def conv(self, input, k_h, k_w, c_o, s_h, s_w, name, relu=True, padding=DEFAULT_PADDING, group=1, biased=True):
# Verify that the padding is acceptable
self.validate_padding(padding)
# Get the number of channels in the input
c_i = input.get_shape()[-1]
# Verify that the grouping parameter is valid
assert c_i % group == 0
assert c_o % group == 0
# Convolution for a given input and kernel
convolve = lambda i, k: tf.nn.conv2d(i, k, [1, s_h, s_w, 1], padding=padding)
with tf.variable_scope(name) as scope:
kernel = self.make_var("weights", shape=[k_h, k_w, c_i / group, c_o])
if group == 1:
# This is the common-case. Convolve the input without any further complications.
output = convolve(input, kernel)
else:
# Split the input into groups and then convolve each of them independently
input_groups = tf.split(3, group, input)
kernel_groups = tf.split(3, group, kernel)
output_groups = [convolve(i, k) for i, k in zip(input_groups, kernel_groups)]
# Concatenate the groups
output = tf.concat(3, output_groups)
# Add the biases
if biased:
biases = self.make_var("biases", [c_o])
output = tf.nn.bias_add(output, biases)
if relu:
# ReLU non-linearity
output = tf.nn.relu(output, name=scope.name)
return output
def compute_IOU(bboxA, bboxB):
"""Compute the Intersection Over Union.
Args:
bboxA: [N X 4 tensor] format = [left, top, right, bottom]
bboxB: [N X 4 tensor]
Return:
IOU: [N X 1 tensor]
"""
x1A, y1A, x2A, y2A = tf.split(1, 4, bboxA)
x1B, y1B, x2B, y2B = tf.split(1, 4, bboxB)
# compute intersection
x1_max = tf.maximum(x1A, x1B)
y1_max = tf.maximum(y1A, y1B)
x2_min = tf.minimum(x2A, x2B)
y2_min = tf.minimum(y2A, y2B)
# overlap_flag = tf.logical_and( tf.less(x1_max, x2_min), tf.less(y1_max, y2_min))
overlap_flag = tf.to_float(tf.less(x1_max, x2_min)) * \
tf.to_float(tf.less(y1_max, y2_min))
overlap_area = tf.mul(overlap_flag, tf.mul(
x2_min - x1_max, y2_min - y1_max))
# compute union
areaA = tf.mul(x2A - x1A, y2A - y1A)
areaB = tf.mul(x2B - x1B, y2B - y1B)
union_area = areaA + areaB - overlap_area
return tf.div(overlap_area, union_area)
def make_example_dict(example_protos, example_weights):
def parse_examples(example_protos):
features = {
"target": tf.FixedLenFeature(shape=[1], dtype=tf.float32, default_value=0),
"age_indices": tf.VarLenFeature(dtype=tf.int64),
"age_values": tf.VarLenFeature(dtype=tf.float32),
"gender_indices": tf.VarLenFeature(dtype=tf.int64),
"gender_values": tf.VarLenFeature(dtype=tf.float32),
}
return tf.parse_example([e.SerializeToString() for e in example_protos], features)
parsed = parse_examples(example_protos)
sparse_features = [
SparseFeatureColumn(
tf.reshape(tf.split(1, 2, parsed["age_indices"].indices)[0], [-1]),
tf.reshape(parsed["age_indices"].values, [-1]),
tf.reshape(parsed["age_values"].values, [-1]),
),
SparseFeatureColumn(
tf.reshape(tf.split(1, 2, parsed["gender_indices"].indices)[0], [-1]),
tf.reshape(parsed["gender_indices"].values, [-1]),
tf.reshape(parsed["gender_values"].values, [-1]),
),
]
return dict(
sparse_features=sparse_features,
dense_features=[],
example_weights=example_weights,
example_labels=tf.reshape(parsed["target"], [-1]),
example_ids=["%d" % i for i in range(0, len(example_protos))],
)
def _composition_function(self, inputs, length, init_state=None):
if self._composition == "GRU":
cell = GRUCell(self._size)
return dynamic_rnn(cell, inputs, sequence_length=length, time_major=True,
initial_state=init_state, dtype=tf.float32)[0]
elif self._composition == "LSTM":
cell = BasicLSTMCell(self._size)
init_state = tf.concat(1, [tf.zeros_like(init_state, tf.float32), init_state]) if init_state else None
outs = dynamic_rnn(cell, inputs, sequence_length=length, time_major=True,
initial_state=init_state, dtype=tf.float32)[0]
return outs
elif self._composition == "BiGRU":
cell = GRUCell(self._size // 2, self._size)
init_state_fw, init_state_bw = tf.split(1, 2, init_state) if init_state else (None, None)
with tf.variable_scope("forward"):
fw_outs = dynamic_rnn(cell, inputs, sequence_length=length, time_major=True,
initial_state=init_state_fw, dtype=tf.float32)[0]
with tf.variable_scope("backward"):
rev_inputs = tf.reverse_sequence(tf.pack(inputs), length, 0, 1)
rev_inputs = [tf.reshape(x, [-1, self._size]) for x in tf.split(0, len(inputs), rev_inputs)]
bw_outs = dynamic_rnn(cell, rev_inputs, sequence_length=length, time_major=True,
initial_state=init_state_bw, dtype=tf.float32)[0]
bw_outs = tf.reverse_sequence(tf.pack(bw_outs), length, 0, 1)
bw_outs = [tf.reshape(x, [-1, self._size]) for x in tf.split(0, len(inputs), bw_outs)]
return [tf.concat(1, [fw_out, bw_out]) for fw_out, bw_out in zip(fw_outs, bw_outs)]
else:
raise NotImplementedError("Other compositions not implemented yet.")
def build(self):
"""None
Build the model graph
:return:
"""
with tf.name_scope('G_'):
self.predict_g = self.__G__()
self.predict_g2 = self.__G2__()
with tf.name_scope('D_'):
# Create reference examples
# Input d holds real&imaginary values. The discriminative decision based on reconstructed image
self.reconstructed_image_reference = self.get_reconstructed_image(real=self.input_d['real'],
imag=self.input_d['imag'], name='Both_gt')
predict_g2_stacked = tf.stack([self.predict_g2['real'][:,0,:,:], self.predict_g2['imag'][:,0,:,:]], axis=1)
self.predict, self.predict_logits = self.__D__([self.reconstructed_image_reference, predict_g2_stacked])
self.predict_d, self.predict_d_for_g = tf.split(value=self.predict, num_or_size_splits=2, axis=0)
self.predict_d_logits, self.predict_d_logits_for_g = tf.split(value=self.predict_logits,
num_or_size_splits=2, axis=0)
self.clip_weights = self.__clip_weights__()
with tf.name_scope('loss'):
# self.loss_g = self.__loss_g__(predict=self.predict_g, self.labels, reg=self.regularization_sum)
self.__loss__()
with tf.name_scope('training'):
self.train_op_d, self.train_op_g = self.__training__(learning_rate=self.FLAGS.learning_rate)
with tf.name_scope('evaluation'):
# Calculate accuracy L2 norm
self.evaluation = self.__evaluation__(predict=self.predict_g, labels=self.labels)
def build(self):
"""None
Build the model graph
:return:
"""
with tf.name_scope('G_'):
self.predict_g = self.__G__()
with tf.name_scope('D_'):
self.predict, self.predict_logits = self.__D__([self.input_d, self.predict_g], input_type="Real")
self.predict_d, self.predict_d_for_g = tf.split(value=self.predict, num_or_size_splits=2, axis=0)
self.predict_d_logits, self.predict_d_logits_for_g = tf.split(value=self.predict_logits,
num_or_size_splits=2, axis=0)
# self.predict_d, self.predict_d_logits
# with tf.variable_scope(tf.get_variable_scope(), reuse=True):
# self.predict_d_for_g, self.predict_d_logits_for_g = self.__D__(self.predict_g, input_type="Gen")
if len(self.regularization_values_d) > 0:
self.regularization_sum_d = sum(self.regularization_values_d)
with tf.name_scope('loss'):
# self.loss_g = self.__loss_g__(predict=self.predict_g, self.labels, reg=self.regularization_sum)
self.__loss__()
with tf.name_scope('training'):
self.train_op_d, self.train_op_g = self.__training__(learning_rate=self.FLAGS.learning_rate)
with tf.name_scope('evaluation'):
# Calculate accuracy L2 norm
self.evaluation = self.__evaluation__(predict=self.predict_g, labels=self.labels)
def _g_recurrence_1(i, x_t, input_x, gen_x, h_tm1, h_tm1_manager, last_goal, real_goal, give_num):
cur_sen = \
tf.split(tf.concat([tf.split(input_x, [i, self.sequence_length - i], 1)[0], self.padding_array], 1),
[self.sequence_length, i], 1)[0]
with tf.variable_scope(self.scope):
feature = self.FeatureExtractor_unit(cur_sen, self.drop_out)
h_t_manager = self.g_manager_recurrent_unit(feature, h_tm1_manager)
sub_goal = self.g_manager_output_unit(h_t_manager)
sub_goal = tf.nn.l2_normalize(sub_goal, 1)
h_t_Worker = tf.cond(i > 0, lambda: self.g_worker_recurrent_unit(x_t, h_tm1),
lambda: h_tm1) # hidden_memory_tuple
real_sub_goal = tf.cond(i > 0, lambda: tf.add(last_goal, sub_goal), lambda: real_goal)
# real_goal_array = real_goal_array.write(i, real_sub_goal)
x_tp1 = tf.cond(i > 0, lambda: ta_emb_x.read(i - 1), lambda: x_t)
# hidden_memory_tuple
with tf.control_dependencies([cur_sen]):
gen_x = tf.cond(i > 0, lambda: gen_x.write(i - 1, ta_x.read(i - 1)), lambda: gen_x)
return i + 1, x_tp1, input_x, gen_x, h_t_Worker, h_t_manager, \
tf.cond(((i) % self.step_size) > 0, lambda: real_sub_goal,
lambda: tf.constant(0.0, shape=[self.batch_size, self.goal_out_size])), \
tf.cond(((i) % self.step_size) > 0, lambda: real_goal, lambda: real_sub_goal), give_num
def testSymbolModalityTargetsFactored(self):
batch_size = 10
num_datashards = 5
length = 6
height = 7
hidden_size = 9
vocab_size = 11
model_hparams = common_hparams.basic_params1()
model_hparams.factored_logits = True
model_hparams.hidden_size = hidden_size
model_hparams.mode = tf.estimator.ModeKeys.TRAIN
body_output = -1 + np.random.random_integers(
100, size=(batch_size, length, height, hidden_size))
targets = -1 + np.random.random_integers(
vocab_size, size=(batch_size, length, height, 1))
m = modalities.SymbolModality(model_hparams, vocab_size)
data_parallelism = expert_utils.Parallelism(
["/device:CPU:0"] * num_datashards)
with self.test_session() as session:
sharded_body_output = tf.split(tf.to_float(body_output), num_datashards)
sharded_targets = tf.split(targets, num_datashards)
sharded_logits = m.top_sharded(sharded_body_output, sharded_targets,
data_parallelism)
train_loss = m.loss_sharded(sharded_logits, sharded_targets,
data_parallelism)
logits = tf.concat(sharded_logits, 0)
session.run(tf.global_variables_initializer())
res1, res2 = session.run((logits, train_loss))
self.assertEqual(res1.shape, (batch_size, length, height, 1, vocab_size))
self.assertEqual(res2.shape, ())
def test_backward_grads_with_nativepy(self):
if not tf.test.is_gpu_available():
self.skipTest("GPU not available")
input_shape = (128, 8, 8)
data_shape = (16,) + input_shape
x = tf.random_normal(shape=data_shape, dtype=tf.float64)
dy = tf.random_normal(shape=data_shape, dtype=tf.float64)
dy1, dy2 = tf.split(dy, num_or_size_splits=2, axis=1)
block = blocks.RevBlock(
n_res=3,
filters=128,
strides=(1, 1),
input_shape=input_shape,
fused=False,
dtype=tf.float64)
with tf.GradientTape() as tape:
tape.watch(x)
x1, x2 = tf.split(x, num_or_size_splits=2, axis=1)
y1, y2 = block((x1, x2), training=True)
y = tf.concat((y1, y2), axis=1)
# Compute true grads
dx_true = tape.gradient(y, x, output_gradients=dy)
# Compute grads from reconstruction
(dx1, dx2), _ = block.backward_grads(
x=(x1, x2), y=(y1, y2), dy=(dy1, dy2), training=True)
dx = tf.concat((dx1, dx2), axis=1)
thres = 1e-5
diff_abs = tf.reshape(abs(dx - dx_true), [-1])
assert all(diff_abs < thres)
def _model_fn(features, labels, mode, params):
model_fn = MODELS[FLAGS.model].model_fn
global_step = tf.train.get_or_create_global_step()
if FLAGS.num_gpus > 0 and mode == learn.ModeKeys.TRAIN:
split_features = {k: tf.split(v, FLAGS.num_gpus)
for k, v in features.iteritems()}
split_labels = {k: tf.split(v, FLAGS.num_gpus)
for k, v in labels.iteritems()}
grads = []
predictions = collections.defaultdict(list)
losses = []
opt = ops.create_optimizer(
params.optimizer, params.learning_rate, params.decay_steps)
for i in range(FLAGS.num_gpus):
with tf.device(tf.DeviceSpec(device_type='GPU', device_index=i)):
with tf.name_scope('tower_%d' % i):
with tf.variable_scope(tf.get_variable_scope(), reuse=i > 0):
device_features = {k: v[i] for k, v in split_features.iteritems()}
device_labels = {k: v[i] for k, v in split_labels.iteritems()}
device_predictions, device_loss = model_fn(
device_features, device_labels, mode, params)
for k, v in device_predictions.iteritems():
predictions[k].append(v)
if device_loss is not None:
losses.append(device_loss)
device_grads = opt.compute_gradients(device_loss)
grads.append(device_grads)
grads = ops.average_gradients(grads)
train_op = opt.apply_gradients(grads, global_step=global_step)
for k, v in predictions.iteritems():
predictions[k] = tf.concat(v, axis=0)
loss = tf.add_n(losses) if losses else None
else:
with tf.device(tf.DeviceSpec(device_type='GPU', device_index=0)):
predictions, loss = model_fn(features, labels, mode, params)
train_op = None
if mode == learn.ModeKeys.TRAIN:
opt = ops.create_optimizer(
params.optimizer, params.learning_rate, params.decay_steps)
train_op = opt.minimize(loss, global_step=global_step)
tf.summary.scalar('loss/loss', loss)
return tf.contrib.learn.ModelFnOps(
mode=mode,
predictions=predictions,
loss=loss,
train_op=train_op)
def backward_grads(self, y, dy, training=True):
"""Manually compute backward gradients given input and output grads."""
dy1, dy2 = tf.split(dy, num_or_size_splits=2, axis=self.axis)
y1, y2 = tf.split(y, num_or_size_splits=2, axis=self.axis)
with tf.GradientTape() as gtape:
gtape.watch(y1)
gy1 = self.g(y1, training=training)
grads_combined = gtape.gradient(
gy1, [y1] + self.g.trainable_variables, output_gradients=dy2)
dg = grads_combined[1:]
dx1 = dy1 + grads_combined[0]
# This doesn't affect eager execution, but improves memory efficiency with
# graphs
with tf.control_dependencies(dg + [dx1]):
x2 = y2 - gy1
with tf.GradientTape() as ftape:
ftape.watch(x2)
fx2 = self.f(x2, training=training)
grads_combined = ftape.gradient(
fx2, [x2] + self.f.trainable_variables, output_gradients=dx1)
df = grads_combined[1:]
dx2 = dy2 + grads_combined[0]
# Same behavior as above
with tf.control_dependencies(df + [dx2]):
x1 = y1 - fx2
x = tf.concat([x1, x2], axis=self.axis)
dx = tf.concat([dx1, dx2], axis=self.axis)
grads = df + dg
return x, dx, grads
def infer(self, features, *args, **kwargs): # pylint: disable=arguments-differ
"""Produce predictions from the model."""
del args, kwargs
# Inputs and features preparation needed to handle edge cases.
if not features:
features = {}
inputs_old = None
if "inputs" in features and len(features["inputs"].shape) < 4:
inputs_old = features["inputs"]
features["inputs"] = tf.expand_dims(features["inputs"], 2)
# Set targets to input size firts.
features["targets"] = tf.zeros_like(features["inputs"])
self._encode_on_predict = True
logits, _ = self(features) # pylint: disable=not-callable
if self.hparams.gan_loss_factor != 0:
logits, _ = tf.split(logits, 2, axis=0) # Remove GAN.
logits, _ = tf.split(logits, 2, axis=0) # Targets and inputs from encoding.
# Uncomment the line below to get reconstructed inputs instead of targets.
# (and comment out the line above at the same time).
# _, logits = tf.split(logits, 2, axis=0)
samples = tf.argmax(logits, axis=-1)
# Restore inputs to not confuse Estimator in edge cases.
if inputs_old is not None:
features["inputs"] = inputs_old
# Return samples.
return samples
def call(self, x, mask=None):
"""Execute this layer on input tensors.
x = [atom_features, atom_mask]
Parameters
----------
x: list
Tensors as listed above
mask: bool, optional
Ignored. Present only to shadow superclass call() method.
Returns
-------
outputs: Tensor
Tensor of concatenated atom features
"""
self.build()
atom_features = x[0]
atom_masks = x[1]
A = tf.split(atom_features, self.batch_size, axis=0)
A_mask = tf.split(
tf.cast(atom_masks, dtype=tf.bool), self.batch_size, axis=0)
outputs = tf.concat(
[tf.boolean_mask(A[i], A_mask[i]) for i in range(len(A))], axis=0)
outputs = tf.matmul(outputs, self.W) + self.b
outputs = self.activation(outputs)
return outputs
def lnlstm(xs, ms, s, scope, nh, init_scale=1.0):
nbatch, nin = [v.value for v in xs[0].get_shape()]
with tf.variable_scope(scope):
wx = tf.get_variable("wx", [nin, nh*4], initializer=ortho_init(init_scale))
gx = tf.get_variable("gx", [nh*4], initializer=tf.constant_initializer(1.0))
bx = tf.get_variable("bx", [nh*4], initializer=tf.constant_initializer(0.0))
wh = tf.get_variable("wh", [nh, nh*4], initializer=ortho_init(init_scale))
gh = tf.get_variable("gh", [nh*4], initializer=tf.constant_initializer(1.0))
bh = tf.get_variable("bh", [nh*4], initializer=tf.constant_initializer(0.0))
b = tf.get_variable("b", [nh*4], initializer=tf.constant_initializer(0.0))
gc = tf.get_variable("gc", [nh], initializer=tf.constant_initializer(1.0))
bc = tf.get_variable("bc", [nh], initializer=tf.constant_initializer(0.0))
c, h = tf.split(axis=1, num_or_size_splits=2, value=s)
for idx, (x, m) in enumerate(zip(xs, ms)):
c = c*(1-m)
h = h*(1-m)
z = _ln(tf.matmul(x, wx), gx, bx) + _ln(tf.matmul(h, wh), gh, bh) + b
i, f, o, u = tf.split(axis=1, num_or_size_splits=4, value=z)
i = tf.nn.sigmoid(i)
f = tf.nn.sigmoid(f)
o = tf.nn.sigmoid(o)
u = tf.tanh(u)
c = f*c + i*u
h = o*tf.tanh(_ln(c, gc, bc))
xs[idx] = h
s = tf.concat(axis=1, values=[c, h])
return xs, s
def add_training_loss(self, final_loss, logits):
"""Computes loss using logits."""
loss_fn = get_loss_fn(final_loss) # Get loss function
task_losses = []
# label_placeholder of shape (batch_size, n_tasks). Split into n_tasks
# tensors of shape (batch_size,)
task_labels = tf.split(
axis=1, num_or_size_splits=self.n_tasks, value=self.label_placeholder)
task_weights = tf.split(
axis=1, num_or_size_splits=self.n_tasks, value=self.weight_placeholder)
for task in range(self.n_tasks):
task_label_vector = task_labels[task]
task_weight_vector = task_weights[task]
# Convert the labels into one-hot vector encodings.
one_hot_labels = tf.to_float(
tf.one_hot(tf.to_int32(tf.squeeze(task_label_vector)), 2))
# Since we use tf.nn.softmax_cross_entropy_with_logits note that we pass in
# un-softmaxed logits rather than softmax outputs.
task_loss = loss_fn(logits[task], one_hot_labels, task_weight_vector)
task_losses.append(task_loss)
# It's ok to divide by just the batch_size rather than the number of nonzero
# examples (effect averages out)
total_loss = tf.add_n(task_losses)
total_loss = tf.div(total_loss, self.batch_size)
return total_loss
def remove_channels(x, data_format='NHWC'):
b, h, w, c = get_conv_shape(x, data_format)
if data_format == 'NCHW':
x, _ = tf.split(x, [3, -1], axis=1)
else:
x, _ = tf.split(x, [3, -1], axis=3)
return x
def testSymbolModalityTargets(self):
batch_size = 10
num_datashards = 5
length = 6
height = 7
hidden_size = 9
vocab_size = 11
model_hparams = tf.contrib.training.HParams(
symbol_modality_num_shards=4,
hidden_size=hidden_size,
label_smoothing=0.2,
shared_embedding_and_softmax_weights=0)
body_output = -1 + np.random.random_integers(
100, size=(batch_size, length, height, hidden_size))
targets = -1 + np.random.random_integers(
vocab_size, size=(batch_size, length, height, 1))
m = modalities.SymbolModality(model_hparams, vocab_size)
data_parallelism = expert_utils.Parallelism(
["/device:CPU:0"] * num_datashards, reuse=True)
with self.test_session() as session:
sharded_body_output = tf.split(tf.to_float(body_output), num_datashards)
sharded_targets = tf.split(targets, num_datashards)
sharded_logits, train_loss = m.top_sharded(
sharded_body_output, sharded_targets, data_parallelism)
logits = tf.concat(sharded_logits, 0)
session.run(tf.global_variables_initializer())
res1, res2 = session.run((logits, train_loss))
self.assertEqual(res1.shape, (batch_size, length, height, 1, vocab_size))
self.assertEqual(res2.shape, ())
def __call__(self, x, state, scope=None):
with tf.variable_scope(scope or type(self).__name__):
c, h = tf.split(state, 2, 1)
x_size = x.get_shape().as_list()[1]
w_init = None # uniform
h_init = lstm_ortho_initializer(1.0)
# Keep W_xh and W_hh separate here as well to use different init methods.
w_xh = tf.get_variable(
'W_xh', [x_size, 4 * self.num_units], initializer=w_init)
w_hh = tf.get_variable(
'W_hh', [self.num_units, 4 * self.num_units], initializer=h_init)
bias = tf.get_variable(
'bias', [4 * self.num_units],
initializer=tf.constant_initializer(0.0))
concat = tf.concat([x, h], 1)
w_full = tf.concat([w_xh, w_hh], 0)
hidden = tf.matmul(concat, w_full) + bias
i, j, f, o = tf.split(hidden, 4, 1)
if self.use_recurrent_dropout:
g = tf.nn.dropout(tf.tanh(j), self.dropout_keep_prob)
else:
g = tf.tanh(j)
new_c = c * tf.sigmoid(f + self.forget_bias) + tf.sigmoid(i) * g
new_h = tf.tanh(new_c) * tf.sigmoid(o)
return new_h, tf.concat([new_c, new_h], 1) # fuk tuples.
def forward_backward(self, obs_prob_seq):
"""
runs forward backward algorithm on observation sequence
Arguments
---------
- obs_seq : matrix of size N by S, where N is number of timesteps and
S is the number of states
Returns
-------
- forward : matrix of size N by S representing
the forward probability of each state at each time step
- backward : matrix of size N by S representing
the backward probability of each state at each time step
- posterior : matrix of size N by S representing
the posterior probability of each state at each time step
"""
obs_prob_list_for = tf.split(0, self.N, obs_prob_seq)
with tf.name_scope('forward_belief_propagation'):
# forward belief propagation
self._forward(obs_prob_list_for)
obs_prob_seq_rev = tf.reverse(obs_prob_seq, [True, False])
obs_prob_list_back = tf.split(0, self.N, obs_prob_seq_rev)
with tf.name_scope('backward_belief_propagation'):
# backward belief propagation
self._backward(obs_prob_list_back)
def minibatch(self, dataset, subset, use_datasets, cache_data,
shift_ratio=-1):
"""Get synthetic image batches.
"""
del subset, use_datasets, cache_data, shift_ratio
input_shape = [self.batch_size, self.height, self.width, self.depth]
images = tf.truncated_normal(
input_shape,
dtype=self.dtype,
stddev=1e-1,
name='synthetic_images')
labels = tf.random_uniform(
[self.batch_size],
minval=0,
maxval=dataset.num_classes - 1,
dtype=tf.int32,
name='synthetic_labels')
# Note: This results in a H2D copy, but no computation
# Note: This avoids recomputation of the random values, but still
# results in a H2D copy.
images = tf.contrib.framework.local_variable(images, name='images')
labels = tf.contrib.framework.local_variable(labels, name='labels')
if self.num_splits == 1:
images_splits = [images]
labels_splits = [labels]
else:
images_splits = tf.split(images, self.num_splits, 0)
labels_splits = tf.split(labels, self.num_splits, 0)
return images_splits, labels_splits
def __call__(self, x, state, timestep=0, scope=None):
with tf.variable_scope(scope or type(self).__name__):
h, c = tf.split(state, 2, 1)
h_size = self.num_units
x_size = x.get_shape().as_list()[1]
batch_size = x.get_shape().as_list()[0]
w_init = None # uniform
h_init = lstm_ortho_initializer(1.0)
w_xh = tf.get_variable(
'W_xh', [x_size, 4 * self.num_units], initializer=w_init)
w_hh = tf.get_variable(
'W_hh', [self.num_units, 4 * self.num_units], initializer=h_init)
concat = tf.concat([x, h], 1) # concat for speed.
w_full = tf.concat([w_xh, w_hh], 0)
concat = tf.matmul(concat, w_full) #+ bias # live life without garbage.
# i = input_gate, j = new_input, f = forget_gate, o = output_gate
concat = layer_norm_all(concat, batch_size, 4, h_size, 'ln_all')
i, j, f, o = tf.split(concat, 4, 1)
if self.use_recurrent_dropout:
g = tf.nn.dropout(tf.tanh(j), self.dropout_keep_prob)
else:
g = tf.tanh(j)
new_c = c * tf.sigmoid(f + self.forget_bias) + tf.sigmoid(i) * g
new_h = tf.tanh(layer_norm(new_c, h_size, 'ln_c')) * tf.sigmoid(o)
return new_h, tf.concat([new_h, new_c], 1)
def _read(self, keys, redundant_states):
read = _comp_mul(keys, redundant_states)
if self._num_copies > 1:
xs_real = tf.split(1, self._num_copies, _comp_real(read))
xs_imag = tf.split(1, self._num_copies, _comp_imag(read))
read = (tf.add_n(xs_real)/self._num_copies, tf.add_n(xs_imag)/self._num_copies)
return read
def decode_bbox_target(box_predictions, anchors):
"""
Args:
box_predictions: (..., 4), logits
anchors: (..., 4), floatbox. Must have the same shape
Returns:
box_decoded: (..., 4), float32. With the same shape.
"""
orig_shape = tf.shape(anchors)
box_pred_txtytwth = tf.reshape(box_predictions, (-1, 2, 2))
box_pred_txty, box_pred_twth = tf.split(box_pred_txtytwth, 2, axis=1)
# each is (...)x1x2
anchors_x1y1x2y2 = tf.reshape(anchors, (-1, 2, 2))
anchors_x1y1, anchors_x2y2 = tf.split(anchors_x1y1x2y2, 2, axis=1)
waha = anchors_x2y2 - anchors_x1y1
xaya = (anchors_x2y2 + anchors_x1y1) * 0.5
clip = np.log(config.PREPROC.MAX_SIZE / 16.)
wbhb = tf.exp(tf.minimum(box_pred_twth, clip)) * waha
xbyb = box_pred_txty * waha + xaya
x1y1 = xbyb - wbhb * 0.5
x2y2 = xbyb + wbhb * 0.5 # (...)x1x2
out = tf.concat([x1y1, x2y2], axis=-2)
return tf.reshape(out, orig_shape)
def __call__(self, inputs, state, scope=None):
"""Long short-term memory cell (LSTM)."""
with tf.variable_scope(self, scope or "basic_lstm_cell", reuse=self._reuse):
# Parameters of gates are concatenated into one multiply for
# efficiency.
if self._state_is_tuple:
c_prev, h_prev = state
else:
c_prev, h_prev = tf.split(
value=state, num_or_size_splits=2, axis=1)
concat = tf.contrib.rnn._linear(
[inputs, h_prev], 4 * self._num_units, True)
# i = input_gate, g = new_input, f = forget_gate, o = output_gate
i, g, f, o = tf.split(value=concat, num_or_size_splits=4, axis=1)
c = (c_prev * tf.sigmoid(f + self._forget_bias) +
tf.sigmoid(i) * tf.tanh(g))
h = tf.tanh(c) * tf.sigmoid(o)
if self._state_is_tuple:
new_state = LSTMStateTuple(c, h)
else:
new_state = tf.concat([c, h], 1)
return h, new_state
def conv(self,
input,
k_h,
k_w,
c_o,
s_h,
s_w,
name,
relu=True,
padding=DEFAULT_PADDING,
group=1):
self.validate_padding(padding)
c_i = input.get_shape()[-1]
assert c_i % group == 0
assert c_o % group == 0
convolve = lambda i, k: tf.nn.conv2d(i, k, [1, s_h, s_w, 1], padding=padding)
with tf.variable_scope(name) as scope:
kernel = self.make_var('weights', shape=[k_h, k_w, c_i / group, c_o])
biases = self.make_var('biases', [c_o])
if group == 1:
conv = convolve(input, kernel)
else:
input_groups = tf.split(3, group, input)
kernel_groups = tf.split(3, group, kernel)
output_groups = [convolve(i, k) for i, k in zip(input_groups, kernel_groups)]
conv = tf.concat(3, output_groups)
if relu:
bias = tf.reshape(tf.nn.bias_add(conv, biases), conv.get_shape().as_list())
return tf.nn.relu(bias, name=scope.name)
return tf.reshape(
tf.nn.bias_add(conv, biases),
conv.get_shape().as_list(),
name=scope.name)
def build_network(self):
net_tensors = self.net_tensors
with self.net_graph.as_default(), tf.device(self.net_device):
logits = tf.placeholder(dtype=tf.float32, shape=(self.batch_size, self.image_classes))
labels = tf.placeholder(dtype=tf.int32, shape=(self.batch_size,))
lambs = tf.placeholder(dtype=tf.float32, shape=(self.image_classes,))
# put a sigfunction on logits and then transpose
logits = tf.transpose(framwork.sig_func(logits))
# according to the labels, erase rows which is not in labels
labels_unique = tf.constant(range(self.image_classes), dtype=tf.int32)
labels_num = self.image_classes
logits = tf.gather(logits, indices=labels_unique)
lambs = tf.gather(lambs, indices=labels_unique)
# set the value of each row to True when it occurs in labels
templete = tf.tile(tf.expand_dims(labels_unique, dim=1), [1, self.batch_size])
labels_expand = tf.tile(tf.expand_dims(labels, dim=0), [labels_num, 1])
indict_logic = tf.equal(labels_expand, templete)
# split the tensor along rows
logit_list = tf.split(0, labels_num, logits)
indict_logic_list = tf.split(0, labels_num, indict_logic)
lamb_list = tf.split(0, self.image_classes, lambs)
logit_list = [tf.squeeze(item) for item in logit_list]
indict_logic_list = [tf.squeeze(item) for item in indict_logic_list]
left_right_tuples = list()
for i in range(self.image_classes):
left_right_tuples.append(framwork.lamb_func(logit_list[i], indict_logic_list[i], lamb=lamb_list[i]))
# func = framwork.lamb_func()
# left_right_tuples = map(func, logit_list, indict_logic_list, lamb_list)
net_tensors.update({'left_right_tuples': left_right_tuples, 'logits': logits, 'labels': labels,
'lambs': lambs})
def build_loss(self, logits, labels, lambs):
# put a sigfunction on logits and then transpose
logits = tf.transpose(framwork.sig_func(logits))
# according to the labels, erase rows which is not in labels
labels_unique = tf.constant(range(self.image_classes), dtype=tf.int32)
labels_num = self.image_classes
logits = tf.gather(logits, indices=labels_unique)
lambs = tf.gather(lambs, indices=labels_unique)
# set the value of each row to True when it occurs in labels
template = tf.tile(tf.expand_dims(labels_unique, dim=1), [1, self.batch_size])
labels_expand = tf.tile(tf.expand_dims(labels, dim=0), [labels_num, 1])
indict_logic = tf.equal(labels_expand, template)
# split the tensor along rows
logit_list = tf.split(0, labels_num, logits)
indict_logic_list = tf.split(0, labels_num, indict_logic)
lambda_list = tf.split(0, self.image_classes, lambs)
# loss_list = list()
# for i in range(self.image_classes):
# loss_list.append(framwork.loss_func(logit_list[i], indict_logic_list[i], lambda_list[i]))
loss_list = map(framwork.loss_func, logit_list, indict_logic_list, lambda_list)
loss = tf.add_n(loss_list)
tensors_dict = {'labels_unique': labels_unique, 'template': template, 'logits_sig_trans': logits,
'loss': loss, 'indict_logic': indict_logic}
self.tensors_names.extend(tensors_dict.keys())
self.net_tensors.update(tensors_dict)
def call(self, x, h):
channels = x.shape[self._feature_axis].value
with tf.variable_scope('gates'):
inputs = tf.concat([x, h], axis=self._feature_axis)
n = channels + self._filters
m = 2 * self._filters if self._filters > 1 else 2
W = tf.get_variable('kernel', self._kernel + [n, m])
y = tf.nn.convolution(inputs, W, 'SAME', data_format=self._data_format)
if self._normalize:
r, u = tf.split(y, 2, axis=self._feature_axis)
r = tf.contrib.layers.layer_norm(r)
u = tf.contrib.layers.layer_norm(u)
else:
y += tf.get_variable('bias', [m], initializer=tf.ones_initializer())
r, u = tf.split(y, 2, axis=self._feature_axis)
r, u = tf.sigmoid(r), tf.sigmoid(u)
# TODO
#tf.summary.histogram('reset_gate', r)
#tf.summary.histogram('update_gate', u)
with tf.variable_scope('candidate'):
inputs = tf.concat([x, r * h], axis=self._feature_axis)
n = channels + self._filters
m = self._filters
W = tf.get_variable('kernel', self._kernel + [n, m])
y = tf.nn.convolution(inputs, W, 'SAME', data_format=self._data_format)
if self._normalize:
y = tf.contrib.layers.layer_norm(y)
else:
y += tf.get_variable('bias', [m], initializer=tf.zeros_initializer())
h = u * h + (1 - u) * self._activation(y)
return h, h
def __call__(self, inputs, state, scope=None):
"""Long short-term memory cell (LSTM)."""
with tf.variable_scope(scope or type(self).__name__): # "BasicLSTMCell"
# Parameters of gates are concatenated into one multiply for efficiency.
c, h = tf.split(1, 2, state)
concat = linear.linear([inputs, h], 4 * self._num_units, True)
fs = []
# This can be made more efficient since we're doing more than needs to be
# done, but for now w/e
for child_state in child_states:
c_k, h_k = tf.split(1, 2, child_state)
concat = linear.linear([inputs, h_k], 4 * self._num_units, True)
i_k, j_k, f_k, o_k = tf.split(1, 4, concat)
fs.append(f_k)
# i = input_gate, j = new_input, f = forget_gate, o = output_gate
# TODO: forget gate for each child, probably need to split by number
# of child states or something
i, j, f, o = tf.split(1, 4, concat)
# If no children just treat it like a regular lstm
if not fs:
fs.append(f)
new_c = sum(c * tf.sigmoid(fs + self._forget_bias)) + tf.sigmoid(i) * tf.tanh(j)
new_h = tf.tanh(new_c) * tf.sigmoid(o)
return new_h, tf.concat(1, [new_c, new_h])
def build_loss(self, out, out_tensor):
"""Build a loss function and accuracy for the model."""
print(' Building loss and accuracy')
with tf.variable_scope('accuracy'):
argmax = tf.to_int32(tf.argmax(out_tensor, 2))
correct = tf.to_float(tf.equal(argmax, self.ts)) * self.t_mask
accuracy = tf.reduce_sum(correct) / tf.reduce_sum(self.t_mask)
with tf.variable_scope('loss'):
with tf.variable_scope('split_t_and_mask'):
split_kwargs = { 'split_dim': 1,
'num_split': self.max_t_seq_len }
ts = tf.split(value=self.ts, **split_kwargs)
t_mask = tf.split(value=self.t_mask, **split_kwargs)
t_mask = [tf.squeeze(weight) for weight in t_mask]
loss = seq2seq.sequence_loss(out, ts, t_mask,
self.max_t_seq_len)
with tf.variable_scope('regularization'):
regularize = tf.contrib.layers.l2_regularizer(self.reg_scale)
params = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)
reg_term = sum([regularize(param) for param in params])
loss += reg_term
return loss, accuracy
def _build_base_rnn(self, inputs, input_seq_lengths, forward_only=True):
"""
Build the Language RNN
Parameters
----------
:param inputs: inputs to the RNN
:param input_seq_lengths: vector containing the length of each input from 'inputs'
:param forward_only: whether the RNN will be used for training or not (if true then add a dropout layer)
Returns
----------
:returns logits: each char probability for each timestep of the input, for each item of the batch
:returns prediction: the best prediction for the input
:returns rnn_keep_state_op: a tensorflow op to save the RNN internal state for the next batch
:returns rnn_state_zero_op: a tensorflow op to reset the RNN internal state to zeros
:returns input_keep_prob_ph: a placeholder for input_keep_prob of the dropout layer
(None if forward_only is True)
:returns output_keep_prob_ph: a placeholder for output_keep_prob of the dropout layer
(None if forward_only is True)
:returns rnn_tuple_state: the RNN internal state
"""
# Define a variable to keep track of the learning process step
global_step = tf.Variable(0, trainable=False, name='global_step')
# If building the RNN for training then create dropout rate placeholders
input_keep_prob_ph = output_keep_prob_ph = None
if not forward_only:
with tf.name_scope('dropout'):
# Create placeholders, used to override values when running on the test set
input_keep_prob_ph = tf.placeholder(tf.float32)
output_keep_prob_ph = tf.placeholder(tf.float32)
# Define cells of language model
with tf.variable_scope('LSTM'):
# Create each layer
layers_list = []
for _ in range(self.num_layers):
cell = tf.contrib.rnn.BasicLSTMCell(self.hidden_size,
state_is_tuple=True)
# If building the RNN for training then add a dropoutWrapper to the cells
if not forward_only:
with tf.name_scope('dropout'):
cell = tf.contrib.rnn.DropoutWrapper(
cell,
input_keep_prob=input_keep_prob_ph,
output_keep_prob=output_keep_prob_ph)
layers_list.append(cell)
# Store the layers in a multi-layer RNN
cell = tf.contrib.rnn.MultiRNNCell(layers_list,
state_is_tuple=True)
# Build the input layer between input and the RNN
with tf.variable_scope('Input_Layer'):
w_i = tf.get_variable(
"input_w", [self.input_dim, self.hidden_size],
tf.float32,
initializer=tf.contrib.layers.xavier_initializer())
b_i = tf.get_variable("input_b", [self.hidden_size],
tf.float32,
initializer=tf.constant_initializer(0.0))
# Apply the input layer to the network input to produce the input for the rnn part of the network
rnn_inputs = [
tf.matmul(tf.squeeze(tf.cast(i, tf.float32), axis=[0]), w_i) + b_i
for i in tf.split(axis=0,
num_or_size_splits=self.max_input_seq_length,
value=inputs)
]
# Switch from a list to a tensor
rnn_inputs = tf.stack(rnn_inputs)
# Define some variables to store the RNN state
# Note : tensorflow keep the state inside a batch but it's necessary to do this in order to keep the state
# between batches, especially when doing live transcript
# Another way would have been to get the state as an output of the session and feed it every time but
# this way is much more efficient
with tf.variable_scope('Hidden_state'):
state_variables = []
for state_c, state_h in cell.zero_state(self.batch_size,
tf.float32):
state_variables.append(
tf.nn.rnn_cell.LSTMStateTuple(
tf.Variable(state_c, trainable=False),
tf.Variable(state_h, trainable=False)))
# Return as a tuple, so that it can be fed to dynamic_rnn as an initial state
rnn_tuple_state = tuple(state_variables)
# Build the RNN
with tf.name_scope('LSTM'):
rnn_output, new_states = tf.nn.dynamic_rnn(
cell,
rnn_inputs,
sequence_length=input_seq_lengths,
initial_state=rnn_tuple_state,
time_major=True)
# Define an op to keep the hidden state between batches
update_ops = []
for state_variable, new_state in zip(rnn_tuple_state, new_states):
# Assign the new state to the state variables on this layer
update_ops.extend([
state_variable[0].assign(new_state[0]),
state_variable[1].assign(new_state[1])
])
# Return a tuple in order to combine all update_ops into a single operation.
# The tuple's actual value should not be used.
rnn_keep_state_op = tf.tuple(update_ops)
# Define an op to reset the hidden state to zeros
update_ops = []
for state_variable in rnn_tuple_state:
# Assign the new state to the state variables on this layer
update_ops.extend([
state_variable[0].assign(tf.zeros_like(state_variable[0])),
state_variable[1].assign(tf.zeros_like(state_variable[1]))
])
# Return a tuple in order to combine all update_ops into a single operation.
# The tuple's actual value should not be used.
rnn_state_zero_op = tf.tuple(update_ops)
# Build the output layer between the RNN and the char_map
with tf.variable_scope('Output_layer'):
w_o = tf.get_variable(
"output_w", [self.hidden_size, self.num_labels],
tf.float32,
initializer=tf.contrib.layers.xavier_initializer())
b_o = tf.get_variable("output_b", [self.num_labels],
tf.float32,
initializer=tf.constant_initializer(0.0))
# Compute the logits (each char probability for each timestep of the input, for each item of the batch)
logits = tf.stack([
tf.matmul(tf.squeeze(i, axis=[0]), w_o) + b_o
for i in tf.split(axis=0,
num_or_size_splits=self.max_input_seq_length,
value=rnn_output)
])
# Compute the prediction which is the best "path" of probabilities for each item of the batch
decoded, _log_prob = tf.nn.ctc_beam_search_decoder(
logits, input_seq_lengths)
# Set the RNN result to the best path found
prediction = tf.to_int32(decoded[0])
return global_step, logits, prediction, rnn_keep_state_op, rnn_state_zero_op, \
input_keep_prob_ph, output_keep_prob_ph, rnn_tuple_state
def learn(make_env, make_policy, *,
n_episodes,
horizon,
delta,
gamma,
max_iters,
sampler=None,
use_natural_gradient=False, #can be 'exact', 'approximate'
fisher_reg=1e-2,
iw_method='is',
iw_norm='none',
bound='J',
line_search_type='parabola',
save_weights=False,
improvement_tol=0.,
center_return=False,
render_after=None,
max_offline_iters=100,
callback=None,
clipping=False,
entropy='none',
positive_return=False,
reward_clustering='none'):
np.set_printoptions(precision=3)
max_samples = horizon * n_episodes
if line_search_type == 'binary':
line_search = line_search_binary
elif line_search_type == 'parabola':
line_search = line_search_parabola
else:
raise ValueError()
# Building the environment
env = make_env()
ob_space = env.observation_space
ac_space = env.action_space
# Building the policy
pi = make_policy('pi', ob_space, ac_space)
oldpi = make_policy('oldpi', ob_space, ac_space)
all_var_list = pi.get_trainable_variables()
var_list = [v for v in all_var_list if v.name.split('/')[1].startswith('pol')]
shapes = [U.intprod(var.get_shape().as_list()) for var in var_list]
n_parameters = sum(shapes)
# Placeholders
ob_ = ob = U.get_placeholder_cached(name='ob')
ac_ = pi.pdtype.sample_placeholder([max_samples], name='ac')
mask_ = tf.placeholder(dtype=tf.float32, shape=(max_samples), name='mask')
rew_ = tf.placeholder(dtype=tf.float32, shape=(max_samples), name='rew')
disc_rew_ = tf.placeholder(dtype=tf.float32, shape=(max_samples), name='disc_rew')
clustered_rew_ = tf.placeholder(dtype=tf.float32, shape=(n_episodes))
gradient_ = tf.placeholder(dtype=tf.float32, shape=(n_parameters, 1), name='gradient')
iter_number_ = tf.placeholder(dtype=tf.int32, name='iter_number')
losses_with_name = []
# Policy densities
target_log_pdf = pi.pd.logp(ac_)
behavioral_log_pdf = oldpi.pd.logp(ac_)
log_ratio = target_log_pdf - behavioral_log_pdf
# Split operations
disc_rew_split = tf.stack(tf.split(disc_rew_ * mask_, n_episodes))
rew_split = tf.stack(tf.split(rew_ * mask_, n_episodes))
log_ratio_split = tf.stack(tf.split(log_ratio * mask_, n_episodes))
target_log_pdf_split = tf.stack(tf.split(target_log_pdf * mask_, n_episodes))
behavioral_log_pdf_split = tf.stack(tf.split(behavioral_log_pdf * mask_, n_episodes))
mask_split = tf.stack(tf.split(mask_, n_episodes))
# Renyi divergence
emp_d2_split = tf.stack(tf.split(pi.pd.renyi(oldpi.pd, 2) * mask_, n_episodes))
emp_d2_cum_split = tf.reduce_sum(emp_d2_split, axis=1)
empirical_d2 = tf.reduce_mean(tf.exp(emp_d2_cum_split))
# Return
ep_return = clustered_rew_ #tf.reduce_sum(mask_split * disc_rew_split, axis=1)
if clipping:
rew_split = tf.clip_by_value(rew_split, -1, 1)
if center_return:
ep_return = ep_return - tf.reduce_mean(ep_return)
rew_split = rew_split - (tf.reduce_sum(rew_split) / (tf.reduce_sum(mask_split) + 1e-24))
discounter = [pow(gamma, i) for i in range(0, horizon)] # Decreasing gamma
discounter_tf = tf.constant(discounter)
disc_rew_split = rew_split * discounter_tf
#tf.add_to_collection('prints', tf.Print(ep_return, [ep_return], 'ep_return_not_clustered', summarize=20))
# Reward clustering
'''
rew_clustering_options = reward_clustering.split(':')
if reward_clustering == 'none':
pass # Do nothing
elif rew_clustering_options[0] == 'global':
assert len(rew_clustering_options) == 2, "Reward clustering: Provide the correct number of parameters"
N = int(rew_clustering_options[1])
tf.add_to_collection('prints', tf.Print(ep_return, [ep_return], 'ep_return', summarize=20))
global_rew_min = tf.Variable(float('+inf'), trainable=False)
global_rew_max = tf.Variable(float('-inf'), trainable=False)
rew_min = tf.reduce_min(ep_return)
rew_max = tf.reduce_max(ep_return)
global_rew_min = tf.assign(global_rew_min, tf.minimum(global_rew_min, rew_min))
global_rew_max = tf.assign(global_rew_max, tf.maximum(global_rew_max, rew_max))
interval_size = (global_rew_max - global_rew_min) / N
ep_return = tf.floordiv(ep_return, interval_size) * interval_size
elif rew_clustering_options[0] == 'batch':
assert len(rew_clustering_options) == 2, "Reward clustering: Provide the correct number of parameters"
N = int(rew_clustering_options[1])
rew_min = tf.reduce_min(ep_return)
rew_max = tf.reduce_max(ep_return)
interval_size = (rew_max - rew_min) / N
ep_return = tf.floordiv(ep_return, interval_size) * interval_size
elif rew_clustering_options[0] == 'manual':
assert len(rew_clustering_options) == 4, "Reward clustering: Provide the correct number of parameters"
N, rew_min, rew_max = map(int, rew_clustering_options[1:])
print("N:", N)
print("Min reward:", rew_min)
print("Max reward:", rew_max)
interval_size = (rew_max - rew_min) / N
print("Interval size:", interval_size)
# Clip to avoid overflow and cluster
ep_return = tf.clip_by_value(ep_return, rew_min, rew_max)
ep_return = tf.cast(tf.floordiv(ep_return, interval_size) * interval_size, tf.float32)
tf.add_to_collection('prints', tf.Print(ep_return, [ep_return], 'ep_return_clustered', summarize=20))
else:
raise Exception('Unrecognized reward clustering scheme.')
'''
return_mean = tf.reduce_mean(ep_return)
return_std = U.reduce_std(ep_return)
return_max = tf.reduce_max(ep_return)
return_min = tf.reduce_min(ep_return)
return_abs_max = tf.reduce_max(tf.abs(ep_return))
return_step_max = tf.reduce_max(tf.abs(rew_split)) # Max step reward
return_step_mean = tf.abs(tf.reduce_mean(rew_split))
positive_step_return_max = tf.maximum(0.0, tf.reduce_max(rew_split))
negative_step_return_max = tf.maximum(0.0, tf.reduce_max(-rew_split))
return_step_maxmin = tf.abs(positive_step_return_max - negative_step_return_max)
losses_with_name.extend([(return_mean, 'InitialReturnMean'),
(return_max, 'InitialReturnMax'),
(return_min, 'InitialReturnMin'),
(return_std, 'InitialReturnStd'),
(empirical_d2, 'EmpiricalD2'),
(return_step_max, 'ReturnStepMax'),
(return_step_maxmin, 'ReturnStepMaxmin')])
if iw_method == 'pdis':
# log_ratio_split cumulative sum
log_ratio_cumsum = tf.cumsum(log_ratio_split, axis=1)
# Exponentiate
ratio_cumsum = tf.exp(log_ratio_cumsum)
# Multiply by the step-wise reward (not episode)
ratio_reward = ratio_cumsum * disc_rew_split
# Average on episodes
ratio_reward_per_episode = tf.reduce_sum(ratio_reward, axis=1)
w_return_mean = tf.reduce_sum(ratio_reward_per_episode, axis=0) / n_episodes
# Get d2(w0:t) with mask
d2_w_0t = tf.exp(tf.cumsum(emp_d2_split, axis=1)) * mask_split # LEAVE THIS OUTSIDE
# Sum d2(w0:t) over timesteps
episode_d2_0t = tf.reduce_sum(d2_w_0t, axis=1)
# Sample variance
J_sample_variance = (1/(n_episodes-1)) * tf.reduce_sum(tf.square(ratio_reward_per_episode - w_return_mean))
losses_with_name.append((J_sample_variance, 'J_sample_variance'))
losses_with_name.extend([(tf.reduce_max(ratio_cumsum), 'MaxIW'),
(tf.reduce_min(ratio_cumsum), 'MinIW'),
(tf.reduce_mean(ratio_cumsum), 'MeanIW'),
(U.reduce_std(ratio_cumsum), 'StdIW')])
losses_with_name.extend([(tf.reduce_max(d2_w_0t), 'MaxD2w0t'),
(tf.reduce_min(d2_w_0t), 'MinD2w0t'),
(tf.reduce_mean(d2_w_0t), 'MeanD2w0t'),
(U.reduce_std(d2_w_0t), 'StdD2w0t')])
elif iw_method == 'is':
iw = tf.exp(tf.reduce_sum(log_ratio_split, axis=1))
if iw_norm == 'none':
iwn = iw / n_episodes
w_return_mean = tf.reduce_sum(iwn * ep_return)
J_sample_variance = (1/(n_episodes-1)) * tf.reduce_sum(tf.square(iw * ep_return - w_return_mean))
losses_with_name.append((J_sample_variance, 'J_sample_variance'))
elif iw_norm == 'sn':
iwn = iw / tf.reduce_sum(iw)
w_return_mean = tf.reduce_sum(iwn * ep_return)
elif iw_norm == 'regression':
iwn = iw / n_episodes
mean_iw = tf.reduce_mean(iw)
beta = tf.reduce_sum((iw - mean_iw) * ep_return * iw) / (tf.reduce_sum((iw - mean_iw) ** 2) + 1e-24)
w_return_mean = tf.reduce_mean(iw * ep_return - beta * (iw - 1))
else:
raise NotImplementedError()
ess_classic = tf.linalg.norm(iw, 1) ** 2 / tf.linalg.norm(iw, 2) ** 2
sqrt_ess_classic = tf.linalg.norm(iw, 1) / tf.linalg.norm(iw, 2)
ess_renyi = n_episodes / empirical_d2
losses_with_name.extend([(tf.reduce_max(iwn), 'MaxIWNorm'),
(tf.reduce_min(iwn), 'MinIWNorm'),
(tf.reduce_mean(iwn), 'MeanIWNorm'),
(U.reduce_std(iwn), 'StdIWNorm'),
(tf.reduce_max(iw), 'MaxIW'),
(tf.reduce_min(iw), 'MinIW'),
(tf.reduce_mean(iw), 'MeanIW'),
(U.reduce_std(iw), 'StdIW'),
(ess_classic, 'ESSClassic'),
(ess_renyi, 'ESSRenyi')])
elif iw_method == 'rbis':
# Get pdfs for episodes
target_log_pdf_episode = tf.reduce_sum(target_log_pdf_split, axis=1)
behavioral_log_pdf_episode = tf.reduce_sum(behavioral_log_pdf_split, axis=1)
# Normalize log_proba (avoid as overflows as possible)
normalization_factor = tf.reduce_mean(tf.stack([target_log_pdf_episode, behavioral_log_pdf_episode]))
target_norm_log_pdf_episode = target_log_pdf_episode - normalization_factor
behavioral_norm_log_pdf_episode = behavioral_log_pdf_episode - normalization_factor
# Exponentiate
target_pdf_episode = tf.clip_by_value(tf.cast(tf.exp(target_norm_log_pdf_episode), tf.float64), 1e-300, 1e+300)
behavioral_pdf_episode = tf.clip_by_value(tf.cast(tf.exp(behavioral_norm_log_pdf_episode), tf.float64), 1e-300, 1e+300)
tf.add_to_collection('asserts', tf.assert_positive(target_pdf_episode, name='target_pdf_positive'))
tf.add_to_collection('asserts', tf.assert_positive(behavioral_pdf_episode, name='behavioral_pdf_positive'))
# Compute the merging matrix (reward-clustering) and the number of clusters
reward_unique, reward_indexes = tf.unique(ep_return)
episode_clustering_matrix = tf.cast(tf.one_hot(reward_indexes, n_episodes), tf.float64)
max_index = tf.reduce_max(reward_indexes) + 1
trajectories_per_cluster = tf.reduce_sum(episode_clustering_matrix, axis=0)[:max_index]
tf.add_to_collection('asserts', tf.assert_positive(tf.reduce_sum(episode_clustering_matrix, axis=0)[:max_index], name='clustering_matrix'))
# Get the clustered pdfs
clustered_target_pdf = tf.matmul(tf.reshape(target_pdf_episode, (1, -1)), episode_clustering_matrix)[0][:max_index]
clustered_behavioral_pdf = tf.matmul(tf.reshape(behavioral_pdf_episode, (1, -1)), episode_clustering_matrix)[0][:max_index]
tf.add_to_collection('asserts', tf.assert_positive(clustered_target_pdf, name='clust_target_pdf_positive'))
tf.add_to_collection('asserts', tf.assert_positive(clustered_behavioral_pdf, name='clust_behavioral_pdf_positive'))
# Compute the J
ratio_clustered = clustered_target_pdf / clustered_behavioral_pdf
#ratio_reward = tf.cast(ratio_clustered, tf.float32) * reward_unique # ---- No cluster cardinality
ratio_reward = tf.cast(ratio_clustered, tf.float32) * reward_unique * tf.cast(trajectories_per_cluster, tf.float32) # ---- Cluster cardinality
#w_return_mean = tf.reduce_sum(ratio_reward) / tf.cast(max_index, tf.float32) # ---- No cluster cardinality
w_return_mean = tf.reduce_sum(ratio_reward) / tf.cast(n_episodes, tf.float32) # ---- Cluster cardinality
# Divergences
ess_classic = tf.linalg.norm(ratio_reward, 1) ** 2 / tf.linalg.norm(ratio_reward, 2) ** 2
sqrt_ess_classic = tf.linalg.norm(ratio_reward, 1) / tf.linalg.norm(ratio_reward, 2)
ess_renyi = n_episodes / empirical_d2
# Summaries
losses_with_name.extend([(tf.reduce_max(ratio_clustered), 'MaxIW'),
(tf.reduce_min(ratio_clustered), 'MinIW'),
(tf.reduce_mean(ratio_clustered), 'MeanIW'),
(U.reduce_std(ratio_clustered), 'StdIW'),
(1-(max_index / n_episodes), 'RewardCompression'),
(ess_classic, 'ESSClassic'),
(ess_renyi, 'ESSRenyi')])
else:
raise NotImplementedError()
if bound == 'J':
bound_ = w_return_mean
elif bound == 'std-d2':
bound_ = w_return_mean - tf.sqrt((1 - delta) / (delta * ess_renyi)) * return_std
elif bound == 'max-d2':
var_estimate = tf.sqrt((1 - delta) / (delta * ess_renyi)) * return_abs_max
bound_ = w_return_mean - tf.sqrt((1 - delta) / (delta * ess_renyi)) * return_abs_max
elif bound == 'max-ess':
bound_ = w_return_mean - tf.sqrt((1 - delta) / delta) / sqrt_ess_classic * return_abs_max
elif bound == 'std-ess':
bound_ = w_return_mean - tf.sqrt((1 - delta) / delta) / sqrt_ess_classic * return_std
elif bound == 'pdis-max-d2':
# Discount factor
if gamma >= 1:
discounter = [float(1+2*(horizon-t-1)) for t in range(0, horizon)]
else:
def f(t):
return pow(gamma, 2*t) + (2*pow(gamma,t)*(pow(gamma, t+1) - pow(gamma, horizon))) / (1-gamma)
discounter = [f(t) for t in range(0, horizon)]
discounter_tf = tf.constant(discounter)
mean_episode_d2 = tf.reduce_sum(d2_w_0t, axis=0) / (tf.reduce_sum(mask_split, axis=0) + 1e-24)
discounted_d2 = mean_episode_d2 * discounter_tf # Discounted d2
discounted_total_d2 = tf.reduce_sum(discounted_d2, axis=0) # Sum over time
bound_ = w_return_mean - tf.sqrt((1-delta) * discounted_total_d2 / (delta*n_episodes)) * return_step_max
elif bound == 'pdis-mean-d2':
# Discount factor
if gamma >= 1:
discounter = [float(1+2*(horizon-t-1)) for t in range(0, horizon)]
else:
def f(t):
return pow(gamma, 2*t) + (2*pow(gamma,t)*(pow(gamma, t+1) - pow(gamma, horizon))) / (1-gamma)
discounter = [f(t) for t in range(0, horizon)]
discounter_tf = tf.constant(discounter)
mean_episode_d2 = tf.reduce_sum(d2_w_0t, axis=0) / (tf.reduce_sum(mask_split, axis=0) + 1e-24)
discounted_d2 = mean_episode_d2 * discounter_tf # Discounted d2
discounted_total_d2 = tf.reduce_sum(discounted_d2, axis=0) # Sum over time
bound_ = w_return_mean - tf.sqrt((1-delta) * discounted_total_d2 / (delta*n_episodes)) * return_step_mean
else:
raise NotImplementedError()
# Policy entropy for exploration
ent = pi.pd.entropy()
meanent = tf.reduce_mean(ent)
losses_with_name.append((meanent, 'MeanEntropy'))
# Add policy entropy bonus
if entropy != 'none':
scheme, v1, v2 = entropy.split(':')
if scheme == 'step':
entcoeff = tf.cond(iter_number_ < int(v2), lambda: float(v1), lambda: float(0.0))
losses_with_name.append((entcoeff, 'EntropyCoefficient'))
entbonus = entcoeff * meanent
bound_ = bound_ + entbonus
elif scheme == 'lin':
ip = tf.cast(iter_number_ / max_iters, tf.float32)
entcoeff_decay = tf.maximum(0.0, float(v2) + (float(v1) - float(v2)) * (1.0 - ip))
losses_with_name.append((entcoeff_decay, 'EntropyCoefficient'))
entbonus = entcoeff_decay * meanent
bound_ = bound_ + entbonus
elif scheme == 'exp':
ent_f = tf.exp(-tf.abs(tf.reduce_mean(iw) - 1) * float(v2)) * float(v1)
losses_with_name.append((ent_f, 'EntropyCoefficient'))
bound_ = bound_ + ent_f * meanent
else:
raise Exception('Unrecognized entropy scheme.')
losses_with_name.append((w_return_mean, 'ReturnMeanIW'))
losses_with_name.append((bound_, 'Bound'))
losses, loss_names = map(list, zip(*losses_with_name))
if use_natural_gradient:
p = tf.placeholder(dtype=tf.float32, shape=[None])
target_logpdf_episode = tf.reduce_sum(target_log_pdf_split * mask_split, axis=1)
grad_logprob = U.flatgrad(tf.stop_gradient(iwn) * target_logpdf_episode, var_list)
dot_product = tf.reduce_sum(grad_logprob * p)
hess_logprob = U.flatgrad(dot_product, var_list)
compute_linear_operator = U.function([p, ob_, ac_, disc_rew_, mask_], [-hess_logprob])
assign_old_eq_new = U.function([], [], updates=[tf.assign(oldv, newv)
for (oldv, newv) in zipsame(oldpi.get_variables(), pi.get_variables())])
assert_ops = tf.group(*tf.get_collection('asserts'))
print_ops = tf.group(*tf.get_collection('prints'))
compute_lossandgrad = U.function([ob_, ac_, rew_, disc_rew_, clustered_rew_, mask_, iter_number_], losses + [U.flatgrad(bound_, var_list), assert_ops, print_ops])
compute_grad = U.function([ob_, ac_, rew_, disc_rew_, clustered_rew_, mask_, iter_number_], [U.flatgrad(bound_, var_list), assert_ops, print_ops])
compute_bound = U.function([ob_, ac_, rew_, disc_rew_, clustered_rew_, mask_, iter_number_], [bound_, assert_ops, print_ops])
compute_losses = U.function([ob_, ac_, rew_, disc_rew_, clustered_rew_, mask_, iter_number_], losses)
#compute_temp = U.function([ob_, ac_, rew_, disc_rew_, mask_], [ratio_cumsum, discounted_ratio])
set_parameter = U.SetFromFlat(var_list)
get_parameter = U.GetFlat(var_list)
if sampler is None:
seg_gen = traj_segment_generator(pi, env, n_episodes, horizon, stochastic=True)
sampler = type("SequentialSampler", (object,), {"collect": lambda self, _: seg_gen.__next__()})()
U.initialize()
# Starting optimizing
episodes_so_far = 0
timesteps_so_far = 0
iters_so_far = 0
tstart = time.time()
lenbuffer = deque(maxlen=n_episodes)
rewbuffer = deque(maxlen=n_episodes)
while True:
iters_so_far += 1
if render_after is not None and iters_so_far % render_after == 0:
if hasattr(env, 'render'):
render(env, pi, horizon)
if callback:
callback(locals(), globals())
if iters_so_far >= max_iters:
print('Finised...')
break
logger.log('********** Iteration %i ************' % iters_so_far)
theta = get_parameter()
with timed('sampling'):
seg = sampler.collect(theta)
add_disc_rew(seg, gamma)
lens, rets = seg['ep_lens'], seg['ep_rets']
lenbuffer.extend(lens)
rewbuffer.extend(rets)
episodes_so_far += len(lens)
timesteps_so_far += sum(lens)
# Get clustered reward
reward_matrix = np.reshape(seg['disc_rew'] * seg['mask'], (n_episodes, horizon))
ep_reward = np.sum(reward_matrix, axis=1)
if reward_clustering == 'none':
pass
elif reward_clustering == 'floor':
ep_reward = np.floor(ep_reward)
elif reward_clustering == 'ceil':
ep_reward = np.ceil(ep_reward)
elif reward_clustering == 'floor10':
ep_reward = np.floor(ep_reward * 10) / 10
elif reward_clustering == 'ceil10':
ep_reward = np.ceil(ep_reward * 10) / 10
elif reward_clustering == 'floor100':
ep_reward = np.floor(ep_reward * 100) / 100
elif reward_clustering == 'ceil100':
ep_reward = np.ceil(ep_reward * 100) / 100
args = ob, ac, rew, disc_rew, clustered_rew, mask, iter_number = seg['ob'], seg['ac'], seg['rew'], seg['disc_rew'], ep_reward, seg['mask'], iters_so_far
assign_old_eq_new()
def evaluate_loss():
loss = compute_bound(*args)
return loss[0]
def evaluate_gradient():
gradient = compute_grad(*args)
return gradient[0]
if use_natural_gradient:
def evaluate_fisher_vector_prod(x):
return compute_linear_operator(x, *args)[0] + fisher_reg * x
def evaluate_natural_gradient(g):
return cg(evaluate_fisher_vector_prod, g, cg_iters=10, verbose=0)
else:
evaluate_natural_gradient = None
with timed('summaries before'):
logger.record_tabular("Iteration", iters_so_far)
logger.record_tabular("InitialBound", evaluate_loss())
logger.record_tabular("EpLenMean", np.mean(lenbuffer))
logger.record_tabular("EpRewMean", np.mean(rewbuffer))
logger.record_tabular("EpThisIter", len(lens))
logger.record_tabular("EpisodesSoFar", episodes_so_far)
logger.record_tabular("TimestepsSoFar", timesteps_so_far)
logger.record_tabular("TimeElapsed", time.time() - tstart)
if save_weights:
logger.record_tabular('Weights', str(get_parameter()))
import pickle
file = open('checkpoint.pkl', 'wb')
pickle.dump(theta, file)
with timed("offline optimization"):
theta, improvement = optimize_offline(theta,
set_parameter,
line_search,
evaluate_loss,
evaluate_gradient,
evaluate_natural_gradient,
max_offline_ite=max_offline_iters)
set_parameter(theta)
with timed('summaries after'):
meanlosses = np.array(compute_losses(*args))
for (lossname, lossval) in zip(loss_names, meanlosses):
logger.record_tabular(lossname, lossval)
logger.dump_tabular()
env.close()
def transpose_coordinates(self):
"""Transpose the coordinate representation in a boxlist."""
with tf.name_scope('transpose_coordinates'):
y_min, x_min, y_max, x_max = tf.split(
value=self.get(), num_or_size_splits=4, axis=1)
self.set(tf.concat([x_min, y_min, x_max, y_max], 1))
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--filelist',
'-t',
help='Path to training set ground truth (.txt)',
required=True)
parser.add_argument('--filelist_val',
'-v',
help='Path to validation set ground truth (.txt)',
required=True)
parser.add_argument('--load_ckpt',
'-l',
help='Path to a check point file for load')
parser.add_argument(
'--save_folder',
'-s',
help='Path to folder for saving check points and summary',
required=True)
parser.add_argument('--model', '-m', help='Model to use', required=True)
parser.add_argument('--setting',
'-x',
help='Setting to use',
required=True)
args = parser.parse_args()
time_string = datetime.now().strftime('%Y-%m-%d-%H-%M-%S')
root_folder = os.path.join(
args.save_folder,
'%s_%s_%d_%s' % (args.model, args.setting, os.getpid(), time_string))
if not os.path.exists(root_folder):
os.makedirs(root_folder)
sys.stdout = open(os.path.join(root_folder, 'log.txt'), 'w')
print('PID:', os.getpid())
print(args)
model = importlib.import_module(args.model)
setting_path = os.path.join(os.path.dirname(__file__), args.model)
sys.path.append(setting_path)
setting = importlib.import_module(args.setting)
num_epochs = setting.num_epochs
batch_size = setting.batch_size
sample_num = setting.sample_num
step_val = 500
num_parts = setting.num_parts
label_weights_list = setting.label_weights
scaling_range = setting.scaling_range
scaling_range_val = setting.scaling_range_val
jitter = setting.jitter
jitter_val = setting.jitter_val
# Prepare inputs
print('{}-Preparing datasets...'.format(datetime.now()))
data_train, _, data_num_train, label_train = data_utils.load_seg(
args.filelist)
data_val, _, data_num_val, label_val = data_utils.load_seg(
args.filelist_val)
# shuffle
data_train, data_num_train, label_train = \
data_utils.grouped_shuffle([data_train, data_num_train, label_train])
num_train = data_train.shape[0]
point_num = data_train.shape[1]
num_val = data_val.shape[0]
print('{}-{:d}/{:d} training/validation samples.'.format(
datetime.now(), num_train, num_val))
batch_num = (num_train * num_epochs + batch_size - 1) // batch_size
print('{}-{:d} training batches.'.format(datetime.now(), batch_num))
######################################################################
# Placeholders
indices = tf.placeholder(tf.int32, shape=(None, None, 2), name="indices")
xforms = tf.placeholder(tf.float32, shape=(None, 3, 3), name="xforms")
rotations = tf.placeholder(tf.float32,
shape=(None, 3, 3),
name="rotations")
jitter_range = tf.placeholder(tf.float32, shape=(1), name="jitter_range")
global_step = tf.Variable(0, trainable=False, name='global_step')
is_training = tf.placeholder(tf.bool, name='is_training')
pts_fts = tf.placeholder(tf.float32,
shape=(None, point_num, setting.data_dim),
name='pts_fts')
labels_seg = tf.placeholder(tf.int32,
shape=(None, point_num),
name='labels_seg')
labels_weights = tf.placeholder(tf.float32,
shape=(None, point_num),
name='labels_weights')
######################################################################
features_augmented = None
if setting.data_dim > 3:
points, features = tf.split(pts_fts, [3, setting.data_dim - 3],
axis=-1,
name='split_points_features')
if setting.use_extra_features:
features_sampled = tf.gather_nd(features,
indices=indices,
name='features_sampled')
if setting.with_normal_feature:
if setting.data_dim < 6:
print('Only 3D normals are supported!')
exit()
elif setting.data_dim == 6:
features_augmented = pf.augment(features_sampled,
rotations)
else:
normals, rest = tf.split(features_sampled,
[3, setting.data_dim - 6])
normals_augmented = pf.augment(normals, rotations)
features_augmented = tf.concat([normals_augmented, rest],
axis=-1)
else:
features_augmented = features_sampled
else:
points = pts_fts
points_sampled = tf.gather_nd(points,
indices=indices,
name='points_sampled')
points_augmented = pf.augment(points_sampled, xforms, jitter_range)
labels_sampled = tf.gather_nd(labels_seg,
indices=indices,
name='labels_sampled')
labels_weights_sampled = tf.gather_nd(labels_weights,
indices=indices,
name='labels_weight_sampled')
net = model.Net(points_augmented, features_augmented, num_parts,
is_training, setting)
logits, probs = net.logits, net.probs
loss_op = tf.losses.sparse_softmax_cross_entropy(
labels=labels_sampled, logits=logits, weights=labels_weights_sampled)
t_1_acc_op = pf.top_1_accuracy(probs, labels_sampled)
_ = tf.summary.scalar('loss/train_seg',
tensor=loss_op,
collections=['train'])
_ = tf.summary.scalar('t_1_acc/train_seg',
tensor=t_1_acc_op,
collections=['train'])
loss_val_avg = tf.placeholder(tf.float32)
t_1_acc_val_avg = tf.placeholder(tf.float32)
_ = tf.summary.scalar('loss/val_seg',
tensor=loss_val_avg,
collections=['val'])
_ = tf.summary.scalar('t_1_acc/val_seg',
tensor=t_1_acc_val_avg,
collections=['val'])
lr_exp_op = tf.train.exponential_decay(setting.learning_rate_base,
global_step,
setting.decay_steps,
setting.decay_rate,
staircase=True)
lr_clip_op = tf.maximum(lr_exp_op, setting.learning_rate_min)
_ = tf.summary.scalar('learning_rate',
tensor=lr_clip_op,
collections=['train'])
reg_loss = setting.weight_decay * tf.losses.get_regularization_loss()
if setting.optimizer == 'adam':
optimizer = tf.train.AdamOptimizer(learning_rate=lr_clip_op,
epsilon=setting.epsilon)
elif setting.optimizer == 'momentum':
optimizer = tf.train.MomentumOptimizer(learning_rate=lr_clip_op,
momentum=0.9,
use_nesterov=True)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
train_op = optimizer.minimize(loss_op + reg_loss,
global_step=global_step)
init_op = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
saver = tf.train.Saver(max_to_keep=None)
# backup this file, model and setting
shutil.copy(__file__, os.path.join(root_folder,
os.path.basename(__file__)))
shutil.copy(os.path.join(os.path.dirname(__file__), args.model + '.py'),
os.path.join(root_folder, args.model + '.py'))
if not os.path.exists(os.path.join(root_folder, args.model)):
os.makedirs(os.path.join(root_folder, args.model))
shutil.copy(os.path.join(setting_path, args.setting + '.py'),
os.path.join(root_folder, args.model, args.setting + '.py'))
folder_ckpt = os.path.join(root_folder, 'ckpts')
if not os.path.exists(folder_ckpt):
os.makedirs(folder_ckpt)
folder_summary = os.path.join(root_folder, 'summary')
if not os.path.exists(folder_summary):
os.makedirs(folder_summary)
parameter_num = np.sum(
[np.prod(v.shape.as_list()) for v in tf.trainable_variables()])
print('{}-Parameter number: {:d}.'.format(datetime.now(), parameter_num))
with tf.Session() as sess:
summaries_op = tf.summary.merge_all('train')
summaries_val_op = tf.summary.merge_all('val')
summary_writer = tf.summary.FileWriter(folder_summary, sess.graph)
sess.run(init_op)
# Load the model
if args.load_ckpt is not None:
saver.restore(sess, args.load_ckpt)
print('{}-Checkpoint loaded from {}!'.format(
datetime.now(), args.load_ckpt))
for batch_idx in range(batch_num):
if (batch_idx != 0 and batch_idx % step_val
== 0) or batch_idx == batch_num - 1:
######################################################################
# Validation
filename_ckpt = os.path.join(folder_ckpt, 'iter')
saver.save(sess, filename_ckpt, global_step=global_step)
print('{}-Checkpoint saved to {}!'.format(
datetime.now(), filename_ckpt))
losses = []
t_1_accs = []
for batch_val_idx in range(math.ceil(num_val / batch_size)):
start_idx = batch_size * batch_val_idx
end_idx = min(start_idx + batch_size, num_val)
batch_size_val = end_idx - start_idx
points_batch = data_val[start_idx:end_idx, ...]
points_num_batch = data_num_val[start_idx:end_idx, ...]
labels_batch = label_val[start_idx:end_idx, ...]
weights_batch = np.array(label_weights_list)[label_val[
start_idx:end_idx, ...]]
xforms_np, rotations_np = pf.get_xforms(
batch_size_val, scaling_range=scaling_range_val)
_, loss_val, t_1_acc_val = \
sess.run([update_ops, loss_op, t_1_acc_op],
feed_dict={
pts_fts: points_batch,
indices: pf.get_indices(batch_size_val, sample_num, points_num_batch, False),
xforms: xforms_np,
rotations: rotations_np,
jitter_range: np.array([jitter_val]),
labels_seg: labels_batch,
labels_weights: weights_batch,
is_training: False,
})
losses.append(loss_val * batch_size_val)
t_1_accs.append(t_1_acc_val * batch_size_val)
print(
'{}-[Val ]-Iter: {:06d} Loss: {:.4f} T-1 Acc: {:.4f}'
.format(datetime.now(), batch_val_idx, loss_val,
t_1_acc_val))
sys.stdout.flush()
loss_avg = sum(losses) / num_val
t_1_acc_avg = sum(t_1_accs) / num_val
summaries_val = sess.run(summaries_val_op,
feed_dict={
loss_val_avg: loss_avg,
t_1_acc_val_avg: t_1_acc_avg,
})
summary_writer.add_summary(summaries_val, batch_idx)
print('{}-[Val ]-Average: Loss: {:.4f} T-1 Acc: {:.4f}'.
format(datetime.now(), loss_avg, t_1_acc_avg))
sys.stdout.flush()
######################################################################
######################################################################
# Training
start_idx = (batch_size * batch_idx) % num_train
end_idx = min(start_idx + batch_size, num_train)
batch_size_train = end_idx - start_idx
points_batch = data_train[start_idx:end_idx, ...]
points_num_batch = data_num_train[start_idx:end_idx, ...]
labels_batch = label_train[start_idx:end_idx, ...]
weights_batch = np.array(label_weights_list)[labels_batch]
if start_idx + batch_size_train == num_train:
data_train, data_num_train, label_train = \
data_utils.grouped_shuffle([data_train, data_num_train, label_train])
offset = int(random.gauss(0, sample_num // 8))
offset = max(offset, -sample_num // 4)
offset = min(offset, sample_num // 4)
sample_num_train = sample_num + offset
xforms_np, rotations_np = pf.get_xforms(
batch_size_train, scaling_range=scaling_range)
_, loss, t_1_acc, summaries = \
sess.run([train_op, loss_op, t_1_acc_op, summaries_op],
feed_dict={
pts_fts: points_batch,
indices: pf.get_indices(batch_size_train, sample_num_train, points_num_batch),
xforms: xforms_np,
rotations: rotations_np,
jitter_range: np.array([jitter]),
labels_seg: labels_batch,
labels_weights: weights_batch,
is_training: True,
})
summary_writer.add_summary(summaries, batch_idx)
print('{}-[Train]-Iter: {:06d} Loss: {:.4f} T-1 Acc: {:.4f}'.
format(datetime.now(), batch_idx, loss, t_1_acc))
sys.stdout.flush()
######################################################################
print('{}-Done!'.format(datetime.now()))
def __init__(self,
embedding_mat,
non_static,
hidden_unit,
sequence_length,
max_pool_size,
num_classes,
embedding_size,
filter_sizes,
num_filters,
l2_reg_lambda=0.0):
self.input_x = tf.placeholder(tf.int32, [None, sequence_length],
name='input_x')
self.input_y = tf.placeholder(tf.float32, [None, num_classes],
name='input_y')
self.dropout_keep_prob = tf.placeholder(tf.float32,
name='dropout_keep_prob')
self.batch_size = tf.placeholder(tf.int32, [])
self.pad = tf.placeholder(tf.float32, [None, 1, embedding_size, 1],
name='pad')
self.real_len = tf.placeholder(tf.int32, [None], name='real_len')
l2_loss = tf.constant(0.0)
with tf.device('/cpu:0'), tf.name_scope('embedding'):
if not non_static:
W = tf.constant(embedding_mat, name='W')
else:
W = tf.Variable(embedding_mat, name='W')
self.embedded_chars = tf.nn.embedding_lookup(W, self.input_x)
emb = tf.expand_dims(self.embedded_chars, -1)
pooled_concat = []
reduced = np.int32(np.ceil((sequence_length) * 1.0 / max_pool_size))
for i, filter_size in enumerate(filter_sizes):
with tf.name_scope('conv-maxpool-%s' % filter_size):
# Zero paddings so that the convolution output have dimension batch x sequence_length x emb_size x channel
num_prio = (filter_size - 1) // 2
num_post = (filter_size - 1) - num_prio
pad_prio = tf.concat([self.pad] * num_prio, 1)
pad_post = tf.concat([self.pad] * num_post, 1)
emb_pad = tf.concat([pad_prio, emb, pad_post], 1)
filter_shape = [filter_size, embedding_size, 1, num_filters]
W = tf.Variable(tf.truncated_normal(filter_shape, stddev=0.1),
name='W')
b = tf.Variable(tf.constant(0.1, shape=[num_filters]),
name='b')
conv = tf.nn.conv2d(emb_pad,
W,
strides=[1, 1, 1, 1],
padding='VALID',
name='conv')
h = tf.nn.relu(tf.nn.bias_add(conv, b), name='relu')
# Maxpooling over the outputs
pooled = tf.nn.max_pool(h,
ksize=[1, max_pool_size, 1, 1],
strides=[1, max_pool_size, 1, 1],
padding='SAME',
name='pool')
pooled = tf.reshape(pooled, [-1, reduced, num_filters])
pooled_concat.append(pooled)
pooled_concat = tf.concat(pooled_concat, 2)
pooled_concat = tf.nn.dropout(pooled_concat, self.dropout_keep_prob)
# lstm_cell = tf.nn.rnn_cell.LSTMCell(num_units=hidden_unit)
# lstm_cell = tf.nn.rnn_cell.GRUCell(num_units=hidden_unit)
lstm_cell = tf.contrib.rnn.GRUCell(num_units=hidden_unit)
# lstm_cell = tf.nn.rnn_cell.DropoutWrapper(lstm_cell, output_keep_prob=self.dropout_keep_prob)
lstm_cell = tf.contrib.rnn.DropoutWrapper(
lstm_cell, output_keep_prob=self.dropout_keep_prob)
self._initial_state = lstm_cell.zero_state(self.batch_size, tf.float32)
# inputs = [tf.squeeze(input_, [1]) for input_ in tf.split(1, reduced, pooled_concat)]
inputs = [
tf.squeeze(input_, [1]) for input_ in tf.split(
pooled_concat, num_or_size_splits=int(reduced), axis=1)
]
# outputs, state = tf.nn.rnn(lstm_cell, inputs, initial_state=self._initial_state, sequence_length=self.real_len)
outputs, state = tf.contrib.rnn.static_rnn(
lstm_cell,
inputs,
initial_state=self._initial_state,
sequence_length=self.real_len)
# Collect the appropriate last words into variable output (dimension = batch x embedding_size)
output = outputs[0]
with tf.variable_scope('Output'):
tf.get_variable_scope().reuse_variables()
one = tf.ones([1, hidden_unit], tf.float32)
for i in range(1, len(outputs)):
ind = self.real_len < (i + 1)
ind = tf.to_float(ind)
ind = tf.expand_dims(ind, -1)
mat = tf.matmul(ind, one)
output = tf.add(tf.multiply(output, mat),
tf.multiply(outputs[i], 1.0 - mat))
with tf.name_scope('output'):
self.W = tf.Variable(tf.truncated_normal(
[hidden_unit, num_classes], stddev=0.1),
name='W')
b = tf.Variable(tf.constant(0.1, shape=[num_classes]), name='b')
l2_loss += tf.nn.l2_loss(W)
l2_loss += tf.nn.l2_loss(b)
self.scores = tf.nn.xw_plus_b(output, self.W, b, name='scores')
self.predictions = tf.argmax(self.scores, 1, name='predictions')
with tf.name_scope('loss'):
losses = tf.nn.softmax_cross_entropy_with_logits(
labels=self.input_y,
logits=self.scores) # only named arguments accepted
self.loss = tf.reduce_mean(losses) + l2_reg_lambda * l2_loss
with tf.name_scope('accuracy'):
correct_predictions = tf.equal(self.predictions,
tf.argmax(self.input_y, 1))
self.accuracy = tf.reduce_mean(tf.cast(correct_predictions,
"float"),
name='accuracy')
with tf.name_scope('num_correct'):
correct = tf.equal(self.predictions, tf.argmax(self.input_y, 1))
self.num_correct = tf.reduce_sum(tf.cast(correct, 'float'))
# a specific format. Read more at: https://www.tensorflow.org/api_docs/python/tf/nn/dynamic_rnn
all_inputs = tf.concat([tf.expand_dims(t, 0) for t in train_inputs], axis=0)
# all_outputs is [seq_length, batch_size, num_nodes]
all_lstm_outputs, state = tf.nn.dynamic_rnn(drop_multi_cell,
all_inputs,
initial_state=tuple(initial_state),
time_major=True,
dtype=tf.float32)
all_lstm_outputs = tf.reshape(all_lstm_outputs,
[batch_size * num_unrollings, num_nodes[-1]])
all_outputs = tf.nn.xw_plus_b(all_lstm_outputs, w, b)
split_outputs = tf.split(all_outputs, num_unrollings, axis=0)
# When calculating the loss you need to be careful about the exact form, because you calculate
# loss of all the unrolled steps at the same time
# Therefore, take the mean error or each batch and get the sum of that over all the unrolled steps
print('Defining training Loss')
loss = 0.0
with tf.control_dependencies(
[tf.assign(c[li], state[li][0]) for li in range(n_layers)] +
[tf.assign(h[li], state[li][1]) for li in range(n_layers)]):
for ui in range(num_unrollings):
loss += tf.reduce_mean(0.5 *
(split_outputs[ui] - train_outputs[ui])**2)
print('Learning rate decay operations')
def G_synthesis_co_mod_gan(
dlatents_in, # Input: Disentangled latents (W) [minibatch, num_layers, dlatent_size].
images_in,
masks_in,
dlatent_size=512, # Disentangled latent (W) dimensionality.
num_channels=3, # Number of output color channels.
resolution=1024, # Output resolution.
fmap_base=16 <<
10, # Overall multiplier for the number of feature maps.
fmap_decay=1.0, # log2 feature map reduction when doubling the resolution.
fmap_min=1, # Minimum number of feature maps in any layer.
fmap_max=512, # Maximum number of feature maps in any layer.
randomize_noise=True, # True = randomize noise inputs every time (non-deterministic), False = read noise inputs from variables.
architecture='skip', # Architecture: 'orig', 'skip', 'resnet'.
nonlinearity='lrelu', # Activation function: 'relu', 'lrelu', etc.
dtype='float32', # Data type to use for activations and outputs.
resample_kernel=[
1, 3, 3, 1
], # Low-pass filter to apply when resampling activations. None = no filtering.
fused_modconv=True, # Implement modulated_conv2d_layer() as a single fused op?
is_training=False, # Network is under training? Enables and disables specific features.
pix2pix=False,
dropout_rate=0.5,
cond_mod=True,
style_mod=True,
noise_injection=True,
**_kwargs): # Ignore unrecognized keyword args.
resolution_log2 = int(np.log2(resolution))
assert resolution == 2**resolution_log2 and resolution >= 4
def nf(stage):
return np.clip(int(fmap_base / (2.0**(stage * fmap_decay))), fmap_min,
fmap_max)
assert architecture in ['orig', 'skip', 'resnet']
act = nonlinearity
num_layers = resolution_log2 * 2 - 2
images_out = None
images_in.set_shape([None, num_channels, resolution, resolution])
masks_in.set_shape([None, 1, resolution, resolution])
images_in = tf.cast(images_in, dtype)
masks_in = tf.cast(masks_in, dtype)
def E_fromrgb(x, y, res): # res = 2..resolution_log2
with tf.variable_scope('FromRGB'):
t = apply_bias_act(conv2d_layer(y, fmaps=nf(res - 1), kernel=1),
act=act)
return t if x is None else x + t
def E_block(x, res, E_features): # res = 2..resolution_log2
with tf.variable_scope('Conv0'):
x = apply_bias_act(conv2d_layer(x, fmaps=nf(res - 1), kernel=3),
act=act)
E_features[res] = x
with tf.variable_scope('Conv1_down'):
x = apply_bias_act(conv2d_layer(x,
fmaps=nf(res - 2),
kernel=3,
down=True,
resample_kernel=resample_kernel),
act=act)
return x
# Primary inputs.
dlatents_in.set_shape([None, num_layers, dlatent_size])
dlatents_in = tf.cast(dlatents_in, dtype)
# Noise inputs.
noise_inputs = []
for layer_idx in range(num_layers - 1):
res = (layer_idx + 5) // 2
shape = [1, 1, 2**res, 2**res]
noise_inputs.append(
tf.get_variable('noise%d' % layer_idx,
shape=shape,
initializer=tf.initializers.random_normal(),
trainable=False))
# Main layers.
E_features = {}
x = None
if pix2pix:
num_channels = num_channels // 2
_, y = tf.split(images_in, 2, axis=1)
cond = y
else:
y = tf.concat([masks_in - 0.5, images_in * masks_in], axis=1)
for res in range(resolution_log2, 2, -1):
with tf.variable_scope('E_%dx%d' % (2**res, 2**res)):
if res == resolution_log2:
x = E_fromrgb(x, y, res)
x = E_block(x, res, E_features)
# Final layers.
with tf.variable_scope('E_4x4'):
with tf.variable_scope('Conv'):
x = apply_bias_act(conv2d_layer(x, fmaps=nf(1), kernel=3), act=act)
E_features[2] = x
with tf.variable_scope('Dense0'):
x = apply_bias_act(dense_layer(x, fmaps=nf(1) * 2), act=act)
# if is_training:
x = tf.nn.dropout(x, dropout_rate)
x_global = x
# Single convolution layer with all the bells and whistles.
def layer(x, layer_idx, fmaps, kernel, up=False):
mod_vector = []
if style_mod:
mod_vector.append(dlatents_in[:, layer_idx])
if cond_mod:
mod_vector.append(x_global)
x = modulated_conv2d_layer(
x,
tf.concat(mod_vector, axis=1) if mod_vector else None,
fmaps=fmaps,
kernel=kernel,
up=up,
resample_kernel=resample_kernel,
fused_modconv=fused_modconv)
if noise_injection:
if randomize_noise:
noise = tf.random_normal(
[tf.shape(x)[0], 1, x.shape[2], x.shape[3]], dtype=x.dtype)
else:
noise = tf.cast(noise_inputs[layer_idx], x.dtype)
noise_strength = tf.get_variable(
'noise_strength',
shape=[],
initializer=tf.initializers.zeros())
x += noise * tf.cast(noise_strength, x.dtype)
return apply_bias_act(x, act=act)
# Building blocks for main layers.
def block(x, res, E_features): # res = 3..resolution_log2
x_skip = E_features[res]
t = x
with tf.variable_scope('Conv0_up'):
x = layer(x,
layer_idx=res * 2 - 5,
fmaps=nf(res - 1),
kernel=3,
up=True)
x = x + x_skip
with tf.variable_scope('Conv1'):
x = layer(x, layer_idx=res * 2 - 4, fmaps=nf(res - 1), kernel=3)
if architecture == 'resnet':
with tf.variable_scope('Skip'):
t = conv2d_layer(t,
fmaps=nf(res - 1),
kernel=1,
up=True,
resample_kernel=resample_kernel)
x = (x + t) * (1 / np.sqrt(2))
return x
def upsample(y):
with tf.variable_scope('Upsample'):
return upsample_2d(y, k=resample_kernel)
def torgb(x, y, res): # res = 2..resolution_log2
mod_vector = []
if style_mod:
mod_vector.append(dlatents_in[:, res * 2 - 3])
if cond_mod:
mod_vector.append(x_global)
with tf.variable_scope('ToRGB'):
t = apply_bias_act(
modulated_conv2d_layer(
x,
tf.concat(mod_vector, axis=1) if mod_vector else None,
fmaps=num_channels,
kernel=1,
demodulate=False,
fused_modconv=fused_modconv))
return t if y is None else y + t
# Early layers.
y = None
with tf.variable_scope('G_4x4'):
with tf.variable_scope('Dense'):
x = apply_bias_act(dense_layer(x, fmaps=nf(1) * 4 * 4), act=act)
x = tf.reshape(x, [-1, nf(1), 4, 4])
x = x + E_features[2]
with tf.variable_scope('Conv'):
x = layer(x, layer_idx=0, fmaps=nf(1), kernel=3)
if architecture == 'skip':
y = torgb(x, y, 2)
# Main layers.
for res in range(3, resolution_log2 + 1):
with tf.variable_scope('G_%dx%d' % (2**res, 2**res)):
x = block(x, res, E_features)
if architecture == 'skip':
y = upsample(y)
if architecture == 'skip' or res == resolution_log2:
y = torgb(x, y, res)
if pix2pix:
images_out = tf.concat([y, cond], axis=1)
else:
images_out = y * (1 - masks_in) + images_in * masks_in
assert images_out.dtype == tf.as_dtype(dtype)
return tf.identity(images_out, name='images_out')
def build(self):
# Setup input placeholders
self._set_up_input_pls()
# Setup feature extractors
self._set_up_feature_extractors()
bev_proposal_input = self.bev_bottleneck
img_proposal_input = self.img_bottleneck
fusion_mean_div_factor = 2.0
# If both img and bev probabilites are set to 1.0, don't do
# path drop.
if not (self._path_drop_probabilities[0] ==
self._path_drop_probabilities[1] == 1.0):
with tf.variable_scope('rpn_path_drop'):
random_values = tf.random_uniform(shape=[3],
minval=0.0,
maxval=1.0)
img_mask, bev_mask = self.create_path_drop_masks(
self._path_drop_probabilities[0],
self._path_drop_probabilities[1], random_values)
img_proposal_input = tf.multiply(img_proposal_input, img_mask)
bev_proposal_input = tf.multiply(bev_proposal_input, bev_mask)
self.img_path_drop_mask = img_mask
self.bev_path_drop_mask = bev_mask
# Overwrite the division factor
fusion_mean_div_factor = img_mask + bev_mask
with tf.variable_scope('proposal_roi_pooling'):
with tf.variable_scope('box_indices'):
def get_box_indices(boxes):
proposals_shape = boxes.get_shape().as_list()
if any(dim is None for dim in proposals_shape):
proposals_shape = tf.shape(boxes)
ones_mat = tf.ones(proposals_shape[:2], dtype=tf.int32)
multiplier = tf.expand_dims(
tf.range(start=0, limit=proposals_shape[0]), 1)
return tf.reshape(ones_mat * multiplier, [-1])
bev_boxes_norm_batches = tf.expand_dims(
self._bev_anchors_norm_pl, axis=0)
# These should be all 0's since there is only 1 image
tf_box_indices = get_box_indices(bev_boxes_norm_batches)
# Do ROI Pooling on BEV
bev_proposal_rois = tf.image.crop_and_resize(
bev_proposal_input, self._bev_anchors_norm_pl, tf_box_indices,
self._proposal_roi_crop_size)
# Do ROI Pooling on image
img_proposal_rois = tf.image.crop_and_resize(
img_proposal_input, self._img_anchors_norm_pl, tf_box_indices,
self._proposal_roi_crop_size)
print("img_proposal_rois shape")
# print(img_proposal_rois.shape)
# for i in range(img_proposal_rois.shape[0]):
# print(img_proposal_rois[i])
####################################################################################
# TODO PROJECT: insert code here to add mixture of experts
# self._moe_model = MoeModel(img_proposal_input, bev_proposal_input)
# self._moe_model._set_up_input_pls()
# moe_prediction = self._moe_model.build()
####################################################################################
with tf.variable_scope('proposal_roi_fusion'):
rpn_fusion_out = None
####################################################################################
# TODO PROJECT: weight the feature before average img and bev
# weighted_img_proposal_rois = tf.multiply(moe_prediction['img_weight'],img_proposal_rois)
# weighted_bev_proposal_rois = tf.multiply(moe_prediction['bev_weight'],bev_proposal_rois)
####################################################################################
if self._fusion_method == 'mean':
tf_features_sum = tf.add(bev_proposal_rois, img_proposal_rois)
rpn_fusion_out = tf.divide(tf_features_sum,
fusion_mean_div_factor)
####################################################################################
# TODO PROJECT: weight the feature before average img and bev
# tf_features_sum = tf.add(weighted_bev_proposal_rois, weighted_img_proposal_rois)
# rpn_fusion_out = tf.divide(tf_features_sum, fusion_mean_div_factor)
####################################################################################
elif self._fusion_method == 'concat':
rpn_fusion_out = tf.concat(
[bev_proposal_rois, img_proposal_rois], axis=3)
####################################################################################
# TODO PROJECT: weight the feature before concatenation
# rpn_fusion_out = tf.concat(
# [weighted_bev_proposal_rois, weighted_img_proposal_rois], axis=3)
####################################################################################
else:
raise ValueError('Invalid fusion method', self._fusion_method)
# TODO: move this section into an separate AnchorPredictor class
with tf.variable_scope('anchor_predictor', 'ap', [rpn_fusion_out]):
tensor_in = rpn_fusion_out
# Parse rpn layers config
layers_config = self._config.layers_config.rpn_config
l2_weight_decay = layers_config.l2_weight_decay
if l2_weight_decay > 0:
weights_regularizer = slim.l2_regularizer(l2_weight_decay)
else:
weights_regularizer = None
with slim.arg_scope([slim.conv2d],
weights_regularizer=weights_regularizer):
# Use conv2d instead of fully_connected layers.
cls_fc6 = slim.conv2d(tensor_in,
layers_config.cls_fc6,
self._proposal_roi_crop_size,
padding='VALID',
scope='cls_fc6')
cls_fc6_drop = slim.dropout(cls_fc6,
layers_config.keep_prob,
is_training=self._is_training,
scope='cls_fc6_drop')
cls_fc7 = slim.conv2d(cls_fc6_drop,
layers_config.cls_fc7, [1, 1],
scope='cls_fc7')
cls_fc7_drop = slim.dropout(cls_fc7,
layers_config.keep_prob,
is_training=self._is_training,
scope='cls_fc7_drop')
cls_fc8 = slim.conv2d(cls_fc7_drop,
2, [1, 1],
activation_fn=None,
scope='cls_fc8')
objectness = tf.squeeze(cls_fc8, [1, 2],
name='cls_fc8/squeezed')
# Use conv2d instead of fully_connected layers.
reg_fc6 = slim.conv2d(tensor_in,
layers_config.reg_fc6,
self._proposal_roi_crop_size,
padding='VALID',
scope='reg_fc6')
reg_fc6_drop = slim.dropout(reg_fc6,
layers_config.keep_prob,
is_training=self._is_training,
scope='reg_fc6_drop')
reg_fc7 = slim.conv2d(reg_fc6_drop,
layers_config.reg_fc7, [1, 1],
scope='reg_fc7')
reg_fc7_drop = slim.dropout(reg_fc7,
layers_config.keep_prob,
is_training=self._is_training,
scope='reg_fc7_drop')
reg_fc8 = slim.conv2d(reg_fc7_drop,
6, [1, 1],
activation_fn=None,
scope='reg_fc8')
offsets = tf.squeeze(reg_fc8, [1, 2], name='reg_fc8/squeezed')
# Histogram summaries
with tf.variable_scope('histograms_feature_extractor'):
with tf.variable_scope('bev_vgg'):
for end_point in self.bev_end_points:
tf.summary.histogram(end_point,
self.bev_end_points[end_point])
with tf.variable_scope('img_vgg'):
for end_point in self.img_end_points:
tf.summary.histogram(end_point,
self.img_end_points[end_point])
with tf.variable_scope('histograms_rpn'):
with tf.variable_scope('anchor_predictor'):
fc_layers = [
cls_fc6, cls_fc7, cls_fc8, objectness, reg_fc6, reg_fc7,
reg_fc8, offsets
]
for fc_layer in fc_layers:
# fix the name to avoid tf warnings
tf.summary.histogram(fc_layer.name.replace(':', '_'),
fc_layer)
# Return the proposals
with tf.variable_scope('proposals'):
anchors = self.placeholders[self.PL_ANCHORS]
# Decode anchor regression offsets
with tf.variable_scope('decoding'):
regressed_anchors = anchor_encoder.offset_to_anchor(
anchors, offsets)
with tf.variable_scope('bev_projection'):
_, bev_proposal_boxes_norm = anchor_projector.project_to_bev(
regressed_anchors, self._bev_extents)
with tf.variable_scope('softmax'):
objectness_softmax = tf.nn.softmax(objectness)
with tf.variable_scope('nms'):
objectness_scores = objectness_softmax[:, 1]
# Do NMS on regressed anchors
top_indices = tf.image.non_max_suppression(
bev_proposal_boxes_norm,
objectness_scores,
max_output_size=self._nms_size,
iou_threshold=self._nms_iou_thresh)
top_anchors = tf.gather(regressed_anchors, top_indices)
top_objectness_softmax = tf.gather(objectness_scores,
top_indices)
# top_offsets = tf.gather(offsets, top_indices)
# top_objectness = tf.gather(objectness, top_indices)
# Get mini batch
all_ious_gt = self.placeholders[self.PL_ANCHOR_IOUS]
all_offsets_gt = self.placeholders[self.PL_ANCHOR_OFFSETS]
all_classes_gt = self.placeholders[self.PL_ANCHOR_CLASSES]
with tf.variable_scope('mini_batch'):
mini_batch_utils = self.dataset.kitti_utils.mini_batch_utils
mini_batch_mask, _ = \
mini_batch_utils.sample_rpn_mini_batch(all_ious_gt)
# ROI summary images
rpn_mini_batch_size = \
self.dataset.kitti_utils.mini_batch_utils.rpn_mini_batch_size
with tf.variable_scope('bev_rpn_rois'):
mb_bev_anchors_norm = tf.boolean_mask(self._bev_anchors_norm_pl,
mini_batch_mask)
mb_bev_box_indices = tf.zeros_like(tf.boolean_mask(
all_classes_gt, mini_batch_mask),
dtype=tf.int32)
# Show the ROIs of the BEV input density map
# for the mini batch anchors
bev_input_rois = tf.image.crop_and_resize(self._bev_preprocessed,
mb_bev_anchors_norm,
mb_bev_box_indices,
(32, 32))
bev_input_roi_summary_images = tf.split(bev_input_rois,
self._bev_depth,
axis=3)
tf.summary.image('bev_rpn_rois',
bev_input_roi_summary_images[-1],
max_outputs=rpn_mini_batch_size)
with tf.variable_scope('img_rpn_rois'):
# ROIs on image input
mb_img_anchors_norm = tf.boolean_mask(self._img_anchors_norm_pl,
mini_batch_mask)
mb_img_box_indices = tf.zeros_like(tf.boolean_mask(
all_classes_gt, mini_batch_mask),
dtype=tf.int32)
# Do test ROI pooling on mini batch
img_input_rois = tf.image.crop_and_resize(self._img_preprocessed,
mb_img_anchors_norm,
mb_img_box_indices,
(32, 32))
tf.summary.image('img_rpn_rois',
img_input_rois,
max_outputs=rpn_mini_batch_size)
# Ground Truth Tensors
with tf.variable_scope('one_hot_classes'):
# Anchor classification ground truth
# Object / Not Object
min_pos_iou = \
self.dataset.kitti_utils.mini_batch_utils.rpn_pos_iou_range[0]
objectness_classes_gt = tf.cast(tf.greater_equal(
all_ious_gt, min_pos_iou),
dtype=tf.int32)
objectness_gt = tf.one_hot(
objectness_classes_gt,
depth=2,
on_value=1.0 - self._config.label_smoothing_epsilon,
off_value=self._config.label_smoothing_epsilon)
# Mask predictions for mini batch
with tf.variable_scope('prediction_mini_batch'):
objectness_masked = tf.boolean_mask(objectness, mini_batch_mask)
offsets_masked = tf.boolean_mask(offsets, mini_batch_mask)
with tf.variable_scope('ground_truth_mini_batch'):
objectness_gt_masked = tf.boolean_mask(objectness_gt,
mini_batch_mask)
offsets_gt_masked = tf.boolean_mask(all_offsets_gt,
mini_batch_mask)
# Specify the tensors to evaluate
predictions = dict()
# Temporary predictions for debugging
# predictions['anchor_ious'] = anchor_ious
# predictions['anchor_offsets'] = all_offsets_gt
if self._train_val_test in ['train', 'val']:
# All anchors
predictions[self.PRED_ANCHORS] = anchors
# Mini-batch masks
predictions[self.PRED_MB_MASK] = mini_batch_mask
# Mini-batch predictions
predictions[self.PRED_MB_OBJECTNESS] = objectness_masked
predictions[self.PRED_MB_OFFSETS] = offsets_masked
# Mini batch ground truth
predictions[self.PRED_MB_OFFSETS_GT] = offsets_gt_masked
predictions[self.PRED_MB_OBJECTNESS_GT] = objectness_gt_masked
# Proposals after nms
predictions[self.PRED_TOP_INDICES] = top_indices
predictions[self.PRED_TOP_ANCHORS] = top_anchors
predictions[
self.PRED_TOP_OBJECTNESS_SOFTMAX] = top_objectness_softmax
else:
# self._train_val_test == 'test'
predictions[self.PRED_TOP_ANCHORS] = top_anchors
predictions[
self.PRED_TOP_OBJECTNESS_SOFTMAX] = top_objectness_softmax
return predictions
def __init__(self,
embedding_mat,
non_static,
hidden_unit,
sequence_length,
max_pool_size,
num_classes,
embedding_size,
filter_sizes,
num_filters,
l2_reg_lambda=0.0):
self.input_x = tf.placeholder(tf.int32, [None, sequence_length],
name='input_x')
self.input_y = tf.placeholder(tf.float32, [None, num_classes],
name='input_y')
self.dropout_keep_prob = tf.placeholder(tf.float32,
name='dropout_keep_prob')
self.batch_size = tf.placeholder(tf.int32)
self.pad = tf.placeholder(tf.float32, [None, 1, embedding_size, 1],
name='pad')
self.real_len = tf.placeholder(tf.int32, [None], name='real_len')
l2_loss = tf.constant(0.0)
with tf.device('/cpu:0'), tf.name_scope('embedding'):
if not non_static:
W = tf.constant(embedding_mat, name='W')
else:
W = tf.Variable(embedding_mat, name='W')
self.embedded_chars = tf.nn.embedding_lookup(W, self.input_x)
# emb = tf.expand_dims(self.embedded_chars, -1)
reduced = sequence_length
lstm_cell = tf.nn.rnn_cell.LSTMCell(num_units=hidden_unit)
lstm_cell = tf.nn.rnn_cell.LSTMCell(num_units=int(hidden_unit / 2))
lstm_cell = tf.nn.rnn_cell.DropoutWrapper(
lstm_cell, output_keep_prob=self.dropout_keep_prob)
lstm_bw_cell = tf.nn.rnn_cell.LSTMCell(num_units=int(hidden_unit / 2))
lstm_bw_cell = tf.nn.rnn_cell.DropoutWrapper(
lstm_bw_cell, output_keep_prob=self.dropout_keep_prob)
self._initial_state = lstm_cell.zero_state(self.batch_size, tf.float32)
self._initial_state_bw = lstm_bw_cell.zero_state(
self.batch_size, tf.float32)
inputs = [
tf.squeeze(input_, [1])
for input_ in tf.split(1, reduced, self.embedded_chars)
]
outputs, _, _ = tf.nn.bidirectional_rnn(
lstm_cell,
lstm_bw_cell,
inputs,
initial_state_fw=self._initial_state,
initial_state_bw=self._initial_state_bw,
sequence_length=self.real_len)
# Collect the appropriate last words into variable output (dimension = batch x embedding_size)
output = outputs[0]
with tf.variable_scope('Output'):
tf.get_variable_scope().reuse_variables()
one = tf.ones([1, hidden_unit], tf.float32)
for i in range(1, len(outputs)):
ind = self.real_len < (i + 1)
ind = tf.to_float(ind)
ind = tf.expand_dims(ind, -1)
mat = tf.matmul(ind, one)
output = tf.add(tf.mul(output, mat),
tf.mul(outputs[i], 1.0 - mat))
with tf.name_scope('output'):
self.W = tf.Variable(tf.truncated_normal(
[hidden_unit, num_classes], stddev=0.1),
name='W')
b = tf.Variable(tf.constant(0.1, shape=[num_classes]), name='b')
l2_loss += tf.nn.l2_loss(W)
l2_loss += tf.nn.l2_loss(b)
self.scores = tf.nn.xw_plus_b(output, self.W, b, name='scores')
self.predictions = tf.argmax(self.scores, 1, name='predictions')
with tf.name_scope('loss'):
losses = tf.nn.softmax_cross_entropy_with_logits(
self.scores, self.input_y)
self.loss = tf.reduce_mean(losses) + l2_reg_lambda * l2_loss
with tf.name_scope('accuracy'):
correct_predictions = tf.equal(self.predictions,
tf.argmax(self.input_y, 1))
self.accuracy = tf.reduce_mean(tf.cast(correct_predictions,
"float"),
name='accuracy')
with tf.name_scope('num_correct'):
correct = tf.equal(self.predictions, tf.argmax(self.input_y, 1))
self.num_correct = tf.reduce_sum(tf.cast(correct, 'float'))
def __call__(self, x, prev_state):
prev_read_vector_list = prev_state.read_vector_list
controller_input = tf.concat([x] + prev_read_vector_list, axis=1)
with tf.compat.v1.variable_scope('controller', reuse=self.reuse):
controller_output, controller_state = self.controller(controller_input, prev_state.controller_state)
num_parameters_per_head = self.memory_vector_dim + 1 + 1 + (self.shift_range * 2 + 1) + 1
num_heads = self.read_head_num + self.write_head_num
total_parameter_num = num_parameters_per_head * num_heads + self.memory_vector_dim * 2 * self.write_head_num
with tf.compat.v1.variable_scope("o2p", reuse=(self.step > 0) or self.reuse):
parameters = tf.compat.v1.layers.dense(
controller_output, total_parameter_num, activation=None,
kernel_initializer=self.o2p_initializer)
parameters = tf.clip_by_value(parameters, -self.clip_value, self.clip_value)
head_parameter_list = tf.split(parameters[:, :num_parameters_per_head * num_heads], num_heads, axis=1)
erase_add_list = tf.split(parameters[:, num_parameters_per_head * num_heads:], 2 * self.write_head_num, axis=1)
prev_w_list = prev_state.w_list
prev_M = prev_state.M
w_list = []
for i, head_parameter in enumerate(head_parameter_list):
k = tf.tanh(head_parameter[:, 0:self.memory_vector_dim])
beta = tf.nn.softplus(head_parameter[:, self.memory_vector_dim])
g = tf.sigmoid(head_parameter[:, self.memory_vector_dim + 1])
s = tf.nn.softmax(
head_parameter[:, self.memory_vector_dim + 2:self.memory_vector_dim + 2 + (self.shift_range * 2 + 1)]
)
gamma = tf.nn.softplus(head_parameter[:, -1]) + 1
with tf.compat.v1.variable_scope('addressing_head_%d' % i):
w = self.addressing(k, beta, g, s, gamma, prev_M, prev_w_list[i])
w_list.append(w)
# Reading (Sec 3.1)
read_w_list = w_list[:self.read_head_num]
read_vector_list = []
for i in range(self.read_head_num):
read_vector = tf.reduce_sum(tf.expand_dims(read_w_list[i], axis=2) * prev_M, axis=1)
read_vector_list.append(read_vector)
# Writing (Sec 3.2)
write_w_list = w_list[self.read_head_num:]
M = prev_M
for i in range(self.write_head_num):
w = tf.expand_dims(write_w_list[i], axis=2)
erase_vector = tf.expand_dims(tf.sigmoid(erase_add_list[i * 2]), axis=1)
add_vector = tf.expand_dims(tf.tanh(erase_add_list[i * 2 + 1]), axis=1)
M = M * (tf.ones(M.get_shape()) - tf.matmul(w, erase_vector)) + tf.matmul(w, add_vector)
if not self.output_dim:
output_dim = x.get_shape()[1]
else:
output_dim = self.output_dim
with tf.compat.v1.variable_scope("o2o", reuse=(self.step > 0) or self.reuse):
NTM_output = tf.compat.v1.layers.dense(
tf.concat([controller_output] + read_vector_list, axis=1), output_dim, activation=None,
kernel_initializer=self.o2o_initializer)
NTM_output = tf.clip_by_value(NTM_output, -self.clip_value, self.clip_value)
self.step += 1
return NTM_output, NTMControllerState(
controller_state=controller_state, read_vector_list=read_vector_list, w_list=w_list, M=M)
def _init_neural_network(self):
"""Initialize the RNNLM network."""
with tf.variable_scope(self.scope_name):
# TODO dropout
# I/O placeholders
self._inputs = tf.placeholder(tf.int32, [None, self.max_sent_len],
name='inputs')
self._targets = tf.placeholder(tf.int32, [None, self.max_sent_len],
name='targets')
# RNN cell type
if self.cell_type.startswith('gru'):
self._cell = tf.nn.rnn_cell.GRUCell(self.emb_size)
else:
self._cell = tf.nn.rnn_cell.BasicLSTMCell(self.emb_size)
if re.match(r'/[0-9]$', self.cell_type):
self._cell = tf.nn.rnn_cell.MultiRNNCell(
[self.cell] * int(self.cell_type[-1]))
self._initial_state = self._cell.zero_state(
tf.shape(self._inputs)[0], tf.float32)
# embeddings
emb_cell = tf.nn.rnn_cell.EmbeddingWrapper(self._cell,
self.vocab_size)
# RNN encoder
inputs = [
tf.squeeze(input_, [1])
for input_ in tf.split(1, self.max_sent_len, self._inputs)
]
outputs, states = tf.nn.rnn(emb_cell,
inputs,
initial_state=self._initial_state)
# output layer
output = tf.reshape(tf.concat(1, outputs), [-1, self.emb_size])
self._logits = (tf.matmul(
output, tf.get_variable("W", [self.emb_size, self.vocab_size]))
+ tf.get_variable("b", [self.vocab_size]))
# cost
targets_1d = tf.reshape(self._targets, [-1])
self._loss = tf.nn.seq2seq.sequence_loss_by_example(
[self._logits], [targets_1d],
[tf.ones_like(targets_1d, dtype=tf.float32)], self.vocab_size)
self._cost = tf.reduce_mean(self._loss)
# optimizer
self._learning_rate = tf.placeholder(tf.float32,
name="learning_rate")
if self.optimizer_type == 'sgd':
opt = tf.train.GradientDescentOptimizer(self._learning_rate)
if self.optimizer_type == 'adagrad':
opt = tf.train.AdagradOptimizer(self._learning_rate)
else:
opt = tf.train.AdamOptimizer(self._learning_rate)
# gradient clipping
grads_tvars = opt.compute_gradients(self._loss,
tf.trainable_variables())
grads, _ = tf.clip_by_global_norm([g for g, _ in grads_tvars],
self.max_grad_norm)
self._train_func = opt.apply_gradients(
zip(grads, [v for _, v in grads_tvars]))
# initialize TF session
session_config = None
if self.max_cores:
session_config = tf.ConfigProto(
inter_op_parallelism_threads=self.max_cores,
intra_op_parallelism_threads=self.max_cores)
self.session = tf.Session(config=session_config)
def _set_up_input_pls(self):
"""Sets up input placeholders by adding them to self._placeholders.
Keys are defined as self.PL_*.
"""
# Combine config data
bev_dims = np.append(self._bev_pixel_size, self._bev_depth)
with tf.variable_scope('bev_input'):
# Placeholder for BEV image input, to be filled in with feed_dict
bev_input_placeholder = self._add_placeholder(
tf.float32, bev_dims, self.PL_BEV_INPUT)
self._bev_input_batches = tf.expand_dims(bev_input_placeholder,
axis=0)
self._bev_preprocessed = \
self._bev_feature_extractor.preprocess_input(
self._bev_input_batches, self._bev_pixel_size)
# Summary Images
bev_summary_images = tf.split(bev_input_placeholder,
self._bev_depth,
axis=2)
tf.summary.image("bev_maps",
bev_summary_images,
max_outputs=self._bev_depth)
with tf.variable_scope('img_input'):
# Take variable size input images
img_input_placeholder = self._add_placeholder(
tf.float32, [None, None, self._img_depth], self.PL_IMG_INPUT)
self._img_input_batches = tf.expand_dims(img_input_placeholder,
axis=0)
self._img_preprocessed = \
self._img_feature_extractor.preprocess_input(
self._img_input_batches, self._img_pixel_size)
# Summary Image
tf.summary.image("rgb_image",
self._img_preprocessed,
max_outputs=2)
with tf.variable_scope('pl_labels'):
self._add_placeholder(tf.float32, [None, 6], self.PL_LABEL_ANCHORS)
self._add_placeholder(tf.float32, [None, 7],
self.PL_LABEL_BOXES_3D)
self._add_placeholder(tf.float32, [None], self.PL_LABEL_CLASSES)
# Placeholders for anchors
with tf.variable_scope('pl_anchors'):
self._add_placeholder(tf.float32, [None, 6], self.PL_ANCHORS)
self._add_placeholder(tf.float32, [None], self.PL_ANCHOR_IOUS)
self._add_placeholder(tf.float32, [None, 6],
self.PL_ANCHOR_OFFSETS)
self._add_placeholder(tf.float32, [None], self.PL_ANCHOR_CLASSES)
with tf.variable_scope('bev_anchor_projections'):
self._add_placeholder(tf.float32, [None, 4],
self.PL_BEV_ANCHORS)
self._bev_anchors_norm_pl = self._add_placeholder(
tf.float32, [None, 4], self.PL_BEV_ANCHORS_NORM)
with tf.variable_scope('img_anchor_projections'):
self._add_placeholder(tf.float32, [None, 4],
self.PL_IMG_ANCHORS)
self._img_anchors_norm_pl = self._add_placeholder(
tf.float32, [None, 4], self.PL_IMG_ANCHORS_NORM)
with tf.variable_scope('sample_info'):
# the calib matrix shape is (3 x 4)
self._add_placeholder(tf.float32, [3, 4], self.PL_CALIB_P2)
self._add_placeholder(tf.int32,
shape=[1],
name=self.PL_IMG_IDX)
self._add_placeholder(tf.float32, [4], self.PL_GROUND_PLANE)
def __init__(self, is_training, args):
self.batch_size = batch_size = np.int32(args['max_group_size'])
self.num_steps = num_steps = np.int32(args['num_steps'])
self.num_features = num_features = args['num_features']
self.dense_units = dense_units = 1
self.hidden_size = size = np.int32(args['hidden_size'])
if self.num_steps < 1:
print 'num_steps cannot be zero -- setting to 1'
self.num_steps = num_steps = 1
if self.hidden_size < 10:
print 'hidden_size should not be less than 10 -- setting to 10'
self.hidden_size = size = 10
if is_training:
print 'Initiating input tensors of shape: {}'.format(
(num_steps, batch_size, num_features))
self._input_data = inputs = tf.placeholder(
tf.float32, [num_steps, batch_size, num_features])
self._targets = tf.placeholder(tf.float32, [num_steps, batch_size])
# Memory cell to use in model
lstm_cell = tf.nn.rnn_cell.BasicLSTMCell(num_units=size,
forget_bias=0.0,
state_is_tuple=True)
if is_training:
print 'Memory Cell: {}'.format(type(lstm_cell))
# Wrap the memory cell in a dropout layer (for outputs)
if is_training and args['keep_prob'] < 1:
lstm_cell = tf.nn.rnn_cell.DropoutWrapper(
cell=lstm_cell, output_keep_prob=args['keep_prob'])
# Create the RNN with 'num_layers' layers
stacked_cell = tf.nn.rnn_cell.MultiRNNCell([lstm_cell] *
args['num_layers'],
state_is_tuple=True)
# Initialize the state -- it will hold the last output, h_t, as well as the memory state, c_t
# Shape will be [batch_size, num_units x 2] -- splitting on dimension 1 will separate output and memory
self._initial_state = stacked_cell.zero_state(
batch_size=tf.constant(batch_size), dtype=tf.float32)
# Split the inputs (by timestep)
inputs = [
tf.squeeze(input, [0]) for input in tf.split(0, num_steps, inputs)
]
# Computes dropout for inputs
if is_training and args['keep_prob'] < 1:
inputs = [tf.nn.dropout(x, args['keep_prob']) for x in inputs]
# Run inputs through the RNN
outputs, state = tf.nn.rnn(stacked_cell,
inputs,
initial_state=self._initial_state)
# Re-joins all output tensors (from each timestep)
output = tf.reshape(tf.concat(1, outputs), shape=[-1, size])
# Add a fully-connected layer
self.dense_w = dense_w = tf.get_variable('dense_w',
shape=[size, dense_units])
self.dense_b = dense_b = tf.get_variable('dense_b',
shape=[dense_units])
# Feed the output from the RNN to the fully-connected layer
self._predictions = predictions = tf.matmul(output, dense_w) + dense_b
self._predictions = predictions = tf.reshape(
self.predictions, shape=[num_steps, batch_size])
# Compute the R^2
numerator = tf.reduce_sum(
tf.square(tf.sub(self.targets, self.predictions)))
denominator = tf.reduce_sum(
tf.square(tf.sub(self.targets, tf.reduce_mean(self.targets))))
self.r2 = r2 = tf.sub(1.0, tf.div(numerator, denominator))
# MSE cost function
self._cost = cost = tf.reduce_mean(
tf.square(tf.sub(self.targets, self.predictions)))
self._final_state = state
# Variable for state (for when saving model)
self.save_state = tf.Variable(
tf.zeros([args['num_layers'], 2, batch_size, size]))
self.save_state.assign(state)
if is_training:
self._lr = tf.Variable(0.0, trainable=False)
# Compute the gradients
tvars = tf.trainable_variables()
grads, _ = tf.clip_by_global_norm(t_list=tf.gradients(cost, tvars),
clip_norm=args['max_grad_norm'])
# Adjust the parameters based on optimizer and gradients
optimizer = args['optimizer'](self.lr)
self._train_op = optimizer.apply_gradients(
grads_and_vars=zip(grads, tvars))
# Summaries for Tensorboard
cost_summ = tf.scalar_summary('mean squared error', cost)
r2_summ = tf.scalar_summary('r-squared', r2)
state_summ = tf.histogram_summary('states', state)
pred_summ = tf.histogram_summary('predictions', predictions)
self.summary = tf.merge_all_summaries()
else:
# Ignore this -- put here so errors are prevented when running model not in training mode
self.summary = predictions
return
def build_graph(reader,
model,
train_data_pattern,
label_loss_fn=losses.CrossEntropyLoss(),
batch_size=1000,
base_learning_rate=0.01,
learning_rate_decay_examples=1000000,
learning_rate_decay=0.95,
optimizer_class=tf.train.AdamOptimizer,
clip_gradient_norm=1.0,
regularization_penalty=1,
num_readers=1,
num_epochs=None):
"""Creates the Tensorflow graph.
This will only be called once in the life of
a training model, because after the graph is created the model will be
restored from a meta graph file rather than being recreated.
Args:
reader: The data file reader. It should inherit from BaseReader.
model: The core model (e.g. logistic or neural net). It should inherit from
BaseModel.
train_data_pattern: glob path to the training data files.
label_loss_fn: What kind of loss to apply to the model. It should inherit
from BaseLoss.
batch_size: How many examples to process at a time.
base_learning_rate: What learning rate to initialize the optimizer with.
optimizer_class: Which optimization algorithm to use.
clip_gradient_norm: Magnitude of the gradient to clip to.
regularization_penalty: How much weight to give the regularization loss
compared to the label loss.
num_readers: How many threads to use for I/O operations.
num_epochs: How many passes to make over the data. 'None' means an unlimited
number of passes.
"""
global_step = tf.Variable(0, trainable=False, name="global_step")
local_device_protos = device_lib.list_local_devices()
gpus = [x.name for x in local_device_protos if x.device_type == "GPU"]
gpus = gpus[:FLAGS.num_gpu]
num_gpus = len(gpus)
if num_gpus > 0:
logging.info("Using the following GPUs to train: " + str(gpus))
num_towers = num_gpus
device_string = "/gpu:%d"
else:
logging.info("No GPUs found. Training on CPU.")
num_towers = 1
device_string = "/cpu:%d"
learning_rate = tf.train.exponential_decay(base_learning_rate,
global_step * batch_size *
num_towers,
learning_rate_decay_examples,
learning_rate_decay,
staircase=True)
tf.summary.scalar("learning_rate", learning_rate)
optimizer = optimizer_class(learning_rate)
input_data_dict = (get_input_data_tensors(reader,
train_data_pattern,
batch_size=batch_size *
num_towers,
num_readers=num_readers,
num_epochs=num_epochs))
model_input_raw = input_data_dict["video_matrix"]
labels_batch = input_data_dict["labels"]
num_frames = input_data_dict["num_frames"]
print("model_input_shape, ", model_input_raw.shape)
tf.summary.histogram("model/input_raw", model_input_raw)
feature_dim = len(model_input_raw.get_shape()) - 1
model_input = tf.nn.l2_normalize(model_input_raw, feature_dim)
tower_inputs = tf.split(model_input, num_towers)
tower_labels = tf.split(labels_batch, num_towers)
tower_num_frames = tf.split(num_frames, num_towers)
tower_gradients = []
tower_predictions = []
tower_label_losses = []
tower_reg_losses = []
for i in range(num_towers):
# For some reason these 'with' statements can't be combined onto the same
# line. They have to be nested.
with tf.device(device_string % i):
with (tf.variable_scope(("tower"), reuse=True if i > 0 else None)):
with (slim.arg_scope(
[slim.model_variable, slim.variable],
device="/cpu:0" if num_gpus != 1 else "/gpu:0")):
result = model.create_model(tower_inputs[i],
num_frames=tower_num_frames[i],
vocab_size=reader.num_classes,
labels=tower_labels[i])
for variable in slim.get_model_variables():
tf.summary.histogram(variable.op.name, variable)
predictions = result["predictions"]
tower_predictions.append(predictions)
if "loss" in result.keys():
label_loss = result["loss"]
else:
label_loss = label_loss_fn.calculate_loss(
predictions, tower_labels[i])
if "regularization_loss" in result.keys():
reg_loss = result["regularization_loss"]
else:
reg_loss = tf.constant(0.0)
reg_losses = tf.losses.get_regularization_losses()
if reg_losses:
reg_loss += tf.add_n(reg_losses)
tower_reg_losses.append(reg_loss)
# Adds update_ops (e.g., moving average updates in batch normalization) as
# a dependency to the train_op.
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
if "update_ops" in result.keys():
update_ops += result["update_ops"]
if update_ops:
with tf.control_dependencies(update_ops):
barrier = tf.no_op(name="gradient_barrier")
with tf.control_dependencies([barrier]):
label_loss = tf.identity(label_loss)
tower_label_losses.append(label_loss)
# Incorporate the L2 weight penalties etc.
final_loss = regularization_penalty * reg_loss + label_loss
gradients = optimizer.compute_gradients(
final_loss, colocate_gradients_with_ops=False)
tower_gradients.append(gradients)
label_loss = tf.reduce_mean(tf.stack(tower_label_losses))
tf.summary.scalar("label_loss", label_loss)
if regularization_penalty != 0:
reg_loss = tf.reduce_mean(tf.stack(tower_reg_losses))
tf.summary.scalar("reg_loss", reg_loss)
merged_gradients = utils.combine_gradients(tower_gradients)
if clip_gradient_norm > 0:
with tf.name_scope("clip_grads"):
merged_gradients = utils.clip_gradient_norms(
merged_gradients, clip_gradient_norm)
train_op = optimizer.apply_gradients(merged_gradients,
global_step=global_step)
tf.add_to_collection("global_step", global_step)
tf.add_to_collection("loss", label_loss)
tf.add_to_collection("predictions", tf.concat(tower_predictions, 0))
tf.add_to_collection("input_batch_raw", model_input_raw)
tf.add_to_collection("input_batch", model_input)
tf.add_to_collection("num_frames", num_frames)
tf.add_to_collection("labels", tf.cast(labels_batch, tf.float32))
tf.add_to_collection("train_op", train_op)
def get_q_values_op(self,
state,
past_a,
seq_len,
h_state,
scope,
reuse=False):
"""
Returns Q values for all actions
Args:
state: (tf tensor)
shape = (batch_size, seq_len, img_w, img_h, nchannel)
goal_state: (tf tensor)
shape = (batch_size, 1, img_w, img_h, nchannel, 4)
past_a: (tf tensor)
shape = (batch_size*seq_len,)
seq_len: (tf tensor)
shape = (batch_size,)
h_state: (tf tensor)
shape = (batch_size, h_size)
scope: (string) scope name, that specifies if target network or not
reuse: (bool) reuse of variables in the scope
Returns:
out: (tf tensor) of shape = (batch_size * seq_len, num_actions)
h_state_out: (tf tensor) of shape = (batch_size, h_size)
"""
num_actions = 2
h_size = self.config.h_size
max_seq_len = tf.shape(state)[1]
state_shape = list(
[4 * 3, 3, len(self.env.state.xmap.item_class_id) + 2])
past_a = tf.reshape(tf.one_hot(past_a, num_actions),
shape=(-1, max_seq_len, 1, num_actions))
past_a = tf.tile(past_a, multiples=[1, 1, 4, 1])
out = tf.reshape(state,
shape=(-1, max_seq_len, 4,
np.int32(state_shape[0] * state_shape[1] *
state_shape[2] / 4)))
with tf.variable_scope(scope, reuse=False):
#### recurrent
out = tf.concat([out, past_a], axis=3)
out = layers.fully_connected(layers.fully_connected(out, 200), 100)
out = tf.reduce_max(out, axis=2)
lstm_cell = tf.nn.rnn_cell.LSTMCell(num_units=h_size)
out, h_state_out = tf.nn.dynamic_rnn(inputs=out,
cell=lstm_cell,
sequence_length=seq_len,
dtype=tf.float32,
initial_state=h_state)
out = tf.reshape(out, shape=[-1, h_size])
#### feed forward
'''
out = layers.fully_connected(layers.fully_connected(out, 200), 100)
out = tf.reduce_max(out, axis=2)
out = tf.reshape(out, shape=[-1,100])
h_state_out = h_state
'''
streamA, streamV = tf.split(out, 2, axis=1)
advantage = layers.fully_connected(
streamA,
num_actions,
activation_fn=None,
weights_initializer=layers.xavier_initializer(),
biases_initializer=tf.zeros_initializer())
value = layers.fully_connected(
streamV,
1,
activation_fn=None,
weights_initializer=layers.xavier_initializer(),
biases_initializer=tf.zeros_initializer())
out = value + tf.subtract(
advantage, tf.reduce_mean(advantage, axis=1, keep_dims=True))
return out, h_state_out
def hsplit(x, n_splits):
return _tf.split(x, num_or_size_splits=n_splits, axis=1)
def _module_fn():
"""
Function building the module
"""
feature_layer = tf.placeholder(
tf.float32,
shape=[None, None, None, None, nchannels],
name='input')
obs_layer = tf.placeholder(tf.float32,
shape=[None, None, None, None, n_y],
name='observations')
# Builds the neural network
net = slim.conv3d(feature_layer,
16,
5,
activation_fn=tf.nn.leaky_relu,
padding='valid')
#net = wide_resnet(feature_layer, 8, activation_fn=tf.nn.leaky_relu, is_training=is_training)
net = wide_resnet(net,
16,
activation_fn=tf.nn.leaky_relu,
keep_prob=dropout,
is_training=is_training)
net = wide_resnet(net,
32,
activation_fn=tf.nn.leaky_relu,
keep_prob=dropout,
is_training=is_training)
net = wide_resnet(net,
32,
activation_fn=tf.nn.leaky_relu,
keep_prob=dropout,
is_training=is_training)
net = slim.conv3d(net, 32, 3, activation_fn=tf.nn.tanh)
# Define the probabilistic layer
net = slim.conv3d(net, n_mixture * 3 * n_y, 1, activation_fn=None)
cube_size = tf.shape(obs_layer)[1]
net = tf.reshape(
net, [-1, cube_size, cube_size, cube_size, n_y, n_mixture * 3])
# net = tf.reshape(net, [None, None, None, None, n_y, n_mixture*3])
loc, unconstrained_scale, logits = tf.split(net,
num_or_size_splits=3,
axis=-1)
scale = tf.nn.softplus(unconstrained_scale)
# Form mixture of discretized logistic distributions. Note we shift the
# logistic distribution by -0.5. This lets the quantization capture "rounding"
# intervals, `(x-0.5, x+0.5]`, and not "ceiling" intervals, `(x-1, x]`.
discretized_logistic_dist = tfd.QuantizedDistribution(
distribution=tfd.TransformedDistribution(
distribution=tfd.Logistic(loc=loc, scale=scale),
bijector=tfb.AffineScalar(shift=-0.5)),
low=0.,
high=2.**3 - 1)
mixture_dist = tfd.MixtureSameFamily(
mixture_distribution=tfd.Categorical(logits=logits),
components_distribution=discretized_logistic_dist)
# Define a function for sampling, and a function for estimating the log likelihood
sample = tf.squeeze(mixture_dist.sample())
loglik = mixture_dist.log_prob(obs_layer)
hub.add_signature(inputs={
'features': feature_layer,
'labels': obs_layer
},
outputs={
'sample': sample,
'loglikelihood': loglik
})
def _create_svgd_update(self):
"""Create a minimization operation for policy update (SVGD)."""
actions = self.policy.actions_for(
observations=self._observations_ph,
n_action_samples=self._kernel_n_particles,
reuse=True)
assert_shape(actions,
[None, self._kernel_n_particles, self._action_dim])
# SVGD requires computing two empirical expectations over actions
# (see Appendix C1.1.). To that end, we first sample a single set of
# actions, and later split them into two sets: `fixed_actions` are used
# to evaluate the expectation indexed by `j` and `updated_actions`
# the expectation indexed by `i`.
n_updated_actions = int(
self._kernel_n_particles * self._kernel_update_ratio)
n_fixed_actions = self._kernel_n_particles - n_updated_actions
fixed_actions, updated_actions = tf.split(
actions, [n_fixed_actions, n_updated_actions], axis=1)
fixed_actions = tf.stop_gradient(fixed_actions)
assert_shape(fixed_actions, [None, n_fixed_actions, self._action_dim])
assert_shape(updated_actions,
[None, n_updated_actions, self._action_dim])
svgd_target_values = self.qf.output_for(
self._observations_ph[:, None, :], fixed_actions, reuse=True)
# Target log-density. Q_soft in Equation 13:
squash_correction = tf.reduce_sum(
tf.log(1 - fixed_actions**2 + EPS), axis=-1)
log_p = svgd_target_values + squash_correction
grad_log_p = tf.gradients(log_p, fixed_actions)[0]
grad_log_p = tf.expand_dims(grad_log_p, axis=2)
grad_log_p = tf.stop_gradient(grad_log_p)
assert_shape(grad_log_p, [None, n_fixed_actions, 1, self._action_dim])
kernel_dict = self._kernel_fn(xs=fixed_actions, ys=updated_actions)
# Kernel function in Equation 13:
kappa = tf.expand_dims(kernel_dict["output"], dim=3)
assert_shape(kappa, [None, n_fixed_actions, n_updated_actions, 1])
# Stein Variational Gradient in Equation 13:
action_gradients = tf.reduce_mean(
kappa * grad_log_p + kernel_dict["gradient"], reduction_indices=1)
assert_shape(action_gradients,
[None, n_updated_actions, self._action_dim])
# Propagate the gradient through the policy network (Equation 14).
gradients = tf.gradients(
updated_actions,
self.policy.get_params_internal(),
grad_ys=action_gradients)
surrogate_loss = tf.reduce_sum([
tf.reduce_sum(w * tf.stop_gradient(g))
for w, g in zip(self.policy.get_params_internal(), gradients)
])
if self._train_policy:
optimizer = tf.train.AdamOptimizer(self._policy_lr)
svgd_training_op = optimizer.minimize(
loss=-surrogate_loss,
var_list=self.policy.get_params_internal())
self._training_ops.append(svgd_training_op)
def _extract_box_classifier_features(self, proposal_feature_maps, scope):
"""Extracts second stage box classifier features.
This function reconstructs the "second half" of the PNASNet
network after the part defined in `_extract_proposal_features`.
Args:
proposal_feature_maps: A 4-D float tensor with shape
[batch_size * self.max_num_proposals, crop_height, crop_width, depth]
representing the feature map cropped to each proposal.
scope: A scope name.
Returns:
proposal_classifier_features: A 4-D float tensor with shape
[batch_size * self.max_num_proposals, height, width, depth]
representing box classifier features for each proposal.
"""
del scope
# Number of used stem cells.
num_stem_cells = 2
# Note that we always feed into 2 layers of equal depth
# where the first N channels corresponds to previous hidden layer
# and the second N channels correspond to the final hidden layer.
hidden_previous, hidden = tf.split(proposal_feature_maps, 2, axis=3)
# Note that what follows is largely a copy of build_pnasnet_large() within
# pnasnet.py. We are copying to minimize code pollution in slim.
# TODO(shlens,skornblith): Determine the appropriate drop path schedule.
# For now the schedule is the default (1.0->0.7 over 250,000 train steps).
hparams = pnasnet.large_imagenet_config()
if not self._is_training:
hparams.set_hparam('drop_path_keep_prob', 1.0)
# Calculate the total number of cells in the network
total_num_cells = hparams.num_cells + num_stem_cells
normal_cell = pnasnet.PNasNetNormalCell(hparams.num_conv_filters,
hparams.drop_path_keep_prob,
total_num_cells,
hparams.total_training_steps)
with arg_scope([slim.dropout, nasnet_utils.drop_path],
is_training=self._is_training):
with arg_scope([slim.batch_norm],
is_training=self._train_batch_norm):
with arg_scope([
slim.avg_pool2d, slim.max_pool2d, slim.conv2d,
slim.batch_norm, slim.separable_conv2d,
nasnet_utils.factorized_reduction,
nasnet_utils.global_avg_pool,
nasnet_utils.get_channel_index,
nasnet_utils.get_channel_dim
],
data_format=hparams.data_format):
# This corresponds to the cell number just past 'Cell_7' used by
# _extract_proposal_features().
start_cell_num = 8
true_cell_num = start_cell_num + num_stem_cells
with slim.arg_scope(pnasnet.pnasnet_large_arg_scope()):
net = _build_pnasnet_base(
hidden_previous,
hidden,
normal_cell=normal_cell,
hparams=hparams,
true_cell_num=true_cell_num,
start_cell_num=start_cell_num)
proposal_classifier_features = net
return proposal_classifier_features
def build_computation_graph(self):
"""
notes on notation:
Symbolic variables have the prefix sy_, to distinguish them from the numerical values
that are computed later in the function
prefixes and suffixes:
ob - observation
ac - action
_no - this tensor should have shape (batch self.size /n/, observation dim)
_na - this tensor should have shape (batch self.size /n/, action dim)
_n - this tensor should have shape (batch self.size /n/)
Note: batch self.size /n/ is defined at runtime, and until then, the shape for that axis
is None
----------------------------------------------------------------------------------
loss: a function of self.sy_lp_n and self.sy_adv_n that we will differentiate
to get the policy gradient.
"""
self.sy_ob_no, self.sy_ac_na, self.sy_adv_n, self.sy_hidden, self.sy_lp_n, self.sy_fixed_lp_n = self.define_placeholders(
)
# The policy takes in an observation and produces a distribution over the action space
policy_outputs = self.policy_forward_pass(self.sy_ob_no,
self.sy_hidden)
self.policy_parameters = policy_outputs[:-1]
# unpack mean and variance
self.policy_parameters = tf.split(self.policy_parameters[0], 2, axis=1)
# We can sample actions from this action distribution.
# This will be called in Agent.sample_trajectory() where we generate a rollout.
self.sy_sampled_ac = self.sample_action(self.policy_parameters)
# We can also compute the logprob of the actions that were actually taken by the policy
# This is used in the loss function.
self.sy_lp_n = self.get_log_prob(self.policy_parameters, self.sy_ac_na)
# PPO critic update
critic_regularizer = tf.contrib.layers.l2_regularizer(
1e-3) if self.l2reg else None
self.critic_prediction = tf.squeeze(
build_critic(self.sy_ob_no,
self.sy_hidden,
1,
'critic_network',
n_layers=self.n_layers,
size=self.size,
gru_size=self.gru_size,
recurrent=self.recurrent,
regularizer=critic_regularizer))
self.sy_target_n = tf.placeholder(shape=[None],
name="critic_target",
dtype=tf.float32)
self.critic_loss = tf.losses.mean_squared_error(
self.sy_target_n, self.critic_prediction)
self.critic_weights = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
scope='critic_network')
self.critic_update_op = tf.train.AdamOptimizer(
self.learning_rate).minimize(self.critic_loss)
# PPO actor update
self.sy_fixed_log_prob_n = tf.placeholder(shape=[None],
name="fixed_log_prob",
dtype=tf.float32)
self.policy_surr_loss = self.ppo_loss(self.sy_lp_n, self.sy_fixed_lp_n,
self.sy_adv_n)
self.policy_weights = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
scope=self.scope)
optimizer = tf.train.AdamOptimizer(self.learning_rate)
self.policy_update_op = minimize_and_clip(optimizer,
self.policy_surr_loss,
var_list=self.policy_weights,
clip_val=40)
def __init__(self, num_words, num_chars, num_classes, num_steps, word_len, embedding_matrix=None):
# Parameter
self.config = Config()
self.dropout_rate = self.config.model_para['dropout_rate']
self.batch_size = self.config.model_para['batch_size']
self.num_layers = self.config.model_para['lstm_layer_num']
self.input_dim = self.config.model_para['input_dim']
self.hidden_dim = self.config.model_para['hidden_dim']
self.char_input_dim = self.config.model_para['char_input_dim']
self.char_hidden_dim = self.config.model_para['char_hidden_dim']
self.use_pa_learning = self.config.model_para['use_pa_learning']
self.embedding_matrix = embedding_matrix
self.word_len = word_len
self.num_steps = num_steps
self.num_words = num_words
self.num_chars = num_chars
self.num_classes = num_classes
self.char_inputs = tf.placeholder(tf.int32, [None, self.word_len])
with tf.variable_scope("character-based-emb"):
# char embedding
self.char_embedding = tf.get_variable("char_emb", [self.num_chars, self.char_input_dim])
self.char_inputs_emb = tf.nn.embedding_lookup(self.char_embedding, self.char_inputs)
self.char_inputs_emb = tf.transpose(self.char_inputs_emb, [1, 0, 2])
self.char_inputs_emb = tf.reshape(self.char_inputs_emb, [-1, self.char_input_dim])
self.char_inputs_emb = tf.split(self.char_inputs_emb, self.word_len, 0)
# char forward and backward
with tf.variable_scope("char-bi-lstm"):
# char lstm cell
char_lstm_cell_fw = rnn.LSTMCell(self.char_hidden_dim)
char_lstm_cell_bw = rnn.LSTMCell(self.char_hidden_dim)
# get the length of each word
self.word_length = tf.reduce_sum(tf.sign(self.char_inputs), reduction_indices=1)
self.word_length = tf.cast(self.word_length, tf.int32)
char_outputs, f_output, r_output = tf.contrib.rnn.static_bidirectional_rnn(
char_lstm_cell_fw,
char_lstm_cell_bw,
self.char_inputs_emb,
dtype=tf.float32,
sequence_length=self.word_length
)
final_word_output = tf.concat([f_output.h, r_output.h], -1)
self.word_lstm_last_output = tf.reshape(final_word_output, [-1, self.num_steps, self.char_hidden_dim*2])
# '''
# word input
# '''
with tf.variable_scope("word-based-emb"):
self.inputs = tf.placeholder(tf.int32, [None, self.num_steps])
# self.targets = tf.placeholder(tf.int32, [None, self.num_steps])
if self.use_pa_learning:
self.targets = tf.placeholder(tf.float32, [None, self.num_steps+2, self.num_classes+1])
else:
self.targets = tf.placeholder(tf.int32, [None, self.num_steps])
self.targets_transition = tf.placeholder(tf.int32, [None])
self.keep_prob = tf.placeholder(tf.float32)
if embedding_matrix is not None:
self.embedding = tf.Variable(embedding_matrix, trainable=True, name="word_emb", dtype=tf.float32)
else:
self.embedding = tf.get_variable("word_emb", [self.num_words, self.input_dim])
self.inputs_emb = tf.nn.embedding_lookup(self.embedding, self.inputs)
self.inputs_emb = tf.concat([self.inputs_emb, self.word_lstm_last_output], -1)
self.inputs_emb = tf.nn.dropout(self.inputs_emb, self.keep_prob)
self.inputs_emb = tf.transpose(self.inputs_emb, [1, 0, 2])
self.inputs_emb = tf.reshape(self.inputs_emb, [-1, self.input_dim+self.char_hidden_dim*2])
self.inputs_emb = tf.split(self.inputs_emb, self.num_steps, 0)
# word lstm cell
lstm_cell_fw = rnn.LSTMCell(self.hidden_dim)
lstm_cell_bw = rnn.LSTMCell(self.hidden_dim)
# get the length of each sample
self.length = tf.reduce_sum(tf.sign(self.inputs), reduction_indices=1)
self.length = tf.cast(self.length, tf.int32)
# forward and backward
with tf.variable_scope("word-bi-lstm"):
self.outputs, _, _ = tf.contrib.rnn.static_bidirectional_rnn(
lstm_cell_fw,
lstm_cell_bw,
self.inputs_emb,
dtype=tf.float32,
sequence_length=self.length
)
# bidirect concat
final_outputs = tf.reshape(tf.concat(self.outputs, 1), [-1, self.hidden_dim * 2])
tanh_layer_w = tf.get_variable("tanh_layer_w", [self.hidden_dim * 2, self.hidden_dim])
tanh_layer_b = tf.get_variable("tanh_layer_b", [self.hidden_dim])
self.final_outputs = tf.tanh(tf.matmul(final_outputs, tanh_layer_w) + tanh_layer_b)
# def add_placeholders(self):
# '''
# char input = sen_batch * sen_len
# '''
# self.char_inputs = tf.placeholder(tf.int32, [None, self.word_len])
# '''
# word input
# '''
# self.inputs = tf.placeholder(tf.int32, [None, self.num_steps])
# self.targets = tf.placeholder(tf.int32, [None, self.num_steps])
# self.targets_transition = tf.placeholder(tf.int32, [None])
# self.keep_prob = tf.placeholder(tf.float32)
# def add_lookup_op(self):
# with tf.variable_scope("character-based-emb"):
# # char embedding
# self.char_embedding = tf.get_variable("char_emb", [self.num_chars, self.char_input_dim])
# self.char_inputs_emb = tf.nn.embedding_lookup(self.char_embedding, self.char_inputs)
# with tf.variable_scope("word-based-emb"):
# if self.embedding_matrix is not None:
# self.embedding = tf.Variable(self.embedding_matrix, trainable=True, name="word_emb", dtype=tf.float32)
# else:
# self.embedding = tf.get_variable("word_emb", [self.num_words, self.input_dim])
# self.inputs_emb = tf.nn.embedding_lookup(self.embedding, self.inputs)
# def add_feature_extractor_op(self):
# with tf.variable_scope("char_bi-lstm"):
# self.char_inputs_emb = tf.transpose(self.char_inputs_emb, [1, 0, 2])
# self.char_inputs_emb = tf.reshape(self.char_inputs_emb, [-1, self.char_input_dim])
# self.char_inputs_emb = tf.split(self.char_inputs_emb, self.word_len, 0)
# # char lstm cell
# char_lstm_cell_fw = rnn.LSTMCell(self.char_hidden_dim)
# char_lstm_cell_bw = rnn.LSTMCell(self.char_hidden_dim)
# # get the length of each word
# self.word_length = tf.reduce_sum(tf.sign(self.char_inputs), reduction_indices=1)
# self.word_length = tf.cast(self.word_length, tf.int32)
# char_outputs, f_output, r_output = tf.contrib.rnn.static_bidirectional_rnn(
# char_lstm_cell_fw,
# char_lstm_cell_bw,
# self.char_inputs_emb,
# dtype=tf.float32,
# sequence_length=self.word_length
# )
# final_word_output = tf.concat([f_output.h, r_output.h], -1)
# self.word_lstm_last_output = tf.reshape(final_word_output, [-1, self.num_steps, self.char_hidden_dim*2])
# with tf.variable_scope("word_bi-lstm"):
# self.inputs_emb = tf.concat([self.inputs_emb, self.word_lstm_last_output], -1)
# self.inputs_emb = tf.nn.dropout(self.inputs_emb, self.keep_prob)
# self.inputs_emb = tf.transpose(self.inputs_emb, [1, 0, 2])
# self.inputs_emb = tf.reshape(self.inputs_emb, [-1, self.input_dim+self.char_hidden_dim*2])
# # self.inputs_emb = tf.reshape(self.inputs_emb, [-1, self.input_dim])
# self.inputs_emb = tf.split(self.inputs_emb, self.num_steps, 0)
# # word lstm cell
# lstm_cell_fw = rnn.LSTMCell(self.hidden_dim)
# lstm_cell_bw = rnn.LSTMCell(self.hidden_dim)
# # get the length of each sample
# self.length = tf.reduce_sum(tf.sign(self.inputs), reduction_indices=1)
# self.length = tf.cast(self.length, tf.int32)
# self.outputs, _, _ = tf.contrib.rnn.static_bidirectional_rnn(
# lstm_cell_fw,
# lstm_cell_bw,
# self.inputs_emb,
# dtype=tf.float32,
# sequence_length=self.length
# )
# with tf.variable_scope("bidirect-concat"):
# final_outputs = tf.reshape(tf.concat(self.outputs, 1), [-1, self.hidden_dim * 2])
# tanh_layer_w = tf.get_variable("tanh_layer_w", [self.hidden_dim * 2, self.hidden_dim])
# tanh_layer_b = tf.get_variable("tanh_layer_b", [self.hidden_dim])
# self.final_outputs = tf.tanh(tf.matmul(final_outputs, tanh_layer_w) + tanh_layer_b)
# def forward(self):
# self.add_placeholders()
# self.add_lookup_op()
# self.add_feature_extractor_op()
# return self.final_outputs, self.length
def run(self,
*in_arrays: Tuple[Union[np.ndarray, None], ...],
input_transform: dict = None,
output_transform: dict = None,
return_as_list: bool = False,
print_progress: bool = False,
minibatch_size: int = None,
num_gpus: int = 1,
assume_frozen: bool = False,
**dynamic_kwargs) -> Union[np.ndarray, Tuple[np.ndarray, ...], List[np.ndarray]]:
"""Run this network for the given NumPy array(s), and return the output(s) as NumPy array(s).
Args:
input_transform: A dict specifying a custom transformation to be applied to the input tensor(s) before evaluating the network.
The dict must contain a 'func' field that points to a top-level function. The function is called with the input
TensorFlow expression(s) as positional arguments. Any remaining fields of the dict will be passed in as kwargs.
output_transform: A dict specifying a custom transformation to be applied to the output tensor(s) after evaluating the network.
The dict must contain a 'func' field that points to a top-level function. The function is called with the output
TensorFlow expression(s) as positional arguments. Any remaining fields of the dict will be passed in as kwargs.
return_as_list: True = return a list of NumPy arrays, False = return a single NumPy array, or a tuple if there are multiple outputs.
print_progress: Print progress to the console? Useful for very large input arrays.
minibatch_size: Maximum minibatch size to use, None = disable batching.
num_gpus: Number of GPUs to use.
assume_frozen: Improve multi-GPU performance by assuming that the trainable parameters will remain changed between calls.
dynamic_kwargs: Additional keyword arguments to be passed into the network build function.
"""
assert len(in_arrays) == self.num_inputs
assert not all(arr is None for arr in in_arrays)
assert input_transform is None or util.is_top_level_function(input_transform["func"])
assert output_transform is None or util.is_top_level_function(output_transform["func"])
output_transform, dynamic_kwargs = _handle_legacy_output_transforms(output_transform, dynamic_kwargs)
num_items = in_arrays[0].shape[0]
if minibatch_size is None:
minibatch_size = num_items
# Construct unique hash key from all arguments that affect the TensorFlow graph.
key = dict(input_transform=input_transform, output_transform=output_transform, num_gpus=num_gpus, assume_frozen=assume_frozen, dynamic_kwargs=dynamic_kwargs)
def unwind_key(obj):
if isinstance(obj, dict):
return [(key, unwind_key(value)) for key, value in sorted(obj.items())]
if callable(obj):
return util.get_top_level_function_name(obj)
return obj
key = repr(unwind_key(key))
# Build graph.
if key not in self._run_cache:
with tfutil.absolute_name_scope(self.scope + "/_Run"), tf.control_dependencies(None):
with tf.device("/cpu:0"):
in_expr = [tf.placeholder(tf.float32, name=name) for name in self.input_names]
in_split = list(zip(*[tf.split(x, num_gpus) for x in in_expr]))
out_split = []
for gpu in range(num_gpus):
with tf.device("/gpu:%d" % gpu):
net_gpu = self.clone() if assume_frozen else self
in_gpu = in_split[gpu]
if input_transform is not None:
in_kwargs = dict(input_transform)
in_gpu = in_kwargs.pop("func")(*in_gpu, **in_kwargs)
in_gpu = [in_gpu] if tfutil.is_tf_expression(in_gpu) else list(in_gpu)
# assert len(in_gpu) == self.num_inputs
out_gpu = net_gpu.get_output_for(*in_gpu, return_as_list=True, **dynamic_kwargs)
if output_transform is not None:
out_kwargs = dict(output_transform)
out_gpu = out_kwargs.pop("func")(*out_gpu, **out_kwargs)
out_gpu = [out_gpu] if tfutil.is_tf_expression(out_gpu) else list(out_gpu)
# assert len(out_gpu) == self.num_outputs
out_split.append(out_gpu)
with tf.device("/cpu:0"):
out_expr = [tf.concat(outputs, axis=0) for outputs in zip(*out_split)]
self._run_cache[key] = in_expr, out_expr
# Run minibatches.
in_expr, out_expr = self._run_cache[key]
out_arrays = [np.empty([num_items] + expr.shape.as_list()[1:], expr.dtype.name) for expr in out_expr]
for mb_begin in range(0, num_items, minibatch_size):
if print_progress:
print("\r%d / %d" % (mb_begin, num_items), end="")
mb_end = min(mb_begin + minibatch_size, num_items)
mb_num = mb_end - mb_begin
mb_in = [src[mb_begin : mb_end] if src is not None else np.zeros([mb_num] + shape[1:]) for src, shape in zip(in_arrays, self.input_shapes)]
mb_out = tf.get_default_session().run(out_expr, dict(zip(in_expr, mb_in)))
for dst, src in zip(out_arrays, mb_out):
dst[mb_begin: mb_end] = src
# Done.
if print_progress:
print("\r%d / %d" % (num_items, num_items))
if not return_as_list:
out_arrays = out_arrays[0] if len(out_arrays) == 1 else tuple(out_arrays)
return out_arrays
def __init__(self, vgg16_npy_path=None, restore_from=None):
# pre-trained parameters
try:
self.data_dict = np.load(vgg16_npy_path, encoding='latin1').item()
except FileNotFoundError:
print(
'Please download VGG16 parameters at here https://mega.nz/#!YU1FWJrA!O1ywiCS2IiOlUCtCpI6HTJOMrneN-Qdv3ywQP5poecM'
)
self.tfx = tf.placeholder(tf.float32, [None, 224, 224, 3])
self.tfy = tf.placeholder(tf.float32, [None, 1])
# Convert RGB to BGR
red, green, blue = tf.split(axis=3,
num_or_size_splits=3,
value=self.tfx * 255.0)
bgr = tf.concat(axis=3,
values=[
blue - self.vgg_mean[0],
green - self.vgg_mean[1],
red - self.vgg_mean[2],
])
# pre-trained VGG layers are fixed in fine-tune
conv1_1 = self.conv_layer(bgr, "conv1_1")
conv1_2 = self.conv_layer(conv1_1, "conv1_2")
pool1 = self.max_pool(conv1_2, 'pool1')
conv2_1 = self.conv_layer(pool1, "conv2_1")
conv2_2 = self.conv_layer(conv2_1, "conv2_2")
pool2 = self.max_pool(conv2_2, 'pool2')
conv3_1 = self.conv_layer(pool2, "conv3_1")
conv3_2 = self.conv_layer(conv3_1, "conv3_2")
conv3_3 = self.conv_layer(conv3_2, "conv3_3")
pool3 = self.max_pool(conv3_3, 'pool3')
conv4_1 = self.conv_layer(pool3, "conv4_1")
conv4_2 = self.conv_layer(conv4_1, "conv4_2")
conv4_3 = self.conv_layer(conv4_2, "conv4_3")
pool4 = self.max_pool(conv4_3, 'pool4')
conv5_1 = self.conv_layer(pool4, "conv5_1")
conv5_2 = self.conv_layer(conv5_1, "conv5_2")
conv5_3 = self.conv_layer(conv5_2, "conv5_3")
pool5 = self.max_pool(conv5_3, 'pool5')
# detach original VGG fc layers and
# reconstruct your own fc layers serve for your own purpose
self.flatten = tf.reshape(pool5, [-1, 7 * 7 * 512])
self.fc6 = tf.layers.dense(self.flatten, 256, tf.nn.relu, name='fc6')
self.out = tf.layers.dense(self.fc6, 1, name='out')
self.sess = tf.Session()
if restore_from:
saver = tf.train.Saver()
saver.restore(self.sess, restore_from)
else: # training graph
self.loss = tf.losses.mean_squared_error(labels=self.tfy,
predictions=self.out)
self.train_op = tf.train.RMSPropOptimizer(0.001).minimize(
self.loss)
self.sess.run(tf.global_variables_initializer())
def extract_features(self, preprocessed_inputs, state_saver=None,
state_name='lstm_state', unroll_length=10, scope=None):
"""Extract features from preprocessed inputs.
The features include the base network features, lstm features and SSD
features, organized in the following name scope:
<scope>/MobilenetV2_1/...
<scope>/MobilenetV2_2/...
<scope>/LSTM/...
<scope>/FeatureMap/...
Args:
preprocessed_inputs: a [batch, height, width, channels] float tensor
representing a batch of consecutive frames from video clips.
state_saver: A state saver object with methods `state` and `save_state`.
state_name: Python string, the name to use with the state_saver.
unroll_length: number of steps to unroll the lstm.
scope: Scope for the base network of the feature extractor.
Returns:
feature_maps: a list of tensors where the ith tensor has shape
[batch, height_i, width_i, depth_i]
Raises:
ValueError: if interleave_method not recognized or large and small base
network output feature maps of different sizes.
"""
preprocessed_inputs = shape_utils.check_min_image_dim(
33, preprocessed_inputs)
preprocessed_inputs = ops.pad_to_multiple(
preprocessed_inputs, self._pad_to_multiple)
batch_size = preprocessed_inputs.shape[0].value / unroll_length
batch_axis = 0
nets = []
# Batch processing of mobilenet features.
with slim.arg_scope(mobilenet_v2.training_scope(
is_training=self._is_training,
bn_decay=0.9997)), \
slim.arg_scope([mobilenet.depth_multiplier],
min_depth=self._min_depth, divisible_by=8):
# Big model.
net, _ = self.extract_base_features_large(preprocessed_inputs)
nets.append(net)
large_base_feature_shape = net.shape
# Small models
net, _ = self.extract_base_features_small(preprocessed_inputs)
nets.append(net)
small_base_feature_shape = net.shape
if not (large_base_feature_shape[1] == small_base_feature_shape[1] and
large_base_feature_shape[2] == small_base_feature_shape[2]):
raise ValueError('Large and Small base network feature map dimension '
'not equal!')
with slim.arg_scope(self._conv_hyperparams_fn()):
with tf.variable_scope('LSTM', reuse=self._reuse_weights) as lstm_scope:
output_size = (large_base_feature_shape[1], large_base_feature_shape[2])
lstm_cell, init_state, step = self.create_lstm_cell(
batch_size, output_size, state_saver, state_name)
nets_seq = [
tf.split(net, unroll_length, axis=batch_axis) for net in nets
]
net_seq, states_out = rnn_decoder.multi_input_rnn_decoder(
nets_seq,
init_state,
lstm_cell,
step,
selection_strategy=self._interleave_method,
is_training=self._is_training,
pre_bottleneck=self._pre_bottleneck,
flatten_state=self._flatten_state,
scope=lstm_scope)
self._states_out = states_out
batcher_ops = None
if state_saver is not None:
self._step = state_saver.state(state_name + '_step')
batcher_ops = [
state_saver.save_state(state_name + '_c', states_out[-1][0]),
state_saver.save_state(state_name + '_h', states_out[-1][1]),
state_saver.save_state(state_name + '_step', self._step + 1)]
image_features = {}
with tf_ops.control_dependencies(batcher_ops):
image_features['layer_19'] = tf.concat(net_seq, 0)
# SSD layers.
with tf.variable_scope('FeatureMap'):
feature_maps = feature_map_generators.multi_resolution_feature_maps(
feature_map_layout=self._feature_map_layout,
depth_multiplier=self._depth_multiplier,
min_depth=self._min_depth,
insert_1x1_conv=True,
image_features=image_features,
pool_residual=True)
return feature_maps.values()
def forward(self, x):
def _NonLocalBlock(input_x,
out_channels,
sub_sample=1,
nltype=0,
is_bn=False,
scope='NonLocalBlock'):
"""
https://github.com/nnUyi/Non-Local_Nets-Tensorflow
"""
batchsize, height, width, in_channels = input_x.get_shape(
).as_list()
typedict = {
0: 'embedded_gaussian',
1: 'gaussian',
2: 'dot_product',
3: 'concat'
}
with tf.variable_scope(scope) as sc:
if nltype <= 2:
with tf.variable_scope('g') as scope:
g = conv2d(input_x,
out_channels,
1,
strides=1,
padding='same',
name='g')
if sub_sample > 1:
g = average_pooling2d(g,
pool_size=sub_sample,
strides=sub_sample,
name='g_pool')
with tf.variable_scope('phi') as scope:
if nltype == 0 or nltype == 2:
phi = conv2d(input_x,
out_channels,
1,
strides=1,
padding='same',
name='phi')
elif nltype == 1:
phi = input_x
if sub_sample > 1:
phi = average_pooling2d(phi,
pool_size=sub_sample,
strides=sub_sample,
name='phi_pool')
with tf.variable_scope('theta') as scope:
if nltype == 0 or nltype == 2:
theta = conv2d(input_x,
out_channels,
1,
strides=1,
padding='same',
name='theta')
elif nltype == 1:
theta = input_x
g_x = tf.reshape(g, [batchsize, -1, out_channels])
theta_x = tf.reshape(theta, [batchsize, -1, out_channels])
# theta_x = tf.reshape(theta, [batchsize, out_channels, -1])
# theta_x = tf.transpose(theta_x, [0,2,1])
phi_x = tf.reshape(phi, [batchsize, -1, out_channels])
phi_x = tf.transpose(phi_x, [0, 2, 1])
#phi_x = tf.reshape(phi_x, [batchsize, out_channels, -1])
f = tf.matmul(theta_x, phi_x)
# ???
if nltype <= 1:
# f_softmax = tf.nn.softmax(f, -1)
f = tf.exp(f)
f_softmax = f / tf.reduce_sum(
f, axis=-1, keepdims=True)
elif nltype == 2:
f = tf.nn.relu(f) #/int(f.shape[-1])
f_mean = tf.reduce_sum(f, axis=[2], keepdims=True)
#print(f.shape,f_mean.shape)
f_softmax = f / f_mean
y = tf.matmul(f_softmax, g_x)
y = tf.reshape(y, [batchsize, height, width, out_channels])
with tf.variable_scope('w') as scope:
w_y = conv2d(y,
in_channels,
1,
strides=1,
padding='same',
name='w')
# if is_bn:
# w_y = slim.batch_norm(w_y)
z = w_y #input_x + w_y
return z
"""
hyper-paras
"""
mf = 64 # output feature map num for most convs
dk = 3 # kernel size for most convs
ds = 1 # stride for most convs
activate = tf.nn.leaky_relu
num_block = self.num_block # progressive fusion block num
ki = tf.contrib.layers.xavier_initializer()
n, nf, w, h, c = x.shape
# print(n, nf, w, h, c)
with tf.variable_scope('network', reuse=tf.AUTO_REUSE) as scope:
conv0 = Conv2D(mf,
5,
strides=ds,
padding='same',
activation=activate,
kernel_initializer=ki,
name='conv0')
conv1 = [
Conv2D(mf,
dk,
strides=ds,
padding='same',
activation=activate,
kernel_initializer=ki,
name='conv1_{}'.format(i)) for i in range(num_block)
]
conv10 = [
Conv2D(mf,
1,
strides=ds,
padding='same',
activation=activate,
kernel_initializer=ki,
name='conv10_{}'.format(i)) for i in range(num_block)
]
conv2 = [
Conv2D(mf,
dk,
strides=ds,
padding='same',
activation=activate,
kernel_initializer=ki,
name='conv2_{}'.format(i)) for i in range(num_block)
]
convout4 = Conv2D(256,
3,
strides=ds,
padding='same',
activation=activate,
kernel_initializer=ki,
name='convout4')
convout3 = Conv2D(128,
3,
strides=ds,
padding='same',
activation=activate,
kernel_initializer=ki,
name='convout3')
convout2 = Conv2D(64,
3,
strides=ds,
padding='same',
activation=activate,
kernel_initializer=ki,
name='convout2')
convout1 = Conv2D(3,
3,
strides=ds,
padding='same',
activation=None,
kernel_initializer=ki,
name='convout1')
"""
center for residual add
"""
center_tensor = x[:,
self.num_frames // 2, :, :, :] # (bs, h, w, c)
# print(center_tensor.get_shape())
"""
nonlocal + res
"""
inp0 = [x[:, i, :, :, :]
for i in range(nf)] # [nf * (bs, h, w, c)]
inp0 = tf.concat(inp0, axis=-1) # (bs, h, w, c*nf)
# print(inp0.get_shape())
if self.refactor > 1:
inp1 = tf.space_to_depth(
inp0, self.refactor
) # space2depth: (h, w, c) -> (h//2, w//2, c*4)
else:
inp1 = inp0
inp1 = _NonLocalBlock(inp1,
int(c) * self.num_frames * self.refactor *
self.refactor,
sub_sample=1,
nltype=1,
scope='nlblock_{}'.format(0))
if self.refactor > 1:
inp1 = tf.depth_to_space(
inp1, self.refactor
) # depth2space: (h//2, w//2, c*4) -> (h, w, c)
inp0 += inp1 # (bs, h, w, c*nf)
"""
5x5 conv
"""
inp0 = tf.split(inp0, num_or_size_splits=self.num_frames,
axis=-1) # [nf * (bs, h, w, c)]
# print(len(inp0))
# print(inp0[0].get_shape())
inp0 = [conv0(f) for f in inp0] # [nf * (bs, h, w, 64)]
# print(inp0[0].get_shape())
"""
progressive fusion blocks
"""
for i in range(num_block):
inp1 = [conv1[i](f) for f in inp0] # [nf * (bs, h, w, 64)]
base = tf.concat(inp1, axis=-1) # (bs, h, w, 64*nf)
base = conv10[i](base)
inp2 = [tf.concat([base, f], -1) for f in inp1]
inp2 = [conv2[i](f) for f in inp2]
inp0 = [tf.add(inp0[j], inp2[j])
for j in range(nf)] # [nf * (bs, h, w, 64)]
"""
merge
"""
merge = tf.concat(inp0, axis=-1) # (bs, h, w, 64*nf)
out = convout4(merge)
out = convout3(out)
out = convout2(out)
out = convout1(out) # (bs, h, w, c)
"""
residual
"""
return tf.stack([out + center_tensor], axis=1,
name='out') # (bs, h, w, c) -> (bs, 1, h, w, c)
def _build_graph(self, inputs):
# ========================== Convert Color Space ==========================
lr_rgb, hr_rgb = inputs
lr_y, hr_y = rgb2y(lr_rgb), rgb2y(hr_rgb)
lr_ycbcr, hr_ycbcr = rgb2ycbcr(lr_rgb), rgb2ycbcr(hr_rgb)
# (b, t, h, w, c) to (b, h, w, c) * t
lr_y = tf.split(lr_y, cfg.frames, axis = 1)
lr_y = [tf.reshape(i, (-1, h, w, 1)) for i in lr_y]
lr_rgb = tf.split(lr_rgb, cfg.frames, axis = 1)
lr_rgb = [tf.reshape(i, (-1, h, w, 3)) for i in lr_rgb]
lr_ycbcr = tf.split(lr_ycbcr, cfg.frames, axis = 1)
lr_ycbcr = [tf.reshape(i, (-1, h, w, 3)) for i in lr_ycbcr]
# ========================== split ==========================
# ========================== Normalization ==========================
lr_y = [i / 255.0 - 0.5 for i in lr_y]
lr_rgb = [i / 255.0 - 0.5 for i in lr_rgb]
lr_ycbcr = [i / 255.0 - 0.5 for i in lr_ycbcr]
hr_y = hr_y / 255.0 - 0.5
referenced_rgb = lr_rgb[cfg.frames // 2]
referenced_y = lr_y[cfg.frames // 2]
ref_ycbcr = lr_ycbcr[cfg.frames // 2]
# ========================== Forward ==========================
hr_sparses = []
flows = []
warped = []
coords = get_coords(h, w)
with tf.variable_scope("ME_SPMC") as scope:
for i, j in zip(lr_y, lr_rgb):
flow_i0 = motion_estimation(referenced_y, i) * h / 2
# freeze in stage 2
if self.stage == 2:
flow_i0 = tf.stop_gradient(flow_i0)
flows.append(flow_i0)
hr_sparses.append(spmc_layer(i, flow_i0))
mapping = coords - flow_i0
backward_warped_img = BackwardWarping('backward_warpped', [referenced_y, mapping], borderMode='constant')
warped.append(backward_warped_img)
scope.reuse_variables()
hr_denses = detail_fusion_net(hr_sparses, ref_ycbcr[:, :, :, :1])
# ========================== Outputs ==========================
flow_after_reshape = [tf.reshape(i, (-1, 1, h, w, 2)) for i in flows]
tf_flows = tf.concat(flow_after_reshape, axis = 1, name = 'flows')
warped_after_reshape = [tf.reshape(i, (-1, 1, h, w, 1)) for i in warped]
after_warp = tf.concat(warped_after_reshape, axis = 1, name = 'after_warp')
padh = int(math.ceil(h / 4.0) * 4.0 - h)
padw = int(math.ceil(w / 4.0) * 4.0 - w)
scale_factor = 2
# Unormalization
output_y = (hr_denses[-1] + 0.5) * 255.
# Unormalization, bicubic插值
output_cbcr = tf.image.resize_images(
(ref_ycbcr + 0.5) * 255.0,
[(h + padh) * scale_factor, (w + padw) * scale_factor],
method = 2)[:, :, :, 1:3]
# Y: 模型输出 Cb&Cr: bicubic插值
prediction = tf.concat([output_y, output_cbcr], axis = -1)
# convert YCbCr to RGB
prediction = tf.identity(ycbcr2rgb(prediction), name = 'predictions')
# ========================== Cost Functions ==========================
k = np.arange(*k_range, 0.5 / cfg.frames)
mask_warped = []
warp_loss = []
mask_warp_loss = []
flow_loss = []
euclidean_loss = []
for i in range(cfg.frames):
mapping = coords - flows[i]
mask1 = tf.greater_equal(mapping[:,:,:,:1], 0.0)
mask2 = tf.less_equal(mapping[:,:,:,:1], h-1)
mask3 = tf.greater_equal(mapping[:,:,:,1:], 0.0)
mask4 = tf.less_equal(mapping[:,:,:,1:], w-1)
mask12 = tf.logical_and(mask1, mask2)
mask34 = tf.logical_and(mask3, mask4)
mask = tf.cast(tf.logical_and(mask12, mask34), tf.float32)
mask_warped.append(self._unorm(warped[i], mask))
mask_warp_loss.append(tf.reduce_sum(mask * tf.abs(lr_y[i] - warped[i])) / tf.reduce_sum(mask) * tf.reduce_sum(tf.ones_like(mask)))
warp_loss.append(tf.reduce_sum(tf.abs(lr_y[i] - warped[i])))
flow_loss.append(tf.reduce_sum(tf.abs(tf.image.total_variation(flows[i]))))
euclidean_loss.append(tf.reduce_sum(tf.square(hr_y - hr_denses[i])))
loss_me_1 = tf.reduce_sum([mask_warp_loss[i] for i in range(cfg.frames)])
loss_me_2 = tf.reduce_sum([flow_loss[i] for i in range(cfg.frames)])
loss_me = loss_me_1 + cfg.lambda1 * loss_me_2
loss_sr = tf.reduce_sum([k[i] * euclidean_loss[i] for i in range(cfg.frames)])
costs = [
loss_me,
loss_sr,
loss_sr + cfg.lambda2 * loss_me
]
self.cost = tf.identity(costs[self.stage - 1], name = 'cost')
# ========================================== Summary ==========================================
tf.summary.image('input', referenced_rgb, max_outputs = 3)
tf.summary.image('groundtruth', hr_rgb, max_outputs=3)
tf.summary.image('frame_pair_1', tf.concat([self._unorm(referenced_y), self._unorm(lr_y[0]), mask_warped[0]], axis=1), max_outputs=3)
tf.summary.image('frame_pair_2', tf.concat([self._unorm(referenced_y), self._unorm(lr_y[1]), mask_warped[1]], axis=1), max_outputs=3)
tf.summary.image('flow', flow_to_color(flows[0]), max_outputs=3)
# tf.summary.image('flow_1', tf.concat([flows[0][:,:,:,:1], flows[0][:,:,:,1:]], axis=1), max_outputs=3)
# tf.summary.image('flow_2', tf.concat([flows[1][:,:,:,:1], flows[1][:,:,:,1:]], axis=1), max_outputs=3)
# tf.summary.image('reference_frame', referenced, max_outputs=3)
tf.summary.image('output', prediction, max_outputs=3)
add_moving_summary(
# tf.identity(loss_me_1, name = 'warp_loss'),
# tf.identity(loss_me_2, name = 'flow_loss'),
tf.identity(loss_me, name = 'loss_me'),
tf.identity(loss_sr, name = 'loss_sr'),
self.cost
)
def _split_tensor(values, num_splits, axis):
if tf.__version__ == '0.12.0':
return tf.split(axis, num_splits, values)
else:
return tf.split(values, num_splits, axis=axis)
def run(
self,
*in_arrays,
return_as_list=False, # True = return a list of NumPy arrays, False = return a single NumPy array, or a tuple if there are multiple outputs.
print_progress=False, # Print progress to the console? Useful for very large input arrays.
minibatch_size=None, # Maximum minibatch size to use, None = disable batching.
num_gpus=1, # Number of GPUs to use.
out_mul=1.0, # Multiplicative constant to apply to the output(s).
out_add=0.0, # Additive constant to apply to the output(s).
out_shrink=1, # Shrink the spatial dimensions of the output(s) by the given factor.
out_dtype=None, # Convert the output to the specified data type.
**dynamic_kwargs
): # Additional keyword arguments to pass into the network construction function.
assert len(in_arrays) == self.num_inputs
num_items = in_arrays[0].shape[0]
if minibatch_size is None:
minibatch_size = num_items
key = str([
list(sorted(dynamic_kwargs.items())), num_gpus, out_mul, out_add,
out_shrink, out_dtype
])
# Build graph.
if key not in self._run_cache:
with absolute_name_scope(self.scope +
'/Run'), tf.control_dependencies(None):
in_split = list(
zip(*[tf.split(x, num_gpus)
for x in self.input_templates]))
out_split = []
for gpu in range(num_gpus):
with tf.device('/gpu:%d' % gpu):
out_expr = self.get_output_for(*in_split[gpu],
return_as_list=True,
**dynamic_kwargs)
if out_mul != 1.0:
out_expr = [x * out_mul for x in out_expr]
if out_add != 0.0:
out_expr = [x + out_add for x in out_expr]
if out_shrink > 1:
ksize = [1, 1, out_shrink, out_shrink]
out_expr = [
tf.nn.avg_pool(x,
ksize=ksize,
strides=ksize,
padding='VALID',
data_format='NCHW')
for x in out_expr
]
if out_dtype is not None:
if tf.as_dtype(out_dtype).is_integer:
out_expr = [tf.round(x) for x in out_expr]
out_expr = [
tf.saturate_cast(x, out_dtype)
for x in out_expr
]
out_split.append(out_expr)
self._run_cache[key] = [
tf.concat(outputs, axis=0) for outputs in zip(*out_split)
]
# Run minibatches.
out_expr = self._run_cache[key]
out_arrays = [
np.empty([num_items] + shape_to_list(expr.shape)[1:],
expr.dtype.name) for expr in out_expr
]
for mb_begin in range(0, num_items, minibatch_size):
if print_progress:
print('\r%d / %d' % (mb_begin, num_items), end='')
mb_end = min(mb_begin + minibatch_size, num_items)
mb_in = [src[mb_begin:mb_end] for src in in_arrays]
mb_out = tf.get_default_session().run(
out_expr, dict(zip(self.input_templates, mb_in)))
for dst, src in zip(out_arrays, mb_out):
dst[mb_begin:mb_end] = src
# Done.
if print_progress:
print('\r%d / %d' % (num_items, num_items))
if not return_as_list:
out_arrays = out_arrays[0] if len(out_arrays) == 1 else tuple(
out_arrays)
return out_arrays
def evaluate(n_token, cutoffs, ps_device):
##### Get input function and model function
eval_input_fn, eval_record_info = data_utils.get_input_fn(
record_info_dir=FLAGS.record_info_dir,
split=FLAGS.eval_split,
per_host_bsz=FLAGS.eval_batch_size,
tgt_len=FLAGS.tgt_len,
num_core_per_host=FLAGS.num_core_per_host,
num_hosts=1,
use_tpu=False)
num_batch = eval_record_info["num_batch"]
if FLAGS.max_eval_batch > 0:
num_batch = FLAGS.max_eval_batch
tf.logging.info("num of batches {}".format(num_batch))
##### Create computational graph
eval_set = eval_input_fn({
"batch_size": FLAGS.eval_batch_size,
"data_dir": FLAGS.data_dir
})
input_feed, label_feed = eval_set.make_one_shot_iterator().get_next()
inputs = tf.split(input_feed, FLAGS.num_core_per_host, 0)
labels = tf.split(label_feed, FLAGS.num_core_per_host, 0)
per_core_bsz = FLAGS.eval_batch_size // FLAGS.num_core_per_host
tower_mems, tower_losses, tower_new_mems = [], [], []
for i in range(FLAGS.num_core_per_host):
with tf.device(assign_to_gpu(i, ps_device)), \
tf.variable_scope(tf.get_variable_scope(), reuse=tf.AUTO_REUSE):
mems_i = [
tf.placeholder(tf.float32,
[FLAGS.mem_len, per_core_bsz, FLAGS.d_model])
for _ in range(FLAGS.n_layer)
]
loss_i, new_mems_i = single_core_graph(n_token=n_token,
cutoffs=cutoffs,
is_training=False,
inp=inputs[i],
tgt=labels[i],
mems=mems_i)
tower_mems.append(mems_i)
tower_losses.append(loss_i)
tower_new_mems.append(new_mems_i)
## sum losses across towers
if len(tower_losses) > 1:
loss = tf.add_n(tower_losses) / len(tower_losses)
else:
loss = tower_losses[0]
##### Evaluation loop
tower_mems_np = [[
np.zeros([FLAGS.mem_len, per_core_bsz, FLAGS.d_model],
dtype=np.float32) for layer in range(FLAGS.n_layer)
] for core in range(FLAGS.num_core_per_host)]
saver = tf.train.Saver()
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True)) as sess:
sess.run(tf.global_variables_initializer())
if FLAGS.eval_ckpt_path is None:
eval_ckpt_path = tf.train.latest_checkpoint(FLAGS.model_dir)
else:
eval_ckpt_path = FLAGS.eval_ckpt_path
tf.logging.info("Evaluate {}".format(eval_ckpt_path))
saver.restore(sess, eval_ckpt_path)
fetches = [loss, tower_new_mems, tf.size(label_feed)]
format_str = " >> processing batch {{:{0}d}}/{{:{0}d}} ..".format(
len(str(num_batch)))
total_loss, total_cnt = 0, 0
for step in range(num_batch):
if step % (num_batch // 10) == 0:
tf.logging.info(format_str.format(step, num_batch))
feed_dict = {}
for i in range(FLAGS.num_core_per_host):
for m, m_np in zip(tower_mems[i], tower_mems_np[i]):
feed_dict[m] = m_np
fetched = sess.run(fetches, feed_dict=feed_dict)
loss_np, tower_mems_np, cnt_np = fetched[:3]
total_loss += loss_np * cnt_np
total_cnt += cnt_np
avg_loss = total_loss / total_cnt
tf.logging.info("| loss {:.2f} | pplx {:>7.2f}, bpc {:>7.4f}".format(
avg_loss, math.exp(avg_loss), avg_loss / math.log(2)))
def pretrain_model(epochs, batch_size, train_steps_per_epoch, save_dir):
# step 1: prepare dataset
train_data, test_data = load_data()
pipeline = fe.Pipeline(
train_data=train_data,
batch_size=batch_size,
ops=[
PadIfNeeded(min_height=40,
min_width=40,
image_in="x",
image_out="x"),
# augmentation 1
RandomCrop(32, 32, image_in="x", image_out="x_aug"),
Sometimes(HorizontalFlip(image_in="x_aug", image_out="x_aug"),
prob=0.5),
Sometimes(ColorJitter(inputs="x_aug",
outputs="x_aug",
brightness=0.8,
contrast=0.8,
saturation=0.8,
hue=0.2),
prob=0.8),
Sometimes(ToGray(inputs="x_aug", outputs="x_aug"), prob=0.2),
Sometimes(GaussianBlur(inputs="x_aug",
outputs="x_aug",
blur_limit=(3, 3),
sigma_limit=(0.1, 2.0)),
prob=0.5),
ToFloat(inputs="x_aug", outputs="x_aug"),
# augmentation 2
RandomCrop(32, 32, image_in="x", image_out="x_aug2"),
Sometimes(HorizontalFlip(image_in="x_aug2", image_out="x_aug2"),
prob=0.5),
Sometimes(ColorJitter(inputs="x_aug2",
outputs="x_aug2",
brightness=0.8,
contrast=0.8,
saturation=0.8,
hue=0.2),
prob=0.8),
Sometimes(ToGray(inputs="x_aug2", outputs="x_aug2"), prob=0.2),
Sometimes(GaussianBlur(inputs="x_aug2",
outputs="x_aug2",
blur_limit=(3, 3),
sigma_limit=(0.1, 2.0)),
prob=0.5),
ToFloat(inputs="x_aug2", outputs="x_aug2")
])
# step 2: prepare network
model_con, model_finetune = fe.build(model_fn=ResNet9,
optimizer_fn=["adam", "adam"])
network = fe.Network(ops=[
LambdaOp(lambda x, y: tf.concat([x, y], axis=0),
inputs=["x_aug", "x_aug2"],
outputs="x_com"),
ModelOp(model=model_con, inputs="x_com", outputs="y_com"),
LambdaOp(lambda x: tf.split(x, 2, axis=0),
inputs="y_com",
outputs=["y_pred", "y_pred2"]),
NTXentOp(arg1="y_pred",
arg2="y_pred2",
outputs=["NTXent", "logit", "label"]),
UpdateOp(model=model_con, loss_name="NTXent")
])
# step 3: prepare estimator
traces = [
Accuracy(true_key="label",
pred_key="logit",
mode="train",
output_name="contrastive_accuracy"),
ModelSaver(model=model_con, save_dir=save_dir),
]
estimator = fe.Estimator(pipeline=pipeline,
network=network,
epochs=epochs,
traces=traces,
train_steps_per_epoch=train_steps_per_epoch)
estimator.fit()
return model_con, model_finetune
