def _noisy_identity_kernel_initializer(shape,
dtype=tf.float32,
partition_info=None):
"""Constructs a noisy identity kernel.
Args:
shape: List of integers. Represents shape of result.
dtype: data type for values in result.
partition_info: Partition information for initializer functions. Ignored.
Returns:
Tensor of desired shape and dtype such that applying it as a convolution
kernel results in a noisy near-identity operation.
Raises:
ValueError: If shape does not define a valid kernel.
If filter width and height differ.
If filter width and height are not odd numbers.
If number of input and output channels are not multiples of
base_num_channels.
"""
if len(shape) != 4:
raise ValueError("Convolution kernels must be rank 4.")
filter_height, filter_width, in_channels, out_channels = shape
if filter_width != filter_height:
raise ValueError(
"Noisy identity initializer only works for square filters.")
if filter_width % 2 != 1:
raise ValueError(
"Noisy identity initializer requires filters have odd height and "
"width.")
if (in_channels % base_num_channels != 0 or
out_channels % base_num_channels != 0):
raise ValueError("in_channels and out_channels must both be multiples of "
"base_num_channels.")
middle_pixel = filter_height // 2
is_middle_pixel = tf.logical_and(
tf.equal(_range_along_dimension(0, shape), middle_pixel),
tf.equal(_range_along_dimension(1, shape), middle_pixel))
is_same_channel_multiple = tf.equal(
tf.floordiv(
_range_along_dimension(2, shape) * base_num_channels, in_channels),
tf.floordiv(
_range_along_dimension(3, shape) * base_num_channels, out_channels))
noise = tf.truncated_normal(shape, stddev=stddev, dtype=dtype)
return tf.where(
tf.logical_and(is_same_channel_multiple, is_middle_pixel),
tf.ones(
shape, dtype=dtype) * (base_num_channels / out_channels),
noise)
def _extract_feature(inputs, idxs):
idxs = tf.expand_dims(idxs,1)
idx_i = tf.floordiv(idxs, map_h)
idx_j = tf.mod(idxs, map_h)
# NOTE:
# calculate the center of input batches
# this depends on coarse layer's architecture
origin_i = 2*(2*idx_i+1)+3
origin_j = 2*(2*idx_j+1)+3
origin_centers = tf.concat(1,[origin_i,origin_j])
origin_centers = tf.to_float(origin_centers)
# NOTE: size also depends on the architecture
patches = tf.image.extract_glimpse(inputs, size=[14,14], offsets=origin_centers,
centered=False, normalized=False)
fine_features = fine_layers(patches)
# reuse variables
tf.get_variable_scope().reuse_variables()
src_idxs = tf.concat(1,[idx_i,idx_j])
return fine_features, src_idxs
def compute_length_after_conv(max_time_steps, ctc_time_steps, input_length):
"""Computes the time_steps/ctc_input_length after convolution.
Suppose that the original feature contains two parts:
1) Real spectrogram signals, spanning input_length steps.
2) Padded part with all 0s.
The total length of those two parts is denoted as max_time_steps, which is
the padded length of the current batch. After convolution layers, the time
steps of a spectrogram feature will be decreased. As we know the percentage
of its original length within the entire length, we can compute the time steps
for the signal after conv as follows (using ctc_input_length to denote):
ctc_input_length = (input_length / max_time_steps) * output_length_of_conv.
This length is then fed into ctc loss function to compute loss.
Args:
max_time_steps: max_time_steps for the batch, after padding.
ctc_time_steps: number of timesteps after convolution.
input_length: actual length of the original spectrogram, without padding.
Returns:
the ctc_input_length after convolution layer.
"""
ctc_input_length = tf.to_float(tf.multiply(
input_length, ctc_time_steps))
return tf.to_int32(tf.floordiv(
ctc_input_length, tf.to_float(max_time_steps)))
def int_to_bit(x_int, nbits):
"""Turn x_int representing numbers into a bitwise (lower-endian) tensor."""
x_l = tf.expand_dims(x_int, axis=-1)
x_labels = []
for i in range(nbits):
x_labels.append(tf.floormod(tf.floordiv(x_l, 2**i), 2))
res = tf.concat(x_labels, axis=-1)
return tf.to_float(res)
def get_staves(self, page):
filt = staff_center_filter(page)
n = self.num_slices
increment = tf.floordiv(tf.shape(filt)[1], tf.constant(n + 1, tf.int32))
staff_sections = []
for i in range(n):
img = filt[:, i*increment:(i+1)*increment if i + 1 < n else None]
staff_sections.append(self.detect(img))
results = tf.py_func(_join_staves, [page.staff_dist] + staff_sections,
[staff_sections[0].dtype])
return results
def ae_latent_softmax(latents_pred, latents_discrete, hparams):
"""Latent prediction and loss."""
vocab_size = 2 ** hparams.z_size
if hparams.num_decode_blocks < 2:
latents_logits = tf.layers.dense(latents_pred, vocab_size,
name="extra_logits")
if hparams.logit_normalization:
latents_logits *= tf.rsqrt(1e-8 +
tf.reduce_mean(tf.square(latents_logits)))
loss = None
if latents_discrete is not None:
if hparams.soft_em:
# latents_discrete is actually one-hot of multinomial samples
assert hparams.num_decode_blocks == 1
loss = tf.nn.softmax_cross_entropy_with_logits_v2(
labels=latents_discrete, logits=latents_logits)
else:
loss = tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=latents_discrete, logits=latents_logits)
sample = multinomial_sample(
latents_logits, vocab_size, hparams.sampling_temp)
return sample, loss
# Multi-block case.
vocab_bits = int(math.log(vocab_size, 2))
assert vocab_size == 2**vocab_bits
assert vocab_bits % hparams.num_decode_blocks == 0
block_vocab_size = 2**(vocab_bits // hparams.num_decode_blocks)
latents_logits = [
tf.layers.dense(
latents_pred, block_vocab_size, name="extra_logits_%d" % i)
for i in range(hparams.num_decode_blocks)
]
loss = None
if latents_discrete is not None:
losses = []
for i in range(hparams.num_decode_blocks):
d = tf.floormod(tf.floordiv(latents_discrete,
block_vocab_size**i), block_vocab_size)
losses.append(tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=d, logits=latents_logits[i]))
loss = sum(losses)
samples = [multinomial_sample(l, block_vocab_size, hparams.sampling_temp)
for l in latents_logits]
sample = sum([s * block_vocab_size**i for i, s in enumerate(samples)])
return sample, loss
def get_angle(page):
img = tf.cast(page.image, tf.float32)
square = get_square(img)
f = tf.complex_abs(tf.fft2d(tf.cast(square, tf.complex64))[:MAX_SIZE//2, :])
x_arr = (
tf.cast(tf.concat(0,
[tf.range(MAX_SIZE // 2),
tf.range(1, MAX_SIZE // 2 + 1)[::-1]]),
tf.float32))[None, :]
y_arr = tf.cast(tf.range(MAX_SIZE // 2), tf.float32)[:, None]
f = tf.select(x_arr * x_arr + y_arr * y_arr < 32 * 32, tf.zeros_like(f), f)
m = tf.argmax(tf.reshape(f, [-1]), dimension=0)
x = tf.cast((m + MAX_SIZE // 4) % (MAX_SIZE // 2) - (MAX_SIZE // 4), tf.float32)
y = tf.cast(tf.floordiv(m, MAX_SIZE // 2), tf.float32)
return(tf.cond(
y > 0, lambda: tf.atan(x / y), lambda: tf.constant(np.nan, tf.float32)),
square)
def compute_progress(current_image_id, stable_stage_num_images,
transition_stage_num_images, num_blocks):
"""Computes the training progress.
The training alternates between stable phase and transition phase.
The `progress` indicates the training progress, i.e. the training is at
- a stable phase p if progress = p
- a transition stage between p and p + 1 if progress = p + fraction
where p = 0,1,2.,...
Note the max value of progress is `num_blocks` - 1.
In terms of LOD (of the original implementation):
progress = `num_blocks` - 1 - LOD
Args:
current_image_id: An scalar integer `Tensor` of the current image id, count
from 0.
stable_stage_num_images: An integer representing the number of images in
each stable stage.
transition_stage_num_images: An integer representing the number of images in
each transition stage.
num_blocks: Number of network blocks.
Returns:
A scalar float `Tensor` of the training progress.
"""
# Note when current_image_id >= min_total_num_images - 1 (which means we
# are already at the highest resolution), we want to keep progress constant.
# Therefore, cap current_image_id here.
capped_current_image_id = tf.minimum(
current_image_id,
min_total_num_images(stable_stage_num_images, transition_stage_num_images,
num_blocks) - 1)
stage_num_images = stable_stage_num_images + transition_stage_num_images
progress_integer = tf.floordiv(capped_current_image_id, stage_num_images)
progress_fraction = tf.maximum(
0.0,
tf.to_float(
tf.mod(capped_current_image_id, stage_num_images) -
stable_stage_num_images) / tf.to_float(transition_stage_num_images))
return tf.to_float(progress_integer) + progress_fraction
def build_network(self): # build the network for one BP iteration
# BP initialization
llr_into_bp_net = tf.Variable(np.ones([self.v_node_num, self.batch_size], dtype=np.float32))
xe_0 = tf.matmul(self.H_x_to_xe0, llr_into_bp_net)
xe_v2c_pre_iter = tf.Variable(np.ones([self.num_all_edges, self.batch_size], dtype=np.float32)) # the v->c messages of the previous iteration
xe_v2c_pre_iter_assign = xe_v2c_pre_iter.assign(xe_0)
# one iteration
H_sumC_to_V = tf.constant(self.H_sumC_to_V, dtype=tf.float32)
H_sumV_to_C = tf.constant(self.H_sumV_to_C, dtype=tf.float32)
xe_v_sumc, xe_c_sumv = self.one_bp_iteration(xe_v2c_pre_iter, H_sumC_to_V, H_sumV_to_C, xe_0)
# start the next iteration
start_next_iteration = xe_v2c_pre_iter.assign(xe_c_sumv)
# get the final marginal probability and decoded results
bp_out_llr = tf.add(llr_into_bp_net, tf.matmul(self.H_xe_v_sumc_to_y, xe_v_sumc))
dec_out = tf.transpose(tf.floordiv(1-tf.to_int32(tf.sign(bp_out_llr)), 2))
return llr_into_bp_net, xe_0, xe_v2c_pre_iter_assign, start_next_iteration, dec_out
def int_to_bit(x_int, num_bits, base=2):
"""Turn x_int representing numbers into a bitwise (lower-endian) tensor.
Args:
x_int: Tensor containing integer to be converted into base notation.
num_bits: Number of bits in the representation.
base: Base of the representation.
Returns:
Corresponding number expressed in base.
"""
x_l = tf.to_int32(tf.expand_dims(x_int, axis=-1))
x_labels = []
for i in range(num_bits):
x_labels.append(
tf.floormod(
tf.floordiv(tf.to_int32(x_l),
tf.to_int32(base)**i), tf.to_int32(base)))
res = tf.concat(x_labels, axis=-1)
return tf.to_float(res)
def beam_step(time, beam_probs, beam_seqs, cand_probs, cand_seqs, *states):
batch_size = tf.shape(beam_probs)[0]
inputs = tf.reshape(tf.slice(beam_seqs, [0, time], [batch_size, 1]), [batch_size])
decoder_input = embedding_ops.embedding_lookup(self.L_dec, inputs)
decoder_output, state_output = self.decoder_graph(decoder_input, states)
with vs.variable_scope("Logistic", reuse=True):
do2d = tf.reshape(decoder_output, [-1, self.size])
logits2d = rnn_cell._linear(do2d, self.vocab_size, True, 1.0)
logprobs2d = tf.nn.log_softmax(logits2d)
total_probs = logprobs2d + tf.reshape(beam_probs, [-1, 1])
total_probs_noEOS = tf.concat(1, [tf.slice(total_probs, [0, 0], [batch_size, nlc_data.EOS_ID]),
tf.tile([[-3e38]], [batch_size, 1]),
tf.slice(total_probs, [0, nlc_data.EOS_ID + 1],
[batch_size, self.vocab_size - nlc_data.EOS_ID - 1])])
flat_total_probs = tf.reshape(total_probs_noEOS, [-1])
beam_k = tf.minimum(tf.size(flat_total_probs), self.beam_size)
next_beam_probs, top_indices = tf.nn.top_k(flat_total_probs, k=beam_k)
next_bases = tf.floordiv(top_indices, self.vocab_size)
next_mods = tf.mod(top_indices, self.vocab_size)
next_states = [tf.gather(state, next_bases) for state in state_output]
next_beam_seqs = tf.concat(1, [tf.gather(beam_seqs, next_bases),
tf.reshape(next_mods, [-1, 1])])
cand_seqs_pad = tf.pad(cand_seqs, [[0, 0], [0, 1]])
beam_seqs_EOS = tf.pad(beam_seqs, [[0, 0], [0, 1]])
new_cand_seqs = tf.concat(0, [cand_seqs_pad, beam_seqs_EOS])
EOS_probs = tf.slice(total_probs, [0, nlc_data.EOS_ID], [batch_size, 1])
new_cand_probs = tf.concat(0, [cand_probs, tf.reshape(EOS_probs, [-1])])
cand_k = tf.minimum(tf.size(new_cand_probs), self.beam_size)
next_cand_probs, next_cand_indices = tf.nn.top_k(new_cand_probs, k=cand_k)
next_cand_seqs = tf.gather(new_cand_seqs, next_cand_indices)
return [time + 1, next_beam_probs, next_beam_seqs, next_cand_probs, next_cand_seqs] + next_states
def ae_latent_softmax(latents_pred, latents_discrete, hparams):
"""Latent prediction and loss."""
vocab_size = hparams.v_size
if hparams.bottleneck_kind == "semhash":
vocab_size = 2**hparams.z_size
if hparams.num_decode_blocks < 2:
latents_logits = tf.layers.dense(latents_pred, vocab_size,
name="extra_logits")
loss = None
if latents_discrete is not None:
loss = tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=latents_discrete, logits=latents_logits)
sample = multinomial_sample(
latents_logits, vocab_size, hparams.sampling_temp)
return sample, loss
# Multi-block case.
vocab_bits = int(math.log(vocab_size, 2))
assert vocab_size == 2**vocab_bits
assert vocab_bits % hparams.num_decode_blocks == 0
block_vocab_size = 2**(vocab_bits // hparams.num_decode_blocks)
latents_logits = [
tf.layers.dense(
latents_pred, block_vocab_size, name="extra_logits_%d" % i)
for i in xrange(hparams.num_decode_blocks)
]
loss = None
if latents_discrete is not None:
losses = []
for i in xrange(hparams.num_decode_blocks):
d = tf.floormod(tf.floordiv(latents_discrete,
block_vocab_size**i), block_vocab_size)
losses.append(tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=d, logits=latents_logits[i]))
loss = sum(losses)
samples = [multinomial_sample(l, block_vocab_size, hparams.sampling_temp)
for l in latents_logits]
sample = sum([s * block_vocab_size**i for i, s in enumerate(samples)])
return sample, loss
def _calc_ctc_input_length(args):
"""Compute the actual input length after convolution for ctc_loss function.
Basically, we need to know the scaled input_length after conv layers.
new_input_length = old_input_length * ctc_time_steps / max_time_steps
Args:
args: the input args to compute ctc input length.
Returns:
ctc_input_length, which is required for ctc loss calculation.
"""
# py2 needs explicit tf import for keras Lambda layer
import tensorflow as tf
input_length, input_data, y_pred = args
max_time_steps = tf.shape(input_data)[1]
ctc_time_steps = tf.shape(y_pred)[1]
ctc_input_length = tf.multiply(
tf.to_float(input_length), tf.to_float(ctc_time_steps))
ctc_input_length = tf.to_int32(tf.floordiv(
ctc_input_length, tf.to_float(max_time_steps)))
return ctc_input_length
def _extract_feature(inputs, idxs):
idxs = tf.expand_dims(idxs,1)
idx_i = tf.floordiv(idxs, map_h)
idx_j = tf.mod(idxs, map_h)
# NOTE: the below origins are starting points, not center!
origin_i = 2*(2*idx_i+1)+3 - 5 + 2
origin_j = 2*(2*idx_j+1)+3 - 5 + 2
origin_centers = tf.concat(1,[origin_i,origin_j])
# NOTE: size also depends on the architecture
#patches = tf.image.extract_glimpse(inputs, size=[14,14], offsets=origin_centers,
# centered=False, normalized=False)
patches = extract_patches(inputs, size=[14,14], offsets=origin_centers)
#fine_features = fine_layers(patches)
fine_features = []
src_idxs = tf.concat(1,[idx_i,idx_j])
return fine_features, src_idxs, patches
def modgrad(op, grad):
x = op.inputs[
0] # the first argument (normally you need those to calculate the gradient, like the gradient of x^2 is 2x. )
y = op.inputs[1] # the second argument
return grad * 1, grad * tf.negative(tf.floordiv(x, y))
def resize_real_images(image, params):
"""Resizes real images to match the GAN's current size.
Args:
image: tensor, original image.
params: dict, user passed parameters.
Returns:
Resized image tensor.
"""
func_name = "resize_real_images"
print_obj("\n" + func_name, "image", image)
# Resize real image for each block.
if len(params["conv_num_filters"]) == 1:
print("\n: NEVER GOING TO GROW, SKIP SWITCH CASE!".format(func_name))
# If we never are going to grow, no sense using the switch case.
# 4x4
resized_image = resize_real_image(image=image,
params=params,
block_idx=0)
else:
if params["growth_idx"] is not None:
block_idx = min((params["growth_idx"] - 1) // 2 + 1,
len(params["conv_num_filters"]) - 1)
resized_image = resize_real_image(image=image,
params=params,
block_idx=block_idx)
else:
# Find growth index based on global step and growth frequency.
growth_index = tf.add(x=tf.floordiv(
x=tf.minimum(x=tf.cast(x=tf.floordiv(
x=tf.train.get_or_create_global_step() - 1,
y=params["num_steps_until_growth"],
name="{}_global_step_floordiv".format(func_name)),
dtype=tf.int32),
y=(len(params["conv_num_filters"]) - 1) * 2) - 1,
y=2),
y=1,
name="{}_growth_index".format(func_name))
# Switch to case based on number of steps for resized image.
resized_image = tf.switch_case(
branch_index=growth_index,
branch_fns=[
# 4x4
lambda: resize_real_image(
image=image, params=params, block_idx=0),
# 8x8
lambda: resize_real_image(
image=image,
params=params,
block_idx=min(1,
len(params["conv_num_filters"]) - 1)),
# 16x16
lambda: resize_real_image(
image=image,
params=params,
block_idx=min(2,
len(params["conv_num_filters"]) - 1)),
# 32x32
lambda: resize_real_image(
image=image,
params=params,
block_idx=min(3,
len(params["conv_num_filters"]) - 1)),
# 64x64
lambda: resize_real_image(
image=image,
params=params,
block_idx=min(4,
len(params["conv_num_filters"]) - 1)),
# 128x128
lambda: resize_real_image(
image=image,
params=params,
block_idx=min(5,
len(params["conv_num_filters"]) - 1)),
# 256x256
lambda: resize_real_image(
image=image,
params=params,
block_idx=min(6,
len(params["conv_num_filters"]) - 1)),
# 512x512
lambda: resize_real_image(
image=image,
params=params,
block_idx=min(7,
len(params["conv_num_filters"]) - 1)),
# 1024x1024
lambda: resize_real_image(
image=image,
params=params,
block_idx=min(8,
len(params["conv_num_filters"]) - 1))
],
name="{}_switch_case_resized_image".format(func_name))
print_obj(func_name, "resized_image", resized_image)
return resized_image
def hier_homography_estimator(inputs,
num_param=8,
num_layer=7,
num_level=3,
dropout_keep_prob=0.8,
reuse=None,
is_training=True,
trainable=True,
final_endpoint=None,
scope='hier_hmg'):
"""A hierarchical VGG-style neural network for homograhy estimation.
Args:
inputs: batch of input image pairs of data type float32 and of shape
[batch_size, height, width, 2]
num_param: the number of parameters for homography (default 8)
num_layer: the number of convolutional layers in the motion feature network
num_level: the number of hierarchical levels
dropout_keep_prob: the percentage of activation values that are kept
reuse: whether to reuse this network weights
is_training: whether used for training or testing
trainable: whether this network is to be trained or not
final_endpoint: specifies the endpoint to construct the network up to
scope: the scope of variables in this function
Returns:
a list of homographies at each level and motion feature maps if
final_endpoint='mfeature'; otherwise a list of images warped by the list of
corresponding homographies
"""
_, h_input, w_input = inputs.get_shape().as_list()[0:3]
hmgs_list = []
warped_list = []
with tf.variable_scope(scope, [inputs], reuse=reuse):
for level_index in range(num_level):
scale = 2**(num_level - 1 - level_index)
h = tf.to_float(tf.floordiv(h_input, scale))
w = tf.to_float(tf.floordiv(w_input, scale))
inputs_il = tf.image.resize_images(inputs, tf.to_int32([h, w]))
if level_index == 0:
mfeature = hier_base_layers(inputs_il,
num_layer + 1 - num_level +
level_index,
level_index,
is_training=is_training,
trainable=trainable)
hmgs_il = homography_regression(
mfeature,
num_param,
level_index,
dropout_keep_prob=dropout_keep_prob,
is_training=is_training,
trainable=trainable)
hmgs_list.append(hmgs_il)
else:
warped, _ = hmg_util.homography_scale_warp_per_batch(
inputs_il[:, :, :, 0], w / 2, h / 2,
hmgs_list[level_index - 1])
pre_warped_inputs_il = tf.stack(
[warped, inputs_il[:, :, :, 1]], -1)
warped_list.append(pre_warped_inputs_il)
if level_index == num_level - 1 and final_endpoint == 'mfeature':
mfeature = hier_base_layers(pre_warped_inputs_il,
num_layer - num_level +
level_index,
level_index,
is_training=is_training,
trainable=trainable)
return hmgs_list, mfeature
else:
mfeature = hier_base_layers(pre_warped_inputs_il,
num_layer + 1 - num_level +
level_index,
level_index,
is_training=is_training,
trainable=trainable)
hmgs_il = homography_regression(
mfeature,
num_param,
level_index,
dropout_keep_prob=dropout_keep_prob,
is_training=is_training,
trainable=trainable)
new_hmgs_il = hmg_util.homography_shift_mult_batch(
hmgs_list[level_index - 1], w / 2, h / 2, hmgs_il, w, h, w,
h)
hmgs_list.append(new_hmgs_il)
return hmgs_list, warped_list
def build_neurosat(d):
# Hyperparameters
learning_rate = 2e-5
parameter_l2norm_scaling = 1e-10
global_norm_gradient_clipping_ratio = 0.65
# Define placeholder for satisfiability statuses (one per problem)
instance_SAT = tf.placeholder(tf.float32, [None], name="instance_SAT")
time_steps = tf.placeholder(tf.int32, shape=(), name='time_steps')
matrix_placeholder = tf.placeholder(tf.float32, [None, None], name="M")
num_vars_on_instance = tf.placeholder(tf.int32, [None], name="instance_n")
# Literals
s = tf.shape(matrix_placeholder)
l = s[0]
m = s[1]
n = tf.floordiv(l, tf.constant(2))
# Compute number of problems
p = tf.shape(instance_SAT)[0]
# Define INV, a tf function to exchange positive and negative literal embeddings
def INV(Lh):
l = tf.shape(Lh)[0]
n = tf.div(l, tf.constant(2))
# Send messages from negated literals to positive ones, and vice-versa
Lh_pos = tf.gather(Lh, tf.range(tf.constant(0), n))
Lh_neg = tf.gather(Lh, tf.range(n, l))
Lh_inverted = tf.concat([Lh_neg, Lh_pos], axis=0)
return Lh_inverted
#end
var = {"L": d, "C": d}
s = tf.shape(matrix_placeholder)
num_vars = {"L": l, "C": m}
initial_embeddings = {
v: tf.get_variable(initializer=tf.random_normal((1, d)),
dtype=tf.float32,
name='{v}_init'.format(v=v))
for (v, d) in var.items()
}
tiled_and_normalized_initial_embeddings = {
v: tf.tile(tf.div(init, tf.sqrt(tf.cast(var[v], tf.float32))),
[num_vars[v], 1])
for v, init in initial_embeddings.items()
}
# Define Graph neural network
gnn = GraphNN(var, {"M": ("L", "C")}, {
"Lmsg": ("L", "C"),
"Cmsg": ("C", "L")
}, {
"L": [{
"fun": INV,
"var": "L"
}, {
"mat": "M",
"msg": "Cmsg",
"var": "C"
}],
"C": [{
"mat": "M",
"transpose?": True,
"msg": "Lmsg",
"var": "L"
}]
},
name="NeuroSAT")
# Define L_vote
L_vote_MLP = Mlp(layer_sizes=[d for _ in range(3)],
activations=[tf.nn.relu for _ in range(3)],
output_size=1,
name="L_vote",
name_internal_layers=True,
kernel_initializer=tf.contrib.layers.xavier_initializer(),
bias_initializer=tf.zeros_initializer())
# Get the last embeddings
L_n = gnn({"M": matrix_placeholder},
tiled_and_normalized_initial_embeddings, time_steps)["L"].h
L_vote = L_vote_MLP(L_n)
# Reorganize votes' result to obtain a prediction for each problem instance
def _vote_while_cond(i, p, n_acc, n, n_var_list, predicted_sat, L_vote):
return tf.less(i, p)
#end _vote_while_cond
def _vote_while_body(i, p, n_acc, n, n_var_list, predicted_SAT, L_vote):
# Helper for the amount of variables in this problem
i_n = n_var_list[i]
# Gather the positive and negative literals for that problem
pos_lits = tf.gather(L_vote, tf.range(n_acc, tf.add(n_acc, i_n)))
neg_lits = tf.gather(
L_vote, tf.range(tf.add(n, n_acc), tf.add(n, tf.add(n_acc, i_n))))
# Concatenate positive and negative literals and average their vote values
problem_predicted_SAT = tf.reduce_mean(
tf.concat([pos_lits, neg_lits], axis=1))
# Update TensorArray
predicted_SAT = predicted_SAT.write(i, problem_predicted_SAT)
return tf.add(i, tf.constant(1)), p, tf.add(
n_acc, i_n), n, n_var_list, predicted_SAT, L_vote
#end _vote_while_body
predicted_SAT = tf.TensorArray(size=p, dtype=tf.float32)
_, _, _, _, _, predicted_SAT, _ = tf.while_loop(
_vote_while_cond, _vote_while_body, [
tf.constant(0, dtype=tf.int32), p,
tf.constant(0, dtype=tf.int32), n, num_vars_on_instance,
predicted_SAT, L_vote
])
predicted_SAT = predicted_SAT.stack()
# Define loss, accuracy
predict_costs = tf.nn.sigmoid_cross_entropy_with_logits(
labels=instance_SAT, logits=predicted_SAT)
predict_cost = tf.reduce_mean(predict_costs)
vars_cost = tf.zeros([])
tvars = tf.trainable_variables()
for var in tvars:
vars_cost = tf.add(vars_cost, tf.nn.l2_loss(var))
#end for
loss = tf.add(predict_cost, tf.multiply(vars_cost,
parameter_l2norm_scaling))
optimizer = tf.train.AdamOptimizer(name="Adam",
learning_rate=learning_rate)
grads, _ = tf.clip_by_global_norm(tf.gradients(loss, tvars),
global_norm_gradient_clipping_ratio)
train_step = optimizer.apply_gradients(zip(grads, tvars))
accuracy = tf.reduce_mean(
tf.cast(
tf.equal(tf.cast(instance_SAT, tf.bool),
tf.cast(tf.round(tf.nn.sigmoid(predicted_SAT)), tf.bool)),
tf.float32))
# Define neurosat dictionary
neurosat = {}
neurosat["M"] = matrix_placeholder
neurosat["time_steps"] = time_steps
neurosat["gnn"] = gnn
neurosat["instance_SAT"] = instance_SAT
neurosat["predicted_SAT"] = predicted_SAT
neurosat["num_vars_on_instance"] = num_vars_on_instance
neurosat["loss"] = loss
neurosat["accuracy"] = accuracy
neurosat["train_step"] = train_step
return neurosat
def VIN(state_input, state_dim, action_dim, config, weights=None):
numactions = action_dim;
numstates = config.numstates;
k = config.k
width = config.width
assert width % 2 == 1
ch_i = config.ch_i
ch_h = config.ch_h
ch_q = config.ch_q
hiddenUnits = config.hidden1
state_batch_size = tf.shape(state_input)[0];
if (weights == None):
#Weights for each action's reward
w = tf.Variable(np.random.randn(width, 1, ch_q) * 0.001, dtype=tf.float32)
# feedback weights from v layer into q layer (~transition probabilities in Bellman equation)
w_fb = tf.Variable(np.random.randn(width, 1, ch_q) * 0.001, dtype=tf.float32)
#Output weights
bias1 = tf.Variable(np.random.randn(1, hiddenUnits) * 0.001, dtype=tf.float32)
w_h1 = tf.Variable(np.random.randn(ch_q + state_dim, hiddenUnits) * 0.001, dtype=tf.float32)
bias_o = tf.Variable(np.random.randn(1, numactions) * 0.001, dtype=tf.float32)
w_o = tf.Variable(np.random.randn(hiddenUnits, numactions) * 0.001, dtype=tf.float32)
#Reward Map
r = tf.Variable(np.random.randn(config.batchsize, config.numstates, config.ch_i) * 0.001, dtype=tf.float32)
else:
w = weights[0]
w_fb = weights[1]
bias1 = weights[2]
w_h1 = weights[3]
bias_o = weights[4]
w_o = weights[5]
r = weights[6]
q = circularConv(r, w)
v = tf.reduce_max(q, axis=2, keep_dims=True, name="v")
wwfb = tf.concat([w, w_fb], 1)
#Value Iteration
for i in range(0, k-1):
rv = tf.concat([r, v], 2)
q = circularConv(rv, wwfb)
v = tf.reduce_max(q, axis=2, keep_dims=True, name="v")
# do one last convolution
q = circularConv(tf.concat([r, v], 2), wwfb) #tf.nn.conv1d(tf.concat([r, v], 2), wwfb, stride=1, padding='SAME', name="q")
# Select the conv-net channels at the state position
position = tf.transpose(state_input, perm=[1,0])
angle = theta(position[0], position[1])
S1 = tf.cast(tf.floordiv(angle, 2*math.pi / numstates), tf.int32)
ins1 = tf.zeros(tf.shape(S1), tf.int32)
idx_in = tf.transpose(tf.stack([ins1, S1]), [1,0])
#Concat action values to observations
q_out = tf.gather_nd(q, idx_in, name="q_out")
inputs = tf.concat([state_input, q_out], axis=1)
#Output action
hiddenLayer1 = tf.nn.relu(tf.matmul(inputs, w_h1) + bias1)
output = tf.nn.tanh(tf.matmul(hiddenLayer1, w_o) + bias_o);
return state_input, output, [w, w_fb, bias1, w_h1, bias_o, w_o, r]
x = tf.add(a, b, name='add')
writer = tf.summary.FileWriter('./graphs/simple', tf.get_default_graph())
with tf.Session() as sess:
# writer = tf.summary.FileWriter('./graphs', sess.graph)
print(sess.run(x))
writer.close() # close the writer when you’re done using it
# Example 2: The wonderful wizard of div
a = tf.constant([2, 2], name='a')
b = tf.constant([[0, 1], [2, 3]], name='b')
with tf.Session() as sess:
print(sess.run(tf.div(b, a)))
print(sess.run(tf.divide(b, a)))
print(sess.run(tf.truediv(b, a)))
print(sess.run(tf.floordiv(b, a)))
# print(sess.run(tf.realdiv(b, a)))
print(sess.run(tf.truncatediv(b, a)))
print(sess.run(tf.floor_div(b, a)))
# Example 3: multiplying tensors
a = tf.constant([10, 20], name='a')
b = tf.constant([2, 3], name='b')
with tf.Session() as sess:
print(sess.run(tf.multiply(a, b)))
print(sess.run(tf.tensordot(a, b, 1)))
# Example 4: Python native type
t_0 = 19
x = tf.zeros_like(t_0) # ==> 0
def mean_shift(X, bandwidth, max_iter):
(m,n) = X.shape
print m,n
graph = tf.Graph()
with graph.as_default():
with tf.name_scope("input") as scope:
data = tf.constant(X, name="data_points")
b = tf.constant(bandwidth,dtype=tf.float32, name="bandwidth")
m = tf.constant(max_iter, name="maximum_iteration")
# n_samples = tf.constant(m, name="no_of_samples")
# n_features = tf.constant(n, name="no_of_features")
# with tf.name_scope("seeding") as scope:
# seed = tf.placeholder(tf.float32, [5], name="seed")
with tf.name_scope("mean_shifting") as scope:
old_mean = tf.placeholder(tf.float32, [n], name="old_mean")
neighbors = tf.placeholder(tf.float32, [None,n], name="neighbors")
new_mean = tf.reduce_mean(neighbors,0)
euclid_dist = tf.sqrt(tf.reduce_sum(tf.pow(tf.sub(old_mean, new_mean), 2)), name="mean_distance")
center_intensity_dict = {}
nbrs = NearestNeighbors(radius=bandwidth).fit(X)
sess = tf.Session()
init = tf.initialize_all_variables()
sess.run(init)
writer = tf.train.SummaryWriter(FLAGS.log_dir, sess.graph_def)
bin_sizes = defaultdict(int)
data_point = tf.placeholder(tf.float32, [n],"data_point")
binned_point = tf.floordiv(data_point,b)
for point in X:
feed={data_point:point}
bp = sess.run(binned_point,feed_dict=feed)
bin_sizes[tuple(bp)] +=1
bin_seeds = np.array([point for point, freq in six.iteritems(bin_sizes) if freq >= 1], dtype=np.float32)
bin_seeds = bin_seeds*bandwidth
print len(bin_seeds)
j=0
for x in bin_seeds:
print "Seed ",j,": ",x
i = 0
o_mean=x
while True:
i_nbrs = nbrs.radius_neighbors([o_mean], bandwidth, return_distance=False)[0]
points_within = X[i_nbrs]
feed = {neighbors: points_within}
n_mean = sess.run(new_mean, feed_dict=feed)
feed = {new_mean: n_mean, old_mean: o_mean}
dist = sess.run(euclid_dist, feed_dict=feed)
if dist < 1e-3*bandwidth or i==max_iter:
center_intensity_dict[tuple(n_mean)] = len(i_nbrs)
break
else:
o_mean = n_mean
print "\t",i,dist,len(i_nbrs)
i+=1
# if j>10:
# break
j+=1
print center_intensity_dict
sorted_by_intensity = sorted(center_intensity_dict.items(),key=lambda tup: tup[1], reverse=True)
sorted_centers = np.array([tup[0] for tup in sorted_by_intensity])
unique = np.ones(len(sorted_centers), dtype=np.bool)
nbrs = NearestNeighbors(radius=bandwidth).fit(sorted_centers)
for i, center in enumerate(sorted_centers):
if unique[i]:
neighbor_idxs = nbrs.radius_neighbors([center],return_distance=False)[0]
unique[neighbor_idxs] = 0
unique[i] = 1 # leave the current point as unique
cluster_centers = sorted_centers[unique]
nbrs = NearestNeighbors(n_neighbors=1).fit(cluster_centers)
labels = np.zeros(154401, dtype=np.int)
distances, idxs = nbrs.kneighbors(X)
labels = idxs.flatten()
return cluster_centers, labels
def test_FloorDiv(self):
t = tf.floordiv(*self.random((3, 5), (3, 5)))
self.check(t)
def build_attention_model(params,
src_vocab,
trg_vocab,
source_placeholders,
target_placeholders,
beam_size=1,
mode=MODE.TRAIN,
burn_in_step=100000,
increment_step=10000,
teacher_rate=1.0,
max_step=100):
"""
Build a model.
:param params: dict.
{encoder: {rnn_cell: {},
...},
decoder: {rnn_cell: {},
...}}
for example:
{'encoder': {'rnn_cell': {'state_size': 512,
'cell_name': 'BasicLSTMCell',
'num_layers': 2,
'input_keep_prob': 1.0,
'output_keep_prob': 1.0},
'attention_key_size': attention_size},
'decoder': {'rnn_cell': {'cell_name': 'BasicLSTMCell',
'state_size': 512,
'num_layers': 1,
'input_keep_prob': 1.0,
'output_keep_prob': 1.0},
'trg_vocab_size': trg_vocab_size}}
:param src_vocab: Vocab of source symbols.
:param trg_vocab: Vocab of target symbols.
:param source_ids: placeholder
:param source_seq_length: placeholder
:param target_ids: placeholder
:param target_seq_length: placeholder
:param beam_size: used in beam inference
:param mode:
:return:
"""
if mode != MODE.TRAIN:
params = sq.disable_dropout(params)
tf.logging.info(json.dumps(params, indent=4))
decoder_params = params['decoder']
# parameters
source_ids = source_placeholders['src']
source_seq_length = source_placeholders['src_len']
source_sample_matrix = source_placeholders['src_sample_matrix']
source_word_seq_length = source_placeholders['src_word_len']
target_ids = target_placeholders['trg']
target_seq_length = target_placeholders['trg_len']
# Because source encoder is different to the target feedback,
# we construct source_embedding_table manually
source_char_embedding_table = sq.LookUpOp(src_vocab.vocab_size,
src_vocab.embedding_dim,
name='source')
source_char_embedded = source_char_embedding_table(source_ids)
# encode char to word
char_encoder = sq.StackRNNEncoder(params['char_encoder'],
params['attention_key_size']['char'],
name='char_rnn',
mode=mode)
# char_encoder_outputs: T_c B F
char_encoded_representation = char_encoder.encode(source_char_embedded,
source_seq_length)
char_encoder_outputs = char_encoded_representation.outputs
#dynamical_batch_size = tf.shape(char_encoder_outputs)[1]
#space_indices = tf.where(tf.equal(tf.transpose(source_ids), src_vocab.space_id))
##space_indices = tf.transpose(tf.gather_nd(tf.transpose(space_indices), [[1], [0]]))
#space_indices = tf.concat(tf.split(space_indices, 2, axis=1)[::-1], axis=1)
#space_indices = tf.transpose(tf.reshape(space_indices, [dynamical_batch_size, -1, 2]),
# [1, 0, 2])
## T_w * B * F
#source_embedded = tf.gather_nd(char_encoder_outputs, space_indices)
# must be time major
char_encoder_outputs = tf.transpose(char_encoder_outputs, perm=(1, 0, 2))
sampled_word_embedded = tf.matmul(source_sample_matrix,
char_encoder_outputs)
source_embedded = tf.transpose(sampled_word_embedded, perm=(1, 0, 2))
encoder = sq.StackBidirectionalRNNEncoder(
params['encoder'],
params['attention_key_size']['word'],
name='stack_rnn',
mode=mode)
encoded_representation = encoder.encode(source_embedded,
source_word_seq_length)
attention_keys = encoded_representation.attention_keys
attention_values = encoded_representation.attention_values
attention_length = encoded_representation.attention_length
encoder_final_states_bw = encoded_representation.final_state[-1][-1].h
# feedback
if mode == MODE.RL:
tf.logging.info('BUILDING RL TRAIN FEEDBACK......')
dynamical_batch_size = tf.shape(attention_keys)[1]
feedback = sq.RLTrainingFeedBack(target_ids,
target_seq_length,
trg_vocab,
dynamical_batch_size,
burn_in_step=burn_in_step,
increment_step=increment_step,
max_step=max_step)
elif mode == MODE.TRAIN:
tf.logging.info('BUILDING TRAIN FEEDBACK WITH {} TEACHER_RATE'
'......'.format(teacher_rate))
feedback = sq.TrainingFeedBack(target_ids,
target_seq_length,
trg_vocab,
teacher_rate,
max_step=max_step)
elif mode == MODE.EVAL:
tf.logging.info('BUILDING EVAL FEEDBACK ......')
feedback = sq.TrainingFeedBack(target_ids,
target_seq_length,
trg_vocab,
0.,
max_step=max_step)
else:
tf.logging.info('BUILDING INFER FEEDBACK WITH BEAM_SIZE {}'
'......'.format(beam_size))
infer_key_size = attention_keys.get_shape().as_list()[-1]
infer_value_size = attention_values.get_shape().as_list()[-1]
infer_states_bw_shape = encoder_final_states_bw.get_shape().as_list(
)[-1]
encoder_final_states_bw = tf.reshape(
tf.tile(encoder_final_states_bw, [1, beam_size]),
[-1, infer_states_bw_shape])
# expand beam
if TIME_MAJOR:
# batch size should be dynamical
dynamical_batch_size = tf.shape(attention_keys)[1]
final_key_shape = [
-1, dynamical_batch_size * beam_size, infer_key_size
]
final_value_shape = [
-1, dynamical_batch_size * beam_size, infer_value_size
]
attention_keys = tf.reshape(
(tf.tile(attention_keys, [1, 1, beam_size])), final_key_shape)
attention_values = tf.reshape(
(tf.tile(attention_values, [1, 1, beam_size])),
final_value_shape)
else:
dynamical_batch_size = tf.shape(attention_keys)[0]
final_key_shape = [
dynamical_batch_size * beam_size, -1, infer_key_size
]
final_value_shape = [
dynamical_batch_size * beam_size, -1, infer_value_size
]
attention_keys = tf.reshape(
(tf.tile(attention_keys, [1, beam_size, 1])), final_key_shape)
attention_values = tf.reshape(
(tf.tile(attention_values, [1, beam_size, 1])),
final_value_shape)
attention_length = tf.reshape(
tf.transpose(tf.tile([attention_length], [beam_size, 1])), [-1])
feedback = sq.BeamFeedBack(trg_vocab,
beam_size,
dynamical_batch_size,
max_step=max_step)
encoder_decoder_bridge = EncoderDecoderBridge(
encoder_final_states_bw.get_shape().as_list()[-1],
decoder_params['rnn_cell'])
decoder_state_size = decoder_params['rnn_cell']['state_size']
# attention
attention = sq.Attention(decoder_state_size, attention_keys,
attention_values, attention_length)
context_size = attention.context_size
with tf.variable_scope('logits_func'):
attention_mix = LinearOp(context_size + feedback.embedding_dim +
decoder_state_size,
decoder_state_size,
name='attention_mix')
attention_mix_middle = LinearOp(decoder_state_size,
decoder_state_size // 2,
name='attention_mix_middle')
logits_trans = LinearOp(decoder_state_size // 2,
feedback.vocab_size,
name='logits_trans')
logits_func = lambda _softmax: logits_trans(
tf.nn.relu(
attention_mix_middle(tf.nn.relu(attention_mix(_softmax)))))
# decoder
decoder = sq.AttentionRNNDecoder(
decoder_params,
attention,
feedback,
logits_func=logits_func,
init_state=encoder_decoder_bridge(encoder_final_states_bw),
mode=mode)
decoder_output, decoder_final_state = sq.dynamic_decode(decoder,
swap_memory=True,
scope='decoder')
# not training
if mode == MODE.EVAL or mode == MODE.INFER:
return decoder_output, decoder_final_state
# bos is added in feedback
# so target_ids is predict_ids
if not TIME_MAJOR:
ground_truth_ids = tf.transpose(target_ids, [1, 0])
else:
ground_truth_ids = target_ids
# construct the loss
if mode == MODE.RL:
# Creates a variable to hold the global_step.
global_step_tensor = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
scope='global_step')[0]
rl_time_steps = tf.floordiv(
tf.maximum(global_step_tensor - burn_in_step, 0), increment_step)
start_rl_step = target_seq_length - rl_time_steps
baseline_states = tf.stop_gradient(decoder_output.baseline_states)
predict_ids = tf.stop_gradient(decoder_output.predicted_ids)
# TODO: bug in tensorflow
ground_or_predict_ids = tf.cond(tf.greater(rl_time_steps,
0), lambda: predict_ids,
lambda: ground_truth_ids)
reward, sequence_length = tf.py_func(
func=_py_func,
inp=[ground_or_predict_ids, ground_truth_ids, trg_vocab.eos_id],
Tout=[tf.float32, tf.int32],
name='reward')
sequence_length.set_shape((None, ))
total_loss_avg, entropy_loss_avg, reward_loss_rmse, reward_predicted \
= rl_sequence_loss(
logits=decoder_output.logits,
predict_ids=predict_ids,
sequence_length=sequence_length,
baseline_states=baseline_states,
start_rl_step=start_rl_step,
reward=reward)
return decoder_output, total_loss_avg, entropy_loss_avg, \
reward_loss_rmse, reward_predicted
else:
total_loss_avg = cross_entropy_sequence_loss(
logits=decoder_output.logits,
targets=ground_truth_ids,
sequence_length=target_seq_length)
return decoder_output, total_loss_avg, total_loss_avg, \
tf.to_float(0.), tf.to_float(0.)
def generate_detections_per_image_op(
cls_outputs, box_outputs, anchor_boxes, image_id, image_info,
num_detections=100, pre_nms_num_detections=1000, nms_threshold=0.3,
bbox_reg_weights=(10., 10., 5., 5.)):
"""Generates detections with model outputs and anchors.
Args:
cls_outputs: a Tensor with shape [N, num_classes], which stacks class
logit outputs on all feature levels. The N is the number of total anchors
on all levels. The num_classes is the number of classes predicted by the
model. Note that the cls_outputs should be the output of softmax().
box_outputs: a Tensor with shape [N, num_classes*4], which stacks
box regression outputs on all feature levels. The N is the number of total
anchors on all levels.
anchor_boxes: a Tensor with shape [N, 4], which stacks anchors on all
feature levels. The N is the number of total anchors on all levels.
image_id: an integer number to specify the image id.
image_info: a tensor of shape [5] which encodes the input image's [height,
width, scale, original_height, original_width]
num_detections: Number of detections after NMS.
pre_nms_num_detections: Number of candidates before NMS.
nms_threshold: a float number to specify the threshold of NMS.
bbox_reg_weights: a list of 4 float scalars, which are default weights on
(dx, dy, dw, dh) for normalizing bbox regression targets.
Returns:
detections: detection results in a tensor with each row representing
[image_id, ymin, xmin, ymax, xmax, score, class]
"""
num_boxes, num_classes = cls_outputs.get_shape().as_list()
# Removes background class scores.
cls_outputs = cls_outputs[:, 1:num_classes]
top_k_scores, top_k_indices_with_classes = tf.nn.top_k(
tf.reshape(cls_outputs, [-1]),
k=pre_nms_num_detections,
sorted=True)
classes = tf.mod(top_k_indices_with_classes, num_classes - 1)
top_k_indices = tf.floordiv(top_k_indices_with_classes, num_classes - 1)
anchor_boxes = tf.gather(anchor_boxes, top_k_indices)
box_outputs = tf.reshape(
box_outputs, [num_boxes, num_classes, 4])[:, 1:num_classes, :]
box_outputs = tf.gather_nd(box_outputs,
tf.stack([top_k_indices, classes], axis=1))
# Applies bounding box regression to anchors.
boxes = box_utils.batch_decode_box_outputs_op(
tf.expand_dims(anchor_boxes, axis=0),
tf.expand_dims(box_outputs, axis=0),
bbox_reg_weights)[0]
boxes = box_utils.clip_boxes(
tf.expand_dims(boxes, axis=0), tf.expand_dims(image_info[:2], axis=0))[0]
classes = tf.tile(tf.reshape(classes, [1, pre_nms_num_detections]),
[num_classes - 1, 1])
scores = tf.tile(tf.reshape(top_k_scores, [1, pre_nms_num_detections]),
[num_classes - 1, 1])
boxes = tf.tile(tf.reshape(boxes, [1, pre_nms_num_detections, 4]),
[num_classes - 1, 1, 1])
class_bitmask = tf.tile(
tf.reshape(tf.range(num_classes-1), [num_classes - 1, 1]),
[1, pre_nms_num_detections])
scores = tf.where(tf.equal(classes, class_bitmask), scores,
tf.zeros_like(scores))
scores = tf.where(tf.greater(scores, 0.05), scores, tf.zeros_like(scores))
# Reshape classes to be compartible with the top_k function.
classes = tf.reshape(classes, [num_classes -1, pre_nms_num_detections, 1])
scores, sorted_tensors = box_utils.top_k(
scores, k=pre_nms_num_detections, tensors=[boxes, classes])
boxes = sorted_tensors[0]
classes = tf.reshape(sorted_tensors[1],
[num_classes - 1, pre_nms_num_detections])
(post_nms_scores,
post_nms_boxes, idx) = non_max_suppression.non_max_suppression_padded(
scores, boxes, max_output_size=num_detections,
iou_threshold=nms_threshold, level=0)
# Sorts all results.
sorted_scores, sorted_indices = tf.nn.top_k(
tf.to_float(tf.reshape(post_nms_scores, [-1])),
k=num_detections,
sorted=True)
post_nms_boxes = tf.gather(tf.reshape(post_nms_boxes, [-1, 4]),
sorted_indices)
classes = tf.batch_gather(classes, idx)
post_nms_classes = tf.gather(tf.reshape(classes, [-1]), sorted_indices) + 1
if isinstance(image_id, int):
image_id = tf.constant(image_id)
image_id = tf.reshape(image_id, [])
detections_result = tf.stack(
[
tf.to_float(tf.fill(tf.shape(sorted_scores), image_id)),
post_nms_boxes[:, 0],
post_nms_boxes[:, 1],
post_nms_boxes[:, 2],
post_nms_boxes[:, 3],
sorted_scores,
tf.to_float(post_nms_classes),
],
axis=1)
return detections_result
def generate_detections_per_image_tpu(cls_outputs,
box_outputs,
anchor_boxes,
image_id,
image_info,
pre_nms_num_detections=1000,
post_nms_num_detections=100,
nms_threshold=0.3,
bbox_reg_weights=(10., 10., 5., 5.)):
"""Generate the final detections per image given the model outputs.
Args:
cls_outputs: a tensor with shape [N, num_classes], which stacks class
logit outputs on all feature levels. The N is the number of total anchors
on all levels. The num_classes is the number of classes predicted by the
model. Note that the cls_outputs should be the output of softmax().
box_outputs: a tensor with shape [N, num_classes*4], which stacks box
regression outputs on all feature levels. The N is the number of total
anchors on all levels.
anchor_boxes: a tensor with shape [N, 4], which stacks anchors on all
feature levels. The N is the number of total anchors on all levels.
image_id: an integer number to specify the image id.
image_info: a tensor of shape [5] which encodes the input image's [height,
width, scale, original_height, original_width]
pre_nms_num_detections: an integer that specifies the number of candidates
before NMS.
post_nms_num_detections: an integer that specifies the number of candidates
after NMS.
nms_threshold: a float number to specify the IOU threshold of NMS.
bbox_reg_weights: a list of 4 float scalars, which are default weights on
(dx, dy, dw, dh) for normalizing bbox regression targets.
Returns:
detections: detection results in a tensor with each row representing
[image_id, ymin, xmin, ymax, xmax, score, class]
"""
num_boxes, num_classes = cls_outputs.get_shape().as_list()
# Remove background class scores.
cls_outputs = cls_outputs[:, 1:num_classes]
top_k_scores, top_k_indices_with_classes = tf.nn.top_k(
tf.reshape(cls_outputs, [-1]), k=pre_nms_num_detections, sorted=False)
classes = tf.mod(top_k_indices_with_classes, num_classes - 1)
top_k_indices = tf.floordiv(top_k_indices_with_classes, num_classes - 1)
anchor_boxes = tf.gather(anchor_boxes, top_k_indices)
box_outputs = tf.reshape(box_outputs,
[num_boxes, num_classes, 4])[:, 1:num_classes, :]
class_indices = classes
box_outputs = tf.gather_nd(
box_outputs, tf.stack([top_k_indices, class_indices], axis=1))
# apply bounding box regression to anchors
boxes = box_utils.decode_boxes(box_outputs, anchor_boxes, bbox_reg_weights)
boxes = box_utils.clip_boxes(boxes, image_info[0], image_info[1])
list_of_all_boxes = []
list_of_all_scores = []
list_of_all_classes = []
# Skip background class.
for class_i in range(num_classes):
# Compute bitmask for the given classes.
class_i_bitmask = tf.cast(tf.equal(classes, class_i),
top_k_scores.dtype)
# This works because score is in [0, 1].
class_i_scores = top_k_scores * class_i_bitmask
# The TPU and CPU have different behaviors for
# tf.image.non_max_suppression_padded (b/116754376).
(class_i_post_nms_indices,
class_i_nms_num_valid) = tf.image.non_max_suppression_padded(
tf.to_float(boxes),
tf.to_float(class_i_scores),
post_nms_num_detections,
iou_threshold=nms_threshold,
score_threshold=0.05,
pad_to_max_output_size=True,
name='nms_detections_' + str(class_i))
class_i_post_nms_boxes = tf.gather(boxes, class_i_post_nms_indices)
class_i_post_nms_scores = tf.gather(class_i_scores,
class_i_post_nms_indices)
mask = tf.less(tf.range(post_nms_num_detections),
[class_i_nms_num_valid])
class_i_post_nms_scores = tf.where(
mask, class_i_post_nms_scores,
tf.zeros_like(class_i_post_nms_scores))
class_i_classes = tf.fill(tf.shape(class_i_post_nms_scores),
class_i + 1)
list_of_all_boxes.append(class_i_post_nms_boxes)
list_of_all_scores.append(class_i_post_nms_scores)
list_of_all_classes.append(class_i_classes)
post_nms_boxes = tf.concat(list_of_all_boxes, axis=0)
post_nms_scores = tf.concat(list_of_all_scores, axis=0)
post_nms_classes = tf.concat(list_of_all_classes, axis=0)
# sort all results.
post_nms_scores, sorted_indices = tf.nn.top_k(tf.to_float(post_nms_scores),
k=post_nms_num_detections,
sorted=True)
post_nms_boxes = tf.gather(post_nms_boxes, sorted_indices)
post_nms_classes = tf.gather(post_nms_classes, sorted_indices)
if isinstance(image_id, int):
image_id = tf.constant(image_id)
image_id = tf.reshape(image_id, [])
detections_result = tf.stack([
tf.to_float(tf.fill(tf.shape(post_nms_scores), image_id)),
post_nms_boxes[:, 0],
post_nms_boxes[:, 1],
post_nms_boxes[:, 2],
post_nms_boxes[:, 3],
post_nms_scores,
tf.to_float(post_nms_classes),
],
axis=1)
return detections_result
import tensorflow as tf
a = tf.constant(0, dtype=tf.float32, name="a")
b = tf.constant(1, dtype=tf.float32, name="b")
c = tf.constant(2, dtype=tf.float32, name="c")
with tf.control_dependencies([a, c]):
d = tf.add(a, b, name="d")
e = tf.subtract(b, c, name="e")
f = tf.multiply(d, e, name="f")
with tf.control_dependencies([d, f]):
g = tf.floordiv(b, d, name="g")
init = tf.global_variables_initializer()
with tf.Session() as sess:
sess.run(init)
writer = tf.summary.FileWriter("logs", sess.graph)
print(g.eval())
writer.close()
print()
import tensorflow as tf
x = tf.placeholder(tf.float32, shape=[None, 16, 16, 3], name='x')
y = tf.placeholder(tf.float32, shape=[None, 16, 16, 3], name='y')
xl = tf.placeholder(tf.bool, shape=[None, 16, 16, 3], name='xl')
yl = tf.placeholder(tf.bool, shape=[None, 16, 16, 3], name='yl')
c = tf.placeholder(tf.float32, shape=[1], name='c')
b = tf.placeholder(tf.bool, shape=[1], name='b')
g = tf.greater(x, y, name='g')
ge = tf.greater_equal(x, y, name='ge')
l = tf.less(x, y, name='l')
le = tf.less_equal(x, y, name='l3')
ne = tf.not_equal(x, y, name='ne')
p = tf.pow(x, y)
fd = tf.floordiv(x, y, name='fd')
lo = tf.logical_or(xl, yl, name='lo')
s = tf.where(b, x, y, name='s')
m = tf.minimum(x, y, name='m')
sess = tf.Session()
sess.run(tf.global_variables_initializer())
c_g = tf.contrib.lite.TFLiteConverter.from_session(sess, [x, y], [g])
c_g_lm = c_g.convert()
open('greater.tflite', 'wb').write(c_g_lm)
c_ge = tf.contrib.lite.TFLiteConverter.from_session(sess, [x, y], [ge])
c_ge_lm = c_ge.convert()
open('greater_equal.tflite', 'wb').write(c_ge_lm)
def hier_homography_fmask_estimator(color_inputs,
num_param=8,
num_layer=7,
num_level=3,
dropout_keep_prob=0.8,
reuse=None,
is_training=True,
trainable=True,
scope='hier_hmg'):
"""A hierarchical neural network with mask for homograhy estimation.
Args:
color_inputs: batch of input image pairs of data type float32 and of shape
[batch_size, height, width, 6]
num_param: the number of parameters for homography (default 8)
num_layer: the number of convolutional layers in the motion feature network
num_level: the number of hierarchical levels
dropout_keep_prob: the percentage of activation values that are kept
reuse: whether to reuse this network weights
is_training: whether used for training or testing
trainable: whether this network is to be trained or not
scope: the scope of variables in this function
Returns:
a list of homographies at each level and motion feature maps if
final_endpoint='mfeature'; otherwise a list of images warped by the list of
corresponding homographies
"""
_, h_input, w_input = color_inputs.get_shape().as_list()[0:3]
vgg_inputs = (color_inputs[Ellipsis, 3:6] * 256 + 128) - VGG_MEANS
with slim.arg_scope([slim.conv2d, slim.max_pool2d], padding='SAME'):
with slim.arg_scope([slim.conv2d, slim.fully_connected],
trainable=False):
with slim.arg_scope([slim.conv2d], normalizer_fn=None):
with slim.arg_scope(vgg.vgg_arg_scope()):
sfeature, _ = vgg.vgg_16(vgg_inputs,
1000,
predictions_fn=slim.softmax,
global_pool=False,
is_training=False,
reuse=reuse,
spatial_squeeze=True,
final_endpoint='pool5',
scope='vgg_16')
gray_image1 = tf.image.rgb_to_grayscale(color_inputs[Ellipsis, 0:3])
gray_image2 = tf.image.rgb_to_grayscale(color_inputs[Ellipsis, 3:6])
inputs = tf.concat([gray_image1, gray_image2], 3)
hmgs_list = []
warped_list = []
with tf.variable_scope(scope, [inputs], reuse=reuse):
for level_index in range(num_level):
scale = 2**(num_level - 1 - level_index)
h = tf.to_float(tf.floordiv(h_input, scale))
w = tf.to_float(tf.floordiv(w_input, scale))
inputs_il = tf.image.resize_images(inputs, tf.to_int32([h, w]))
if level_index == 0:
mfeature = hier_base_layers(inputs_il,
num_layer + 1 - num_level +
level_index,
level_index,
is_training=is_training,
trainable=trainable)
hmgs_il = homography_regression(
mfeature,
num_param,
level_index,
dropout_keep_prob=dropout_keep_prob,
is_training=is_training,
trainable=trainable)
hmgs_list.append(hmgs_il)
else:
warped, _ = hmg_util.homography_scale_warp_per_batch(
inputs_il[:, :, :, 0], w / 2, h / 2,
hmgs_list[level_index - 1])
pre_warped_inputs_il = tf.stack(
[warped, inputs_il[:, :, :, 1]], -1)
warped_list.append(pre_warped_inputs_il)
mfeature = hier_base_layers(pre_warped_inputs_il,
num_layer + 1 - num_level +
level_index,
level_index,
is_training=is_training,
trainable=trainable)
if level_index == num_level - 1:
mfeature = fmask_layers_semantic(mfeature,
sfeature,
level_index,
is_training=is_training,
trainable=trainable)
hmgs_il = homography_regression(
mfeature,
num_param,
level_index,
dropout_keep_prob=dropout_keep_prob,
is_training=is_training,
trainable=trainable)
new_hmgs_il = hmg_util.homography_shift_mult_batch(
hmgs_list[level_index - 1], w / 2, h / 2, hmgs_il, w, h, w,
h)
hmgs_list.append(new_hmgs_il)
return hmgs_list, warped_list
def beam_step(beam_seqs, beam_probs, cand_seqs, cand_probs, states,
time):
'''
beam_seqs : [beam_size, time]
beam_probs: [beam_size, ]
cand_seqs : [beam_size, time]
cand_probs: [beam_size, ]
states : [beam_size * hidden_size, beam_size * hidden_size]
'''
inputs = tf.reshape(tf.slice(beam_seqs, [0, time], [beam_size, 1]),
[beam_size])
# print inputs.get_shape().as_list()
x_t = tf.nn.embedding_lookup(self.embedding, inputs)
# print(x_t.get_shape().as_list())
o_t, s_nt = self.dec_lstm(x_t, states)
o_t, w_t = self.att_layer(o_t)
o_t = self.dec_out(o_t)
logprobs2d = tf.nn.log_softmax(o_t)
print logprobs2d.get_shape().as_list()
total_probs = logprobs2d + tf.reshape(beam_probs, [-1, 1])
print total_probs.get_shape().as_list()
total_probs_noEOS = tf.concat([
tf.slice(total_probs, [0, 0], [beam_size, self.stop_token]),
tf.tile([[-3e38]], [beam_size, 1]),
tf.slice(total_probs, [0, self.stop_token + 1],
[beam_size, self.target_vocab - self.stop_token - 1])
], 1)
print total_probs_noEOS.get_shape().as_list()
flat_total_probs = tf.reshape(total_probs_noEOS, [-1])
print flat_total_probs.get_shape().as_list()
beam_k = tf.minimum(tf.size(flat_total_probs), beam_size)
next_beam_probs, top_indices = tf.nn.top_k(flat_total_probs,
k=beam_k)
print next_beam_probs.get_shape().as_list()
next_bases = tf.floordiv(top_indices, self.target_vocab)
next_mods = tf.mod(top_indices, self.target_vocab)
print next_mods.get_shape().as_list()
next_beam_seqs = tf.concat([
tf.gather(beam_seqs, next_bases),
tf.reshape(next_mods, [-1, 1])
], 1)
next_states = (tf.gather(s_nt[0], next_bases),
tf.gather(s_nt[1], next_bases))
print next_beam_seqs.get_shape().as_list()
cand_seqs_pad = tf.pad(cand_seqs, [[0, 0], [0, 1]])
beam_seqs_EOS = tf.pad(beam_seqs, [[0, 0], [0, 1]])
new_cand_seqs = tf.concat([cand_seqs_pad, beam_seqs_EOS], 0)
print new_cand_seqs.get_shape().as_list()
EOS_probs = tf.slice(total_probs, [0, self.stop_token],
[beam_size, 1])
new_cand_probs = tf.concat(
[cand_probs, tf.reshape(EOS_probs, [-1])], 0)
cand_k = tf.minimum(tf.size(new_cand_probs), self.beam_size)
next_cand_probs, next_cand_indices = tf.nn.top_k(new_cand_probs,
k=cand_k)
next_cand_seqs = tf.gather(new_cand_seqs, next_cand_indices)
return next_beam_seqs, next_beam_probs, next_cand_seqs, next_cand_probs, next_states, time + 1
def __call__(self, inputs, state, scope=None):
(
past_cand_symbols, # [batch_size, max_len]
past_cand_logprobs,# [batch_size]
past_beam_symbols, # [batch_size*self.beam_size, max_len], right-aligned!!!
past_beam_logprobs,# [batch_size*self.beam_size]
past_cell_state,
) = state
batch_size = tf.shape(past_cand_logprobs)[0] # TODO: get as int, if possible
full_size = batch_size * self.beam_size
cell_inputs = inputs
cell_outputs, raw_cell_state = self.cell(cell_inputs, past_cell_state)
logprobs = tf.nn.log_softmax(cell_outputs)
logprobs_batched = tf.reshape(logprobs + tf.expand_dims(past_beam_logprobs, 1),
[-1, self.beam_size * self.num_classes])
logprobs_batched.set_shape((None, self.beam_size * self.num_classes))
# prints and asserts
tf.assert_less_equal(logprobs, 0.0)
tf.assert_less_equal(past_beam_logprobs, 0.0)
masked_logprobs = tf.reshape(logprobs_batched, [-1, self.beam_size * self.num_classes])
# print masked_logprobs.get_shape()
beam_logprobs, indices = tf.nn.top_k(
masked_logprobs,
self.beam_size
)
beam_logprobs = tf.reshape(beam_logprobs, [-1])
# For continuing to the next symbols
symbols = indices % self.num_classes # [batch_size, self.beam_size]
parent_refs = tf.reshape(indices // self.num_classes, [-1]) # [batch_size*self.beam_size]
# TODO: this technically doesn't need to be recalculated every loop
parent_refs_offsets = tf.mul(tf.floordiv(tf.range(full_size), self.beam_size), self.beam_size)
parent_refs = parent_refs + parent_refs_offsets
if past_beam_symbols is not None:
symbols_history = tf.gather(past_beam_symbols, parent_refs)
beam_symbols = tf.concat(1, [tf.reshape(symbols, [-1, 1]), symbols_history])
else:
beam_symbols = tf.reshape(symbols, [-1, 1])
# Above ends up outputting reversed. Below doesn't work though because tf doesn't support negative indexing.
# last = past_beam_symbols.get_shape()[1]
# symbols_history = tf.gather(past_beam_symbols[:,last - 1], parent_refs)
# beam_symbols = tf.concat(1, [past_beam_symbols[:,:last-1], tf.reshape(symbols_history, [-1, 1]), tf.reshape(symbols, [-1, 1]), ])
# Handle the output and the cell state shuffling
outputs = tf.reshape(symbols, [-1]) # [batch_size*beam_size, 1]
cell_state = nest_map(
lambda element: tf.gather(element, parent_refs),
raw_cell_state
)
# Handling for getting a done token
# logprobs_done = tf.reshape(logprobs_batched, [-1, self.beam_size, self.num_classes])[:,:,self.stop_token]
# done_parent_refs = tf.to_int32(tf.argmax(logprobs_done, 1))
# done_parent_refs_offsets = tf.range(batch_size) * self.beam_size
# done_symbols = tf.gather(past_beam_symbols, done_parent_refs + done_parent_refs_offsets)
# logprobs_done_max = tf.reduce_max(logprobs_done, 1)
# cand_symbols = tf.select(logprobs_done_max > past_cand_logprobs,
# done_symbols,
# past_cand_symbols)
# cand_logprobs = tf.maximum(logprobs_done_max, past_cand_logprobs)
cand_symbols = past_cand_symbols # current last symbol in the beam [batch_size*self.beam_size]
cand_logprobs = past_cand_logprobs
return outputs, (
cand_symbols,
cand_logprobs,
beam_symbols,
beam_logprobs,
cell_state,
)
def beam_init():
# return beam_seqs_1 beam_probs_1 cand_seqs_1 cand_prob_1 next_states time
time_1 = tf.constant(1, dtype=tf.int32)
beam_seqs_0 = tf.constant([[self.start_token]] * beam_size)
beam_probs_0 = tf.constant([0.] * beam_size)
cand_seqs_0 = tf.constant([[self.start_token]])
cand_probs_0 = tf.constant([-3e38])
beam_seqs_0._shape = tf.TensorShape((None, None))
beam_probs_0._shape = tf.TensorShape((None, ))
cand_seqs_0._shape = tf.TensorShape((None, None))
cand_probs_0._shape = tf.TensorShape((None, ))
inputs = [self.start_token]
x_t = tf.nn.embedding_lookup(self.embedding, inputs)
print(x_t.get_shape().as_list())
o_t, s_nt = self.dec_lstm(x_t, initial_state)
o_t, w_t = self.att_layer(o_t)
o_t = self.dec_out(o_t)
print(s_nt[0].get_shape().as_list())
# initial_state = tf.reshape(initial_state, [1,-1])
logprobs2d = tf.nn.log_softmax(o_t)
total_probs = logprobs2d + tf.reshape(beam_probs_0, [-1, 1])
total_probs_noEOS = tf.concat([
tf.slice(total_probs, [0, 0], [1, self.stop_token]),
tf.tile([[-3e38]], [1, 1]),
tf.slice(total_probs, [0, self.stop_token + 1],
[1, self.target_vocab - self.stop_token - 1])
], 1)
flat_total_probs = tf.reshape(total_probs_noEOS, [-1])
print flat_total_probs.get_shape().as_list()
beam_k = tf.minimum(tf.size(flat_total_probs), beam_size)
next_beam_probs, top_indices = tf.nn.top_k(flat_total_probs,
k=beam_k)
next_bases = tf.floordiv(top_indices, self.target_vocab)
next_mods = tf.mod(top_indices, self.target_vocab)
next_beam_seqs = tf.concat([
tf.gather(beam_seqs_0, next_bases),
tf.reshape(next_mods, [-1, 1])
], 1)
cand_seqs_pad = tf.pad(cand_seqs_0, [[0, 0], [0, 1]])
beam_seqs_EOS = tf.pad(beam_seqs_0, [[0, 0], [0, 1]])
new_cand_seqs = tf.concat([cand_seqs_pad, beam_seqs_EOS], 0)
print new_cand_seqs.get_shape().as_list()
EOS_probs = tf.slice(total_probs, [0, self.stop_token],
[beam_size, 1])
new_cand_probs = tf.concat(
[cand_probs_0, tf.reshape(EOS_probs, [-1])], 0)
cand_k = tf.minimum(tf.size(new_cand_probs), self.beam_size)
next_cand_probs, next_cand_indices = tf.nn.top_k(new_cand_probs,
k=cand_k)
next_cand_seqs = tf.gather(new_cand_seqs, next_cand_indices)
part_state_0 = tf.reshape(tf.stack([s_nt[0]] * beam_size),
[beam_size, self.hidden_size])
part_state_1 = tf.reshape(tf.stack([s_nt[1]] * beam_size),
[beam_size, self.hidden_size])
part_state_0._shape = tf.TensorShape((None, None))
part_state_1._shape = tf.TensorShape((None, None))
next_states = (part_state_0, part_state_1)
print next_states[0].get_shape().as_list()
return next_beam_seqs, next_beam_probs, next_cand_seqs, next_cand_probs, next_states, time_1
def build_attention_model(params,
src_vocab,
trg_vocab,
source_ids,
source_seq_length,
target_ids,
target_seq_length,
beam_size=1,
mode=MODE.TRAIN,
burn_in_step=100000,
increment_step=10000,
teacher_rate=1.0,
max_step=100):
"""
Build a model.
:param params: dict.
{encoder: {rnn_cell: {},
...},
decoder: {rnn_cell: {},
...}}
for example:
{'encoder': {'rnn_cell': {'state_size': 512,
'cell_name': 'BasicLSTMCell',
'num_layers': 2,
'input_keep_prob': 1.0,
'output_keep_prob': 1.0},
'attention_key_size': attention_size},
'decoder': {'rnn_cell': {'cell_name': 'BasicLSTMCell',
'state_size': 512,
'num_layers': 1,
'input_keep_prob': 1.0,
'output_keep_prob': 1.0},
'trg_vocab_size': trg_vocab_size}}
:param src_vocab: Vocab of source symbols.
:param trg_vocab: Vocab of target symbols.
:param source_ids: placeholder
:param source_seq_length: placeholder
:param target_ids: placeholder
:param target_seq_length: placeholder
:param beam_size: used in beam inference
:param mode:
:return:
"""
if mode != MODE.TRAIN:
params = sq.disable_dropout(params)
tf.logging.info(json.dumps(params, indent=4))
# parameters
encoder_params = params['encoder']
decoder_params = params['decoder']
# Because source encoder is different to the target feedback,
# we construct source_embedding_table manually
source_embedding_table = sq.LookUpOp(src_vocab.vocab_size,
src_vocab.embedding_dim,
name='source')
source_embedded = source_embedding_table(source_ids)
encoder = sq.StackBidirectionalRNNEncoder(encoder_params,
name='stack_rnn',
mode=mode)
encoded_representation = encoder.encode(source_embedded, source_seq_length)
attention_keys = encoded_representation.attention_keys
attention_values = encoded_representation.attention_values
attention_length = encoded_representation.attention_length
# feedback
if mode == MODE.RL:
tf.logging.info('BUILDING RL TRAIN FEEDBACK......')
dynamical_batch_size = tf.shape(attention_keys)[1]
feedback = sq.RLTrainingFeedBack(target_ids,
target_seq_length,
trg_vocab,
dynamical_batch_size,
burn_in_step=burn_in_step,
increment_step=increment_step,
max_step=max_step)
elif mode == MODE.TRAIN:
tf.logging.info('BUILDING TRAIN FEEDBACK WITH {} TEACHER_RATE'
'......'.format(teacher_rate))
feedback = sq.TrainingFeedBack(target_ids,
target_seq_length,
trg_vocab,
teacher_rate,
max_step=max_step)
elif mode == MODE.EVAL:
tf.logging.info('BUILDING EVAL FEEDBACK ......')
feedback = sq.TrainingFeedBack(target_ids,
target_seq_length,
trg_vocab,
0.,
max_step=max_step)
else:
tf.logging.info('BUILDING INFER FEEDBACK WITH BEAM_SIZE {}'
'......'.format(beam_size))
infer_key_size = attention_keys.get_shape().as_list()[-1]
infer_value_size = attention_values.get_shape().as_list()[-1]
# expand beam
if TIME_MAJOR:
# batch size should be dynamical
dynamical_batch_size = tf.shape(attention_keys)[1]
final_key_shape = [
-1, dynamical_batch_size * beam_size, infer_key_size
]
final_value_shape = [
-1, dynamical_batch_size * beam_size, infer_value_size
]
attention_keys = tf.reshape(
(tf.tile(attention_keys, [1, 1, beam_size])), final_key_shape)
attention_values = tf.reshape(
(tf.tile(attention_values, [1, 1, beam_size])),
final_value_shape)
else:
dynamical_batch_size = tf.shape(attention_keys)[0]
final_key_shape = [
dynamical_batch_size * beam_size, -1, infer_key_size
]
final_value_shape = [
dynamical_batch_size * beam_size, -1, infer_value_size
]
attention_keys = tf.reshape(
(tf.tile(attention_keys, [1, beam_size, 1])), final_key_shape)
attention_values = tf.reshape(
(tf.tile(attention_values, [1, beam_size, 1])),
final_value_shape)
attention_length = tf.reshape(
tf.transpose(tf.tile([attention_length], [beam_size, 1])), [-1])
feedback = sq.BeamFeedBack(trg_vocab,
beam_size,
dynamical_batch_size,
max_step=max_step)
# attention
attention = sq.Attention(decoder_params['rnn_cell']['state_size'],
attention_keys, attention_values,
attention_length)
# decoder
decoder = sq.AttentionRNNDecoder(decoder_params,
attention,
feedback,
mode=mode)
decoder_output, decoder_final_state = sq.dynamic_decode(decoder,
swap_memory=True,
scope='decoder')
# not training
if mode == MODE.EVAL or mode == MODE.INFER:
return decoder_output, decoder_final_state
# bos is added in feedback
# so target_ids is predict_ids
if not TIME_MAJOR:
ground_truth_ids = tf.transpose(target_ids, [1, 0])
else:
ground_truth_ids = target_ids
# construct the loss
if mode == MODE.RL:
# Creates a variable to hold the global_step.
global_step_tensor = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
scope='global_step')[0]
rl_time_steps = tf.floordiv(
tf.maximum(global_step_tensor - burn_in_step, 0), increment_step)
start_rl_step = target_seq_length - rl_time_steps
baseline_states = tf.stop_gradient(decoder_output.baseline_states)
predict_ids = tf.stop_gradient(decoder_output.predicted_ids)
# TODO: bug in tensorflow
ground_or_predict_ids = tf.cond(tf.greater(rl_time_steps,
0), lambda: predict_ids,
lambda: ground_truth_ids)
reward, sequence_length = tf.py_func(
func=_py_func,
inp=[ground_or_predict_ids, ground_truth_ids, trg_vocab.eos_id],
Tout=[tf.float32, tf.int32],
name='reward')
sequence_length.set_shape((None, ))
total_loss_avg, entropy_loss_avg, reward_loss_rmse, reward_predicted \
= rl_sequence_loss(
logits=decoder_output.logits,
predict_ids=predict_ids,
sequence_length=sequence_length,
baseline_states=baseline_states,
start_rl_step=start_rl_step,
reward=reward)
return decoder_output, total_loss_avg, entropy_loss_avg, \
reward_loss_rmse, reward_predicted
else:
total_loss_avg = cross_entropy_sequence_loss(
logits=decoder_output.logits,
targets=ground_truth_ids,
sequence_length=target_seq_length)
return decoder_output, total_loss_avg, total_loss_avg, \
tf.to_float(0.), tf.to_float(0.)
#Readout Layer
with tf.name_scope('fc2') as scope:
W_fc2 = weight_variable([1024, NUM_CLASSES])
b_fc2 = bias_variable([NUM_CLASSES])
with tf.name_scope('RELU') as scope:
y_conv =tf.nn.relu(tf.matmul(h_fc1_drop, W_fc2)+b_fc2)
#Train and Evaluate the Model############################################################
# 評価系の関数を用意
cross_entropy = -tf.reduce_sum(tf.square(y_ - y_conv))#-tf.reduce_sum(y_*tf.log(y_conv))←うまく行かず
train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
divide = tf.constant([0.7])
correct_prediction = tf.equal(tf.floordiv(y_conv, divide), y_ )#01に修正divideでわって切り捨て
accuracy = tf.reduce_sum(tf.cast(correct_prediction, tf.float32), name = "accuracy")
sess.run(tf.initialize_all_variables())
#tensorboard 書き出し用
tf.scalar_summary("accuracy" , accuracy)
print "befor merge all"
# 全ての要約をマージしてそれらを /tmp/mnist_logs に書き出します。
merged = tf.merge_all_summaries()
writer = tf.train.SummaryWriter("/tmp/logs",sess.graph)
print "NetWorkMaking Fin"
def _model_fn(self, model_inputs, model_targets, mode):
inputs = {}
if self._hparams.context_actions:
inputs['context_frames'] = model_inputs['images'][:, :self._hparams
.context_actions]
inputs['start_images'] = model_inputs['images'][:, self._hparams.
context_actions]
inputs['goal_images'] = model_inputs['images'][:, -1]
n_xy = (len(self._hparams.pivots[0]) +
1) * (len(self._hparams.pivots[1]) + 1)
n_z = len(self._hparams.pivots[2]) + 1
n_theta = len(self._hparams.pivots[3]) + 1
if mode == tf.estimator.ModeKeys.TRAIN:
one_hot_actions = _binarize(model_targets['actions'],
self._hparams.pivots)
if self._hparams.context_actions:
inputs['context_actions'] = tf.concat([
x[:, :self._hparams.context_actions]
for x in one_hot_actions
], -1)
real_pred_actions = [
x[:, self._hparams.context_actions:] for x in one_hot_actions
]
inputs['real_actions'] = tf.concat(real_pred_actions, -1)
inputs['T'] = model_targets['actions'].get_shape().as_list(
)[1] - self._hparams.context_actions
else:
assert model_inputs[
'adim'] == 4, "only supports [x,y,z,theta] action space for now!"
inputs['T'] = model_inputs['T'] - self._hparams.context_actions
if self._hparams.context_actions:
one_hot_actions = _binarize(model_inputs['context_actions'],
self._hparams.pivots)
inputs['context_actions'] = tf.concat([
x[:, :self._hparams.context_actions]
for x in one_hot_actions
], -1)
inputs['adim'] = (len(self._hparams.pivots[0]) + 1) * (
len(self._hparams.pivots[1]) + 1) + sum(
[len(arr) + 1 for arr in self._hparams.pivots[2:]])
# build the graph
self._model_graph = model_graph = self._graph_class()
if self._num_gpus <= 1:
outputs = model_graph.build_graph(mode, inputs, self._hparams,
self._graph_scope)
else:
# TODO: add multi-gpu support
raise NotImplementedError
# train
if mode == tf.estimator.ModeKeys.TRAIN:
global_step = tf.train.get_or_create_global_step()
lr, optimizer = tf_utils.build_optimizer(self._hparams.lr,
self._hparams.beta1,
self._hparams.beta2,
global_step=global_step)
pred_xy = outputs['pred_actions'][:, :, :n_xy]
pred_z = outputs['pred_actions'][:, :, n_xy:n_z + n_xy]
pred_theta = outputs['pred_actions'][:, :, n_z + n_xy:]
pred_one_hots = [pred_xy, pred_z, pred_theta]
losses = [
tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits_v2(real, pred))
for real, pred in zip(real_pred_actions, pred_one_hots)
]
loss = sum(losses)
print('computing gradient and train_op')
g_train_op = optimizer.minimize(loss, global_step=global_step)
est = tf.estimator.EstimatorSpec(mode,
loss=loss,
train_op=g_train_op)
scalar_summaries = {}
if 'ground_truth_sampling_mean' in outputs:
scalar_summaries['ground_truth_sampling_mean'] = outputs[
'ground_truth_sampling_mean']
for k, loss in zip(['xy_loss', 'z_loss', 'theta_loss'], losses):
scalar_summaries[k] = loss
return est, scalar_summaries, {}
#test
means = tf.convert_to_tensor(self._hparams.means)
pred_xy = outputs['pred_actions'][:, :, :n_xy]
pred_z = outputs['pred_actions'][:, :, n_xy:n_z + n_xy]
pred_theta = outputs['pred_actions'][:, :, n_z + n_xy:]
pred_xy = tf.reshape(
tf.random.categorical(tf.reshape(pred_xy, (-1, n_xy)),
1,
dtype=tf.int32), (-1, inputs['T']))
pred_x, pred_y = tf.mod(pred_xy,
len(self._hparams.pivots[0]) + 1), tf.floordiv(
pred_xy,
len(self._hparams.pivots[0]) + 1)
pred_z = tf.reshape(
tf.random.categorical(tf.reshape(pred_z, (-1, n_z)),
1,
dtype=tf.int32), (-1, inputs['T']))
pred_theta = tf.reshape(
tf.random.categorical(tf.reshape(pred_theta, (-1, n_theta)),
1,
dtype=tf.int32), (-1, inputs['T']))
outputs['pred_actions'] = tf.concat([
tf.gather(means[i], indices)[:, :, None]
for i, indices in enumerate([pred_x, pred_y, pred_z, pred_theta])
],
axis=-1)
return outputs['pred_actions']
x = tf.add(a, b, name='add')
writer = tf.summary.FileWriter('./graphs/simple', tf.get_default_graph())
with tf.Session() as sess:
# writer = tf.summary.FileWriter('./graphs', sess.graph)
print(sess.run(x))
writer.close() # close the writer when you’re done using it
# Example 2: The wonderful wizard of div
a = tf.constant([2, 2], name='a')
b = tf.constant([[0, 1], [2, 3]], name='b')
with tf.Session() as sess:
print(sess.run(tf.div(b, a)))
print(sess.run(tf.divide(b, a)))
print(sess.run(tf.truediv(b, a)))
print(sess.run(tf.floordiv(b, a)))
# print(sess.run(tf.realdiv(b, a)))
print(sess.run(tf.truncatediv(b, a)))
print(sess.run(tf.floor_div(b, a)))
# Example 3: multiplying tensors
a = tf.constant([10, 20], name='a')
b = tf.constant([2, 3], name='b')
with tf.Session() as sess:
print(sess.run(tf.multiply(a, b)))
print(sess.run(tf.tensordot(a, b, 1)))
# Example 4: Python native type
t_0 = 19
x = tf.zeros_like(t_0) # ==> 0
def _pad_top_bottom():
pad_top = tf.floordiv(width - height, 2)
pad_bottom = width - height - pad_top
return [[pad_top, pad_bottom], [0, 0], [0, 0]]
import tensorflow as tf
import numpy as np
# 计算图的其他操作
sess = tf.Session()
#div()函数返回值的数据类型与输入数据类型一致
print(sess.run(tf.div(3, 4)))
#truediv()函数会将计算结果强制转换为浮点数类型
print(sess.run(tf.truediv(3, 4)))
#floordiv()函数，会将计算结果向下取整
print(sess.run(tf.floordiv(5, 4)))
#mod()取模运算，返回除法的余数
print(sess.run(tf.mod(22.0, 5.0)))
#通过cross()函数计算两个张量级的点积。只为三维向量定义。
print(sess.run(tf.cross([1., 0., 0.], [0., 1., 0.])))
# 组合预处理函数生成自定义函数
print(sess.run(tf.div(tf.sin(3.1416 / 4.), tf.cos(3.1416 / 4.))))
# 延伸学习
#创建一个自定义二次多项式函数，3x^2-x+10
def custom_polynomial(value):
return (tf.subtract(3 * tf.square(value), value) + 10)
def _expand_to_binary_form(self, x, input_bits): """Expands input into binary representation.""" # This operation is inverse of self._pack_binary_form, except padded zeros # are not removed. expand_vector = tf.constant([2**i for i in range(input_bits)], tf.int32) return tf.reshape(tf.mod(tf.floordiv(x, expand_vector), 2), [-1])
def mean_shift(X, bandwidth, max_iter):
(m, n) = X.shape
print m, n
graph = tf.Graph()
with graph.as_default():
with tf.name_scope("input") as scope:
data = tf.constant(X, name="data_points")
b = tf.constant(bandwidth, dtype=tf.float32, name="bandwidth")
m = tf.constant(max_iter, name="maximum_iteration")
# n_samples = tf.constant(m, name="no_of_samples")
# n_features = tf.constant(n, name="no_of_features")
# with tf.name_scope("seeding") as scope:
# seed = tf.placeholder(tf.float32, [5], name="seed")
with tf.name_scope("mean_shifting") as scope:
old_mean = tf.placeholder(tf.float32, [n], name="old_mean")
neighbors = tf.placeholder(tf.float32, [None, n], name="neighbors")
new_mean = tf.reduce_mean(neighbors, 0)
euclid_dist = tf.sqrt(tf.reduce_sum(
tf.pow(tf.sub(old_mean, new_mean), 2)),
name="mean_distance")
center_intensity_dict = {}
nbrs = NearestNeighbors(radius=bandwidth).fit(X)
sess = tf.Session()
init = tf.initialize_all_variables()
sess.run(init)
writer = tf.train.SummaryWriter(FLAGS.log_dir, sess.graph_def)
bin_sizes = defaultdict(int)
data_point = tf.placeholder(tf.float32, [n], "data_point")
binned_point = tf.floordiv(data_point, b)
for point in X:
feed = {data_point: point}
bp = sess.run(binned_point, feed_dict=feed)
bin_sizes[tuple(bp)] += 1
bin_seeds = np.array(
[point for point, freq in six.iteritems(bin_sizes) if freq >= 1],
dtype=np.float32)
bin_seeds = bin_seeds * bandwidth
print len(bin_seeds)
j = 0
for x in bin_seeds:
print "Seed ", j, ": ", x
i = 0
o_mean = x
while True:
i_nbrs = nbrs.radius_neighbors([o_mean],
bandwidth,
return_distance=False)[0]
points_within = X[i_nbrs]
feed = {neighbors: points_within}
n_mean = sess.run(new_mean, feed_dict=feed)
feed = {new_mean: n_mean, old_mean: o_mean}
dist = sess.run(euclid_dist, feed_dict=feed)
if dist < 1e-3 * bandwidth or i == max_iter:
center_intensity_dict[tuple(n_mean)] = len(i_nbrs)
break
else:
o_mean = n_mean
print "\t", i, dist, len(i_nbrs)
i += 1
# if j>10:
# break
j += 1
print center_intensity_dict
sorted_by_intensity = sorted(center_intensity_dict.items(),
key=lambda tup: tup[1],
reverse=True)
sorted_centers = np.array([tup[0] for tup in sorted_by_intensity])
unique = np.ones(len(sorted_centers), dtype=np.bool)
nbrs = NearestNeighbors(radius=bandwidth).fit(sorted_centers)
for i, center in enumerate(sorted_centers):
if unique[i]:
neighbor_idxs = nbrs.radius_neighbors([center],
return_distance=False)[0]
unique[neighbor_idxs] = 0
unique[i] = 1 # leave the current point as unique
cluster_centers = sorted_centers[unique]
nbrs = NearestNeighbors(n_neighbors=1).fit(cluster_centers)
labels = np.zeros(154401, dtype=np.int)
distances, idxs = nbrs.kneighbors(X)
labels = idxs.flatten()
return cluster_centers, labels
def network(batch_size=1,
class_num=5990,
character_height=32,
character_width=32,
character_step=8,
is_train=False):
network = {}
network["inputs"] = tf.placeholder(tf.float32, [batch_size, 32, None, 1],
name='inputs')
network["seq_len"] = tf.multiply(
tf.ones(shape=[batch_size], dtype=tf.int32),
tf.floordiv(
tf.shape(network["inputs"])[2] - character_width, character_step))
network["inputs_batch"] = tf.py_func(
sliding_generate_batch_layer,
[network["inputs"], character_width, character_step], tf.float32)
network["inputs_batch"] = tf.reshape(
network["inputs_batch"], [-1, character_height, character_width, 1])
network["conv1"] = tf.layers.conv2d(inputs=network["inputs_batch"],
filters=50,
kernel_size=(3, 3),
padding="same",
activation=None)
network["batch_norm1"] = tf.contrib.layers.batch_norm(network["conv1"],
decay=0.9,
center=True,
scale=True,
epsilon=0.001,
is_training=is_train)
network["batch_norm1"] = tf.nn.relu(network["batch_norm1"])
network["conv2"] = tf.layers.conv2d(inputs=network["batch_norm1"],
filters=100,
kernel_size=(3, 3),
padding="same",
activation=tf.nn.relu)
network["dropout2"] = tf.layers.dropout(inputs=network["conv2"], rate=0.1)
network["conv3"] = tf.layers.conv2d(inputs=network["dropout2"],
filters=100,
kernel_size=(3, 3),
padding="same",
activation=None)
network["dropout3"] = tf.layers.dropout(inputs=network["conv3"], rate=0.1)
network["batch_norm3"] = tf.contrib.layers.batch_norm(network["dropout3"],
decay=0.9,
center=True,
scale=True,
epsilon=0.001,
is_training=is_train)
network["batch_norm3"] = tf.nn.relu(network["batch_norm3"])
network["pool3"] = tf.layers.max_pooling2d(inputs=network["batch_norm3"],
pool_size=[2, 2],
strides=2)
network["conv4"] = tf.layers.conv2d(inputs=network["pool3"],
filters=150,
kernel_size=(3, 3),
padding="same",
activation=None)
network["dropout4"] = tf.layers.dropout(inputs=network["conv4"], rate=0.2)
network["batch_norm4"] = tf.contrib.layers.batch_norm(network["dropout4"],
decay=0.9,
center=True,
scale=True,
epsilon=0.001,
is_training=is_train)
network["batch_norm4"] = tf.nn.relu(network["batch_norm4"])
network["conv5"] = tf.layers.conv2d(inputs=network["batch_norm4"],
filters=200,
kernel_size=(3, 3),
padding="same",
activation=tf.nn.relu)
network["dropout5"] = tf.layers.dropout(inputs=network["conv5"], rate=0.2)
network["conv6"] = tf.layers.conv2d(inputs=network["dropout5"],
filters=200,
kernel_size=(3, 3),
padding="same",
activation=None)
network["dropout6"] = tf.layers.dropout(inputs=network["conv6"], rate=0.2)
network["batch_norm6"] = tf.contrib.layers.batch_norm(network["dropout6"],
decay=0.9,
center=True,
scale=True,
epsilon=0.001,
is_training=is_train)
network["batch_norm6"] = tf.nn.relu(network["batch_norm6"])
network["pool6"] = tf.layers.max_pooling2d(inputs=network["batch_norm6"],
pool_size=[2, 2],
strides=2)
network["conv7"] = tf.layers.conv2d(inputs=network["pool6"],
filters=250,
kernel_size=(3, 3),
padding="same",
activation=None)
network["dropout7"] = tf.layers.dropout(inputs=network["conv7"], rate=0.3)
network["batch_norm7"] = tf.contrib.layers.batch_norm(network["dropout7"],
decay=0.9,
center=True,
scale=True,
epsilon=0.001,
is_training=is_train)
network["batch_norm7"] = tf.nn.relu(network["batch_norm7"])
network["conv8"] = tf.layers.conv2d(inputs=network["batch_norm7"],
filters=300,
kernel_size=(3, 3),
padding="same",
activation=tf.nn.relu)
network["dropout8"] = tf.layers.dropout(inputs=network["conv8"], rate=0.3)
network["conv9"] = tf.layers.conv2d(inputs=network["dropout8"],
filters=300,
kernel_size=(3, 3),
padding="same",
activation=None)
network["dropout9"] = tf.layers.dropout(inputs=network["conv9"], rate=0.3)
network["batch_norm9"] = tf.contrib.layers.batch_norm(network["dropout9"],
decay=0.9,
center=True,
scale=True,
epsilon=0.001,
is_training=is_train)
network["batch_norm9"] = tf.nn.relu(network["batch_norm9"])
network["pool9"] = tf.layers.max_pooling2d(inputs=network["batch_norm9"],
pool_size=[2, 2],
strides=2)
network["conv10"] = tf.layers.conv2d(inputs=network["pool9"],
filters=350,
kernel_size=(3, 3),
padding="same",
activation=None)
network["dropout10"] = tf.layers.dropout(inputs=network["conv10"],
rate=0.4)
network["batch_norm10"] = tf.contrib.layers.batch_norm(
network["dropout10"],
decay=0.9,
center=True,
scale=True,
epsilon=0.001,
is_training=is_train)
network["batch_norm10"] = tf.nn.relu(network["batch_norm10"])
network["conv11"] = tf.layers.conv2d(inputs=network["batch_norm10"],
filters=400,
kernel_size=(3, 3),
padding="same",
activation=tf.nn.relu)
network["dropout11"] = tf.layers.dropout(inputs=network["conv11"],
rate=0.4)
network["conv12"] = tf.layers.conv2d(inputs=network["dropout11"],
filters=400,
kernel_size=(3, 3),
padding="same",
activation=None)
network["dropout12"] = tf.layers.dropout(inputs=network["conv12"],
rate=0.4)
network["batch_norm12"] = tf.contrib.layers.batch_norm(
network["dropout12"],
decay=0.9,
center=True,
scale=True,
epsilon=0.001,
is_training=is_train)
network["batch_norm12"] = tf.nn.relu(network["batch_norm12"])
#2*2*400
network["pool12"] = tf.layers.max_pooling2d(inputs=network["batch_norm12"],
pool_size=[2, 2],
strides=2)
network["flatten"] = tf.contrib.layers.flatten(network["pool12"])
network["fc1"] = tf.contrib.layers.fully_connected(
inputs=network["flatten"], num_outputs=900, activation_fn=tf.nn.relu)
network["dropout_fc1"] = tf.layers.dropout(inputs=network["fc1"], rate=0.5)
network["fc2"] = tf.contrib.layers.fully_connected(
inputs=network["dropout_fc1"],
num_outputs=200,
activation_fn=tf.nn.relu)
if is_train:
network["fc3"] = tf.contrib.layers.fully_connected(
inputs=network["fc2"], num_outputs=class_num, activation_fn=None)
else:
network["fc3"] = tf.contrib.layers.fully_connected(
inputs=network["fc2"],
num_outputs=class_num,
activation_fn=tf.nn.sigmoid)
network["outputs"] = tf.reshape(network["fc3"],
[batch_size, -1, class_num])
network["outputs"] = tf.transpose(network["outputs"], (1, 0, 2))
return network
def test_FloorDiv(self):
t = tf.floordiv(*self.random((3, 5), (3, 5)))
self.check(t)
def multilevel_crop_and_resize(features,
boxes,
output_size=7,
is_gpu_inference=False):
"""Crop and resize on multilevel feature pyramid.
Generate the (output_size, output_size) set of pixels for each input box
by first locating the box into the correct feature level, and then cropping
and resizing it using the correspoding feature map of that level.
Args:
features: A dictionary with key as pyramid level and value as features. The
features are in shape of [batch_size, height_l, width_l, num_filters].
boxes: A 3-D Tensor of shape [batch_size, num_boxes, 4]. Each row represents
a box with [y1, x1, y2, x2] in un-normalized coordinates.
output_size: A scalar to indicate the output crop size.
is_gpu_inference: whether to build the model for GPU inference.
Returns:
A 5-D tensor representing feature crop of shape
[batch_size, num_boxes, output_size, output_size, num_filters].
"""
with tf.name_scope('multilevel_crop_and_resize'):
levels = features.keys()
min_level = min(levels)
max_level = max(levels)
_, max_feature_height, max_feature_width, _ = (
features[min_level].get_shape().as_list())
# Stack feature pyramid into a features_all of shape
# [batch_size, levels, height, width, num_filters].
features_all = []
for level in range(min_level, max_level + 1):
features_all.append(
tf.image.pad_to_bounding_box(features[level], 0, 0,
max_feature_height,
max_feature_width))
features_all = tf.stack(features_all, axis=1)
# Assign boxes to the right level.
box_width = tf.squeeze(boxes[:, :, 3:4] - boxes[:, :, 1:2], axis=-1)
box_height = tf.squeeze(boxes[:, :, 2:3] - boxes[:, :, 0:1], axis=-1)
areas_sqrt = tf.sqrt(box_height * box_width)
levels = tf.floordiv(tf.log(tf.div(areas_sqrt, 224.0)),
tf.log(2.0)) + 4.0
if not is_gpu_inference:
levels = tf.cast(levels, dtype=tf.int32)
# Map levels between [min_level, max_level].
levels = tf.minimum(
float(max_level) if is_gpu_inference else max_level,
tf.maximum(levels,
float(min_level) if is_gpu_inference else min_level))
# Project box location and sizes to corresponding feature levels.
scale_to_level = tf.cast(tf.pow(
tf.constant(2.0),
levels if is_gpu_inference else tf.cast(levels, tf.float32)),
dtype=boxes.dtype)
boxes /= tf.expand_dims(scale_to_level, axis=2)
box_width /= scale_to_level
box_height /= scale_to_level
boxes = tf.concat([
boxes[:, :, 0:2],
tf.expand_dims(box_height, -1),
tf.expand_dims(box_width, -1)
],
axis=-1)
# Map levels to [0, max_level-min_level].
levels -= min_level
level_strides = tf.pow(
[[2.0]],
levels if is_gpu_inference else tf.cast(levels, tf.float32))
boundary = tf.cast(
tf.concat([
tf.expand_dims([[tf.cast(max_feature_height, tf.float32)]] /
level_strides - 1,
axis=-1),
tf.expand_dims([[tf.cast(max_feature_width, tf.float32)]] /
level_strides - 1,
axis=-1),
],
axis=-1), boxes.dtype)
return selective_crop_and_resize(features_all, boxes, levels, boundary,
output_size, is_gpu_inference)
def get_updates(self, loss, params):
grads = self.get_gradients(loss, params)
self.updates = [K.update_add(self.iterations, 1)]
lr = self.lr
completed_updates = K.cast(
tf.floordiv(self.iterations, self.accum_iters), K.floatx())
if self.initial_decay > 0:
lr = lr * (1. / (1. + self.decay * completed_updates))
t = completed_updates + 1
lr_t = lr * (K.sqrt(1. - K.pow(self.beta_2, t)) /
(1. - K.pow(self.beta_1, t)))
# self.iterations incremented after processing a batch
# batch: 1 2 3 4 5 6 7 8 9
# self.iterations: 0 1 2 3 4 5 6 7 8
# update_switch = 1: x x (if accum_iters=4)
update_switch = K.equal((self.iterations + 1) % self.accum_iters, 0)
update_switch = K.cast(update_switch, K.floatx())
ms = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
vs = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
gs = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
if self.amsgrad:
vhats = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
else:
vhats = [K.zeros(1) for _ in params]
self.weights = [self.iterations] + ms + vs + vhats
for p, g, m, v, vhat, tg in zip(params, grads, ms, vs, vhats, gs):
sum_grad = tg + g
avg_grad = sum_grad / self.accum_iters_float
m_t = (self.beta_1 * m) + (1. - self.beta_1) * avg_grad
v_t = (self.beta_2 * v) + (1. - self.beta_2) * K.square(avg_grad)
if self.amsgrad:
vhat_t = K.maximum(vhat, v_t)
p_t = p - lr_t * m_t / (K.sqrt(vhat_t) + self.epsilon)
self.updates.append(
K.update(vhat, (1 - update_switch) * vhat +
update_switch * vhat_t))
else:
p_t = p - lr_t * m_t / (K.sqrt(v_t) + self.epsilon)
self.updates.append(
K.update(m, (1 - update_switch) * m + update_switch * m_t))
self.updates.append(
K.update(v, (1 - update_switch) * v + update_switch * v_t))
self.updates.append(K.update(tg, (1 - update_switch) * sum_grad))
new_p = p_t
# Apply constraints.
if getattr(p, 'constraint', None) is not None:
new_p = p.constraint(new_p)
self.updates.append(
K.update(p, (1 - update_switch) * p + update_switch * new_p))
return self.updates
def __rfloordiv__(self, other):
return tf.floordiv(other, self)
def body(x_in, y_in, domain_in, i_in, cond_in):
# Create graph for model logits and predictions
logits = model.get_logits(x_in)
preds = tf.nn.softmax(logits)
preds_onehot = tf.one_hot(tf.argmax(preds, axis=-1), depth=nb_classes)
# create the Jacobian graph
list_derivatives = []
for class_ind in xrange(nb_classes):
derivatives = tf.gradients(logits[:, class_ind], x_in)
list_derivatives.append(derivatives[0])
grads = tf.reshape(tf.stack(list_derivatives), shape=[nb_classes, -1, nb_features])
# Compute the Jacobian components
# To help with the computation later, reshape the target_class and other_class to [nb_classes, -1, 1].
# The last dimention is added to allow broadcasting later.
target_class = tf.reshape(tf.transpose(y_in, perm=[1, 0]), shape=[nb_classes, -1, 1])
other_classes = tf.cast(tf.not_equal(target_class, 1), tf_dtype)
grads_target = reduce_sum(grads * target_class, axis=0)
grads_other = reduce_sum(grads * other_classes, axis=0)
# Remove the already-used input features from the search space
# Subtract 2 times the maximum value from those value so that
# they won't be picked later
increase_coef = (4 * int(increase) - 2) * tf.cast(tf.equal(domain_in, 0), tf_dtype)
target_tmp = grads_target
target_tmp -= increase_coef * reduce_max(tf.abs(grads_target), axis=-1, keepdims=True)
target_sum = tf.reshape(target_tmp, shape=[-1, nb_features, 1]) + tf.reshape(target_tmp, shape=[-1, 1, nb_features])
other_tmp = grads_other
other_tmp += increase_coef * reduce_max(tf.abs(grads_other), axis=-1, keepdims=True)
other_sum = tf.reshape(other_tmp, shape=[-1, nb_features, 1]) + tf.reshape(other_tmp, shape=[-1, 1, nb_features])
# Create a mask to only keep features that match conditions
if increase:
scores_mask = ((target_sum > 0) & (other_sum < 0))
else:
scores_mask = ((target_sum < 0) & (other_sum > 0))
# Create a 2D numpy array of scores for each pair of candidate features
scores = tf.cast(scores_mask, tf_dtype) * (-target_sum * other_sum) * zero_diagonal
# Extract the best two pixels
best = tf.argmax(tf.reshape(scores, shape=[-1, nb_features * nb_features]), axis=-1)
p1 = tf.mod(best, nb_features)
p2 = tf.floordiv(best, nb_features)
p1_one_hot = tf.one_hot(p1, depth=nb_features)
p2_one_hot = tf.one_hot(p2, depth=nb_features)
# Check if more modification is needed for each sample
mod_not_done = tf.equal(reduce_sum(y_in * preds_onehot, axis=-1), 0)
cond = mod_not_done & (reduce_sum(domain_in, axis=-1) >= 2)
# Update the search domain
cond_float = tf.reshape(tf.cast(cond, tf_dtype), shape=[-1, 1])
to_mod = (p1_one_hot + p2_one_hot) * cond_float
domain_out = domain_in - to_mod
# Apply the modification to the images
to_mod_reshape = tf.reshape(to_mod, shape=([-1] + x_in.shape[1:].as_list()))
if increase:
x_out = tf.minimum(clip_max, x_in + to_mod_reshape * theta)
else:
x_out = tf.maximum(clip_min, x_in - to_mod_reshape * theta)
# Increase the iterator, and check if all misclassifications are done
i_out = tf.add(i_in, 1)
cond_out = reduce_any(cond)
return x_out, y_in, domain_out, i_out, cond_out
def _pad_left_right():
pad_left = tf.floordiv(height - width, 2)
pad_right = height - width - pad_left
return [[0, 0], [pad_left, pad_right], [0, 0]]
print('*9*'.center(50, '-'))
print(sess.run(A))
print(sess.run(A + B))
print(sess.run(tf.add(A, B)))
print(sess.run(A - B))
print(sess.run(tf.subtract(A, B)))
print(sess.run(tf.matmul(B, identity_matrix)))
print(sess.run(tf.transpose(C)))
print(sess.run(tf.matrix_determinant(D)))
print(sess.run(tf.matrix_inverse(D)))
print(sess.run(tf.cholesky(identity_matrix)))
print(sess.run(tf.self_adjoint_eig(D)))
print('*10*'.center(50, '-'))
# print(sess.run(tf.div(3,4)))
print(sess.run(tf.truediv(3, 4)))
print(sess.run(tf.floordiv(3.0, 4.0)))
print(sess.run(tf.mod(22.0, 5.0)))
print(sess.run(tf.cross([1., 0., 0.], [0., 1., 0.])))
print('*12*'.center(50, '-'))
print(sess.run(output, feed_dict={input1: [3.], input2: [5]}))
"""
TensorFlow是采用数据流图（data flow graphs）来计算, 所以首先我们得创建一个数据流流图,
然后再将我们的数据（数据以张量(tensor)的形式存在）放在数据流图中计算. 节点（Nodes）在图
中表示数学操作,图中的线（edges）则表示在节点间相互联系的多维数据数组, 即张量（tensor).
训练模型时tensor会不断的从数据流图中的一个节点flow到另一节点, 这就是TensorFlow名字的由来.
Tensor 张量意义
张量（Tensor):
def __floordiv__(self, other):
return tf.floordiv(self, other)
# in Tensorflow # Declaring Operations import matplotlib.pyplot as plt import numpy as np import tensorflow as tf from tensorflow.python.framework import ops ops.reset_default_graph() # Open graph session sess = tf.Session() # div() vs truediv() vs floordiv() print(sess.run(tf.div(3,4))) print(sess.run(tf.truediv(3,4))) print(sess.run(tf.floordiv(3.0,4.0))) # Mod function print(sess.run(tf.mod(22.0,5.0))) # Cross Product print(sess.run(tf.cross([1.,0.,0.],[0.,1.,0.]))) # Trig functions print(sess.run(tf.sin(3.1416))) print(sess.run(tf.cos(3.1416))) # Tangemt print(sess.run(tf.div(tf.sin(3.1416/4.), tf.cos(3.1416/4.)))) # Custom operation test_nums = range(15)
#--------------1.3基本运算------------------- # Declaring Operations import matplotlib.pyplot as plt '''基本运算 add()加法 subtract() multiply() matmul()矩阵相乘 div()除法 ''' # div() vs truediv() vs floordiv() print(sess.run(tf.div(3, 4))) print(sess.run(tf.truediv(3, 4))) #浮点运算 print(sess.run(tf.floordiv(3.0, 4.0))) #对浮点数进行整数除法运算 # Mod function取模运算 print(sess.run(tf.mod(22.0, 5.0))) # Cross Product，点积运算，只有三维向量可用 print(sess.run(tf.cross([1., 0., 0.], [0., 1., 0.]))) #三角函数运算. # Trig functions print(sess.run(tf.sin(3.1416))) print(sess.run(tf.cos(3.1416))) # Tangemt print(sess.run(tf.div(tf.sin(3.1416 / 4.), tf.cos(3.1416 / 4.)))) #自定义函数运算 # Custom operation test_nums = range(15)
