n_layers = len(self.n_hidden)
all_weights['Coef'] = tf.Variable(
0 * tf.ones([self.batch_size, self.batch_size], tf.float32),
name='Coef')
all_weights['enc_w0'] = tf.get_variable(
"enc_w0",
shape=[
self.kernel_size[0], self.kernel_size[0], 1, self.n_hidden[0]
],
initializer=layers.xavier_initializer_conv2d(),
regularizer=self.reg)
all_weights['enc_b0'] = tf.Variable(
tf.zeros([self.n_hidden[0]],
dtype=tf.float32)) # , name = 'enc_b0'
iter_i = 1
while iter_i < n_layers:
enc_name_wi = 'enc_w' + str(iter_i)
all_weights[enc_name_wi] = tf.get_variable(enc_name_wi, shape=[self.kernel_size[iter_i], self.kernel_size[iter_i], self.n_hidden[iter_i-1], \
self.n_hidden[iter_i]], initializer=layers.xavier_initializer_conv2d(),regularizer = self.reg)
enc_name_bi = 'enc_b' + str(iter_i)
all_weights[enc_name_bi] = tf.Variable(
tf.zeros([self.n_hidden[iter_i]],
dtype=tf.float32)) # , name = enc_name_bi
iter_i = iter_i + 1
iter_i = 1
while iter_i < n_layers:
dec_name_wi = 'dec_w' + str(iter_i - 1)
all_weights[dec_name_wi] = tf.get_variable(
dec_name_wi,
shape=[
self.kernel_size[n_layers - iter_i],
self.kernel_size[n_layers - iter_i],
self.n_hidden[n_layers - iter_i - 1],
self.n_hidden[n_layers - iter_i]
],
initializer=layers.xavier_initializer_conv2d(),
regularizer=self.reg)
dec_name_bi = 'dec_b' + str(iter_i - 1)
all_weights[dec_name_bi] = tf.Variable(
tf.zeros([self.n_hidden[n_layers - iter_i - 1]],
dtype=tf.float32)) # , name = dec_name_bi
iter_i = iter_i + 1
dec_name_wi = 'dec_w' + str(iter_i - 1)
all_weights[dec_name_wi] = tf.get_variable(
dec_name_wi,
shape=[
self.kernel_size[0], self.kernel_size[0], 1, self.n_hidden[0]
],
initializer=layers.xavier_initializer_conv2d(),
regularizer=self.reg)
dec_name_bi = 'dec_b' + str(iter_i - 1)
all_weights[dec_name_bi] = tf.Variable(tf.zeros(
[1], dtype=tf.float32)) # , name = dec_name_bi
return all_weights
def __init__(self, params=None):
self.log_dis = {}
self.log_dis[0] = [0, 0]
for i in range(1, 10):
self.log_dis[i] = [self.log_dis[i-1][1] + 1, self.log_dis[i-1][1] + 2**i]
self.params = params
self.keep_prob = params["keep_prob"]
self.nb_epochs = params["nb_epochs"]
# load word and char dictionaries
dicts = pickle.load(open(params["dicts_file"], "rb"))
self.w2i = dicts["w2i"]
self.i2w = dicts["i2w"]
self.c2i = dicts["c2i"]
self.i2c = dicts["i2c"]
self.word_dim = params["word_dim"]
self.word_vocab_size = len(self.w2i)
self.char_dim = params["char_dim"]
self.char_vocab_size = len(self.c2i)
print("Sizes of word and char dictionaries: {}, {}".format(self.word_vocab_size, self.char_vocab_size))
self.word_emb = np.zeros(shape=(self.word_vocab_size, self.word_dim))
# load word embedding
if "word_emb" in params:
print("pre-trained word embedding {} is being loaded ...".format(params["word_emb"]))
self.load_word_emb(params["word_emb"])
tf.reset_default_graph()
# [sent, word]
self.tf_word_ids = tf.placeholder(dtype=tf.int32, shape=[None, None], name="word_ids")
# real length sents
self.tf_sentence_lengths= tf.placeholder(dtype=tf.int32, shape=[None], name="sentence_lengths")
# [sent, word, char]
self.tf_char_ids = tf.placeholder(dtype=tf.int32, shape=[None, None, None], name="char_ids")
# binary matrix representing the relationship between sents: [sent, sent]
self.tf_target_matrix = tf.placeholder(dtype=tf.int32, shape=[None, None], name="target_matrix")
# keep_prob
self.tf_keep_prob = tf.placeholder(dtype=tf.float32, shape=[], name="keep_prob")
# learning rate
self.tf_learning_rate= tf.placeholder(dtype=tf.float32, shape=[], name="learning_rate")
# distant sentences
self.tf_distant_sents = tf.placeholder(dtype=tf.int32, shape=[None, None], name="tf_distant_sents")
# load word embedding
with tf.variable_scope("word_embedding"):
tf_word_embeddings = tf.Variable(self.word_emb, dtype=tf.float32,
trainable=True, name="word_embedding")
embedded_words = tf.nn.embedding_lookup(tf_word_embeddings, self.tf_word_ids, name="embedded_words")
self.input = embedded_words # sent, word, word_dim
# CNN network to capture character-level features
with tf.variable_scope("char_cnn"):
tf_char_embeddings = tf.get_variable(name="char_embeddings",
dtype=tf.float32,
shape=[self.char_vocab_size, self.char_dim],
trainable=True,
initializer=xavier_initializer())
conv = tf.nn.embedding_lookup(tf_char_embeddings,
self.tf_char_ids,
name="embedded_cnn_chars")
for i, (ks, fil) in enumerate(self.params["conv"]):
conv = tf.layers.conv2d(inputs=conv, # sent, word, char, feature
filters=fil,
kernel_size=(1, ks),
strides=(1, 1),
padding="same",
name="conv_{}".format(i),
kernel_initializer=xavier_initializer_conv2d())
# sent, word, char, cnn_feature
# shape = tf.shape(conv)
# conv = tf.nn.dropout(x=conv, keep_prob=self.tf_keep_prob, noise_shape=[shape[0], 1, shape[2], shape[3]])
# conv = tf.nn.dropout(x=conv, keep_prob=self.tf_keep_prob)
self.char_cnn = tf.reduce_max(conv, axis=2) # sent, word, cnn_feature
self.input = tf.nn.dropout(tf.concat([self.input, self.char_cnn], axis=-1), self.tf_keep_prob) # [sents, words, word_dim + cnn_features]
# Bi-LSTM to generate final input representation in combination with both left and right contexts
with tf.variable_scope("bi_lstm_words"):
cell_fw = tf.contrib.rnn.LSTMCell(self.params["word_lstm_units"])
cell_bw = tf.contrib.rnn.LSTMCell(self.params["word_lstm_units"])
(output_fw, output_bw), _ = tf.nn.bidirectional_dynamic_rnn(cell_fw, cell_bw, self.input,
sequence_length=self.tf_sentence_lengths,
dtype=tf.float32)
bilstm_output = tf.concat([output_fw, output_bw], axis=-1) # [sents, words, 2*lstm_units]
mask = tf.where(condition=tf.equal(self.tf_word_ids, self.w2i["<PAD>"]), x=-1e10*tf.ones_like(self.tf_word_ids, dtype=tf.float32), y=tf.zeros_like(self.tf_word_ids, dtype=tf.float32))
mask = tf.tile(tf.expand_dims(mask, -1), (1, 1, 2*self.params["word_lstm_units"]))
bilstm_output = bilstm_output + mask
# bilstm_output = tf.nn.dropout(tf.reduce_max(bilstm_output, axis=1), self.tf_keep_prob) # [sents, 2*word_lstm_units]
bilstm_output = tf.reduce_max(bilstm_output, axis=1) # [sents, 2*word_lstm_units]
# represent sents in their contexts
with tf.variable_scope("sent_representation"):
# sent_rep = tf.layers.conv1d(inputs=sent_rep[None, :, :], filters=2*self.lstm_num_units, kernel_size=3, padding='same')
# sent_rep = tf.squeeze(sent_rep, axis=0)
sent_fw = tf.contrib.rnn.LSTMCell(self.params["sent_lstm_units"])
sent_bw = tf.contrib.rnn.LSTMCell(self.params["sent_lstm_units"])
(output_sent_fw, output_sent_bw), _ = tf.nn.bidirectional_dynamic_rnn(sent_fw, sent_bw, bilstm_output[None, :, :],
# sequence_length=self.tf_sentence_lengths,
dtype=tf.float32)
sent_bilstm = tf.concat([output_sent_fw, output_sent_bw], axis=-1)
sent_rep = tf.squeeze(sent_bilstm, axis=0) # [sents, 2*sent_lstm_units]
with tf.variable_scope("self_att"):
distant_embeddings = tf.get_variable(name="distant_embeddings",
dtype=tf.float32,
shape=[10, 30],
trainable=True,
initializer=xavier_initializer()
)
embedded_disant_sent = tf.nn.embedding_lookup(distant_embeddings, self.tf_distant_sents, name="embedded_disant_sent") # [sent, sent, 20]
x = sent_rep
de = x.get_shape()[-1]
w = tf.get_variable(dtype=tf.float32, shape=[de], trainable=True, name="w")
w1 = tf.get_variable(dtype=tf.float32, shape=[de, de], trainable=True, name="W1")
w2 = tf.get_variable(dtype=tf.float32, shape=[de, de], trainable=True, name="W2")
b1 = tf.get_variable(dtype=tf.float32, shape=[de], trainable=True, name="b1")
b = tf.get_variable(dtype=tf.float32, shape=[], trainable=True, name="b")
x_w1 = tf.matmul(x, w1) # n, de
s = tf.map_fn(lambda xj: x_w1 + tf.tensordot(w2, xj, axes=1) + b1, x) # n, n, de
s = s + tf.layers.dense(embedded_disant_sent, de, use_bias=False, reuse=tf.AUTO_REUSE)
f = tf.tensordot(tf.tanh(s), w, axes=1) + b # n, n
weight = tf.nn.softmax(f, axis=1) # n, n
h = tf.transpose(tf.map_fn(lambda weight_j: tf.transpose(x)*weight_j, weight), [0, 2, 1]) # [sents, sents, 4*sent_lstm_units]
with tf.variable_scope("loss_and_opt"):
self.logits = tf.nn.dropout(tf.layers.dense(inputs=h, units=2, activation=None), self.tf_keep_prob)
self.logits = 0.5*(self.logits + tf.transpose(self.logits, [1, 0, 2]))
self.loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(labels=tf.one_hot(self.tf_target_matrix, 2),
logits=self.logits,
name="loss_function"))
self.pred = tf.argmax(self.logits, axis=-1)
eq = tf.cast(tf.equal(tf.cast(self.pred, tf.int32), self.tf_target_matrix), tf.float32)
self.acc = tf.reduce_mean(eq)
self.opt = tf.train.AdamOptimizer(learning_rate=self.tf_learning_rate).minimize(self.loss)
def redSimple(input):
out = tf.reshape(input, [-1, 128, 128, 1])
out = conv2d(out,
64,
3,
activation_fn=tf.nn.relu,
weights_initializer=xavier_initializer_conv2d())
out = batch_normalization(out)
out = conv2d(out,
64,
3,
activation_fn=tf.nn.relu,
weights_initializer=xavier_initializer_conv2d())
out = batch_normalization(out)
out = max_pool2d(out, 3, stride=2)
out = conv2d(out,
128,
3,
activation_fn=tf.nn.relu,
weights_initializer=xavier_initializer_conv2d())
out = batch_normalization(out)
out = conv2d(out,
128,
3,
activation_fn=tf.nn.relu,
weights_initializer=xavier_initializer_conv2d())
out = batch_normalization(out)
out = max_pool2d(out, 3, stride=2)
out = conv2d(out,
128,
3,
activation_fn=tf.nn.relu,
weights_initializer=xavier_initializer_conv2d())
out = batch_normalization(out)
out = conv2d(out,
128,
3,
activation_fn=tf.nn.relu,
weights_initializer=xavier_initializer_conv2d())
out = batch_normalization(out)
out = max_pool2d(out, 3, stride=2)
out = conv2d(out,
256,
3,
activation_fn=tf.nn.relu,
weights_initializer=xavier_initializer_conv2d())
out = batch_normalization(out)
out = conv2d(out,
256,
3,
activation_fn=tf.nn.relu,
weights_initializer=xavier_initializer_conv2d())
out = batch_normalization(out)
out = max_pool2d(out, 3, stride=2)
out = tf.contrib.layers.flatten(out)
out = fully_connected(out,
1024,
activation_fn=tf.nn.relu,
weights_initializer=xavier_initializer())
out = fully_connected(out,
100,
activation_fn=None,
weights_initializer=xavier_initializer())
return out
def generator_real(self, emg_data, z, reuse=False):
# (None, 300, 16) -> (1, 100)
# input = self.lstm(emg_data)
_ = self.lstm(emg_data, reuse)
input = z
print(input.shape)
# input = tf.concat([input, z], 1)
# print(input.shape)
test = True
# Generator
with tf.variable_scope('generator_real', reuse=reuse):
net = slim.fully_connected(
input,
64 * 8 * 8,
activation_fn=tf.nn.relu,
weights_initializer=tflayers.xavier_initializer(),
reuse=reuse)
print(net)
net = tf.reshape(net, [-1, 8, 8, 64])
net = slim.conv2d(net, num_outputs=128, kernel_size=1, stride=1)
print(net)
if not test:
with slim.arg_scope(
[slim.conv2d_transpose],
kernel_size=3,
stride=2,
weights_initializer=tflayers.xavier_initializer()):
with slim.arg_scope([slim.batch_norm],
activation_fn=tf.nn.relu,
is_training=(self.mode == 'train')):
net = slim.conv2d_transpose(net, num_outputs=128)
net = slim.batch_norm(net)
print(net)
net = slim.conv2d_transpose(
net, num_outputs=256) # output : 32 x 32
net = slim.batch_norm(net)
print(net)
net = slim.conv2d_transpose(
net, num_outputs=512) # output : 64 x 64
net = slim.batch_norm(net)
print(net)
net = slim.conv2d_transpose(
net, num_outputs=256) # output : 128 x 128
net = slim.batch_norm(net)
print(net)
net = slim.conv2d_transpose(
net,
num_outputs=3,
kernel_size=1,
stride=1,
weights_initializer=tflayers.xavier_initializer())
print(net)
else:
with slim.arg_scope([slim.conv2d_transpose],
kernel_size=3,
stride=1,
padding='SAME',
weights_initializer=tflayers.
xavier_initializer_conv2d()):
with slim.arg_scope([slim.batch_norm],
activation_fn=tf.nn.relu):
# 2x Upsampling -> Conv2D
# Xavier -> Truncated normal
# 16 x 16
net = tf.image.resize_images(net, [16, 16])
net = slim.conv2d_transpose(net, num_outputs=128)
net = slim.batch_norm(net)
print(net)
# 32 x 32
net = tf.image.resize_images(net, [32, 32])
net = slim.conv2d_transpose(net, num_outputs=256)
net = slim.batch_norm(net)
print(net)
# 64 x 64
net = tf.image.resize_images(net, [64, 64])
net = slim.conv2d_transpose(net, num_outputs=512)
net = slim.batch_norm(net)
print(net)
# 128 x 128
net = tf.image.resize_images(net, [128, 128])
net = slim.conv2d_transpose(net, num_outputs=256)
net = slim.batch_norm(net)
print(net)
net = slim.conv2d_transpose(
net,
num_outputs=1,
kernel_size=1,
stride=1,
padding='SAME',
activation_fn=tf.nn.tanh,
weights_initializer=tflayers.xavier_initializer_conv2d())
print(net)
return net
