def build(self, guidence, newNet):
with tf.variable_scope("training_variable"):
inputEmb = tf.nn.embedding_lookup(self.embedding, self.X)
initFw = tf.nn.rnn_cell.LSTMStateTuple(
tf.nn.relu(
tf.matmul(guidence, self.weights["Fw1"]) +
self.biases["Fw1"]),
tf.nn.relu(
tf.matmul(guidence, self.weights["Fw2"]) +
self.biases["Fw2"]))
initBw = tf.nn.rnn_cell.LSTMStateTuple(
tf.nn.relu(
tf.matmul(guidence, self.weights["Bw1"]) +
self.biases["Bw1"]),
tf.nn.relu(
tf.matmul(guidence, self.weights["Bw2"]) +
self.biases["Bw2"]))
rnnCellFw = tf.compat.v1.nn.rnn_cell.DropoutWrapper(
tf.nn.rnn_cell.BasicLSTMCell(self.nHidden),
input_keep_prob=self.pKeep,
output_keep_prob=1.0)
rnnCellBw = tf.compat.v1.nn.rnn_cell.DropoutWrapper(
tf.nn.rnn_cell.BasicLSTMCell(self.nHidden),
input_keep_prob=self.pKeep,
output_keep_prob=1.0)
outputs, state = tf.nn.bidirectional_dynamic_rnn(
cell_fw=rnnCellFw,
cell_bw=rnnCellBw,
inputs=inputEmb,
initial_state_fw=initFw,
initial_state_bw=initBw,
dtype=tf.float32)
outputsConcat = tf.concat(outputs, axis=2)
self.outputs = outputsConcat
self.RNNState = tf.reduce_mean(outputsConcat, axis=1)
def build_layer(s, c_names, n_l1, n_l2, w_initializer, b_initializer):
with tf.variable_scope('l1'):
w1 = tf.get_variable('w1', [self.n_features, n_l1],
initializer=w_initializer,
collections=c_names)
b1 = tf.get_variable('b1', [1, n_l1],
initializer=b_initializer,
collections=c_names)
l1 = tf.nn.relu(tf.matmul(s, w1) + b1)
with tf.variable_scope('l2'):
w2 = tf.get_variable('w2', [n_l1, n_l2],
initializer=w_initializer,
collections=c_names)
b2 = tf.get_variable('b2', [1, n_l2],
initializer=b_initializer,
collections=c_names)
l2 = tf.nn.relu(tf.matmul(l1, w2) + b2)
with tf.variable_scope('l3'):
w3 = tf.get_variable('w3', [n_l2, self.n_actions],
initializer=w_initializer,
collections=c_names)
b3 = tf.get_variable('b3', [1, self.n_actions],
initializer=b_initializer,
collections=c_names)
l3 = tf.nn.relu(tf.matmul(l2, w3) + b3)
return l3
def generator(x, m):
# Concatenate Mask and Data
inputs = tf.concat(values=[x, m], axis=1)
G_h1 = tf.nn.relu(tf.matmul(inputs, G_W1) + G_b1)
G_h2 = tf.nn.relu(tf.matmul(G_h1, G_W2) + G_b2)
# MinMax normalized output
G_prob = tf.nn.sigmoid(tf.matmul(G_h2, G_W3) + G_b3)
return G_prob
def discriminator(x, h):
# Concatenate Data and Hint
inputs = tf.concat(values=[x, h], axis=1)
D_h1 = tf.nn.relu(tf.matmul(inputs, D_W1) + D_b1)
D_h2 = tf.nn.relu(tf.matmul(D_h1, D_W2) + D_b2)
D_logit = tf.matmul(D_h2, D_W3) + D_b3
D_prob = tf.nn.sigmoid(D_logit)
return D_prob
def multihead_attn(q, k, v):
# q, k, v have shape [batch, heads, sequence, features]
w = tf.matmul(q, k, transpose_b=True)
# w = w * tf.rsqrt(tf.cast(v.shape[-1].value, w.dtype))
w = w * tf.rsqrt(tf.cast(v.shape[-1], w.dtype))
w = mask_attn_weights(w)
w = softmax(w)
a = tf.matmul(w, v)
return a
def model(hparams, X, past=None, scope='model', reuse=tf.AUTO_REUSE):
with tf.variable_scope(scope, reuse=reuse):
results = {}
batch, sequence = shape_list(X)
wpe = tf.get_variable('wpe', [hparams.n_ctx, hparams.n_embd],
initializer=tf.random_normal_initializer(stddev=0.01))
wte = tf.get_variable('wte', [hparams.n_vocab, hparams.n_embd],
initializer=tf.random_normal_initializer(stddev=0.02))
past_length = 0 if past is None else tf.shape(past)[-2]
h = tf.gather(wte, X) + tf.gather(wpe, positions_for(X, past_length))
# Transformer
presents = []
pasts = tf.unstack(past, axis=1) if past is not None else [None] * hparams.n_layer
assert len(pasts) == hparams.n_layer
for layer, past in enumerate(pasts):
h, present = block(h, 'h%d' % layer, past=past, hparams=hparams)
if layer == 10:
tf.add_to_collection('checkpoints', h)
presents.append(present)
results['present'] = tf.stack(presents, axis=1)
h = norm(h, 'ln_f')
# Language model loss. Do tokens <n predict token n?
h_flat = tf.reshape(h, [batch * sequence, hparams.n_embd])
logits = tf.matmul(h_flat, wte, transpose_b=True)
logits = tf.reshape(logits, [batch, sequence, hparams.n_vocab])
results['logits'] = logits
return results
def pca(x, dim=2):
'''
x:输入矩阵
dim:降维之后的维度数
'''
with tf.name_scope("PCA"):
m, n = tf.to_float(x.get_shape()[0]), tf.to_int32(x.get_shape()[1])
assert not tf.assert_less(dim, n)
mean = tf.reduce_mean(x, axis=1)
x_new = x - tf.reshape(mean, (-1, 1))
cov = tf.matmul(x_new, x_new, transpose_a=True) / (m - 1)
e, v = tf.linalg.eigh(cov, name="eigh")
e_index_sort = tf.math.top_k(e, sorted=True, k=dim)[1]
v_new = tf.gather(v, indices=e_index_sort)
pca = tf.matmul(x_new, v_new, transpose_b=True)
return pca
def conv1d(x, scope, nf, *, w_init_stdev=0.02):
with tf.variable_scope(scope):
*start, nx = shape_list(x)
w = tf.get_variable('w', [1, nx, nf],
initializer=tf.random_normal_initializer(stddev=w_init_stdev))
b = tf.get_variable('b', [nf], initializer=tf.constant_initializer(0))
c = tf.reshape(tf.matmul(tf.reshape(x, [-1, nx]), tf.reshape(w, [-1, nf])) + b,
start + [nf])
return c
def predicting(self, rate):
hidden = tf.nn.relu(
tf.matmul(self.concatInput, self.weights["MLP1"]) +
self.biases["MLP1"])
logits = tf.matmul(hidden, self.weights["MLP2"]) + self.biases["MLP2"]
predictPossibility = tf.nn.sigmoid(logits)
accuracy = tf.reduce_mean(
tf.cast(
tf.equal(tf.cast(predictPossibility > 0.5, tf.float32),
self.y), tf.float32))
loss = tf.reduce_mean(
tf.nn.weighted_cross_entropy_with_logits(targets=self.y,
logits=logits,
pos_weight=rate))
tv = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
'training_variable')
l2_loss = self.l2_para * tf.reduce_sum([tf.nn.l2_loss(v) for v in tv])
loss += l2_loss
return loss, accuracy, predictPossibility
def _build_layer(input_dim,
output_dim,
input_data,
c_name,
layer_name,
w_name='w',
bias_name='b',
has_activate=True):
with tf.variable_scope(layer_name):
w = tf.get_variable(w_name, [input_dim, output_dim],
dtype=tf.float32,
initializer=tf.truncated_normal_initializer(),
collections=c_name)
b = tf.get_variable(bias_name, [output_dim],
dtype=tf.float32,
initializer=tf.truncated_normal_initializer(),
collections=c_name)
if has_activate:
l = tf.nn.relu(tf.add(tf.matmul(input_data, w), b))
else:
l = tf.add(tf.matmul(input_data, w), b)
return l
def choose_action(self, state, epsilon=1.):
# 计算q值. epsilon为随机度. 为1时, 则为全随机选择action
M_s = np.zeros([self.n_action, self.fai_s_size])
for i in range(self.n_action):
M_s[i] = self.sess.run(self.eval_M[i],
feed_dict={self.state, state})
w = self.sess.graph.get_tensor_by_name('eval_net/R_s/r/w:0')
# fai = self.sess.graph.get_tensor_by_name('eval_net/f_theta/fai_s:0')
q_ = tf.matmul(w, tf.Variable(M_s))
if np.random.uniform(0, 1) < epsilon:
action = np.random.randint(self.n_action)
else:
action = np.argmax(q_)
return action
def __init__(self):
self.embedding = self.getEmb()
self.embSize = self.embedding.shape[1]
self.vocabSize = self.embedding.shape[0]
self.x = tf.placeholder(tf.int32, [None, 5])
with tf.variable_scope("training_variable"):
self.weights = {
"MLP1":
tf.Variable(
tf.truncated_normal(
shape=[self.embSize,
int(self.embSize / 2)],
stddev=0.08)),
"MLP2":
tf.Variable(
tf.truncated_normal(shape=[int(self.embSize / 2), 1],
stddev=0.08))
}
self.biases = {
"MLP1":
tf.Variable(
tf.constant(0.01,
shape=[int(self.embSize / 2)],
dtype=tf.float32)),
"MLP2":
tf.Variable(tf.constant(0.01, shape=[1], dtype=tf.float32))
}
self.inputEmb = tf.nn.embedding_lookup(self.embedding, self.x)
p1 = tf.matmul(tf.reshape(self.inputEmb, [-1, self.embSize]),
self.weights["MLP1"]) + self.biases["MLP1"]
p1 = tf.matmul(tf.nn.relu(p1),
self.weights["MLP2"]) + self.biases["MLP2"]
p1 = tf.reshape(p1, [-1, 5])
p1 = tf.reshape(tf.nn.softmax(p1), [-1, 1, 5])
self.finalState = tf.reshape(tf.matmul(p1, self.inputEmb),
[-1, self.embSize])
def __fully_connected(self,
input,
inputs_count,
outputs_count,
relu=True,
init_biases_with_the_constant_1=False,
name='fully_connected'):
with tf.name_scope(name):
wights = tf.Variable(
self.__random_values(shape=[inputs_count, outputs_count]),
name='wights')
if init_biases_with_the_constant_1:
biases = tf.Variable(tf.ones(shape=[outputs_count],
dtype=tf.float32),
name='biases')
else:
biases = tf.Variable(tf.zeros(shape=[outputs_count],
dtype=tf.float32),
name='biases')
preactivations = tf.nn.bias_add(tf.matmul(input, wights),
biases,
name='preactivations')
if relu:
activations = tf.nn.relu(preactivations, name='activations')
with tf.name_scope('wight_summaries'):
self.__variable_summaries(wights)
with tf.name_scope('bias_summaries'):
self.__variable_summaries(biases)
with tf.name_scope('preactivations_histogram'):
tf.summary.histogram('preactivations', preactivations)
if relu:
with tf.name_scope('activations_histogram'):
tf.summary.histogram('activations', activations)
if relu:
return activations
else:
return preactivations
def train(x_train, y_train):
n_samples, n_features = x_train.shape
w = tf.Variable(np.random.rand(input_dim, 1).astype(dtype='float32'),
name="weight")
b = tf.Variable(0.0, dtype=tf.float32, name="bias")
x = tf.placeholder(dtype=tf.float32, name='x')
y = tf.placeholder(dtype=tf.float32, name='y')
predictions = tf.matmul(x, w) + b
loss = tf.reduce_mean(
tf.log(1 + tf.exp(tf.multiply(-1.0 * y, predictions))))
# optimizer = tf.train.GradientDescentOptimizer(learn_rate).minimize(loss)
optimizer = tf.train.ProximalGradientDescentOptimizer(
learning_rate=learn_rate,
l1_regularization_strength=0.1).minimize(loss)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for epoch in range(n_epochs):
for idx in range(0, n_samples, batch_size):
iE = min(n_samples, idx + batch_size)
x_batch = x_train[idx:iE, :]
y_batch = y_train[idx:iE, :]
sess.run([optimizer], feed_dict={x: x_batch, y: y_batch})
curr_w, curr_b = sess.run([w, b])
for idx in range(len(curr_w)):
if curr_w[idx] < threshold * -1:
curr_w[idx] += threshold
else:
curr_w[idx] -= threshold
sess.run([tf.assign(w, curr_w)])
return curr_w, curr_b
def Attention(self, g0, sequence, g1, g2, AttStr, AttStr2, length, flag):
tmpLS = []
tmpLS2 = []
tmpLS3 = []
if not flag:
seq = tf.transpose(sequence, [1, 0, 2])
else:
seq = sequence
for i in range(length):
nHidden = tf.tanh(
tf.matmul(tf.concat([seq[i], g1], axis=1), self.weights[
AttStr + "_1"]) + self.biases[AttStr + "_1"])
nHidden2 = tf.tanh(
tf.matmul(tf.concat([seq[i], g2], axis=1), self.weights[
AttStr + "_2"]) + self.biases[AttStr + "_2"])
nHidden3 = tf.tanh(
tf.matmul(tf.concat([seq[i], g0], axis=1), self.weights[
AttStr + "_3"]) + self.biases[AttStr + "_3"])
tmpLS.append(
tf.matmul(nHidden, self.weights[AttStr2 + "_1"]) +
self.biases[AttStr2 + "_1"])
tmpLS2.append(
tf.matmul(nHidden2, self.weights[AttStr2 + "_2"]) +
self.biases[AttStr2 + "_2"])
tmpLS3.append(
tf.matmul(nHidden3, self.weights[AttStr2 + "_3"]) +
self.biases[AttStr2 + "_3"])
tmpLS = tf.nn.softmax(tmpLS, dim=0)
tmpLS2 = tf.nn.softmax(tmpLS2, dim=0)
tmpLS3 = tf.nn.softmax(tmpLS3, dim=0)
self.representation_score[AttStr] = (tmpLS + tmpLS2 + tmpLS3) / 3
ret = tmpLS[0] * seq[0] / 3 + tmpLS2[0] * seq[0] / 3 + tmpLS3[0] * seq[
0] / 3
for i in range(1, length):
ret += tmpLS[i] * seq[i] / 3 + tmpLS2[i] * seq[i] / 3 + tmpLS3[
i] * seq[i] / 3
return ret
def main(trainModel=True,
buildConfusionMatrix=True,
restore=False,
buildClassifiedMatrix=True):
tf.disable_v2_behavior()
input_images = tf.placeholder(tf.float32, [None, 28, 28], name="Input")
real = tf.placeholder(tf.float32, [None, CLASSES], name="real_classes")
layer1 = create_conv_layer(tf.reshape(input_images, [-1, 28, 28, 1]),
1,
28, [5, 5], [2, 2],
name="conv_no_pool")
layer2 = create_conv_layer(layer1,
28,
56, [5, 5], [2, 2],
name='conv_with_pool')
conv_result = tf.reshape(layer2, [-1, 7 * 7 * 56])
relu_layer_weight = tf.Variable(tf.truncated_normal([7 * 7 * 56, 1000],
stddev=STDDEV * 2),
name='relu_layer_weight')
rely_layer_bias = tf.Variable(tf.truncated_normal([1000],
stddev=STDDEV / 2),
name='rely_layer_bias')
relu_layer = tf.matmul(conv_result, relu_layer_weight) + rely_layer_bias
relu_layer = tf.nn.relu(relu_layer)
relu_layer = tf.nn.dropout(relu_layer, DROPOUT)
final_layer_weight = tf.Variable(tf.truncated_normal([1000, CLASSES],
stddev=STDDEV * 2),
name='final_layer_weight')
final_layer_bias = tf.Variable(tf.truncated_normal([CLASSES],
stddev=STDDEV / 2),
name='final_layer_bias')
final_layer = tf.matmul(relu_layer, final_layer_weight) + final_layer_bias
predicts = tf.nn.softmax(final_layer)
predicts_for_log = tf.clip_by_value(predicts, 1e-9, 0.999999999)
#crossEntropy = -tf.reduce_mean(tf.reduce_sum(y * tf.log(y_clipped) + (1 - y) * tf.log(1 - y_clipped), axis=1))
loss = -tf.reduce_mean(
tf.reduce_sum(real * tf.log(predicts_for_log) +
(1 - real) * tf.log(1 - predicts_for_log),
axis=1),
axis=0)
#test = tf.reduce_sum(real * tf.log(predicts_for_log) + (1 - real) * tf.log(1 - predicts_for_log), axis=1)
#loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=final_layer, labels=real))
optimiser = tf.train.GradientDescentOptimizer(
learning_rate=LEARNING_RATE).minimize(loss)
correct_prediction = tf.equal(tf.argmax(real, axis=1),
tf.argmax(predicts, axis=1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
confusion_matrix = tf.confusion_matrix(labels=tf.argmax(real, axis=1),
predictions=tf.argmax(predicts,
axis=1),
num_classes=CLASSES)
saver = tf.train.Saver()
# dataset = get_mnist_dataset()
dataset = get_fashion_dataset()
with tf.Session() as session:
session.run(tf.global_variables_initializer())
if restore:
saver.restore(session, SAVE_PATH)
if trainModel:
train(input_images, real, session, optimiser, loss, accuracy,
saver, dataset)
if buildConfusionMatrix:
test_cm = session.run(confusion_matrix,
feed_dict={
input_images: dataset.test_x,
real: dataset.test_y
})
draw_confusion_matrix(test_cm)
if buildClassifiedMatrix:
all_probs = session.run(predicts,
feed_dict={
input_images: dataset.test_x,
real: dataset.test_y
})
max_failure_picture_index = [[(-1, -1.0)] * CLASSES
for _ in range(CLASSES)]
for i in range(len(all_probs)):
real = np.argmax(dataset.test_y[i])
for j in range(CLASSES):
if max_failure_picture_index[real][j][1] < all_probs[i][j]:
max_failure_picture_index[real][j] = (i,
all_probs[i][j])
draw_max_failure_pictures(dataset.test_x,
max_failure_picture_index)
import tensorflow._api.v2.compat.v1 as tf
import numpy as np
tf.reset_default_graph()
tf.compat.v1.disable_eager_execution()
tf.compat.v1.disable_v2_behavior()
tf.global_variables_initializer()
x = tf.placeholder(tf.float32, shape=[None, 4])
y = tf.placeholder(tf.float32, shape=[None, 1])
w = tf.Variable(tf.random_normal([4, 1]), name="weight")
b = tf.Variable(tf.random_normal([1]), name="bias")
hypo = tf.matmul(x, w) + b
saver = tf.train.Saver()
test_arr = [[12, 6.5, 15.7, 10.8]]
sess2 = tf.Session()
saver.restore(sess2, "./saved.ckpt")
predict = sess2.run(hypo, feed_dict={x: test_arr})
print(predict[0])
# Import data
# from tensorflow.examples.tutorials.mnist import input_data
import input_data
mnist = input_data.read_data_sets("tmp/data/", one_hot=True)
g = tf.Graph()
with g.as_default():
# Create the model
# x = tf.placeholder("float", [None, 784])
x = tf.compat.v1.placeholder("float", [None, 784])
W = tf.Variable(tf.zeros([784, 10]), name="vaiable_W")
b = tf.Variable(tf.zeros([10]), name="variable_b")
y = tf.nn.softmax(tf.matmul(x, W) + b)
# Define loss and optimizer
y_ = tf.compat.v1.placeholder("float", [None, 10])
cross_entropy = -tf.reduce_sum(y_ * tf.log(y))
train_step = tf.train.GradientDescentOptimizer(0.01).minimize(
cross_entropy)
sess = tf.Session()
# Train
init = tf.initialize_all_variables()
sess.run(init)
for i in range(1000):
batch_xs, batch_ys = mnist.train.next_batch(100)
# Create a simple TF Graph # By Omid Alemi - Jan 2017 # Works with TF r1.0 # import tensorflow as tf import tensorflow._api.v2.compat.v1 as tf tf.disable_v2_behavior() I = tf.placeholder(tf.float32, shape=[None,3], name='I') # input W = tf.Variable(tf.zeros(shape=[3,2]), dtype=tf.float32, name='W') # weights b = tf.Variable(tf.zeros(shape=[2]), dtype=tf.float32, name='b') # biases O = tf.nn.relu(tf.matmul(I, W) + b, name='O') # activation / output saver = tf.train.Saver() init_op = tf.global_variables_initializer() with tf.Session() as sess: sess.run(init_op) # save the graph tf.train.write_graph(sess.graph_def, '.', 'tfdroid.pbtxt') # normally you would do some training here # but fornow we will just assign something to W sess.run(tf.assign(W, [[1, 2],[4,5],[7,8]])) sess.run(tf.assign(b, [1,1])) #save a checkpoint file, which will store the above assignment saver.save(sess, 'tfdroid.ckpt')
def __init__(self, nHidden, seqLen):
self.representation_score = {}
self.y = tf.placeholder(tf.float32, shape=[None, 1])
self.extractFeature = ExtractFeature.ExtractFeature()
self.imageFeature = ImageFeature.ImageFeature()
newNet = tf.reduce_mean(self.imageFeature.outputLS, axis=0)
self.textFeature = TextFeature.TextFeature(
nHidden, seqLen, self.extractFeature.finalState, newNet)
self.l2_para = 1e-7
with tf.variable_scope("training_variable"):
self.weights = {
"MLP1":
tf.Variable(
tf.truncated_normal(shape=[512, 256],
stddev=0.08,
name="MLP1_W")),
"MLP2":
tf.Variable(
tf.truncated_normal(shape=[256, 1],
stddev=0.08,
name="MLP2_W")),
"ATT_attr1_1":
tf.Variable(
tf.truncated_normal(shape=[
self.imageFeature.defaultFeatureSize +
self.extractFeature.embSize,
int(self.imageFeature.defaultFeatureSize / 2 +
self.extractFeature.embSize / 2)
],
stddev=0.08,
name="ATT_attr1_1")),
"ATT_attr1_2":
tf.Variable(
tf.truncated_normal(shape=[
self.textFeature.nHidden * 2 +
self.extractFeature.embSize,
int(self.textFeature.nHidden +
self.extractFeature.embSize / 2)
],
stddev=0.08,
name="ATT_attr1_2")),
"ATT_attr1_3":
tf.Variable(
tf.truncated_normal(shape=[
2 * self.extractFeature.embSize,
self.extractFeature.embSize
],
stddev=0.08,
name="ATT_attr1_3")),
"ATT_attr2_1":
tf.Variable(
tf.truncated_normal(shape=[
int(self.imageFeature.defaultFeatureSize / 2 +
self.extractFeature.embSize / 2), 1
],
stddev=0.08,
name="ATT_attr2_1")),
"ATT_attr2_2":
tf.Variable(
tf.truncated_normal(shape=[
int(self.textFeature.nHidden +
self.extractFeature.embSize / 2), 1
],
stddev=0.08,
name="ATT_attr2_2")),
"ATT_attr2_3":
tf.Variable(
tf.truncated_normal(shape=[self.extractFeature.embSize, 1],
stddev=0.08,
name="ATT_attr2_3")),
"ATT_img1_1":
tf.Variable(
tf.truncated_normal(shape=[
self.imageFeature.defaultFeatureSize +
self.textFeature.nHidden * 2,
int(self.imageFeature.defaultFeatureSize / 2 +
self.textFeature.nHidden)
],
stddev=0.08,
name="ATT_image1_1")),
"ATT_img1_2":
tf.Variable(
tf.truncated_normal(shape=[
self.imageFeature.defaultFeatureSize +
self.extractFeature.embSize,
int(self.imageFeature.defaultFeatureSize / 2 +
self.extractFeature.embSize / 2)
],
stddev=0.08,
name="ATT_image1_2")),
"ATT_img1_3":
tf.Variable(
tf.truncated_normal(shape=[
self.imageFeature.defaultFeatureSize * 2,
self.imageFeature.defaultFeatureSize
],
stddev=0.08,
name="ATT_image1_3")),
"ATT_img2_1":
tf.Variable(
tf.truncated_normal(shape=[
int(self.imageFeature.defaultFeatureSize / 2 +
self.textFeature.nHidden), 1
],
stddev=0.08,
name="ATT_image2_1")),
"ATT_img2_2":
tf.Variable(
tf.truncated_normal(shape=[
int(self.imageFeature.defaultFeatureSize / 2 +
self.extractFeature.embSize / 2), 1
],
stddev=0.08,
name="ATT_image2_2")),
"ATT_img2_3":
tf.Variable(
tf.truncated_normal(
shape=[self.imageFeature.defaultFeatureSize, 1],
stddev=0.08,
name="ATT_image2_3")),
"ATT_text1_1":
tf.Variable(
tf.truncated_normal(shape=[
self.imageFeature.defaultFeatureSize +
self.textFeature.nHidden * 2,
int(self.imageFeature.defaultFeatureSize / 2 +
self.textFeature.nHidden)
],
stddev=0.08,
name="ATT_text1_1")),
"ATT_text1_2":
tf.Variable(
tf.truncated_normal(shape=[
self.textFeature.nHidden * 2 +
self.extractFeature.embSize,
int(self.textFeature.nHidden +
self.extractFeature.embSize / 2)
],
stddev=0.08,
name="ATT_text1_2")),
"ATT_text1_3":
tf.Variable(
tf.truncated_normal(shape=[
self.textFeature.nHidden * 4,
self.textFeature.nHidden * 2
],
stddev=0.08,
name="ATT_text1_3")),
"ATT_text2_1":
tf.Variable(
tf.truncated_normal(shape=[
int(self.imageFeature.defaultFeatureSize / 2 +
self.textFeature.nHidden), 1
],
stddev=0.08,
name="ATT_text2_1")),
"ATT_text2_2":
tf.Variable(
tf.truncated_normal(shape=[
int(self.textFeature.nHidden +
self.extractFeature.embSize / 2), 1
],
stddev=0.08,
name="ATT_text2_2")),
"ATT_text2_3":
tf.Variable(
tf.truncated_normal(
shape=[self.textFeature.nHidden * 2, 1],
stddev=0.08,
name="ATT_text2_3")),
"ATT_WI1":
tf.Variable(
tf.truncated_normal(
shape=[self.imageFeature.defaultFeatureSize, 512],
stddev=0.08,
name="ATT_WI")),
"ATT_WT1":
tf.Variable(
tf.truncated_normal(shape=[2 * nHidden, 512],
stddev=0.08,
name="ATT_WT")),
"ATT_WA1":
tf.Variable(
tf.truncated_normal(shape=[200, 512],
stddev=0.08,
name="ATT_WA")),
"ATT_WI2":
tf.Variable(
tf.truncated_normal(
shape=[self.imageFeature.defaultFeatureSize, 512],
stddev=0.08,
name="ATT_WI2")),
"ATT_WT2":
tf.Variable(
tf.truncated_normal(shape=[2 * nHidden, 512],
stddev=0.08,
name="ATT_WT2")),
"ATT_WA2":
tf.Variable(
tf.truncated_normal(shape=[200, 512],
stddev=0.08,
name="ATT_WA2")),
"ATT_WF_1":
tf.Variable(
tf.truncated_normal(shape=[512, 1],
stddev=0.08,
name="ATT_WF_1")),
"ATT_WF_2":
tf.Variable(
tf.truncated_normal(shape=[512, 1],
stddev=0.08,
name="ATT_WF_2")),
"ATT_WF_3":
tf.Variable(
tf.truncated_normal(shape=[512, 1],
stddev=0.08,
name="ATT_WF_3")),
}
self.biases = {
"MLP1":
tf.Variable(
tf.constant(0.01,
shape=[256],
dtype=tf.float32,
name="MLP1_b")),
"MLP2":
tf.Variable(
tf.constant(0.01,
shape=[1],
dtype=tf.float32,
name="MLP2_b")),
"ATT_attr1_1":
tf.Variable(
tf.constant(
0.01,
shape=[
int(self.imageFeature.defaultFeatureSize / 2 +
self.extractFeature.embSize / 2)
],
name="ATT_attr1_1")),
"ATT_attr1_2":
tf.Variable(
tf.constant(0.01,
shape=[
int(self.textFeature.nHidden +
self.extractFeature.embSize / 2)
],
name="ATT_attr1_2")),
"ATT_attr1_3":
tf.Variable(
tf.constant(0.01,
shape=[self.extractFeature.embSize],
name="ATT_attr1_3")),
"ATT_attr2_1":
tf.Variable(tf.constant(0.01, shape=[1], name="ATT_attr2_1")),
"ATT_attr2_2":
tf.Variable(tf.constant(0.01, shape=[1], name="ATT_attr2_2")),
"ATT_attr2_3":
tf.Variable(tf.constant(0.01, shape=[1], name="ATT_attr2_3")),
"ATT_img1_1":
tf.Variable(
tf.constant(
0.01,
shape=[
int(self.imageFeature.defaultFeatureSize / 2 +
self.textFeature.nHidden)
],
name="ATT_image1_1")),
"ATT_img1_2":
tf.Variable(
tf.constant(
0.01,
shape=[
int(self.imageFeature.defaultFeatureSize / 2 +
self.extractFeature.embSize / 2)
],
name="ATT_image1_2")),
"ATT_img1_3":
tf.Variable(
tf.constant(0.01,
shape=[self.imageFeature.defaultFeatureSize],
name="ATT_image1_3")),
"ATT_img2_1":
tf.Variable(tf.constant(0.01, shape=[1], name="ATT_image2_1")),
"ATT_img2_2":
tf.Variable(tf.constant(0.01, shape=[1], name="ATT_image2_2")),
"ATT_img2_3":
tf.Variable(tf.constant(0.01, shape=[1], name="ATT_image2_3")),
"ATT_text1_1":
tf.Variable(
tf.constant(
0.01,
shape=[
int(self.imageFeature.defaultFeatureSize / 2 +
self.textFeature.nHidden)
],
name="ATT_text1_1")),
"ATT_text1_2":
tf.Variable(
tf.constant(0.01,
shape=[
int(self.textFeature.nHidden +
self.extractFeature.embSize / 2)
],
name="ATT_text1_2")),
"ATT_text1_3":
tf.Variable(
tf.constant(0.01,
shape=[self.textFeature.nHidden * 2],
name="ATT_text1_3")),
"ATT_text2_1":
tf.Variable(tf.constant(0.01, shape=[1], name="ATT_text2_1")),
"ATT_text2_2":
tf.Variable(tf.constant(0.01, shape=[1], name="ATT_text2_2")),
"ATT_text2_3":
tf.Variable(tf.constant(0.01, shape=[1], name="ATT_text2_3")),
"ATT_WW":
tf.Variable(tf.constant(0.01, shape=[512], name="ATT_WW")),
"ATT_WI":
tf.Variable(tf.constant(0.01, shape=[512], name="ATT_WI")),
"ATT_WT":
tf.Variable(tf.constant(0.01, shape=[512], name="ATT_WT")),
"ATT_WI1":
tf.Variable(tf.constant(0.01, shape=[512], name="ATT_WI1")),
"ATT_WT1":
tf.Variable(tf.constant(0.01, shape=[512], name="ATT_WT1")),
"ATT_WA":
tf.Variable(tf.constant(0.01, shape=[512], name="ATT_WA")),
"ATT_WF_1":
tf.Variable(tf.constant(0.01, shape=[1], name="ATT_WF_1")),
"ATT_WF_2":
tf.Variable(tf.constant(0.01, shape=[1], name="ATT_WF_2")),
"ATT_WF_3":
tf.Variable(tf.constant(0.01, shape=[1], name="ATT_WF_3")),
}
print("newnet dimension :", newNet)
imageVec = self.Attention(newNet, self.imageFeature.outputLS,
self.textFeature.RNNState,
self.extractFeature.finalState, "ATT_img1",
"ATT_img2", 196, True)
textVec = self.Attention(self.textFeature.RNNState,
self.textFeature.outputs, newNet,
self.extractFeature.finalState, "ATT_text1",
"ATT_text2", self.textFeature.seqLen, False)
attrVec = self.Attention(self.extractFeature.finalState,
self.extractFeature.inputEmb, newNet,
self.textFeature.RNNState, "ATT_attr1",
"ATT_attr2", 5, False)
attHidden = tf.tanh(
tf.matmul(imageVec, self.weights["ATT_WI1"]) +
self.biases["ATT_WI1"])
attHidden2 = tf.tanh(
tf.matmul(textVec, self.weights["ATT_WT1"]) +
self.biases["ATT_WT1"])
attHidden3 = tf.tanh(
tf.matmul(attrVec, self.weights["ATT_WA1"]) +
self.biases["ATT_WW"])
scores1 = tf.matmul(attHidden,
self.weights["ATT_WF_1"]) + self.biases["ATT_WF_1"]
scores2 = tf.matmul(attHidden2,
self.weights["ATT_WF_2"]) + self.biases["ATT_WF_2"]
scores3 = tf.matmul(attHidden3,
self.weights["ATT_WF_3"]) + self.biases["ATT_WF_3"]
scoreLS = [scores1, scores2, scores3]
scoreLS = tf.nn.softmax(scoreLS, dim=0)
imageVec = tf.tanh(
tf.matmul(imageVec, self.weights["ATT_WI2"]) +
self.biases["ATT_WI"])
textVec = tf.tanh(
tf.matmul(textVec, self.weights["ATT_WT2"]) +
self.biases["ATT_WT"])
attrVec = tf.tanh(
tf.matmul(attrVec, self.weights["ATT_WA2"]) +
self.biases["ATT_WA"])
self.concatInput = scoreLS[0] * imageVec + scoreLS[
1] * textVec + scoreLS[2] * attrVec
W_conv1 = weight_variable([5, 5, 1, 32])
b_conv1 = bias_variable([32])
x_image = tf.reshape(x, [-1, 28, 28, 1])
h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)
h_pool1 = max_pool_2x2(h_conv1)
W_conv2 = weight_variable([5, 5, 32, 64])
b_conv2 = bias_variable([64])
h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)
h_pool2 = max_pool_2x2(h_conv2)
W_fc1 = weight_variable([7 * 7 * 64, 1024])
b_fc1 = bias_variable([1024])
h_pool2_flat = tf.reshape(h_pool2, [-1, 7 * 7 * 64])
h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)
keep_prob = tf.placeholder("float")
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)
W_fc2 = weight_variable([1024, 10])
b_fc2 = bias_variable([10])
y_conv = tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2)
cross_entropy = -tf.reduce_sum(y_ * tf.log(y_conv))
train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
correct_prediction = tf.equal(tf.argmax(y_conv, 1), tf.argmax(y_, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))
sess = tf.Session()
# Convolution Layer 5 | ReLU | Max Pooling
c_layer_5 = tf.nn.conv2d(c_layer_4,
conv_weights["c5_weights"],
strides=[1, 1, 1, 1],
padding="SAME",
name="c_layer_5")
c_layer_5 += conv_biases["c5_biases"]
c_layer_5 = tf.nn.relu(c_layer_5)
c_layer_5 = tf.nn.max_pool(c_layer_5,
ksize=[1, 3, 3, 1],
strides=[1, 1, 1, 1],
padding="VALID")
# Flatten the multi-dimensional outputs to feed fully connected layers
feature_map = tf.reshape(c_layer_5, [-1, 13, 13, 256])
# Fully Connected Layer 1 | Dropout
fc_layer_1 = tf.matmul(feature_map,
conv_weights["f1_weights"]) + conv_biases["f1_biases"]
fc_layer_1 = tf.nn.dropout(fc_layer_1, keep_prob=DROPOUT_KEEP_PROB)
# Fully Connected Layer 2 | Dropout
fc_layer_2 = tf.matmul(fc_layer_1,
conv_weights["f2_weights"]) + conv_biases["f2_biases"]
fc_layer_2 = tf.nn.dropout(fc_layer_2, keep_prob=DROPOUT_KEEP_PROB)
# Fully Connected Layer 3 | Softmax
fc_layer_3 = tf.matmul(fc_layer_2,
conv_weights["f3_weights"]) + conv_biases["f3_biases"]
cnn_output = tf.nn.softmax(fc_layer_3)
