class LandmarkDataLayer(caffe.Layer):
def setup(self, bottom, top):
param = eval(self.param_str)
self.batch = int(param['batch'])
self.img_size = config.IMG_SIZE
self.batch_loader = BatchLoader(param, "train")
self.ohem_batch_loader = BatchLoader(param, 'ohem')
self.train_ratio = 1.
self.ohem_ratio = 0
if self.ohem_batch_loader.is_loaded():
self.train_ratio = 7 / 8.
self.ohem_ratio = 1. - self.train_ratio
top[0].reshape(self.batch, 3, self.img_size, self.img_size) # data
top[1].reshape(self.batch, config.LANDMARK_SIZE * 2) # landmark
top[2].reshape(self.batch, 1) # landmark
def reshape(self, bottom, top):
pass
def forward(self, bottom, top):
batch_data = self.batch_loader.next_batch(
self.batch * self.train_ratio, '')
batch_data += self.ohem_batch_loader.next_batch(
self.batch * self.ohem_ratio, '')
for i, datum in enumerate(batch_data):
img, pts, eye_dist = datum
top[0].data[i, ...] = img
top[1].data[i, ...] = pts
top[2].data[i, ...] = eye_dist
def backward(self, bottom, top):
pass
class LandmarkDataLayer(caffe.Layer):
def setup(self, bottom, top):
param = eval(self.param_str)
self.batch = int(param['batch'])
self.batch_loader = BatchLoader(param, "train")
self.ohem_batch_loader = BatchLoader(param, 'ohem')
self.train_ratio = 1.
self.ohem_ratio = 0
if self.ohem_batch_loader.is_loaded():
self.train_ratio = 7 / 8.
self.ohem_ratio = 1. - self.train_ratio
top[0].reshape(self.batch, 3, self.img_size, self.img_size) # data
top[1].reshape(self.batch, 72) # landmark
def reshape(self, bottom, top):
pass
def forward(self, bottom, top):
batch_data = self.batch_loader.next_batch(
self.batch * self.train_ratio, '')
batch_data += self.ohem_batch_loader.next_batch(
self.batch * self.ohem_ratio, '')
random.shuffle(batch_data)
for i, datum in enumerate(batch_data):
img, label, bbox, landm5 = datum
top[0].data[i, ...] = img
top[1].data[i, ...] = label
top[2].data[i, ...] = bbox
if self.net != 'pnet':
top[3].data[i, ...] = landm5
def backward(self, bottom, top):
pass
class GanNetTrain(caffe.Layer):
"""Data layer for training"""
def setup(self, bottom, top):
layer_params = yaml.load(self.param_str)
self.batch_size = layer_params['batch_size']
self.image_file = layer_params['image_file']
self.batch_loader = BatchLoader(self.image_file, self.batch_size)
def forward(self, bottom, top):
# assign output
top[0].data[...] = self.images_a
top[1].data[...] = self.images_b
top[2].data[...] = self.label_true
top[3].data[...] = self.label_false
def backward(self, top, propagate_down, bottom):
"""This layer does not propagate gradients."""
pass
def reshape(self, bottom, top):
images_A, images_B = self.batch_loader.next_batch()
images_A = np.array(images_A)
images_B = np.array(images_B)
images_A = images_A.transpose((0, 3, 1, 2))
images_B = images_B.transpose((0, 3, 1, 2))
self.images_a = images_A
self.images_b = images_B
self.label_true = np.ones((self.batch_size, 1), dtype='float32')
self.label_false = np.zeros((self.batch_size, 1), dtype='float32')
top[0].reshape(*self.images_a.shape)
top[1].reshape(*self.images_b.shape)
top[2].reshape(*self.label_true.shape)
top[3].reshape(*self.label_false.shape)
class DataLayer(caffe.Layer):
def setup(self, bottom, top):
param = eval(self.param_str)
self.batch = int(param['batch'])
self.net = param['net']
self.img_size = config.NET_IMG_SIZES[self.net]
self.batch_loader = BatchLoader(param, "train")
self.ohem_batch_loader = BatchLoader(param, 'ohem')
self.train_ratio = 1.
self.ohem_ratio = 0
if self.ohem_batch_loader.is_loaded():
self.train_ratio = 7 / 8.
self.ohem_ratio = 1 / 8.
top[0].reshape(self.batch, 3, self.img_size, self.img_size) # data
top[1].reshape(self.batch, 1) # label
top[2].reshape(self.batch, 4) # bbox
if self.net != 'pnet':
top[3].reshape(self.batch, config.LANDMARK_SIZE * 2)
def reshape(self, bottom, top):
pass
def forward(self, bottom, top):
task = random.choice(config.TRAIN_TASKS[self.net])
batch_data = self.batch_loader.next_batch(
self.batch * self.train_ratio, '')
batch_data += self.ohem_batch_loader.next_batch(
self.batch * self.ohem_ratio, '')
random.shuffle(batch_data)
for i, datum in enumerate(batch_data):
img, label, bbox, landm5 = datum
top[0].data[i, ...] = img
top[1].data[i, ...] = label
top[2].data[i, ...] = bbox
if self.net != 'pnet':
top[3].data[i, ...] = landm5
def backward(self, bottom, top):
pass
class LSTMTDNN(Model):
"""Time-delayed Nueral Network (cf. http://arxiv.org/abs/1508.06615v4)
"""
def __init__(self,
sess,
batch_size=100,
rnn_size=650,
layer_depth=2,
word_embed_dim=650,
char_embed_dim=15,
feature_maps=[50, 100, 150, 200, 200, 200, 200],
kernels=[1, 2, 3, 4, 5, 6, 7],
seq_length=35,
max_word_length=65,
use_word=False,
use_char=True,
hsm=0,
max_grad_norm=5,
highway_layers=2,
dropout_prob=0.5,
use_batch_norm=True,
checkpoint_dir="checkpoint",
forward_only=False,
data_dir="data",
dataset_name="pdb",
use_progressbar=False):
"""Initialize the parameters for LSTM TDNN
Args:
rnn_size: the dimensionality of hidden layers
layer_depth: # of depth in LSTM layers
batch_size: size of batch per epoch
word_embed_dim: the dimensionality of word embeddings
char_embed_dim: the dimensionality of character embeddings
feature_maps: list of feature maps (for each kernel width)
kernels: list of kernel widths
seq_length: max length of a word
use_word: whether to use word embeddings or not
use_char: whether to use character embeddings or not
highway_layers: # of highway layers to use
dropout_prob: the probability of dropout
use_batch_norm: whether to use batch normalization or not
hsm: whether to use hierarchical softmax
"""
self.sess = sess
self.batch_size = batch_size
self.seq_length = seq_length
# RNN
self.rnn_size = rnn_size
self.layer_depth = layer_depth
# CNN
self.use_word = use_word
self.use_char = use_char
self.word_embed_dim = word_embed_dim
self.char_embed_dim = char_embed_dim
self.feature_maps = feature_maps
self.kernels = kernels
# General
self.highway_layers = highway_layers
self.dropout_prob = dropout_prob
self.use_batch_norm = use_batch_norm
# Training
self.max_grad_norm = max_grad_norm
self.max_word_length = max_word_length
self.hsm = hsm
self.data_dir = data_dir
self.dataset_name = dataset_name
self.checkpoint_dir = checkpoint_dir
self.forward_only = forward_only
self.use_progressbar = use_progressbar
self.loader = BatchLoader(self.data_dir, self.dataset_name,
self.batch_size, self.seq_length,
self.max_word_length)
print('Word vocab size: %d, Char vocab size: %d, Max word length (incl. padding): %d' % \
(len(self.loader.idx2word), len(self.loader.idx2char), self.loader.max_word_length))
self.max_word_length = self.loader.max_word_length
self.char_vocab_size = len(self.loader.idx2char)
self.word_vocab_size = len(self.loader.idx2word)
# build LSTMTDNN model
self.prepare_model()
# load checkpoints
if self.forward_only == True:
if self.load(self.checkpoint_dir, self.dataset_name):
print(" [*] SUCCESS to load model for %s." % self.dataset_name)
else:
print(" [!] Failed to load model for %s." % self.dataset_name)
sys.exit(1)
def prepare_model(self):
with tf.variable_scope("LSTMTDNN"):
self.char_inputs = []
self.word_inputs = []
self.cnn_outputs = []
if self.use_char:
char_W = tf.get_variable(
"char_embed", [self.char_vocab_size, self.char_embed_dim])
else:
word_W = tf.get_variable(
"word_embed", [self.word_vocab_size, self.word_embed_dim])
with tf.variable_scope("CNN") as scope:
self.char_inputs = tf.placeholder(
tf.int32,
[self.batch_size, self.seq_length, self.max_word_length])
self.word_inputs = tf.placeholder(
tf.int32, [self.batch_size, self.seq_length])
char_indices = tf.split(1, self.seq_length, self.char_inputs)
word_indices = tf.split(1, self.seq_length,
tf.expand_dims(self.word_inputs, -1))
for idx in xrange(self.seq_length):
char_index = tf.reshape(char_indices[idx],
[-1, self.max_word_length])
word_index = tf.reshape(word_indices[idx], [-1, 1])
if idx != 0:
scope.reuse_variables()
if self.use_char:
# [batch_size x word_max_length, char_embed]
char_embed = tf.nn.embedding_lookup(char_W, char_index)
char_cnn = TDNN(char_embed, self.char_embed_dim,
self.feature_maps, self.kernels)
if self.use_word:
word_embed = tf.nn.embedding_lookup(
word_W, word_index)
cnn_output = tf.concat(1, char_cnn.output,
word_embed)
else:
cnn_output = char_cnn.output
else:
cnn_output = tf.squeeze(
tf.nn.embedding_lookup(word_W, word_index))
if self.use_batch_norm:
bn = batch_norm()
norm_output = bn(
tf.expand_dims(tf.expand_dims(cnn_output, 1), 1))
cnn_output = tf.squeeze(norm_output)
if highway:
#cnn_output = highway(input_, input_dim_length, self.highway_layers, 0)
cnn_output = highway(cnn_output,
cnn_output.get_shape()[1],
self.highway_layers, 0)
self.cnn_outputs.append(cnn_output)
with tf.variable_scope("LSTM") as scope:
self.cell = rnn_cell.BasicLSTMCell(self.rnn_size)
self.stacked_cell = rnn_cell.MultiRNNCell([self.cell] *
self.layer_depth)
outputs, _ = rnn.rnn(self.stacked_cell,
self.cnn_outputs,
dtype=tf.float32)
self.lstm_outputs = []
self.true_outputs = tf.placeholder(
tf.float32,
[self.batch_size, self.seq_length, self.word_vocab_size])
loss = 0
true_outputs = tf.split(1, self.seq_length, self.true_outputs)
for idx, (top_h,
true_output) in enumerate(zip(outputs,
true_outputs)):
if self.dropout_prob > 0:
top_h = tf.nn.dropout(top_h, self.dropout_prob)
if self.hsm > 0:
self.lstm_outputs.append(top_h)
else:
if idx != 0:
scope.reuse_variables()
proj = rnn_cell.linear(top_h, self.word_vocab_size, 0)
log_softmax = tf.log(tf.nn.softmax(proj))
self.lstm_outputs.append(log_softmax)
loss += tf.nn.softmax_cross_entropy_with_logits(
self.lstm_outputs[idx], tf.squeeze(true_output))
self.loss = tf.reduce_mean(loss) / self.seq_length
tf.scalar_summary("loss", self.loss)
tf.scalar_summary("perplexity", tf.exp(self.loss))
def train(self, epoch):
cost = 0
target = np.zeros(
[self.batch_size, self.seq_length, self.word_vocab_size])
N = self.loader.sizes[0]
for idx in xrange(N):
target.fill(0)
x, y, x_char = self.loader.next_batch(0)
for b in xrange(self.batch_size):
for t, w in enumerate(y[b]):
target[b][t][w] = 1
feed_dict = {
self.word_inputs: x,
self.char_inputs: x_char,
self.true_outputs: target,
}
_, loss, step, summary_str = self.sess.run(
[self.optim, self.loss, self.global_step, self.merged_summary],
feed_dict=feed_dict)
self.writer.add_summary(summary_str, step)
if idx % 50 == 0:
if self.use_progressbar:
progress(
idx / N, "epoch: [%2d] [%4d/%4d] loss: %2.6f" %
(epoch, idx, N, loss))
else:
print("epoch: [%2d] [%4d/%4d] loss: %2.6f" %
(epoch, idx, N, loss))
cost += loss
return cost / N
def test(self, split_idx, max_batches=None):
if split_idx == 1:
set_name = 'Valid'
else:
set_name = 'Test'
N = self.loader.sizes[split_idx]
if max_batches != None:
N = min(max_batches, N)
self.loader.reset_batch_pointer(split_idx)
target = np.zeros(
[self.batch_size, self.seq_length, self.word_vocab_size])
cost = 0
for idx in xrange(N):
target.fill(0)
x, y, x_char = self.loader.next_batch(split_idx)
for b in xrange(self.batch_size):
for t, w in enumerate(y[b]):
target[b][t][w] = 1
feed_dict = {
self.word_inputs: x,
self.char_inputs: x_char,
self.true_outputs: target,
}
loss = self.sess.run(self.loss, feed_dict=feed_dict)
if idx % 50 == 0:
if self.use_progressbar:
progress(
idx / N, "> %s: loss: %2.6f, perplexity: %2.6f" %
(set_name, loss, np.exp(loss)))
else:
print(" > %s: loss: %2.6f, perplexity: %2.6f" %
(set_name, loss, np.exp(loss)))
cost += loss
cost = cost / N
return cost
def run(self, epoch=25, learning_rate=1, learning_rate_decay=0.5):
self.current_lr = learning_rate
self.lr = tf.Variable(learning_rate, trainable=False)
self.opt = tf.train.GradientDescentOptimizer(self.lr)
#self.opt = tf.train.AdamOptimizer(learning_rate, beta1=0.5).minimize(self.loss)
# clip gradients
params = tf.trainable_variables()
grads = []
for grad in tf.gradients(self.loss, params):
if grad:
grads.append(tf.clip_by_norm(grad, self.max_grad_norm))
else:
grads.append(grad)
self.global_step = tf.Variable(0, name="global_step", trainable=False)
self.optim = self.opt.apply_gradients(zip(grads, params),
global_step=self.global_step)
# ready for train
tf.initialize_all_variables().run()
if self.load(self.checkpoint_dir, self.dataset_name):
print(" [*] SUCCESS to load model for %s." % self.dataset_name)
else:
print(" [!] Failed to load model for %s." % self.dataset_name)
self.saver = tf.train.Saver()
self.merged_summary = tf.merge_all_summaries()
self.writer = tf.train.SummaryWriter("./logs", self.sess.graph_def)
self.log_loss = []
self.log_perp = []
if not self.forward_only:
for idx in xrange(epoch):
train_loss = self.train(idx)
valid_loss = self.test(1)
# Logging
self.log_loss.append([train_loss, valid_loss])
self.log_perp.append([np.exp(train_loss), np.exp(valid_loss)])
state = {
'perplexity': np.exp(train_loss),
'epoch': idx,
'learning_rate': self.current_lr,
'valid_perplexity': np.exp(valid_loss)
}
print(state)
# Learning rate annealing
if len(self.log_loss) > 1 and self.log_loss[idx][
1] > self.log_loss[idx - 1][1] * 0.9999:
self.current_lr = self.current_lr * learning_rate_decay
self.lr.assign(self.current_lr).eval()
if self.current_lr < 1e-5: break
if idx % 2 == 0:
self.save(self.checkpoint_dir, self.dataset_name)
test_loss = self.test(2)
print(" [*] Test loss: %2.6f, perplexity: %2.6f" %
(test_loss, np.exp(test_loss)))
summary_op = tf.summary.merge_all()
# In[12]:
sess = tf.Session()
sess.run(tf.global_variables_initializer())
writer = tf.summary.FileWriter('/tmp/mnist_log', sess.graph)
# ## Train
# In[14]:
saver = tf.train.Saver()
step = sess.run(global_step)
while step <= 80000:
batch_images, batch_labels = train_batch_loader.next_batch()
# print batch_images.shape
# print batch_labels.shape
_, summary_str, train_acc, Center_loss, Softmax_loss = sess.run(
[train_op, summary_op, accuracy, center_loss, softmax_loss],
feed_dict={
input_images: (batch_images - 127.5) * 0.0078125, # - mean_data,
labels: batch_labels,
})
step += 1
if step % 100 == 0:
print "********* Step %s: ***********" % str(step)
print "center loss: %s" % str(Center_loss)
print "softmax_loss: %s" % str(Softmax_loss)
print "train_acc: %s" % str(train_acc)
print "*******************************"
class LSTMTDNN(Model):
"""
Time-delayed Neural Network (cf. http://arxiv.org/abs/1508.06615v4)
"""
def __init__(self, sess,
batch_size=100, rnn_size=650, layer_depth=2,
word_embed_dim=650, char_embed_dim=15,
feature_maps=[50, 100, 150, 200, 200, 200, 200],
kernels=[1,2,3,4,5,6,7], seq_length=35, max_word_length=65,
use_word=False, use_char=True, hsm=0, max_grad_norm=5,
highway_layers=2, dropout_prob=0.5, use_batch_norm=True,
checkpoint_dir="checkpoint", forward_only=False,
data_dir="data", dataset_name="pdb", use_progressbar=False):
"""
Initialize the parameters for LSTM TDNN
Args:
rnn_size: the dimensionality of hidden layers
layer_depth: # of depth in LSTM layers
batch_size: size of batch per epoch
word_embed_dim: the dimensionality of word embeddings
char_embed_dim: the dimensionality of character embeddings
feature_maps: list of feature maps (for each kernel width)
kernels: list of kernel widths
seq_length: max length of a word
use_word: whether to use word embeddings or not
use_char: whether to use character embeddings or not
highway_layers: # of highway layers to use
dropout_prob: the probability of dropout
use_batch_norm: whether to use batch normalization or not
hsm: whether to use hierarchical softmax
"""
self.sess = sess
self.batch_size = batch_size
self.seq_length = seq_length
# RNN
self.rnn_size = rnn_size
self.layer_depth = layer_depth
# CNN
self.use_word = use_word
self.use_char = use_char
self.word_embed_dim = word_embed_dim
self.char_embed_dim = char_embed_dim
self.feature_maps = feature_maps
self.kernels = kernels
# General
self.highway_layers = highway_layers
self.dropout_prob = dropout_prob
self.use_batch_norm = use_batch_norm
# Training
self.max_grad_norm = max_grad_norm
self.max_word_length = max_word_length
self.hsm = hsm
self.data_dir = data_dir
self.dataset_name = dataset_name
self.checkpoint_dir = checkpoint_dir
self.forward_only = forward_only
self.use_progressbar = use_progressbar
self.loader = BatchLoader(self.data_dir, self.dataset_name, self.batch_size, self.seq_length, self.max_word_length)
print('Word vocab size: %d, Char vocab size: %d, Max word length (incl. padding): %d' % \
(len(self.loader.idx2word), len(self.loader.idx2char), self.loader.max_word_length))
self.max_word_length = self.loader.max_word_length
self.char_vocab_size = len(self.loader.idx2char)
self.word_vocab_size = len(self.loader.idx2word)
# build LSTMTDNN model
self.prepare_model()
# load checkpoints
if self.forward_only == True:
if self.load(self.checkpoint_dir, self.dataset_name):
print("[*] SUCCESS to load model for %s." % self.dataset_name)
else:
print("[!] Failed to load model for %s." % self.dataset_name)
sys.exit(1)
def prepare_model(self):
with tf.variable_scope("LSTMTDNN"):
self.char_inputs = []
self.word_inputs = []
self.cnn_outputs = []
if self.use_char:
char_W = tf.get_variable("char_embed",
[self.char_vocab_size, self.char_embed_dim])
if self.use_word:
word_W = tf.get_variable("word_embed",
[self.word_vocab_size, self.word_embed_dim])
with tf.variable_scope("CNN") as scope:
self.char_inputs = tf.placeholder(tf.int32, [self.batch_size, self.seq_length, self.max_word_length])
self.word_inputs = tf.placeholder(tf.int32, [self.batch_size, self.seq_length])
char_indices = tf.split(1, self.seq_length, self.char_inputs)
word_indices = tf.split(1, self.seq_length, tf.expand_dims(self.word_inputs, -1))
for idx in xrange(self.seq_length):
char_index = tf.reshape(char_indices[idx], [-1, self.max_word_length])
word_index = tf.reshape(word_indices[idx], [-1, 1])
if idx != 0:
scope.reuse_variables()
if self.use_char:
# [batch_size x word_max_length, char_embed]
char_embed = tf.nn.embedding_lookup(char_W, char_index)
char_cnn = TDNN(char_embed, self.char_embed_dim, self.feature_maps, self.kernels)
if self.use_word:
word_embed = tf.nn.embedding_lookup(word_W, word_index)
cnn_output = tf.concat(1, [char_cnn.output, tf.squeeze(word_embed, [1])])
else:
cnn_output = char_cnn.output
else:
cnn_output = tf.squeeze(tf.nn.embedding_lookup(word_W, word_index))
if self.use_batch_norm:
bn = batch_norm()
norm_output = bn(tf.expand_dims(tf.expand_dims(cnn_output, 1), 1))
cnn_output = tf.squeeze(norm_output)
if highway:
#cnn_output = highway(input_, input_dim_length, self.highway_layers, 0)
cnn_output = highway(cnn_output, cnn_output.get_shape()[1], self.highway_layers, 0)
self.cnn_outputs.append(cnn_output)
with tf.variable_scope("LSTM") as scope:
self.cell = tf.nn.rnn_cell.BasicLSTMCell(self.rnn_size)
self.stacked_cell = tf.nn.rnn_cell.MultiRNNCell([self.cell] * self.layer_depth)
outputs, _ = tf.nn.rnn(self.stacked_cell,
self.cnn_outputs,
dtype=tf.float32)
self.lstm_outputs = []
self.true_outputs = tf.placeholder(tf.int64,
[self.batch_size, self.seq_length])
loss = 0
true_outputs = tf.split(1, self.seq_length, self.true_outputs)
for idx, (top_h, true_output) in enumerate(zip(outputs, true_outputs)):
if self.dropout_prob > 0:
top_h = tf.nn.dropout(top_h, self.dropout_prob)
if self.hsm > 0:
self.lstm_outputs.append(top_h)
else:
if idx != 0:
scope.reuse_variables()
proj = tf.nn.rnn_cell._linear(top_h, self.word_vocab_size, 0)
self.lstm_outputs.append(proj)
loss += tf.nn.sparse_softmax_cross_entropy_with_logits(self.lstm_outputs[idx], tf.squeeze(true_output))
self.loss = tf.reduce_mean(loss) / self.seq_length
tf.scalar_summary("loss", self.loss)
tf.scalar_summary("perplexity", tf.exp(self.loss))
def train(self, epoch):
cost = 0
target = np.zeros([self.batch_size, self.seq_length])
N = self.loader.sizes[0]
for idx in xrange(N):
target.fill(0)
x, y, x_char = self.loader.next_batch(0)
for b in xrange(self.batch_size):
for t, w in enumerate(y[b]):
target[b][t] = w
feed_dict = {
self.word_inputs: x,
self.char_inputs: x_char,
self.true_outputs: target,
}
_, loss, step, summary_str = self.sess.run(
[self.optim, self.loss, self.global_step, self.merged_summary], feed_dict=feed_dict)
self.writer.add_summary(summary_str, step)
if idx % 50 == 0:
if self.use_progressbar:
progress(idx/N, "epoch: [%2d] [%4d/%4d] loss: %2.6f" % (epoch, idx, N, loss))
else:
print("epoch: [%2d] [%4d/%4d] loss: %2.6f" % (epoch, idx, N, loss))
cost += loss
return cost / N
def test(self, split_idx, max_batches=None):
if split_idx == 1:
set_name = 'Valid'
else:
set_name = 'Test'
N = self.loader.sizes[split_idx]
if max_batches != None:
N = min(max_batches, N)
self.loader.reset_batch_pointer(split_idx)
target = np.zeros([self.batch_size, self.seq_length])
cost = 0
for idx in xrange(N):
target.fill(0)
x, y, x_char = self.loader.next_batch(split_idx)
for b in xrange(self.batch_size):
for t, w in enumerate(y[b]):
target[b][t] = w
feed_dict = {
self.word_inputs: x,
self.char_inputs: x_char,
self.true_outputs: target,
}
loss = self.sess.run(self.loss, feed_dict=feed_dict)
if idx % 50 == 0:
if self.use_progressbar:
progress(idx/N, "> %s: loss: %2.6f, perplexity: %2.6f" % (set_name, loss, np.exp(loss)))
else:
print(" > %s: loss: %2.6f, perplexity: %2.6f" % (set_name, loss, np.exp(loss)))
cost += loss
cost = cost / N
return cost
def run(self, epoch=25,
learning_rate=1, learning_rate_decay=0.5):
self.current_lr = learning_rate
self.lr = tf.Variable(learning_rate, trainable=False)
self.opt = tf.train.GradientDescentOptimizer(self.lr)
#self.opt = tf.train.AdamOptimizer(learning_rate, beta1=0.5).minimize(self.loss)
# clip gradients
params = tf.trainable_variables()
grads = []
for grad in tf.gradients(self.loss, params):
if grad is not None:
grads.append(tf.clip_by_norm(grad, self.max_grad_norm))
else:
grads.append(grad)
self.global_step = tf.Variable(0, name="global_step", trainable=False)
self.optim = self.opt.apply_gradients(zip(grads, params),
global_step=self.global_step)
# ready for train
tf.initialize_all_variables().run()
if self.load(self.checkpoint_dir, self.dataset_name):
print("[*] SUCCESS to load model for %s." % self.dataset_name)
else:
print("[!] Failed to load model for %s." % self.dataset_name)
self.saver = tf.train.Saver()
self.merged_summary = tf.merge_all_summaries()
self.writer = tf.train.SummaryWriter("./logs", self.sess.graph_def)
self.log_loss = []
self.log_perp = []
if not self.forward_only:
for idx in xrange(epoch):
train_loss = self.train(idx)
valid_loss = self.test(1)
# Logging
self.log_loss.append([train_loss, valid_loss])
self.log_perp.append([np.exp(train_loss), np.exp(valid_loss)])
state = {
'perplexity': np.exp(train_loss),
'epoch': idx,
'learning_rate': self.current_lr,
'valid_perplexity': np.exp(valid_loss)
}
print(state)
# Learning rate annealing
if len(self.log_loss) > 1 and self.log_loss[idx][1] > self.log_loss[idx-1][1] * 0.9999:
self.current_lr = self.current_lr * learning_rate_decay
self.lr.assign(self.current_lr).eval()
if self.current_lr < 1e-5: break
if idx % 2 == 0:
self.save(self.checkpoint_dir, self.dataset_name)
test_loss = self.test(2)
print("[*] Test loss: %2.6f, perplexity: %2.6f" % (test_loss, np.exp(test_loss)))
def main():
LAMBDA = 0.0
num_class = 526
checkpoint_dir = "../model/"
with tf.name_scope('input'):
input_images = tf.placeholder(tf.float32,
shape=(None, 100, 100, 3),
name='input_images')
labels = tf.placeholder(tf.int64, shape=(None), name='labels')
logits, features, total_loss, accuracy, centers_update_op, center_loss, softmax_loss = build_network(
input_images, labels, num_class, ratio=LAMBDA)
global_step = tf.Variable(0, trainable=False, name='global_step')
train_batch_loader = BatchLoader("../data/facescrub_train.list", 16)
test_batch_loader = BatchLoader("../data/facescrub_val.list", 16)
optimizer = tf.train.AdamOptimizer(0.001)
with tf.control_dependencies([centers_update_op]):
train_op = optimizer.minimize(total_loss, global_step=global_step)
summary_op = tf.summary.merge_all()
with tf.Session(config=tf.ConfigProto(log_device_placement=True)) as sess:
sess.run(tf.global_variables_initializer())
writer = tf.summary.FileWriter('../tmp/face_log', sess.graph)
saver = tf.train.Saver()
step = sess.run(global_step)
while step <= 80000:
batch_images, batch_labels = train_batch_loader.next_batch()
# print batch_images.shape
# print batch_labels.shape
_, summary_str, train_acc, Center_loss, Softmax_loss = sess.run(
[train_op, summary_op, accuracy, center_loss, softmax_loss],
feed_dict={
input_images:
(batch_images - 127.5) * 0.0078125, # - mean_data,
labels: batch_labels,
})
step += 1
print("step", step)
if step % 100 == 0:
print("********* Step %s: ***********" % str(step))
print("center loss: %s" % str(Center_loss))
print("softmax_loss: %s" % str(Softmax_loss))
print("train_acc: %s" % str(train_acc))
print("*******************************")
if step % 10000 == 0:
saver.save(sess,
checkpoint_dir + 'model.ckpt',
global_step=step)
writer.add_summary(summary_str, global_step=step)
if step % 2000 == 0:
batch_images, batch_labels = test_batch_loader.next_batch()
vali_image = (batch_images - 127.5) * 0.0078125
vali_acc = sess.run(accuracy,
feed_dict={
input_images: vali_image,
labels: batch_labels
})
print(("step: {}, train_acc:{:.4f}, vali_acc:{:.4f}".format(
step, train_acc, vali_acc)))
sess.close()
def main():
params = args()
model_prefix = params.model_prefix
load_epoch = params.load_epoch
batch_size = params.batch_size
img1_path = params.image1
img2_path = params.image2
img_shape = params.image_shape
img_dir = params.img_dir
txt_path = params.file_txt
# in out files
file_ = open(txt_path, 'r')
lines_ = file_.readlines()
result_ = open("result.txt", 'w')
#
if Test_Img == 'True':
img1 = cv2.imread(img1_path)
img2 = cv2.imread(img2_path)
img1 = cv2.resize(img1, (img_shape, img_shape))
img2 = cv2.resize(img2, (img_shape, img_shape))
img1 = np.expand_dims(img1, 0)
img2 = np.expand_dims(img2, 0)
test_batch_loader = BatchLoader("../data/facescrub_val.list", batch_size,
img_shape)
tf.reset_default_graph()
with tf.name_scope('input'):
input_images = tf.placeholder(tf.float32,
shape=(batch_size, img_shape, img_shape,
3),
name='input_images')
labels = tf.placeholder(tf.int32, shape=(batch_size), name='labels')
features, accuracy, pred_class, res1 = build_network(
input_images, labels, 526, 'test')
check_ckpt(model_prefix + '-' + str(load_epoch))
#detect
detector = face_dect()
#
with tf.Session() as sess:
restore_model = tf.train.Saver()
#restore_model = tf.train.import_meta_graph(model_prefix+'-'+str(load_epoch) +'.meta')
restore_model.restore(sess, model_prefix + '-' + str(load_epoch))
print("face model restore over")
'''
all_vars = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES)
key_list = []
var_dic = dict()
for v_name in tf.global_variables():
i=0
print("name : ",v_name.name[:-2],v_name.shape)
print("shape",all_vars[i])
key_list.append(v_name.name[:-2])
i+=1
#print(tf.get_variable_scope())
#all_vars = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES)
vas= sess.run([all_vars])
print(len(vas))
for i in range(len(vas)):
cur_name = key_list[i]
cur_var = vas[i]
print("name ,shape : ",cur_name,np.shape(cur_var))
var_dic[cur_name] = cur_var
'''
#restore_model = tf.train.import_meta_graph(model_prefix+'-'+str(load_epoch)+'.meta')
#restore_model.restore(sess,model_prefix+'-'+str(load_epoch))
if Test_Img == 'False':
iter_num = 0
accuracy_sum = 0
while iter_num < test_batch_loader.batch_num:
batch_images, batch_labels = test_batch_loader.next_batch()
images_in = (batch_images - 127.5) * 0.0078125
feat, batch_accuracy = sess.run([features, accuracy],
feed_dict={
input_images: images_in,
labels: batch_labels
})
accuracy_sum += batch_accuracy
iter_num += 1
if iter_num % 10 == 0:
print("step ", iter_num, batch_accuracy)
print("image num: ", test_batch_loader.data_num,
"the test accuracy: ",
accuracy_sum / (test_batch_loader.batch_num))
elif Test_Img == 'True':
with tf.name_scope('valdata'):
label_t = np.zeros([1])
feat1 = sess.run([features],
feed_dict={
input_images: img1,
labels: label_t
})
feat2 = sess.run([features],
feed_dict={
input_images: img2,
labels: label_t
})
distance = L2_distance(feat1, feat2, 512)
print("the 2 image dis: ", distance)
for rot, fdir, fname in os.walk(img_dir):
if len(fname) != 0:
break
img_list = []
print(fname)
for i in range(len(fname)):
org_img = cv2.imread(os.path.join(rot, fname[i]))
img_org = cv2.resize(org_img, (img_shape, img_shape))
img_org = np.expand_dims(img_org, 0)
img_list.append(img_org)
for i in range(len(fname)):
img1 = img_list[i]
feat1 = sess.run([features],
feed_dict={
input_images: img1,
labels: label_t
})
j = i + 1
while j < len(fname):
img2 = img_list[j]
t1 = time.time()
feat2 = sess.run([features],
feed_dict={
input_images: img2,
labels: label_t
})
t2 = time.time()
print("one image time ", t2 - t1)
distance = L2_distance(feat1, feat2, 512)
t3 = time.time()
print("compere time ", t3 - t2)
print(i, j, distance)
j += 1
elif Test_Img == 'File':
label_t = np.ones([1])
for i in range(len(lines_)):
feat_vec = []
feat1_fg = 0
feat2_fg = 0
line_1 = lines_[i]
line_1 = string.strip(line_1)
line_s = line_1.split(',')
dir_path_save = line_s[0][:-4]
dir_path_save = "../cropface/" + dir_path_save
mkdir_(dir_path_save)
for j in range(len(line_s)):
feat_vec2 = []
if j == 0:
#print("line ",line_s)
img1_pic = line_s[0]
img1_path = os.path.join(img_dir, img1_pic)
img1 = cv2.imread(img1_path)
bboxes_1 = detector.get_face(img1)
if bboxes_1 is not None:
for k in range(bboxes_1.shape[0]):
crop_img1 = get_img(img1, bboxes_1[k],
img_shape)
if k == 0 and SV_IMG:
img_save_path = dir_path_save + '/' + line_s[
0][:-4] + ".jpg"
save_image(img_save_path, crop_img1)
crop_img1 = (crop_img1 - 127.5) * 0.0078125
crop_img1 = np.expand_dims(crop_img1, 0)
feat1 = sess.run([features],
feed_dict={
input_images: crop_img1,
labels: label_t
})
print("a feature shape ", np.shape(feat1))
feat_vec.append(feat1)
feat_fg = 1
else:
print("a no face detect ")
break
else:
img2_pic = line_s[j]
img2_path = os.path.join(img_dir, img2_pic)
img2 = cv2.imread(img2_path)
bboxes_2 = detector.get_face(img2)
if bboxes_2 is not None:
for k in range(bboxes_2.shape[0]):
crop_img2 = get_img(img2, bboxes_2[k],
img_shape)
if SV_IMG:
img_save_path = dir_path_save + '/' + line_s[
j][:-4] + "-" + str(k) + ".jpg"
save_image(img_save_path, crop_img2)
crop_img2 = (crop_img2 - 127.5) * 0.0078125
crop_img2 = np.expand_dims(crop_img2, 0)
feat2 = sess.run([features],
feed_dict={
input_images: crop_img2,
labels: label_t
})
print("b feature shape ", np.shape(feat2))
feat_vec2.append(feat2)
feat2_fg = 1
else:
print("b no face detect ")
continue
if j > 0:
t2 = time.time()
distance = L2_distance(feat_vec[0], feat_vec2[0], 512)
print("distance is ", distance)
t3 = time.time()
print("compere time ", t3 - t2)
result_.write("{} ".format(distance))
result_.write("\n")
print(feat2)
file_.close()
result_.close()
def main():
LAMBDA = 1e-8
center_alpha = 0.9
num_class = 10000
embd_size = 512
args = argument()
checkpoint_dir = args.save_model_name
lr = args.lr
batch_size = args.batch_size
epoch_num = args.epoch_num
sta = args.sta
img_shape = args.img_size
train_file = args.train_file
val_file = args.val_file
train_batch_loader = BatchLoader(train_file, batch_size, img_shape)
test_batch_loader = BatchLoader(val_file, batch_size, img_shape)
#(Height,Width) = (train_batch_loader.height,train_batch_loader.width)
#train_batch_loader = mnist_data(batch_size)
print("img shape", img_shape)
with tf.name_scope('input'):
input_images = tf.placeholder(tf.float32,
shape=(batch_size, img_shape[0],
img_shape[1], 3),
name='input_images')
labels = tf.placeholder(tf.int32, shape=(batch_size), name='labels')
learn_rate = tf.placeholder(tf.float32,
shape=(None),
name='learn_rate')
with tf.name_scope('var'):
global_step = tf.Variable(0, trainable=False, name='global_step')
loss_op = CenterLoss(center_alpha, num_class, embd_size)
#with tf.device('/gpu:0'):
total_loss, accuracy, centers_update_op, center_loss, softmax_loss, pred_class, res1 = build_network(
input_images, labels, num_class, sta, loss_op, ratio=LAMBDA)
optimizer = tf.train.AdamOptimizer(learn_rate)
#optimizer = tf.train.GradientDescentOptimizer(learn_rate)
with tf.control_dependencies([centers_update_op]):
train_op = optimizer.minimize(total_loss, global_step=global_step)
#train_op = optimizer.minimize(total_loss, global_step=global_step)
summary_op = tf.summary.merge_all()
with tf.Session(config=tf.ConfigProto(log_device_placement=False)) as sess:
#sess.run(tf.global_variables_initializer())
writer = tf.summary.FileWriter('../tmp/face_log', sess.graph)
saver = tf.train.Saver()
if args.pretrained is not None:
model_path = args.save_model_name + '-' + str(args.pretrained)
#saver.restore(sess,'./face_model/high_score-60')
saver.restore(sess, model_path)
else:
init = tf.global_variables_initializer()
sess.run(init)
step = sess.run(global_step)
epoch_idx = 0
graph_step = 0
while epoch_idx <= epoch_num:
step = 0
ckpt_fg = 'True'
ps_loss = 0.0
pc_loss = 0.0
acc_sum = 0.0
while step < train_batch_loader.batch_num:
batch_images, batch_labels = train_batch_loader.next_batch()
#batch_images, batch_labels = train_batch_loader.get_batchdata()
in_imgs = (batch_images - 127.5) * 0.0078125
#print("data in ",in_img[0,:2,:2,0])
_, summary_str, train_acc, Center_loss, Softmax_loss, Pred_class, res1_o = sess.run(
[
train_op, summary_op, accuracy, center_loss,
softmax_loss, pred_class, res1
],
feed_dict={
input_images: in_imgs,
labels: batch_labels,
learn_rate: lr
})
step += 1
#print("step",step, str(Softmax_loss),str(Center_loss))
#print("step label",step, str(batch_labels))
graph_step += 1
if step % 100 == 0:
writer.add_summary(summary_str, global_step=graph_step)
pc_loss += Center_loss
ps_loss += Softmax_loss
acc_sum += train_acc
if step % 1000 == 0:
#lr = lr*0.1
#c_loss+=c_loss
#s_loss+=s_loss
print("****** Epoch {} Step {}: ***********".format(
str(epoch_idx), str(step)))
print("center loss: {}".format(pc_loss / 1000.0))
print("softmax_loss: {}".format(ps_loss / 1000.0))
print("train_acc: {}".format(acc_sum / 1000.0))
print("centers", res1_o[0, :5])
print("*******************************")
#if (acc_sum/100.0) >= 0.98 and (pc_loss/100.0)<40 and (ps_loss/100.0) <0.1 and ckpt_fg=='True':
if ckpt_fg == 'True':
print(
"******************************************************************************"
)
saver.save(sess, checkpoint_dir, global_step=epoch_idx)
ckpt_fg = 'False'
ps_loss = 0.0
pc_loss = 0.0
acc_sum = 0.0
epoch_idx += 1
if epoch_idx % 10 == 0:
print(
"******************************************************************************"
)
saver.save(sess, checkpoint_dir, global_step=epoch_idx)
#writer.add_summary(summary_str, global_step=step)
if epoch_idx % 5 == 0:
lr = lr * 0.5
if epoch_idx:
batch_images, batch_labels = test_batch_loader.next_batch()
#batch_images,batch_labels = train_batch_loader.get_valdata()
vali_image = (batch_images - 127.5) * 0.0078125
vali_acc = sess.run(accuracy,
feed_dict={
input_images: vali_image,
labels: batch_labels
})
print(("epoch: {}, train_acc:{:.4f}, vali_acc:{:.4f}".format(
epoch_idx, train_acc, vali_acc)))
sess.close()
from batch_loader import BatchLoader
import config
import random
if __name__ == '__main__':
param = {"net": "pnet", "batch": 64}
batch_loader = BatchLoader(param)
task = random.choice(config.TRAIN_TASKS[param['net']])
data = batch_loader.next_batch(64, task)
for datum in data:
print(datum)
