def chamfer_loss(A,B):
r=tf.reduce_sum(A*A,2)
r=tf.reshape(r,[int(r.shape[0]),int(r.shape[1]),1])
r2=tf.reduce_sum(B*B,2)
r2=tf.reshape(r2,[int(r.shape[0]),int(r.shape[1]),1])
t=(r-2*tf.matmul(A, tf.transpose(B,perm=[0, 2, 1])) + tf.transpose(r2,perm=[0, 2, 1]))
return tf.reduce_mean((tf.reduce_min(t, axis=1)+tf.reduce_min(t,axis=2))/2.0)
def sg_quasi_rnn(tensor, opt):
# Split
if opt.att:
H, Z, F, O = tf.split(tensor, 4, axis=0) # (16, 150, 320) for all
else:
Z, F, O = tf.split(tensor, 3, axis=0) # (16, 150, 320) for all
# step func
def step(z, f, o, c):
'''
Runs fo-pooling at each time step
'''
c = f * c + (1 - f) * z
if opt.att: # attention
a = tf.nn.softmax(tf.einsum("ijk,ik->ij", H,
c)) # alpha. (16, 150)
k = (a.sg_expand_dims() * H).sg_sum(
axis=1) # attentional sum. (16, 320)
h = o * (k.sg_dense(act="linear") + \
c.sg_dense(act="linear"))
else:
h = o * c
return h, c # hidden states, (new) cell memories
# Do rnn loop
c, hs = 0, []
timesteps = tensor.get_shape().as_list()[1]
for t in range(timesteps):
z = Z[:, t, :] # (16, 320)
f = F[:, t, :] # (16, 320)
o = O[:, t, :] # (16, 320)
# apply step function
h, c = step(z, f, o, c) # (16, 320), (16, 320)
# save result
hs.append(h.sg_expand_dims(axis=1))
# Concat to return
H = tf.concat(hs, 1) # (16, 150, 320)
seqlen = tf.to_int32(
tf.reduce_sum(tf.sign(tf.abs(tf.reduce_sum(H, axis=-1))),
1)) # (16,) float32
h = tf.reverse_sequence(input=H, seq_lengths=seqlen,
seq_dim=1)[:, 0, :] # last hidden state vector
if opt.is_enc:
H_z = tf.tile((h.sg_dense(act="linear").sg_expand_dims(axis=1)),
[1, timesteps, 1])
H_f = tf.tile((h.sg_dense(act="linear").sg_expand_dims(axis=1)),
[1, timesteps, 1])
H_o = tf.tile((h.sg_dense(act="linear").sg_expand_dims(axis=1)),
[1, timesteps, 1])
concatenated = tf.concat([H, H_z, H_f, H_o], 0) # (16*4, 150, 320)
return concatenated
else:
return H # (16, 150, 320)
def penalize_loss(gamma, lambd, tensor, tensor_n):
#gamma * (vector-vector_d)**2 - lamada * (vector dot vector_d)/(nor(vector)*nor(vector_d))
with tf.sg_context(name='penalize'):
square = tf.reduce_sum(tf.reduce_sum(tf.square(tensor - tensor_n), 2),
1)
cosine = tf.reduce_sum(
tf.reduce_sum(
tf.multiply(tf.nn.l2_normalize(tensor, 2),
tf.nn.l2_normalize(tensor_n, 2)), 2), 1)
return gamma * square - lambd * cosine
def rnn_classify(x, num_classes, is_test=False):
with tf.sg_context(name='rnn_classify'):
fw_cell = tf.nn.rnn_cell.MultiRNNCell(
[lstm_cell(is_test) for _ in range(num_blocks)],
state_is_tuple=True)
bw_cell = tf.nn.rnn_cell.MultiRNNCell(
[lstm_cell(is_test) for _ in range(num_blocks)],
state_is_tuple=True)
words_used_in_sent = tf.sign(
tf.reduce_max(tf.abs(x), reduction_indices=2))
length = tf.cast(
tf.reduce_sum(words_used_in_sent, reduction_indices=1), tf.int32)
outputs, _ = tf.nn.bidirectional_dynamic_rnn(fw_cell,
bw_cell,
x,
dtype=tf.float32,
sequence_length=length)
output = tf.concat(outputs, 2).sg_reshape(shape=[-1, 2 * latent_dim])
prediction = output.sg_dense(dim=num_classes, name='dense')
res = tf.reshape(prediction,
[x.get_shape().as_list()[0], -1, num_classes])
return res
def ner_cost(tensor, opt):
one_hot_labels = tf.one_hot(opt.target - 1, opt.num_classes, dtype=tf.float32)
cross_entropy = one_hot_labels * tf.log(tensor)
cross_entropy = -tf.reduce_sum(cross_entropy, reduction_indices=2)
mask = tf.sign(tf.abs(opt.target))
cross_entropy *= tf.cast(mask, tf.float32)
cross_entropy = tf.reduce_sum(cross_entropy, reduction_indices=1)
length = tf.cast(tf.reduce_sum(tf.sign(opt.target), reduction_indices=1), tf.int32)
cross_entropy /= tf.cast(length, tf.float32)
out = tf.reduce_mean(cross_entropy, name='ner_cost')
# add summary
tf.sg_summary_loss(out, name=opt.name)
return out
def ner_cost(tensor, opt):
one_hot_labels = tf.one_hot(opt.target - 1, opt.num_classes, dtype=tf.float32)
cross_entropy = one_hot_labels * tf.log(tensor)
cross_entropy = -tf.reduce_sum(cross_entropy, reduction_indices=2)
mask = tf.sign(tf.reduce_max(tf.abs(one_hot_labels), reduction_indices=2))
cross_entropy *= tf.cast(mask, tf.float32)
cross_entropy = tf.reduce_sum(cross_entropy, reduction_indices=1)
length = tf.cast(tf.reduce_sum(tf.sign(opt.target), reduction_indices=1), tf.int32)
cross_entropy /= tf.cast(length, tf.float32)
out = tf.reduce_mean(cross_entropy, name='ner_cost')
# add summary
tf.sg_summary_loss(out, name=opt.name)
return out
def rnn_classify(x, num_classes, is_test=False):
with tf.sg_context(name='rnn_classify'):
fw_cell = tf.nn.rnn_cell.MultiRNNCell([lstm_cell(is_test) for _ in range(num_blocks)], state_is_tuple=True)
bw_cell = tf.nn.rnn_cell.MultiRNNCell([lstm_cell(is_test) for _ in range(num_blocks)], state_is_tuple=True)
words_used_in_sent = tf.sign(tf.reduce_max(tf.abs(x), reduction_indices=2))
length = tf.cast(tf.reduce_sum(words_used_in_sent, reduction_indices=1), tf.int32)
outputs, _ = tf.nn.bidirectional_dynamic_rnn(fw_cell, bw_cell, x, dtype=tf.float32, sequence_length=length)
output = tf.concat(outputs, 2).sg_reshape(shape=[-1, 2 * latent_dim])
prediction = output.sg_dense(dim=num_classes, name='dense')
res = tf.reshape(prediction, [x.get_shape().as_list()[0], -1, num_classes])
return res
def sg_sum(tensor, opt):
r"""Computes the sum of elements across axis of a tensor.
See `tf.reduce_sum()` in tensorflow.
Args:
tensor: A `Tensor` with zero-padding (automatically given by chain).
opt:
axis: A tuple/list of integers or an integer. The axis to reduce.
keep_dims: If true, retains reduced dimensions with length 1.
name: If provided, replace current tensor's name.
Returns:
A `Tensor`.
"""
return tf.reduce_sum(tensor, axis=opt.axis, keep_dims=opt.keep_dims, name=opt.name)
def sg_quasi_conv1d(tensor, opt):
'''
Args:
tensor: A 3-D tensor of either [batch size, time steps, embedding size] for original
X or [batch size * 4, time steps, embedding size] for the others.
'''
opt += tf.sg_opt(is_enc=False)
# Split into H and H_zfo
H = tensor[:Hp.batch_size]
H_z = tensor[Hp.batch_size:2 * Hp.batch_size]
H_f = tensor[2 * Hp.batch_size:3 * Hp.batch_size]
H_o = tensor[3 * Hp.batch_size:]
if opt.is_enc:
H_z, H_f, H_o = 0, 0, 0
# Convolution and merging
with tf.sg_context(size=opt.size, act="linear", causal=(not opt.is_enc)):
Z = H.sg_aconv1d() + H_z # (16, 150, 320)
F = H.sg_aconv1d() + H_f # (16, 150, 320)
O = H.sg_aconv1d() + H_o # (16, 150, 320)
# Activation
with tf.sg_context(ln=True):
Z = Z.sg_bypass(act="tanh") # (16, 150, 320)
F = F.sg_bypass(act="sigmoid") # (16, 150, 320)
O = O.sg_bypass(act="sigmoid") # (16, 150, 320)
# Masking
M = tf.sign(tf.abs(tf.reduce_sum(
H, axis=-1, keep_dims=True))) # (16, 150, 1) float32. 0 or 1
Z *= M # broadcasting
F *= M # broadcasting
O *= M # broadcasting
# Concat
ZFO = tf.concat([Z, F, O], 0)
return ZFO # (16*3, 150, 320)
def sg_sum(tensor, opt):
return tf.reduce_sum(tensor,
reduction_indices=opt.dims,
keep_dims=opt.keep_dims,
name=opt.name)
bn=False)
d_p4 = ops.upconv_and_scale(d_p3,
dim=1,
size=size,
stride=stride,
act='linear',
bn=False)
disc = d_p4
#
# pull-away term ( PT ) regularizer
#
sample = gen.sg_flatten()
nom = tf.matmul(sample, tf.transpose(sample, perm=[1, 0]))
denom = tf.reduce_sum(tf.square(sample), reduction_indices=[1], keep_dims=True)
pt = tf.square(nom / denom)
pt -= tf.diag(tf.diag_part(pt))
pt = tf.reduce_sum(pt) / (batch_size * (batch_size - 1))
#
# loss & train ops
#
# mean squared errors
mse = tf.reduce_mean(tf.square(disc - xx), reduction_indices=[1, 2, 3])
mse_real, mse_fake = mse[:batch_size], mse[batch_size:]
loss_disc = mse_real + tf.maximum(margin - mse_fake, 0) # discriminator loss
loss_gen = mse_fake + pt * pt_weight # generator loss + PT regularizer
tf.random_normal([n_hidden_units_two, n_hidden_units_three],
mean=0,
stddev=sd))
b_3 = tf.Variable(tf.random_normal([n_hidden_units_three], mean=0, stddev=sd))
h_3 = tf.nn.sigmoid(tf.matmul(h_2, W_3) + b_3)
W = tf.Variable(
tf.random_normal([n_hidden_units_three, num_classes], mean=0, stddev=sd))
b = tf.Variable(tf.random_normal([num_classes], mean=0, stddev=sd))
with tf.name_scope('out'):
y_ = tf.nn.softmax(tf.matmul(h_3, W) + b, name="out")
init = tf.global_variables_initializer()
cost_function = tf.reduce_mean(
-tf.reduce_sum(Y * tf.log(y_), reduction_indices=[1]))
#optimizer = tf.train.RMSPropOptimizer(learning_rate,decay=0.9,momentum=0.9,centered=True).minimize(cost_function)
optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(
cost_function)
correct_prediction = tf.equal(tf.argmax(y_, 1), tf.argmax(Y, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
cost_history = np.empty(shape=[1], dtype=float)
acc_history = np.empty(shape=[1], dtype=float)
t_cost_history = np.empty(shape=[1], dtype=float)
t_acc_history = np.empty(shape=[1], dtype=float)
y_true, y_pred = None, None
with tf.Session() as session:
session.run(init)
def train(self):
''' Network '''
batch_pred_feats, batch_pred_coords, batch_pred_confs, self.final_state = self.LSTM(
'lstm', self.x)
iou_predict_truth, intersection = self.iou(batch_pred_coords,
self.y[:, 0:4])
should_exist = I = tf.cast(
tf.reduce_sum(self.y[:, 0:4], axis=1) > 0., tf.float32)
no_I = tf.ones_like(I, dtype=tf.float32) - I
object_loss = tf.nn.l2_loss(
I * (batch_pred_confs - iou_predict_truth)) * self.object_scale
noobject_loss = tf.nn.l2_loss(
no_I *
(batch_pred_confs - iou_predict_truth)) * self.noobject_scale
p_sqrt_w = tf.sqrt(
tf.minimum(1.0, tf.maximum(0.0, batch_pred_coords[:, 2])))
p_sqrt_h = tf.sqrt(
tf.minimum(1.0, tf.maximum(0.0, batch_pred_coords[:, 3])))
sqrt_w = tf.sqrt(tf.abs(self.y[:, 2]))
sqrt_h = tf.sqrt(tf.abs(self.y[:, 3]))
loss = (tf.nn.l2_loss(I * (batch_pred_coords[:, 0] - self.y[:, 0])) +
tf.nn.l2_loss(I * (batch_pred_coords[:, 1] - self.y[:, 1])) +
tf.nn.l2_loss(I * (p_sqrt_w - sqrt_w)) +
tf.nn.l2_loss(I * (p_sqrt_h - sqrt_h))) * self.coord_scale
#max_iou = tf.nn.l2_loss(I*(tf.ones_like(iou_predict_truth, dtype=tf.float32) - iou_predict_truth))
total_loss = loss + object_loss + noobject_loss #+ max_iou
''' Optimizer '''
optimizer = tf.train.AdamOptimizer(
learning_rate=self.learning_rate).minimize(
total_loss) # Adam Optimizer
''' Summary for tensorboard analysis '''
dataset_loss = -1
dataset_loss_best = 100
test_writer = tf.summary.FileWriter('summary/test')
tf.summary.scalar('dataset_loss', dataset_loss)
summary_op = tf.summary.merge_all()
''' Initializing the variables '''
self.saver = tf.train.Saver()
batch_states = np.zeros((self.batchsize, 2 * self.len_vec))
# TODO: make this a command line argument, etc.
# training set loader
batch_loader = BatchLoader(
"./DATA/TRAINING/",
seq_len=self.nsteps,
batch_size=self.batchsize,
step_size=1,
folders_to_use=[
"GOPR0005", "GOPR0006", "GOPR0008", "GOPR0008_2", "GOPR0009",
"GOPR0009_2", "GOPR0010", "GOPR0011", "GOPR0012", "GOPR0013",
"GOPR0014", "GOPR0015", "GOPR0016", "MVI_8607", "MVI_8609",
"MVI_8610", "MVI_8612", "MVI_8614", "MVI_8615", "MVI_8616"
])
validation_set_loader = BatchLoader(
"./DATA/VALID/",
seq_len=self.nsteps,
batch_size=self.batchsize,
step_size=1,
folders_to_use=[
"bbd_2017__2017-01-09-21-40-02_cam_flimage_raw",
"bbd_2017__2017-01-09-21-44-31_cam_flimage_raw",
"bbd_2017__2017-01-09-21-48-46_cam_flimage_raw",
"bbd_2017__2017-01-10-16-07-49_cam_flimage_raw",
"bbd_2017__2017-01-10-16-21-01_cam_flimage_raw",
"bbd_2017__2017-01-10-16-31-57_cam_flimage_raw",
"bbd_2017__2017-01-10-21-43-03_cam_flimage_raw",
"bbd_2017__2017-01-11-20-21-32_cam_flimage_raw",
"bbd_2017__2017-01-11-21-02-37_cam_flimage_raw"
])
print("%d available training batches" % len(batch_loader.batches))
print("%d available validation batches" %
len(validation_set_loader.batches))
''' Launch the graph '''
with tf.Session() as sess:
if self.restore_weights == True and os.path.isfile(
self.rolo_current_save + ".index"):
# sess.run(init)
tf.sg_init(sess)
self.saver.restore(sess, self.rolo_current_save)
print("Weight loaded, finetuning")
else:
# sess.run(init)
tf.sg_init(sess)
print("Training from scratch")
epoch_loss = []
for self.iter_id in range(self.n_iters):
''' Load training data & ground truth '''
batch_id = self.iter_id - self.batch_offset
batch_xs, batch_ys, _ = batch_loader.load_batch(batch_id)
''' Update weights by back-propagation '''
sess.run(optimizer,
feed_dict={
self.x: batch_xs,
self.y: batch_ys
})
if self.iter_id % self.display_step == 0:
''' Calculate batch loss '''
batch_loss = sess.run(total_loss,
feed_dict={
self.x: batch_xs,
self.y: batch_ys
})
epoch_loss.append(batch_loss)
print("Total Batch loss for iteration %d: %.9f" %
(self.iter_id, batch_loss))
if self.iter_id % self.display_step == 0:
''' Calculate batch loss '''
batch_loss = sess.run(loss,
feed_dict={
self.x: batch_xs,
self.y: batch_ys
})
print(
"Bounding box coord error loss for iteration %d: %.9f"
% (self.iter_id, batch_loss))
if self.display_object_loss and self.iter_id % self.display_step == 0:
''' Calculate batch object loss '''
batch_o_loss = sess.run(object_loss,
feed_dict={
self.x: batch_xs,
self.y: batch_ys
})
print("Object loss for iteration %d: %.9f" %
(self.iter_id, batch_o_loss))
if self.display_object_loss and self.iter_id % self.display_step == 0:
''' Calculate batch object loss '''
batch_noo_loss = sess.run(noobject_loss,
feed_dict={
self.x: batch_xs,
self.y: batch_ys
})
print("No Object loss for iteration %d: %.9f" %
(self.iter_id, batch_noo_loss))
if self.iou_with_ground_truth and self.iter_id % self.display_step == 0:
''' Calculate batch object loss '''
batch_o_loss = sess.run(tf.reduce_mean(iou_predict_truth),
feed_dict={
self.x: batch_xs,
self.y: batch_ys
})
print("Average IOU with ground for iteration %d: %.9f" %
(self.iter_id, batch_o_loss))
if self.display_coords is True and self.iter_id % self.display_step == 0:
''' Caculate predicted coordinates '''
coords_predict = sess.run(batch_pred_coords,
feed_dict={
self.x: batch_xs,
self.y: batch_ys
})
print("predicted coords:" + str(coords_predict[0]))
print("ground truth coords:" + str(batch_ys[0]))
''' Save model '''
if self.iter_id % self.save_step == 1:
self.saver.save(sess, self.rolo_current_save)
print("\n Model saved in file: %s" %
self.rolo_current_save)
''' Validation '''
if self.validate == True and self.iter_id % self.validate_step == 0 and self.iter_id > 0:
# Run validation set
dataset_loss = self.test(sess, total_loss,
validation_set_loader,
batch_pred_feats,
batch_pred_coords,
batch_pred_confs,
self.final_state)
''' Early-stop regularization '''
if dataset_loss <= dataset_loss_best:
dataset_loss_best = dataset_loss
self.saver.save(sess, self.rolo_weights_file)
print("\n Better Model saved in file: %s" %
self.rolo_weights_file)
''' Write summary for tensorboard '''
summary = sess.run(summary_op,
feed_dict={
self.x: batch_xs,
self.y: batch_ys
})
test_writer.add_summary(summary, self.iter_id)
print("Average total loss %f" % np.mean(epoch_loss))
return
