def _testMultipleReduceJoin(self, input_array, reduction_indices, separator=" "):
"""Tests reduce_join for one input and multiple reduction_indices.
Does so by comparing the output to that from nested reduce_string_joins.
The correctness of single-dimension reduce_join is verified by other
tests below using _testReduceJoin.
Args:
input_array: The input to test.
reduction_indices: The indices to reduce.
separator: The separator to use when joining.
"""
num_dims = len(input_array.shape)
truth_red_indices = reduction_indices or list(reversed(xrange(num_dims)))
with self.test_session():
output = tf.reduce_join(
inputs=input_array, reduction_indices=reduction_indices, keep_dims=False, separator=separator
)
output_keep_dims = tf.reduce_join(
inputs=input_array, reduction_indices=reduction_indices, keep_dims=True, separator=separator
)
truth = input_array
for index in truth_red_indices:
truth = tf.reduce_join(inputs=truth, reduction_indices=index, keep_dims=True, separator=separator)
truth_squeezed = tf.squeeze(truth, squeeze_dims=truth_red_indices)
output_array = output.eval()
output_keep_dims_array = output_keep_dims.eval()
truth_array = truth.eval()
truth_squeezed_array = truth_squeezed.eval()
self.assertAllEqualUnicode(truth_array, output_keep_dims_array)
self.assertAllEqualUnicode(truth_squeezed_array, output_array)
self.assertAllEqual(truth.get_shape(), output_keep_dims.get_shape())
self.assertAllEqual(truth_squeezed.get_shape(), output.get_shape())
def create_metric_ops(self, _inputs, labels, predictions):
"""Creates (value, update_op) tensors
"""
with tf.variable_scope(self._name):
# Join tokens into single strings
predictions_flat = tf.reduce_join(
predictions["predicted_tokens"], 1, separator=self._separator)
labels_flat = tf.reduce_join(
labels["target_tokens"], 1, separator=self._separator)
sources_value, sources_update = accumulate_strings(
values=predictions_flat, name="sources")
targets_value, targets_update = accumulate_strings(
values=labels_flat, name="targets")
metric_value = tf.py_func(
func=self._py_func,
inp=[sources_value, targets_value],
Tout=tf.float32,
name="value")
with tf.control_dependencies([sources_update, targets_update]):
update_op = tf.identity(metric_value, name="update_op")
return metric_value, update_op
def testZeroDims(self):
valid_truth_shape = [0]
with self.test_session():
inputs = np.zeros([0, 1], dtype=str)
with self.assertRaisesRegexp(ValueError, "dimension 0 with size 0"):
tf.reduce_join(inputs=inputs, reduction_indices=0)
valid = tf.reduce_join(inputs=inputs, reduction_indices=1)
valid_array_shape = valid.eval().shape
self.assertAllEqualUnicode(valid_truth_shape, valid_array_shape)
def testInvalidArgsUnknownShape(self):
with self.test_session():
placeholder = tf.placeholder(tf.string, name="placeholder")
index_too_high = tf.reduce_join(placeholder, reduction_indices=1)
duplicate_index = tf.reduce_join(placeholder, reduction_indices=[-1, 1])
with self.assertRaisesOpError("Invalid reduction dimension 1"):
index_too_high.eval(feed_dict={placeholder.name: [""]})
with self.assertRaisesOpError("Duplicate reduction dimension 1"):
duplicate_index.eval(feed_dict={placeholder.name: [[""]]})
def main(_):
path_to_image_file = FLAGS.image
path_to_restore_checkpoint_file = FLAGS.restore_checkpoint
image = tf.image.decode_jpeg(tf.read_file(path_to_image_file), channels=3)
image = tf.reshape(image, [64, 64, 3])
image = tf.image.convert_image_dtype(image, dtype=tf.float32)
image = tf.multiply(tf.subtract(image, 0.5), 2)
image = tf.image.resize_images(image, [54, 54])
images = tf.reshape(image, [1, 54, 54, 3])
length_logits, digits_logits = Model.inference(images, drop_rate=0.0)
length_predictions = tf.argmax(length_logits, axis=1)
digits_predictions = tf.argmax(digits_logits, axis=2)
digits_predictions_string = tf.reduce_join(tf.as_string(digits_predictions), axis=1)
with tf.Session() as sess:
restorer = tf.train.Saver()
restorer.restore(sess, path_to_restore_checkpoint_file)
length_predictions_val, digits_predictions_string_val = sess.run([length_predictions, digits_predictions_string])
length_prediction_val = length_predictions_val[0]
digits_prediction_string_val = digits_predictions_string_val[0]
print 'length: %d' % length_prediction_val
print 'digits: %s' % digits_prediction_string_val
def lowercase(self, raw_post):
split_chars = tf.string_split(tf.reshape(raw_post, [-1]), delimiter="").values
upchar_inds = self.upchars_lut.lookup(split_chars)
return tf.reduce_join(tf.map_fn(lambda x: tf.cond(x[0] > 25,
lambda: x[1],
lambda: self.lchars[x[0]]),
(upchar_inds, split_chars), dtype=tf.string))
def get_text(self, ids):
"""Returns a string corresponding to a sequence of character ids.
Args:
ids: a tensor with shape [batch_size, max_sequence_length]
"""
return tf.reduce_join(
self.table.lookup(tf.to_int64(ids)), reduction_indices=1)
def testUnknownShape(self):
input_array = [["a"], ["b"]]
truth = ["ab"]
with self.test_session():
placeholder = tf.placeholder(tf.string, name="placeholder")
reduced = tf.reduce_join(placeholder, reduction_indices=0)
output_array = reduced.eval(feed_dict={placeholder.name: input_array})
self.assertAllEqualUnicode(truth, output_array)
def testInvalidArgsUnknownIndices(self):
with self.test_session():
placeholder = tf.placeholder(tf.int32, name="placeholder")
reduced = tf.reduce_join(["test", "test2"], reduction_indices=placeholder)
with self.assertRaisesOpError("reduction dimension -2"):
reduced.eval(feed_dict={placeholder.name: -2})
with self.assertRaisesOpError("reduction dimension 2"):
reduced.eval(feed_dict={placeholder.name: 2})
def resize_sen(self, raw, mxlen):
"""
Splits and rejoins a string to ensure that tokens meet
the required max len.
"""
raw_tokens = tf.string_split(tf.reshape(raw, [-1])).values
# sentence length > mxlen
raw_post = tf.reduce_join(raw_tokens[:mxlen], separator=" ")
return raw_post
def testUnknownIndices(self):
input_array = [["this", "is", "a", "test"],
["please", "do", "not", "panic"]]
truth_dim_zero = ["thisplease", "isdo", "anot", "testpanic"]
truth_dim_one = ["thisisatest", "pleasedonotpanic"]
with self.test_session():
placeholder = tf.placeholder(tf.int32, name="placeholder")
reduced = tf.reduce_join(input_array, reduction_indices=placeholder)
output_array_dim_zero = reduced.eval(feed_dict={placeholder.name: [0]})
output_array_dim_one = reduced.eval(feed_dict={placeholder.name: [1]})
self.assertAllEqualUnicode(truth_dim_zero, output_array_dim_zero)
self.assertAllEqualUnicode(truth_dim_one, output_array_dim_one)
def _testReduceJoin(self, input_array, truth, truth_shape, reduction_indices, keep_dims=False, separator=""):
"""Compares the output of reduce_join to an expected result.
Args:
input_array: The string input to be joined.
truth: An array or np.array of the expected result.
truth_shape: An array or np.array of the expected shape.
reduction_indices: The indices to reduce over.
keep_dims: Whether or not to retain reduced dimensions.
separator: The separator to use for joining.
"""
with self.test_session():
output = tf.reduce_join(
inputs=input_array, reduction_indices=reduction_indices, keep_dims=keep_dims, separator=separator
)
output_array = output.eval()
self.assertAllEqualUnicode(truth, output_array)
self.assertAllEqual(truth_shape, output.get_shape())
def markdown_table(step):
# The text summary can also contain Markdown, including Markdown
# tables. Markdown tables look like this:
#
# | hello | there |
# |-------|-------|
# | this | is |
# | a | table |
#
# The leading and trailing pipes in each row are optional, and the text
# doesn't actually have to be neatly aligned, so we can create these
# pretty easily. Let's do so.
header_row = 'Pounds of chocolate | Happiness'
chocolate = tf.range(step)
happiness = tf.square(chocolate + 1)
chocolate_column = tf.as_string(chocolate)
happiness_column = tf.as_string(happiness)
table_rows = tf.string_join([chocolate_column, " | ", happiness_column])
table_body = tf.reduce_join(table_rows, separator='\n')
table = tf.string_join([header_row, "---|---", table_body], separator='\n')
preamble = 'We conducted an experiment and found the following data:\n\n'
result = tf.string_join([preamble, table])
tf.summary.text('chocolate_study', result)
def bpe2word(tensor):
joined_tensor = tf.reduce_join(tensor)
replaced_tensor = tf.regex_replace(joined_tensor, "▁", " ")
return tf.string_split([replaced_tensor]).values
tf.sparse_accumulator_take_gradient() tf.sparse_add() tf.sparse_concat() tf.sparse_conditional_accumulator() tf.sparse_mask() tf.sparse_matmul() tf.sparse_maximum() tf.sparse_merge() tf.sparse_minimum() tf.sparse_reduce_max() tf.sparse_reduce_max_sparse() tf.reduce_all() tf.reduce_any() tf.reduce_join() tf.reduce_logsumexp() tf.reduce_max() tf.reduce_mean() tf.reduce_min() tf.reduce_prod() tf.reduce_sum() tf.reduced_shape() tf.random_crop() tf.random_gamma() tf.random_normal() tf.random_poisson() tf.random_poisson_v2() tf.random_shuffle() tf.random_uniform()
def get_text(self,ids):
return tf.reduce_join(
self.table.lookup(tf.to_int64(ids)),reduction_indices=1
)
def from_tokens(raw, lookup_): gathered = tf.gather(lookup_, tf.cast(raw, tf.int32)) joined = tf.regex_replace(tf.reduce_join(gathered, axis=1), b"<EOS>.*", b"") cleaned = tf.regex_replace(joined, b"_", b" ") tokens = tf.string_split(cleaned, " ") return tokens
outputs = []
for i in range(paral_nets):
h = tf.matmul(x, W1[i]) + b1[i]
# h = tf.matmul(x, W1[i])
# h = tf.sign(h)
h = tf.nn.relu(h)
h2 = tf.matmul(h, W2[i]) + b2[i]
# h2 = tf.sign(h2)
h2 = tf.nn.relu(h2)
logits = tf.matmul(h2, W3[i]) + b3[i]
# logits = tf.matmul(h, W2[i]) + b2[i]
# logits = tf.matmul(h, W2[i])
o = tf.sign(logits)
# outputs.append((o+1)/2)
outputs.append(tf.reduce_join(tf.reduce_join(tf.as_string(tf.cast((o+1)//2,tf.int32)), 0),0))
session = tf.Session()
inputs = [[float(xx) for xx in "{0:07b}".format(i)] for i in range(2**7)]
# N=10
cnt = Counter()
#weights = {}
for i in range(N):
# if i%(N/100) == 0:
print(i)
session.run(tf.global_variables_initializer())
fs = session.run(outputs, feed_dict={x:inputs})
varss = session.run(variables,feed_dict={x:inputs})
#for i,f in enumerate(fs):
def build_network(self,
names,
sources,
targets,
masks,
angles,
views,
is_training=False,
is_validation=False,
is_testing=False,
is_encoding=False):
"""
input:
names : n x String shape names
sources : n x H x W x C source images
targets : (n*m) x H x W x C target images in m views (ground-truth)
masks : (n*m) x H x W x 1 target boolean masks in m views (ground-truth)
angles : (n*m) x 4 viewing angle parameters (m=1 for continuous view prediction)
views : vw.Views view points information
is_training : boolean whether it is in training routine
is_validation : boolean whether it is handling validation data set
is_testing : boolean whether it is in testing routine
is_encoding : boolean whether it is encoding input
"""
print('Building network...')
source_size = sources.get_shape().as_list()
if self.config.continuous_view:
num_output_views = 1
else:
num_output_views = views.num_views
# scope names
var_scope_G = 'G_net'
var_scope_D = 'D_net'
bn_scope_G = 'G_bn'
bn_scope_D = 'D_bn'
train_summary_G_name = 'train_summary_G'
train_summary_D_name = 'train_summary_D'
valid_summary_name = 'valid_summary'
# generator
num_channels = targets.get_shape()[3].value
if not self.config.continuous_view:
with tf.variable_scope(var_scope_G):
with tf_framework.arg_scope(layer.unet_scopes(bn_scope_G)):
preds, features = network.generateUNet(
sources, num_output_views,
num_channels) # (n*m) x H x W x C ; n x D
else:
with tf.variable_scope(var_scope_G):
with tf_framework.arg_scope(layer.cnet_scopes(bn_scope_G)):
preds, features = network.generateCNet(
sources, angles, num_channels) # n x H x W x C ; n x D
if is_encoding:
self.encode_names = names
self.encode_features = features
return # all stuffs below are irrelevant to encoding pass
# extract prediction contents
preds_content = tf.slice(preds, [0, 0, 0, 0],
[-1, -1, -1, num_channels - 1])
preds_mask = tf.slice(preds, [0, 0, 0, num_channels - 1],
[-1, -1, -1, 1])
preds = image.apply_mask(preds_content, preds_mask)
targets_content = tf.slice(targets, [0, 0, 0, 0],
[-1, -1, -1, num_channels - 1])
targets_mask = tf.slice(targets, [0, 0, 0, num_channels - 1],
[-1, -1, -1, 1])
targets = image.apply_mask(targets_content, targets_mask)
if self.config.predict_normal:
preds_normal = tf.slice(preds_content, [0, 0, 0, 0],
[-1, -1, -1, 3])
preds_depth = tf.slice(preds_content, [0, 0, 0, 3],
[-1, -1, -1, 1])
targets_normal = tf.slice(targets_content, [0, 0, 0, 0],
[-1, -1, -1, 3])
targets_depth = tf.slice(targets_content, [0, 0, 0, 3],
[-1, -1, -1, 1])
else:
preds_depth = preds_content
preds_normal = tf.tile(tf.zeros_like(preds_depth), [1, 1, 1, 3])
targets_depth = targets_content
targets_normal = tf.tile(tf.zeros_like(targets_depth),
[1, 1, 1, 3])
# expand tensors
sources_expanded = tf.reshape(
tf.tile(sources, [1, num_output_views, 1, 1]),
[-1, source_size[1], source_size[2], source_size[3]
]) # (n*m) x H x W x C
names_expanded = tf.reshape(
tf.tile(tf.expand_dims(names, 1), [1, num_output_views]), [-1])
names_suffix = [
"--%d" % view for batch in range(source_size[0])
for view in range(num_output_views)
]
names_expanded = tf.reduce_join([names_expanded, names_suffix], 0)
self.names = names_expanded
# discriminator
if not self.config.no_adversarial:
with tf.variable_scope(var_scope_D):
with tf_framework.arg_scope(layer.unet_scopes(bn_scope_D)):
disc_data = tf.concat([targets, preds], 0)
disc_data = tf.concat([
tf.concat([sources_expanded, sources_expanded], 0),
disc_data
], 3) # HACK: insert input data for discrimination in UNet
probs = network.discriminate(disc_data) # (n*m*2)
# losses
# NOTE: learning hyper-parameters
lambda_p = 1.0 # image loss
lambda_a = 0.01 # adversarial loss
dl = loss.compute_depth_loss(preds_depth, targets_depth, masks)
nl = loss.compute_normal_loss(preds_normal, targets_normal, masks)
ml = loss.compute_mask_loss(preds_mask, targets_mask)
loss_g_p = dl + nl + ml
if self.config.no_adversarial:
loss_g_a = 0.0
loss_d_r = 0.0
loss_d_f = 0.0
else:
probs_targets, probs_preds = tf.split(probs, 2, axis=0) # (n*m)
loss_g_a = tf.reduce_sum(-tf.log(tf.maximum(probs_preds, 1e-6)))
loss_d_r = tf.reduce_sum(-tf.log(tf.maximum(probs_targets, 1e-6)))
loss_d_f = tf.reduce_sum(
-tf.log(tf.maximum(1.0 - probs_preds, 1e-6)))
loss_G = loss_g_p * lambda_p + loss_g_a * lambda_a
loss_D = loss_d_r + loss_d_f
if is_validation:
self.valid_losses = tf.stack(
[loss_G, loss_g_p, loss_g_a, loss_D, loss_d_r, loss_d_f])
self.valid_images = tf.stack([
image.encode_raw_batch_images(preds),
image.encode_raw_batch_images(targets),
image.encode_raw_batch_images(preds_normal),
image.encode_raw_batch_images(preds_depth),
image.encode_raw_batch_images(preds_mask)
])
self.valid_summary_losses = tf.placeholder(
tf.float32, shape=self.valid_losses.get_shape())
vG_all, vG_p, vG_a, vD_all, vD_r, vD_f = tf.unstack(
self.valid_summary_losses)
tf.summary.scalar('vG_all',
vG_all,
collections=[valid_summary_name])
tf.summary.scalar('vG_p', vG_p, collections=[valid_summary_name])
tf.summary.scalar('vG_a', vG_a, collections=[valid_summary_name])
tf.summary.scalar('vD_all',
vD_all,
collections=[valid_summary_name])
tf.summary.scalar('vD_r', vD_r, collections=[valid_summary_name])
tf.summary.scalar('vD_f', vD_f, collections=[valid_summary_name])
self.valid_summary_op = tf.summary.merge_all(valid_summary_name)
return # all stuffs below are irrelevant to validation pass
self.train_losses_G = tf.stack([loss_G, loss_g_p, loss_g_a])
self.train_losses_D = tf.stack([loss_D, loss_d_r, loss_d_f])
tf.summary.scalar('G_all', loss_G, collections=[train_summary_G_name])
tf.summary.scalar('G_p', loss_g_p, collections=[train_summary_G_name])
tf.summary.scalar('G_a', loss_g_a, collections=[train_summary_G_name])
tf.summary.scalar('D_all', loss_D, collections=[train_summary_D_name])
tf.summary.scalar('D_r', loss_d_r, collections=[train_summary_D_name])
tf.summary.scalar('D_f', loss_d_f, collections=[train_summary_D_name])
# statistics on variables
all_vars = tf.trainable_variables()
all_vars_G = [var for var in all_vars if var_scope_G in var.name]
all_vars_D = [var for var in all_vars if var_scope_D in var.name]
#print('Num all vars: %d' % len(all_vars))
#print('Num vars on G net: %d' % len(all_vars_G))
#print('Num vars on D net: %d' % len(all_vars_D))
num_params_G = 0
num_params_D = 0
# print('G vars:')
for var in all_vars_G:
num_params_G += np.prod(var.get_shape().as_list())
# print(var.name, var.get_shape().as_list())
# print('D vars:')
for var in all_vars_D:
num_params_D += np.prod(var.get_shape().as_list())
# print(var.name, var.get_shape().as_list())
#print('Num all params: %d + %d = %d' % (num_params_G, num_params_D, num_params_G+num_params_D))
#input('pause')
# optimization
# NOTE: learning hyper-parameters
init_learning_rate = 0.0001
adam_beta1 = 0.9
adam_beta2 = 0.999
opt_step = tf.Variable(0, trainable=False)
learning_rate = tf.train.exponential_decay(init_learning_rate,
global_step=opt_step,
decay_steps=10000,
decay_rate=0.96,
staircase=True)
opt_G = tf.train.AdamOptimizer(learning_rate=learning_rate,
beta1=adam_beta1,
beta2=adam_beta2,
name='ADAM_G')
opt_D = tf.train.AdamOptimizer(learning_rate=learning_rate,
beta1=adam_beta1,
beta2=adam_beta2,
name='ADAM_D')
# opt_G = tf.train.GradientDescentOptimizer(learning_rate=learning_rate, name='SGD_G')
# opt_D = tf.train.GradientDescentOptimizer(learning_rate=learning_rate, name='SGD_D')
grad_G = opt_G.compute_gradients(loss_G,
var_list=all_vars_G,
colocate_gradients_with_ops=True)
self.grad_G_placeholder = [(tf.placeholder(tf.float32,
shape=grad[1].get_shape()),
grad[1]) for grad in grad_G
if grad[0] is not None]
self.grad_G_list = [grad[0] for grad in grad_G if grad[0] is not None]
self.update_G_op = opt_G.apply_gradients(
self.grad_G_placeholder,
global_step=opt_step) # only update opt_step in G net
if not self.config.no_adversarial:
grad_D = opt_D.compute_gradients(loss_D,
var_list=all_vars_D,
colocate_gradients_with_ops=True)
self.grad_D_placeholder = [
(tf.placeholder(tf.float32,
shape=grad[1].get_shape()), grad[1])
for grad in grad_D if grad[0] is not None
]
self.grad_D_list = [
grad[0] for grad in grad_D if grad[0] is not None
]
self.update_D_op = opt_D.apply_gradients(self.grad_D_placeholder)
# visualization stuffs
sources_original, sources_flipped = tf.split(sources_expanded,
2,
axis=3)
if len(self.config.sketch_views) == 1: # single input
sources_front = sources_original
sources_side = tf.ones_like(sources_front) # fake side sketch
sources_top = tf.ones_like(sources_front) # fake top sketch
elif len(self.config.sketch_views) == 2: # double input
sources_front, sources_side = tf.split(sources_original, 2, axis=3)
sources_top = tf.ones_like(sources_front) # fake top sketch
elif len(self.config.sketch_views) == 3: # triple input
sources_front, sources_side, sources_top = tf.split(
sources_original, 3, axis=3)
if sources_front.get_shape()[3].value == 1 and targets.get_shape(
)[3].value == 4:
alpha_front = tf.ones_like(sources_front)
alpha_side = tf.ones_like(sources_side)
alpha_top = tf.ones_like(sources_top)
rgb_front = image.convert_to_rgb(sources_front, channels=3)
rgb_side = image.convert_to_rgb(sources_side, channels=3)
rgb_top = image.convert_to_rgb(sources_top, channels=3)
sources_front = tf.concat([rgb_front, alpha_front], 3)
sources_side = tf.concat([rgb_side, alpha_side], 3)
sources_top = tf.concat([rgb_top, alpha_top], 3)
input_row = tf.concat([sources_front, sources_side], 2)
output_row = tf.concat([targets, preds], 2)
result_tile = tf.concat([input_row, output_row], 1)
result_tile = image.saturate_image(
image.unnormalize_image(result_tile))
tf.summary.image('result', result_tile, 12, [train_summary_G_name])
self.train_summary_G_op = tf.summary.merge_all(train_summary_G_name)
self.train_summary_D_op = tf.summary.merge_all(train_summary_D_name)
# output images
num_sketch_views = len(self.config.sketch_views)
if num_sketch_views == 1:
all_input_row = sources_front
elif num_sketch_views == 2:
all_input_row = tf.concat([sources_front, sources_side], 2)
elif num_sketch_views == 3:
all_input_row = tf.concat(
[sources_front, sources_side, sources_top], 2)
img_input = image.saturate_image(image.unnormalize_image(
all_input_row, maxval=65535.0),
dtype=tf.uint16)
img_gt = image.saturate_image(image.unnormalize_image(targets,
maxval=65535.0),
dtype=tf.uint16)
img_output = image.saturate_image(image.unnormalize_image(
preds, maxval=65535.0),
dtype=tf.uint16)
png_input = image.encode_batch_images(img_input)
png_gt = image.encode_batch_images(img_gt)
png_output = image.encode_batch_images(img_output)
img_normal = image.saturate_image(image.unnormalize_image(
preds_normal, maxval=65535.0),
dtype=tf.uint16)
img_depth = image.saturate_image(image.unnormalize_image(
preds_depth, maxval=65535.0),
dtype=tf.uint16)
img_mask = image.saturate_image(image.unnormalize_image(
preds_mask, maxval=65535.0),
dtype=tf.uint16)
png_normal = image.encode_batch_images(img_normal)
png_depth = image.encode_batch_images(img_depth)
png_mask = image.encode_batch_images(img_mask)
self.pngs = tf.stack(
[png_input, png_gt, png_output, png_normal, png_depth, png_mask])
# output results
pixel_shape = preds.get_shape().as_list()
num_pixels = np.prod(pixel_shape[1:])
self.errors = tf.reduce_sum(
tf.abs(preds - targets),
[1, 2, 3]) / num_pixels # just a quick check
self.results = preds
# batch normalization
bn_G_collection = tf.get_collection(bn_scope_G)
bn_D_collection = tf.get_collection(bn_scope_D)
self.bn_G_op = tf.group(*bn_G_collection)
self.bn_D_op = tf.group(*bn_D_collection)
def coords_single_sequence():
return tf.reduce_join(characters_list, keep_dims=True)
# Tensorflow workshop with Jan Idziak #------------------------------------- # #String operations # tensor `a` is [["a", "b"], ["c", "d"]] import tensorflow as tf sess = tf.Session() a = tf.convert_to_tensor([["a", "b"], ["c", "d"]]) print(sess.run(a)) tf.reduce_join(a, 0) #==> ["ac", "bd"] tf.reduce_join(a, 1) #==> ["ab", "cd"] tf.reduce_join(a, -2) # ==> ["ac", "bd"] tf.reduce_join(a, -1) # ==> ["ab", "cd"] tf.reduce_join(a, 0, keep_dims=True) #==> [["ac", "bd"]] tf.reduce_join(a, 1, keep_dims=True) #==> [["ab"], ["cd"]] tf.reduce_join(a, 0, separator=".") #==> ["a.c", "b.d"] tf.reduce_join(a, [0, 1]) #==> ["acbd"] tf.reduce_join(a, [1, 0]) #==> ["abcd"] tf.reduce_join(a, []) #==> ["abcd"] b = tf.convert_to_tensor(["ac"]) c = tf.convert_to_tensor(["bd"]) d = tf.string_join([b, c], separator=" ", name=None) print(sess.run(d)) e = tf.reduce_join(a, 0) print(tf.string_to_hash_bucket(e, 2)) print(sess.run(tf.string_to_hash_bucket(e, 5)))
h5 = tf.matmul(h4, W5[i]) + b5[i]
h4 = tf.nn.relu(h5)
h6 = tf.matmul(h5, W6[i]) + b6[i]
h6 = tf.nn.relu(h6)
h7 = tf.matmul(h6, W7[i]) + b7[i]
h7 = tf.nn.relu(h7)
h8 = tf.matmul(h7, W8[i]) + b8[i]
h8 = tf.nn.relu(h8)
logits = tf.matmul(h8, W9[i]) + b9[i]
# logits = tf.matmul(h, W2[i]) + b2[i]
# logits = tf.matmul(h, W2[i])
o = tf.sign(logits)
# outputs.append((o+1)/2)
outputs.append(
tf.reduce_join(
tf.reduce_join(tf.as_string(tf.cast((o + 1) // 2, tf.int32)), 0),
0))
session = tf.Session()
inputs = [[float(xx) for xx in "{0:07b}".format(i)] for i in range(2**7)]
# N=10
cnt = Counter()
#weights = {}
for i in range(N):
# if i%(N/100) == 0:
print(i)
session.run(tf.global_variables_initializer())
fs = session.run(outputs, feed_dict={x: inputs})
varss = session.run(variables, feed_dict={x: inputs})
def _predict_labels(self, examples):
"""Builds tf graph for prediction.
Args:
examples: dict of input tensors keyed by name.
Returns:
predictions: dict of prediction results keyed by name.
"""
options = self._model_proto
is_training = self._is_training
# Extract input data fields.
(image, image_id, num_captions, caption_strings,
caption_lengths) = (examples[InputDataFields.image],
examples[InputDataFields.image_id],
examples[InputDataFields.num_captions],
examples[InputDataFields.caption_strings],
examples[InputDataFields.caption_lengths])
class_act_map_predictions = self._calc_class_act_map(examples)
(class_act_map,
logits) = (class_act_map_predictions[VOCPredictions.class_act_map],
class_act_map_predictions[VOCPredictions.logits])
# Load the vocabulary.
vocabulary_list = self._read_vocabulary(options.vocabulary_file)
tf.logging.info("Read a vocabulary with %i words.", len(vocabulary_list))
# Encode labels, shape=[batch, num_classes].
class_labels = self._encode_labels(num_captions, caption_strings,
caption_lengths, vocabulary_list)
# visualize
with tf.name_scope("visualize"):
image_vis = tf.cast(image, tf.uint8)
image_vis = plotlib.draw_caption(
image_vis,
tf.reduce_join(caption_strings[:, 0, :], axis=-1, separator=','),
org=(5, 5),
fontscale=1.0,
color=(255, 0, 0),
thickness=1)
image_vis = plotlib.draw_caption(
image_vis,
tf.gather(vocabulary_list, tf.argmax(logits, axis=-1)),
org=(5, 25),
fontscale=1.0,
color=(255, 0, 0),
thickness=1)
class_act_map_list = []
batch_size, height, width, _ = utils.get_tensor_shape(image_vis)
for i, x in enumerate(tf.unstack(class_act_map, axis=-1)):
x = plotlib.convert_to_heatmap(x, normalize=True, normalize_to=[-4, 4])
#x = tf.image.resize_images(x, [height, width], tf.image.ResizeMethod.NEAREST_NEIGHBOR)
x = tf.image.resize_images(x, [height, width])
x = imgproc.gaussian_filter(x, ksize=32)
x = tf.image.convert_image_dtype(x, tf.uint8)
x = plotlib.draw_caption(
x,
tf.tile(tf.expand_dims(vocabulary_list[i], axis=0), [batch_size]),
org=(5, 5),
fontscale=1.0,
color=(255, 0, 0),
thickness=1)
class_act_map_list.append(x)
tf.summary.image(
"image",
tf.concat([image_vis] + class_act_map_list, axis=2),
max_outputs=1)
predictions = {
VOCPredictions.image_id: image_id,
VOCPredictions.class_labels: class_labels,
VOCPredictions.class_act_map: class_act_map,
VOCPredictions.logits: logits,
}
return predictions
def get_infer_iterator(hparams,
src_dataset,
src_vocab_table,
batch_size,
eos,
ctx_dataset=None,
annot_dataset=None,
trie_exclude_dataset=None,
src_max_len=None,
tgt_vocab_table=None):
src_eos_id = tf.cast(src_vocab_table.lookup(tf.constant(eos)), tf.int32)
# even in append mode, this will be only the source, without attached context.
src_string_dataset = src_dataset
# Create a fake dataset for weights.
wgt_dataset = src_dataset.map(lambda src: tf.constant(1.0))
# Create a fake context dataset.
if ctx_dataset is None:
ctx_dataset = src_dataset.map(lambda src: tf.constant("no context"))
# Create a fake annotations dataset.
if annot_dataset is None:
annot_dataset = src_dataset.map(lambda src: tf.constant("1\t1"))
# Create a fake trie exclude dataset.
if trie_exclude_dataset is None:
trie_exclude_dataset = src_dataset.map(lambda src: tf.constant(""))
if tgt_vocab_table is None:
tgt_vocab_table = src_vocab_table
if hparams.context_feed == "append":
src_dataset = tf.data.Dataset.zip((ctx_dataset, src_dataset)).map(
lambda ctx, src: ctx + " " + hparams.context_delimiter + " " + src)
src_dataset = src_dataset.map(lambda src: tf.string_split([src]).values)
ctx_dataset = ctx_dataset.map(lambda ctx: tf.string_split([ctx]).values)
annot_dataset = annot_dataset.map(
# We only need the first column, which contains the doc id that will be
# later passed to the environment.
lambda annot: tf.string_split([annot], delimiter="\t").values[0])
trie_exclude_dataset = trie_exclude_dataset.map(
lambda tex: tf.string_split([tex]).values)
if src_max_len:
src_dataset = src_dataset.map(lambda src: src[:src_max_len])
# Convert the word strings to ids.
src_dataset = src_dataset.map(
lambda src: tf.cast(src_vocab_table.lookup(src), tf.int32))
ctx_dataset = ctx_dataset.map(
lambda ctx: tf.cast(src_vocab_table.lookup(ctx), tf.int32))
trie_exclude_dataset = trie_exclude_dataset.map(
lambda tex: tf.cast(tgt_vocab_table.lookup(tex), tf.int32))
# Append context with <eos> so the length will be at least 1.
ctx_dataset = ctx_dataset.map(lambda ctx: tf.concat(
(ctx, [src_eos_id]), 0))
trie_exclude_dataset = trie_exclude_dataset.map(
lambda tex: tf.reduce_join(tf.as_string(tex), separator=" "))
src_dataset = tf.data.Dataset.zip(
(src_string_dataset, src_dataset, wgt_dataset, ctx_dataset,
annot_dataset, trie_exclude_dataset))
# Add in the annotations and word counts.
src_dataset = src_dataset.map(
lambda src_string, src, wgt, ctx, annot, tex:
(src_string, src, wgt, ctx, annot, tex, tf.size(src), tf.size(ctx)))
def batching_func(x):
return x.padded_batch(
batch_size,
# The entry is the source line rows;
# this has unknown-length vectors. The last entry is
# the source row size; this is a scalar.
padded_shapes=(
tf.TensorShape([]), # src_string
tf.TensorShape([None]), # src
tf.TensorShape([]), # wgt
tf.TensorShape([None]), # ctx
tf.TensorShape([]), # annot
tf.TensorShape([]), # tex
tf.TensorShape([]), # src_len
tf.TensorShape([])), # ctx_len
# Pad the source and context sequences with eos tokens.
# (Though notice we don't generally need to do this since
# later on we will be masking out calculations past the true sequence.
padding_values=(
"", # src_string
src_eos_id, # src
1.0, # wgt
src_eos_id, # ctx
"", # annot --unused
"", # tex --unused
0, # src_len -- unused
0)) # ctx_len -- unused
batched_dataset = batching_func(src_dataset)
batched_iter = batched_dataset.make_initializable_iterator()
(src_raw, src_ids, weights, ctx_ids, annot_strs, tex_strs, src_seq_len,
ctx_seq_len) = batched_iter.get_next()
return BatchedInput(initializer=batched_iter.initializer,
source_string=src_raw,
source=src_ids,
target_input=None,
target_output=None,
weights=weights,
context=ctx_ids,
annotation=annot_strs,
trie_exclude=tex_strs,
source_sequence_length=src_seq_len,
target_sequence_length=None,
context_sequence_length=ctx_seq_len)
accuracy = tf.reduce_mean(tf.reduce_min(accuracy_internal, 2))
accuracy_letter = tf.reduce_mean(tf.reshape(accuracy_internal, [-1]))
predict1 = tf.argmax(letter1, 1)
predict2 = tf.argmax(letter2, 1)
predict3 = tf.argmax(letter3, 1)
predict4 = tf.argmax(letter4, 1)
#生成最终结果
base_str = tf.constant("BCEFGHJKMPQRTVWXY2346789")
poses = tf.stack([predict1, predict2, predict3, predict4], axis=1)
length = tf.constant([1, 1, 1, 1], tf.int64)
predicts = tf.map_fn(lambda pos: tf.substr(base_str, pos, length), poses,
tf.string)
predict_join = tf.reduce_join(predicts, axis=1)
initer = tf.global_variables_initializer()
saver = tf.train.Saver()
sess = tf.Session()
sess.run(initer)
saver.restore(sess)
pickup = "BCEFGHJKMPQRTVWXY2346789"
reader = pd.read_source(souce_id=802522, iterator=True)
identity = np.identity(24)
for i in range(15000):
#获取图片数据
df = reader.get_chunk(500)
def evaluate(path_to_checkpoint, ds, val_data, val_labels, num_examples,
global_step):
batch_size = 128
num_batches = num_examples // batch_size
needs_include_length = False
with tf.Graph().as_default():
with tf.name_scope('test_inputs'):
xs = tf.placeholder(shape=[None, 54, 54, 3], dtype=tf.float32)
ys1 = tf.placeholder(shape=[
None,
], dtype=tf.int32)
ys2 = tf.placeholder(shape=[None, 5], dtype=tf.int32)
length_logits, digits_logits = Model.layers(xs, drop_rate=0.0)
length_predictions = tf.argmax(length_logits, axis=1)
digits_predictions = tf.argmax(digits_logits, axis=2)
if needs_include_length:
labels = tf.concat([tf.reshape(ys1, [-1, 1]), ys2], axis=1)
predictions = tf.concat(
[tf.reshape(length_predictions, [-1, 1]), digits_predictions],
axis=1)
else:
labels = ys2
predictions = digits_predictions
labels_string = tf.reduce_join(tf.as_string(labels), axis=1)
predictions_string = tf.reduce_join(tf.as_string(predictions), axis=1)
accuracy, update_accuracy = tf.metrics.accuracy(
labels=labels_string, predictions=predictions_string)
tf.summary.image('image', xs)
tf.summary.scalar('accuracy', accuracy)
tf.summary.histogram(
'variables',
tf.concat(
[tf.reshape(var, [-1]) for var in tf.trainable_variables()],
axis=0))
summary = tf.summary.merge_all()
with tf.Session() as sess:
sess.run([
tf.global_variables_initializer(),
tf.local_variables_initializer()
])
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
restorer = tf.train.Saver()
restorer.restore(sess, path_to_checkpoint)
for _ in range(num_batches):
image_batch, label = ds.build_batch(val_data,
val_labels,
batch_size,
is_train=False,
shuffle=False)
length_batch = label[:, 0]
digits_batch = label[:, 1:6]
acc, update = sess.run([accuracy, update_accuracy],
feed_dict={
xs: image_batch,
ys1: length_batch,
ys2: digits_batch
})
coord.request_stop()
coord.join(threads)
return acc
def build_graph(checkpoint_dir,
log_dir,
batch_size,
steps,
dropout_rate=0,
mode=tf.estimator.ModeKeys.TRAIN,
learning_rate=None,
decay_rate=None,
decay_steps=None):
g = tf.Graph()
with g.as_default():
y_usenow3 = tf.placeholder(tf.int32,
shape=[batch_size, 1],
name='usenow3')
y_ecignow = tf.placeholder(tf.int32,
shape=[batch_size, 1],
name='ecignow')
global_step = tf.train.get_or_create_global_step()
if mode == ModeKeys.TRAIN:
if learning_rate is None:
raise Exception('learning_rate must be set in train mode')
if (decay_steps is None and decay_rate is not None) or (
decay_rate is None and decay_steps is not None):
raise Exception(
'decay_steps and decay_rate must both be set if one is set'
)
if decay_steps is not None and decay_rate is not None:
learning_rate = tf.train.exponential_decay(
learning_rate,
global_step=global_step,
decay_steps=decay_steps,
decay_rate=decay_rate)
examples = (float(batch_size) * tf.to_float(global_step))
else:
examples = float(batch_size)
def fill(val):
return tf.fill([
batch_size,
], val)
features = {
'WEIGHT2': fill(9999),
'HEIGHT3': fill(9999),
'SEX': fill(9),
'EMPLOY1': fill(-1),
'INCOME2': fill(-1),
'MARITAL': fill(-1),
'EDUCA': fill(-1),
'CHILDREN': fill(-1),
'_AGEG5YR': fill(-1)
}
input_layer = tf.feature_column.input_layer(features, columns)
model = MultitaskDNN(input_layer=input_layer,
labels_ndims=2,
hidden_units=[64, 32],
dropout_rate=dropout_rate)
loss = model.loss([y_usenow3, y_ecignow])
with tf.variable_scope('predictions_usenow3'):
logistic_usenow3, probabilities_usenow3, class_ids_usenow3 = _prediction_variables(
model.logits_layers[0])
with tf.variable_scope('predictions_ecignow'):
logistic_ecignow, probabilities_ecignow, class_ids_ecignow = _prediction_variables(
model.logits_layers[1])
predictions_usenow3_str = tf.reduce_join(
tf.as_string(class_ids_usenow3))
predictions_ecignow_str = tf.reduce_join(
tf.as_string(class_ids_ecignow))
labels_usenow3 = tf.reduce_join(tf.as_string(y_usenow3), axis=0)
(usenow3_accuracy, usenow3_accuracy_update_op) = tf.metrics.accuracy(
labels=labels_usenow3, predictions=predictions_usenow3_str)
(ecignow_accuracy, ecignow_accuracy_update_op) = tf.metrics.accuracy(
labels=tf.reduce_join(tf.as_string(y_ecignow), axis=0),
predictions=predictions_ecignow_str)
summary_ops = []
summary_ops += _add_summaries(y_usenow3,
class_ids_usenow3,
family='usenow3',
n_examples=examples)
summary_ops += _add_summaries(y_ecignow,
class_ids_ecignow,
family='ecignow',
n_examples=examples)
(usenow3_precisions, usenow3_precisions_update_op) = \
tf.metrics.precision_at_thresholds(labels=y_usenow3,
predictions=logistic_usenow3,
thresholds=[0.1, 0.5, 0.75])
tf.summary.scalar('precision_at_0.1',
usenow3_precisions[0],
family='usenow3')
tf.summary.scalar('precision_at_0.5',
usenow3_precisions[1],
family='usenow3')
tf.summary.scalar('precision_at_0.75',
usenow3_precisions[2],
family='usenow3')
tf.summary.scalar('loss', loss)
tf.summary.scalar('usenow3_accuracy',
usenow3_accuracy,
family='usenow3')
tf.summary.scalar('ecignow_accuracy',
ecignow_accuracy,
family='ecignow')
tf.summary.histogram('probabilities',
probabilities_usenow3,
family='usenow3')
tf.summary.histogram('probabilities',
probabilities_ecignow,
family='ecignow')
if mode == ModeKeys.TRAIN:
tf.summary.scalar('learning_rate', learning_rate)
summary = tf.summary.merge_all()
saver = tf.train.Saver()
if mode == ModeKeys.TRAIN:
save_increment = 1000
optimizer = tf.train.GradientDescentOptimizer(learning_rate)
train_op = optimizer.minimize(loss, global_step=global_step)
checkpoint_saver_hook = tf.train.CheckpointSaverHook(
checkpoint_dir=checkpoint_dir,
save_steps=save_increment,
saver=saver)
summary_saver_hook = tf.train.SummarySaverHook(
save_steps=save_increment,
output_dir=log_dir,
summary_op=summary)
profiler_hook = tf.train.ProfilerHook(save_steps=save_increment,
output_dir=log_dir)
stop_at_step_hook = tf.train.StopAtStepHook(num_steps=steps)
logging_hook = tf.train.LoggingTensorHook(
{
'loss': loss,
'usenow3_accuracy': usenow3_accuracy,
'ecignow_accuracy': ecignow_accuracy,
'usenow3_precision_at_thresholds': usenow3_precisions
},
every_n_iter=save_increment)
hooks = [
checkpoint_saver_hook, summary_saver_hook, profiler_hook,
stop_at_step_hook, logging_hook
]
ops = [
global_step, train_op, loss, usenow3_accuracy_update_op,
ecignow_accuracy_update_op, usenow3_precisions_update_op
] + summary_ops
else:
summary_saver_hook = tf.train.SummarySaverHook(save_steps=1,
output_dir=log_dir,
summary_op=summary)
hooks = [summary_saver_hook]
ops = [
summary, usenow3_accuracy, ecignow_accuracy,
usenow3_accuracy_update_op, ecignow_accuracy_update_op
] + summary_ops
return g, saver, (features, y_usenow3, y_ecignow), hooks, ops
def evaluate(self, path_to_checkpoint, path_to_tfrecords_file,
num_examples, global_step):
batch_size = 128
num_batches = num_examples // batch_size
needs_include_length = False
with tf.Graph().as_default():
image_batch, length_batch, digits_batch = build_batch(
path_to_tfrecords_file,
num_examples=num_examples,
batch_size=batch_size,
shuffled=False)
length_logits, digits_logits = Model.inference(image_batch,
drop_rate=0.0)
length_predictions = tf.argmax(length_logits, axis=1)
digits_predictions = tf.argmax(digits_logits, axis=2)
if needs_include_length:
labels = tf.concat(
[tf.reshape(length_batch, [-1, 1]), digits_batch], axis=1)
predictions = tf.concat([
tf.reshape(length_predictions, [-1, 1]), digits_predictions
],
axis=1)
else:
labels = digits_batch
predictions = digits_predictions
labels_string = tf.reduce_join(tf.as_string(labels), axis=1)
predictions_string = tf.reduce_join(tf.as_string(predictions),
axis=1)
accuracy, update_accuracy = tf.metrics.accuracy(
labels=labels_string, predictions=predictions_string)
tf.summary.image('image', image_batch)
tf.summary.scalar('accuracy', accuracy)
tf.summary.histogram(
'variables',
tf.concat([
tf.reshape(var, [-1]) for var in tf.trainable_variables()
],
axis=0))
summary = tf.summary.merge_all()
with tf.Session() as sess:
sess.run([
tf.global_variables_initializer(),
tf.local_variables_initializer()
])
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
restorer = tf.train.Saver()
restorer.restore(sess, path_to_checkpoint)
for _ in range(num_batches):
sess.run(update_accuracy)
accuracy_val, summary_val = sess.run([accuracy, summary])
self.summary_writer.add_summary(summary_val,
global_step=global_step)
coord.request_stop()
coord.join(threads)
return accuracy_val
def cnn_model_fn():
x = tf.placeholder(tf.string, name='x')
input_layer = _parse_function(x)
input_layer = tf.reshape(input_layer, [-1, 224, 224, 1])
net, _ = resnet_v2.resnet_v2_50(input_layer)
y1 = tf.placeholder(tf.float32, shape=[None, RESNET_CLASSES])
y2 = tf.placeholder(tf.float32, shape=[None, RESNET_CLASSES])
y3 = tf.placeholder(tf.float32, shape=[None, RESNET_CLASSES])
y4 = tf.placeholder(tf.float32, shape=[None, RESNET_CLASSES])
y5 = tf.placeholder(tf.float32, shape=[None, RESNET_CLASSES])
y6 = tf.placeholder(tf.float32, shape=[None, RESNET_CLASSES])
net = tf.squeeze(net, axis=[1, 2])
letter1 = slim.fully_connected(net, num_outputs=33, activation_fn=None, scope='train')
letter2 = slim.fully_connected(net, num_outputs=33, activation_fn=None, scope='train1')
letter3 = slim.fully_connected(net, num_outputs=33, activation_fn=None, scope='train2')
letter4 = slim.fully_connected(net, num_outputs=33, activation_fn=None, scope='train3')
letter5 = slim.fully_connected(net, num_outputs=33, activation_fn=None, scope='train4')
letter6 = slim.fully_connected(net, num_outputs=33, activation_fn=None, scope='train5')
letter1_cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y1, logits=letter1))
letter2_cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y2, logits=letter2))
letter3_cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y3, logits=letter3))
letter4_cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y4, logits=letter4))
letter5_cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y5, logits=letter5))
letter6_cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y6, logits=letter6))
loss = letter1_cross_entropy + letter2_cross_entropy + letter3_cross_entropy + letter4_cross_entropy + letter5_cross_entropy + letter6_cross_entropy
optimizer = tf.train.AdamOptimizer(learning_rate=0.0001)
train_op = slim.learning.create_train_op(loss, optimizer,
summarize_gradients=True)
tf.summary.scalar('loss', loss)
predict_concat = tf.stack([tf.argmax(letter1, 1),
tf.argmax(letter2, 1),
tf.argmax(letter3, 1),
tf.argmax(letter4, 1),
tf.argmax(letter5, 1),
tf.argmax(letter6, 1)],
1)
y_concat = tf.stack([tf.argmax(y1, 1),
tf.argmax(y2, 1),
tf.argmax(y3, 1),
tf.argmax(y4, 1),
tf.argmax(y5, 1),
tf.argmax(y6, 1)],
1)
accuracy_internal = tf.cast(tf.equal(predict_concat, y_concat), tf.float32),
accuracy = tf.reduce_mean(tf.reduce_min(accuracy_internal, 2))
tf.summary.scalar("accuracy", accuracy)
merged = tf.summary.merge_all()
accuracy_letter = tf.reduce_mean(tf.reshape(accuracy_internal, [-1]))
length = tf.constant([1, 1, 1, 1, 1, 1], tf.int64)
predicts = tf.map_fn(lambda pos: tf.substr(pickup, pos, length), predict_concat, tf.string)
predict_join = tf.reduce_join(predicts, axis=1)
# tf.print(predict_join)
initer = tf.global_variables_initializer()
saver = tf.train.Saver()
sess = tf.Session()
sess.run(initer)
ckpt = tf.train.get_checkpoint_state(model_path_path)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
train_writer = tf.summary.FileWriter(user_home + '/50logs')
# saver.restore(sess, model_path)
images = datasets.images #get_train_dataset(dataset_path)
labels = datasets.labels
test_images = datasets.test_images
test_labels = datasets.test_labels
n = 1
ret_status = False
for j in range(15000):
if ret_status:
break
for i in range(len(images)):
if ret_status:
break
image = images[i]
label = labels[i]
batch_y_1 = [identity[pickup.find(label[0])]]
batch_y_2 = [identity[pickup.find(label[1])]]
batch_y_3 = [identity[pickup.find(label[2])]]
batch_y_4 = [identity[pickup.find(label[3])]]
batch_y_5 = [identity[pickup.find(label[4])]]
batch_y_6 = [identity[pickup.find(label[5])]]
sess.run(train_op,
feed_dict={x: image, y1: batch_y_1, y2: batch_y_2, y3: batch_y_3, y4: batch_y_4, y5: batch_y_5,
y6: batch_y_6})
n = n + 1
if n % 100 == 0:
saver.save(sess, model_path, global_step=n)
# if i % 10 == 0:
summary = sess.run(merged,feed_dict={x: image, y1: batch_y_1, y2: batch_y_2, y3: batch_y_3, y4: batch_y_4, y5: batch_y_5, y6: batch_y_6})
train_writer.add_summary(summary, n)
count = 0
for k in range(len(test_images)):
image = test_images[k]
label = test_labels[k]
print(image)
print(label)
batch_y_1 = [identity[pickup.find(label[0])]]
batch_y_2 = [identity[pickup.find(label[1])]]
batch_y_3 = [identity[pickup.find(label[2])]]
batch_y_4 = [identity[pickup.find(label[3])]]
batch_y_5 = [identity[pickup.find(label[4])]]
batch_y_6 = [identity[pickup.find(label[5])]]
accuracy_letter_, accuracy_, predict_ = sess.run([accuracy_letter, accuracy, predict_join],
feed_dict={x: image, y1: batch_y_1, y2: batch_y_2,
y3: batch_y_3,
y4: batch_y_4, y5: batch_y_5,
y6: batch_y_6})
print(accuracy_letter_)
print("accuracy is ====>%f" % accuracy_)
print(predict_)
if accuracy_ == 1.00:
count += 1
accuracy_ = count / len(test_labels)
print(accuracy_)
if accuracy_ > 0.9:
ret_status = True
print(j, n)
print("-" * 20)
prediction_signature = tf.saved_model.signature_def_utils.predict_signature_def({'image_path': x},
{'label': predict_join})
legacy_init_op = tf.group(tf.tables_initializer(), name='legacy_init_op')
builder = tf.saved_model.builder.SavedModelBuilder(user_home + '/model')
builder.add_meta_graph_and_variables(
sess, [tf.saved_model.tag_constants.SERVING],
signature_def_map={
tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY:
prediction_signature,
},
legacy_init_op=legacy_init_op)
builder.save()
def get_accuracy(self, path_to_checkpoint, path_to_tfrecords_file, num_examples, global_step):
"""
input: path_to_checkpoint => model checkpoint path
path_to_tfrecords_file => tfrecords path
num_samples => number of samples to be measured
funct: evaluates the accuracy of the predicted values of the samples
output: returns the total accuracy of the model on the sample data given
"""
batch_size = 128
num_batches = num_examples / batch_size
needs_include_length = False
with tf.Graph().as_default():
# gets the batches of the evaluting data
image_batch, length_batch, digits_batch = DataPreprocessor.build_batch(path_to_tfrecords_file,
num_examples=num_examples,
batch_size=batch_size,
shuffled=False)
length_logits, digits_logits = CNNModel.get_inference(image_batch, drop_rate=0.0)
length_predictions = tf.argmax(length_logits, axis=1)
digits_predictions = tf.argmax(digits_logits, axis=2)
soft = tf.nn.softmax(digits_logits)
coverage = tf.reduce_max(soft, reduction_indices=2)
proba = tf.reduce_mean(coverage, axis=1)
ones = 0.8*tf.ones_like(proba)
mask = tf.greater(proba, ones)
# if length and batch to be concatenated then concatenates
if needs_include_length:
labels = tf.concat([tf.reshape(length_batch, [-1, 1]), digits_batch], axis=1)
predictions = tf.concat([tf.reshape(length_predictions, [-1, 1]), digits_predictions], axis=1)
else:
labels = digits_batch
predictions = digits_predictions
labels_string = tf.reduce_join(tf.as_string(labels), axis=1)
predictions_string = tf.reduce_join(tf.as_string(predictions), axis=1)
labels_mask_string = tf.boolean_mask(labels_string, mask)
predictions_mask_string = tf.boolean_mask(predictions_string, mask)
coverage_size = tf.size(predictions_mask_string)/tf.size(predictions_string)
# determining the accuracy of the evaluating data
accuracy, update_accuracy = tf.metrics.accuracy(
labels=labels_string,
predictions=predictions_string
)
# determining the accuracy mask of the evaluating data
accuracy_mask, update_accuracy_mask = tf.metrics.accuracy(
labels=labels_mask_string,
predictions=predictions_mask_string
)
tf.summary.image('image', image_batch)
tf.summary.scalar('accuracy', accuracy)
tf.summary.histogram('variables',
tf.concat([tf.reshape(var, [-1]) for var in tf.trainable_variables()], axis=0))
summary = tf.summary.merge_all()
with tf.Session() as sess:
sess.run([tf.global_variables_initializer(), tf.local_variables_initializer()])
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
restorer = tf.train.Saver()
restorer.restore(sess, path_to_checkpoint)
for _ in xrange(num_batches):
sess.run([update_accuracy, update_accuracy_mask])
accuracy_val, summary_val = sess.run([accuracy, summary])
accuracy_mask_val, coverage_size_val = sess.run([accuracy_mask, coverage_size])
self.summary_writer.add_summary(summary_val, global_step=global_step)
coord.request_stop()
coord.join(threads)
return accuracy_val, accuracy_mask_val, coverage_size_val
def cnn_model_fn():
x = tf.placeholder(tf.string, name='x')
input_layer = _parse_function(x)
input_layer = tf.reshape(input_layer, [-1, 224, 224, 1])
net, _ = resnet_v2.resnet_v2_50(input_layer)
y1 = tf.placeholder(tf.float32, shape=[None, RESNET_CLASSES])
y2 = tf.placeholder(tf.float32, shape=[None, RESNET_CLASSES])
y3 = tf.placeholder(tf.float32, shape=[None, RESNET_CLASSES])
y4 = tf.placeholder(tf.float32, shape=[None, RESNET_CLASSES])
y5 = tf.placeholder(tf.float32, shape=[None, RESNET_CLASSES])
y6 = tf.placeholder(tf.float32, shape=[None, RESNET_CLASSES])
net = tf.squeeze(net, axis=[1, 2])
letter1 = slim.fully_connected(net,
num_outputs=33,
activation_fn=None,
scope='train')
letter2 = slim.fully_connected(net,
num_outputs=33,
activation_fn=None,
scope='train1')
letter3 = slim.fully_connected(net,
num_outputs=33,
activation_fn=None,
scope='train2')
letter4 = slim.fully_connected(net,
num_outputs=33,
activation_fn=None,
scope='train3')
letter5 = slim.fully_connected(net,
num_outputs=33,
activation_fn=None,
scope='train4')
letter6 = slim.fully_connected(net,
num_outputs=33,
activation_fn=None,
scope='train5')
letter1_cross_entropy = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(labels=y1, logits=letter1))
letter2_cross_entropy = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(labels=y2, logits=letter2))
letter3_cross_entropy = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(labels=y3, logits=letter3))
letter4_cross_entropy = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(labels=y4, logits=letter4))
letter5_cross_entropy = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(labels=y5, logits=letter5))
letter6_cross_entropy = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(labels=y6, logits=letter6))
loss = letter1_cross_entropy + letter2_cross_entropy + letter3_cross_entropy + letter4_cross_entropy + letter5_cross_entropy + letter6_cross_entropy
optimizer = tf.train.AdamOptimizer(learning_rate=0.0001)
train_op = slim.learning.create_train_op(loss,
optimizer,
summarize_gradients=True)
predict_concat = tf.stack([
tf.argmax(letter1, 1),
tf.argmax(letter2, 1),
tf.argmax(letter3, 1),
tf.argmax(letter4, 1),
tf.argmax(letter5, 1),
tf.argmax(letter6, 1)
], 1)
y_concat = tf.stack([
tf.argmax(y1, 1),
tf.argmax(y2, 1),
tf.argmax(y3, 1),
tf.argmax(y4, 1),
tf.argmax(y5, 1),
tf.argmax(y6, 1)
], 1)
accuracy_internal = tf.cast(tf.equal(predict_concat, y_concat),
tf.float32),
accuracy = tf.reduce_mean(tf.reduce_min(accuracy_internal, 2))
accuracy_letter = tf.reduce_mean(tf.reshape(accuracy_internal, [-1]))
length = tf.constant([1, 1, 1, 1, 1, 1], tf.int64)
predicts = tf.map_fn(lambda pos: tf.substr(pickup, pos, length),
predict_concat, tf.string)
predict_join = tf.reduce_join(predicts, axis=1)
# tf.print(predict_join)
initer = tf.global_variables_initializer()
saver = tf.train.Saver()
sess = tf.Session()
sess.run(initer)
ckpt = tf.train.get_checkpoint_state(model_path_path)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
# saver.restore(sess, model_path)
# """
image = '/Users/geu/testimg/IDO4N4.jpeg'
label = 'IDO4N4'
batch_y_1 = [identity[pickup.find(label[0])]]
batch_y_2 = [identity[pickup.find(label[1])]]
batch_y_3 = [identity[pickup.find(label[2])]]
batch_y_4 = [identity[pickup.find(label[3])]]
batch_y_5 = [identity[pickup.find(label[4])]]
batch_y_6 = [identity[pickup.find(label[5])]]
accuracy_letter_, accuracy_, predict_ = sess.run(
[accuracy_letter, accuracy, predict_join],
feed_dict={
x: image,
y1: batch_y_1,
y2: batch_y_2,
y3: batch_y_3,
y4: batch_y_4,
y5: batch_y_5,
y6: batch_y_6
})
print(accuracy_letter_)
print("accuracy is ====>%f" % accuracy_)
print(predict_)
def join_charcaters_fn(coords):
return tf.reduce_join(characters_list[coords[0]:coords[1]])
def main(_):
path = FLAGS.path
path_to_restore_checkpoint_file = FLAGS.restore_checkpoint
files = getTestFiles(path)
num = 0
correct = 0
image_paths = []
number_predictions = []
labels = []
for key in files:
print(key)
# print files['1']
class_num = len(files[key])
num = num + class_num
images = [tf.image.decode_jpeg(tf.read_file(image_dir), channels=3) for image_dir in files[key]]
# print images[0], images[1]
# image = tf.image.resize_image_with_crop_or_pad(image, 64, 64)
for i in range(0, len(files[key])):
im = Image.open(files[key][i])
image_paths += [files[key][i]]
width, height = im.size
# resize method 1
# images[i] = tf.image.crop_to_bounding_box(images[i], height / 5, width / 2 - 32, 64, 64)
# resize method 2 (float image)
# images[i] = tf.image.resize_image_with_crop_or_pad(images[i], int(min(width, height)/1.5), int(min(width, height)/1.5))
# images[i] = tf.image.resize_images(images[i], [64, 64])/255.0
# images[i] = tf.image.resize_images(images[i], [64, 64])
# resize method 3 for cropped images
images[i] = tf.image.resize_image_with_crop_or_pad(images[i], 54, 54)
# print str(i) + ' ' + files[key][i]
# stack multiple images
images = tf.stack(images)
# image = tf.image.resize_images(image, [64, 64])
# images = tf.reshape(images, [-1, 64, 64, 3])
images = tf.image.convert_image_dtype(images, dtype=tf.float32)
images = tf.multiply(tf.subtract(images, 0.5), 2)
#images = tf.image.resize_images(images, [54, 54])
images = tf.reshape(images, [-1, 54, 54, 3])
#display(images[124])
#figures = {'img' + str(i): images[i] for i in range(len(files[key]))}
# print figures
#plot_figures(figures)
# images = tf.unstack(images)
# print images
# for image in images:
length_logits, digits_logits = Model.inference(images, drop_rate=0.0)
length_predictions = tf.argmax(length_logits, axis=1)
digits_predictions = tf.argmax(digits_logits, axis=2)
digits_predictions_string = tf.reduce_join(tf.as_string(digits_predictions), axis=1)
with tf.Session() as sess:
restorer = tf.train.Saver()
restorer.restore(sess, path_to_restore_checkpoint_file)
length_predictions_val, digits_predictions_string_val = sess.run(
[length_predictions, digits_predictions_string])
# print 'length: ', length_predictions_val
# print 'digits: ', digits_predictions_string_val
predictions = [int(d[:l]) for l, d in zip(length_predictions_val, digits_predictions_string_val)]
# print predictions
number_predictions += predictions
labels += [int(key)] * class_num
# class_correct = sum([1.0 for p in predictions if p == key])
# correct = correct + class_correct
# print 'Accuracy: ', class_correct / class_num
# break
acc, acc_op = tf.metrics.accuracy(labels = tf.Variable(labels), predictions = tf.Variable(number_predictions))
# auc, auc_op = tf.metrics.auc(labels = tf.Variable(labels), predictions = tf.Variable(number_predictions))
pre, pre_op = tf.metrics.precision(labels = tf.Variable(labels), predictions = tf.Variable(number_predictions))
recall, recall_op = tf.metrics.recall(labels = tf.Variable(labels), predictions = tf.Variable(number_predictions))
sess = tf.InteractiveSession()
tf.global_variables_initializer().run()
tf.local_variables_initializer().run()
sess.run(acc_op)
#sess.run(auc_op)
sess.run(pre_op)
sess.run(recall_op)
print(labels)
print(number_predictions)
# sess.run(init)
print("Number of test images: " + str(num) + ".")
print("Accuracy: " + str(sess.run(acc)))
print("Precision: " + str(sess.run(pre)))
print("Recall: " + str(sess.run(recall)))
confusion_matrix(image_paths, labels, number_predictions)
def evaluate(self, path_to_restore_model_checkpoint_file,
path_to_restore_defender_checkpoint_file,
path_to_tfrecords_file, num_examples, global_step,
defend_layer, attacker_type):
batch_size = 32
num_batches = num_examples // batch_size
needs_include_length = False
with tf.Graph().as_default():
image_batch, length_batch, digits_batch = Donkey.build_batch(
path_to_tfrecords_file,
num_examples=num_examples,
batch_size=batch_size,
shuffled=False)
with tf.variable_scope('model'):
length_logits, digits_logits, hidden_out = Model.inference(
image_batch,
drop_rate=0.0,
is_training=False,
defend_layer=defend_layer)
with tf.variable_scope('defender'):
recovered = Attacker.recover_hidden(attacker_type,
hidden_out,
is_training=False,
defend_layer=defend_layer)
ssim = tf.reduce_mean(
tf.abs(tf.image.ssim(image_batch, recovered, max_val=2)))
length_predictions = tf.argmax(length_logits, axis=1)
digits_predictions = tf.argmax(digits_logits, axis=2)
if needs_include_length:
labels = tf.concat(
[tf.reshape(length_batch, [-1, 1]), digits_batch], axis=1)
predictions = tf.concat([
tf.reshape(length_predictions, [-1, 1]), digits_predictions
],
axis=1)
else:
labels = digits_batch
predictions = digits_predictions
labels_string = tf.reduce_join(tf.as_string(labels), axis=1)
predictions_string = tf.reduce_join(tf.as_string(predictions),
axis=1)
accuracy, update_accuracy = tf.metrics.accuracy(
labels=labels_string, predictions=predictions_string)
tf.summary.image('image', image_batch, max_outputs=20)
tf.summary.image('recovered', recovered, max_outputs=20)
tf.summary.scalar('ssim', ssim)
tf.summary.scalar('accuracy', accuracy)
tf.summary.histogram(
'variables',
tf.concat([
tf.reshape(var, [-1]) for var in tf.trainable_variables()
],
axis=0))
summary = tf.summary.merge_all()
with tf.Session() as sess:
sess.run([
tf.global_variables_initializer(),
tf.local_variables_initializer()
])
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
model_saver = tf.train.Saver(var_list=tf.get_collection(
tf.GraphKeys.TRAINABLE_VARIABLES, scope='model'))
defender_saver = tf.train.Saver(var_list=tf.get_collection(
tf.GraphKeys.TRAINABLE_VARIABLES, scope='defender'))
model_saver.restore(sess,
path_to_restore_model_checkpoint_file)
print("Evaluation model restored from {}".format(
path_to_restore_model_checkpoint_file))
defender_saver.restore(
sess, path_to_restore_defender_checkpoint_file)
for _ in range(num_batches):
sess.run(update_accuracy)
accuracy_val, summary_val = sess.run([accuracy, summary])
self.summary_writer.add_summary(summary_val,
global_step=global_step)
coord.request_stop()
coord.join(threads)
return accuracy_val
def join_beams(tokens, tokens_lengths, separator='\n'):
beams = join_tokens_beam(tokens, tokens_lengths)
joined = tf.reduce_join(beams, axis=1, separator=separator)
joined = tf.reshape(joined, [-1, 1])
return joined
def testInvalidReductionIndices(self):
with self.test_session():
with self.assertRaisesRegexp(ValueError, "scalar"):
tf.reduce_join(inputs="", reduction_indices=0)
with self.assertRaisesRegexp(ValueError,
"Invalid reduction dimension -3"):
tf.reduce_join(inputs=[[""]], reduction_indices=-3)
with self.assertRaisesRegexp(ValueError, "Invalid reduction dimension 2"):
tf.reduce_join(inputs=[[""]], reduction_indices=2)
with self.assertRaisesRegexp(ValueError,
"Invalid reduction dimension -3"):
tf.reduce_join(inputs=[[""]], reduction_indices=[0, -3])
with self.assertRaisesRegexp(ValueError, "Invalid reduction dimension 2"):
tf.reduce_join(inputs=[[""]], reduction_indices=[0, 2])
with self.assertRaisesRegexp(ValueError, "Duplicate reduction index 0"):
tf.reduce_join(inputs=[[""]], reduction_indices=[0, 0])
def read_file(self, example):
example_fmt = {
"imagename":
tf.FixedLenFeature((), tf.string, ""),
"pts2d_68":
tf.FixedLenFeature(
(self.args.gt_num_lmks * 2),
tf.float32,
default_value=list(np.zeros(self.args.gt_num_lmks * 2)))
}
parsed = tf.parse_single_example(example, example_fmt)
label = tf.reshape(parsed["pts2d_68"], [-1, 2])
img_string = tf.read_file(
tf.reduce_join(
[tf.constant(self.args.data_eval_dir), parsed["imagename"]]))
image = tf.image.decode_image(tf.reshape(img_string, shape=[]),
channels=3)
image = tf.image.convert_image_dtype(image, dtype=tf.float32)
h = tf.shape(image)[0]
w = tf.shape(image)[1]
label = label - 1
h = tf.shape(image)[0]
w = tf.shape(image)[1]
xmin = tf.reduce_min(label[:, 0], axis=0)
ymin = tf.reduce_min(label[:, 1], axis=0)
xmax = tf.reduce_max(label[:, 0], axis=0)
ymax = tf.reduce_max(label[:, 1], axis=0)
bbox_h = ymax - ymin
bbox_w = xmax - xmin
scale_factor = 0.01
y1 = (ymin - bbox_h * scale_factor) / tf.cast(h - 1, tf.float32)
x1 = (xmin - bbox_w * scale_factor) / tf.cast(w - 1, tf.float32)
y2 = (ymax + bbox_h * scale_factor) / tf.cast(h - 1, tf.float32)
x2 = (xmax + bbox_w * scale_factor) / tf.cast(w - 1, tf.float32)
image = tf.squeeze(tf.image.crop_and_resize(
tf.expand_dims(image, 0),
boxes=[[y1, x1, y2, x2]],
box_ind=[0],
crop_size=[self.args.IMAGE_SIZE[0], self.args.IMAGE_SIZE[1]],
method='bilinear'),
axis=0)
corner = tf.stack([
x1 * tf.cast(w - 1, tf.float32) * tf.ones(self.args.num_lmks),
y1 * tf.cast(h - 1, tf.float32) * tf.ones(self.args.num_lmks)
],
axis=1)
scale = tf.stack([tf.ones(self.args.num_lmks, dtype=tf.float32) * (self.args.IMAGE_SIZE[1]-1) \
/ ((x2 - x1) * tf.cast(w - 1, tf.float32)),
tf.ones(self.args.num_lmks, dtype=tf.float32) * (self.args.IMAGE_SIZE[0]-1) \
/ ((y2 - y1) * tf.cast(h - 1, tf.float32))], axis=1)
image.set_shape([self.args.IMAGE_SIZE[0], self.args.IMAGE_SIZE[1], 3])
return image, label + 1, corner, scale
def testInvalidReductionIndices(self):
with self.test_session():
with self.assertRaisesRegexp(ValueError, "scalar"):
tf.reduce_join(inputs="", reduction_indices=0)
with self.assertRaisesRegexp(ValueError,
"Invalid reduction dimension -3"):
tf.reduce_join(inputs=[[""]], reduction_indices=-3)
with self.assertRaisesRegexp(ValueError,
"Invalid reduction dimension 2"):
tf.reduce_join(inputs=[[""]], reduction_indices=2)
with self.assertRaisesRegexp(ValueError,
"Invalid reduction dimension -3"):
tf.reduce_join(inputs=[[""]], reduction_indices=[0, -3])
with self.assertRaisesRegexp(ValueError,
"Invalid reduction dimension 2"):
tf.reduce_join(inputs=[[""]], reduction_indices=[0, 2])
with self.assertRaisesRegexp(ValueError,
"Duplicate reduction index 0"):
tf.reduce_join(inputs=[[""]], reduction_indices=[0, 0])
for grad, var in grads:
if grad is not None:
tf.summary.histogram(var.op.name + '/grads', grad)
# Evaluate model accuracy
with tf.name_scope('predict'):
pred_correct = tf.equal(tf.argmax(pred, 1), tf.argmax(y_, 1))
accuracy = tf.reduce_mean(tf.cast(pred_correct, tf.float32))
accuracy_summary = tf.summary.scalar("accuracy_summary", accuracy)
# Cloud ML online prediction
with tf.name_scope('serving'):
# Input is a base64 string
image_bytes = tf.placeholder(tf.string)
rgb_image = tf.image.decode_jpeg(tf.reduce_join(image_bytes, []),
channels=FLAGS.image_channels)
rgb_image = tf.image.convert_image_dtype(rgb_image, dtype=tf.float32)
rgb_image = tf.image.resize_images(
rgb_image, [FLAGS.image_size, FLAGS.image_size])
image_batch = tf.expand_dims(rgb_image, 0)
# Define tensor inputs to use during prediction
inputs = {'image_bytes': image_bytes.name}
tf.add_to_collection('inputs', json.dumps(inputs))
softmax_pred = tf.nn.softmax(
conv_net(image_batch, weights, biases, FLAGS.image_size, 1.0))
# Define tensor outputs for prediction
outputs = {'scores': softmax_pred.name}
def main(_):
# Data File
train_file_dir = FLAGS.train_file_dir
test_file_dir = FLAGS.test_file_dir
item_tower_file = FLAGS.item_tower_file
# Hyper Parameters
learning_rate = FLAGS.learning_rate
batch_size = FLAGS.batch_size
item_embedding_size = FLAGS.item_embedding_size
cate_embedding_size = FLAGS.cate_embedding_size
tag_embedding_size = FLAGS.tag_embedding_size
is_train = FLAGS.is_train
output_table = FLAGS.output_table
saved_model_dir = FLAGS.saved_model_dir
checkpoint_dir = FLAGS.checkpointDir
oss_bucket_dir = FLAGS.buckets
local = FLAGS.local
recall_cnt_file = FLAGS.recall_cnt_file
top_k_num = int(FLAGS.top_k_num)
neg_sample_num = int(FLAGS.neg_sample_num)
print("train_file_dir: %s" % train_file_dir)
print("test_file_dir: %s" % test_file_dir)
print("is_train: %d" % is_train)
print("learning_rate: %f" % learning_rate)
print("item_embedding_size: %d" % item_embedding_size)
print("cate_embedding_size: %d" % cate_embedding_size)
print("tag_embedding_size: %d" % tag_embedding_size)
print("batch_size: %d" % batch_size)
print("output table name: %s " % output_table)
print("checkpoint_dir: %s " % checkpoint_dir)
print("oss bucket dir: %s" % oss_bucket_dir)
print("recall_cnt_file: %s" % recall_cnt_file)
if local:
# summary_dir = "../summary/"
# recall_cnt_file = "../data/youtube_recall_item_cnt*"
pass
else:
# oss_bucket_dir = "oss://ivwen-recsys.oss-cn-shanghai-internal.aliyuncs.com/"
# summary_dir = oss_bucket_dir + "experiment/summary/"
train_file_dir = oss_bucket_dir + train_file_dir
test_file_dir = oss_bucket_dir + test_file_dir
recall_cnt_file = oss_bucket_dir + recall_cnt_file
item_tower_file = oss_bucket_dir + item_tower_file
saved_model_dir = oss_bucket_dir + saved_model_dir
# get item cnt
# item_count, cate_count, tag_count = utils.get_item_cnt(recall_cnt_file)
# item_tower_file = [utils.get_file_name(item_tower_file)]
# print("item tower file: %s" % item_tower_file)
# print("item_count: ", item_count)
# print("cate_count: ", cate_count)
# print("tag_count: ", tag_count)
print("saved_model_dir: %s " % saved_model_dir)
# GPU config
# gpu_config = tf.ConfigProto()
# gpu_config.gpu_options.allow_growth = True
#
with tf.Session() as sess:
train_file_name = utils.get_file_name(train_file_dir)
test_file_dir = utils.get_file_name(test_file_dir)
two_tower_model = TwoTowerModelFRV2(
train_file_dir=train_file_name,
test_file_dir=test_file_dir,
# item_tower_file=item_tower_file,
is_train=is_train,
item_embedding_size=item_embedding_size,
cate_embedding_size=cate_embedding_size,
tag_embedding_size=tag_embedding_size,
batch_size=batch_size,
learning_rate=learning_rate,
local=local,
# item_count=item_count,
# cate_count=cate_count,
# tag_count=tag_count,
output_table=output_table,
top_k_num=top_k_num,
neg_sample_num=neg_sample_num,
sess=sess)
try:
loss_sum = 0.0
pred_step = 0
if is_train == 1:
train, losses = two_tower_model.train_model()
sess.run(tf.global_variables_initializer())
# sess.run(tf.tables_initializer())
sess.run(tf.local_variables_initializer())
sess.run(two_tower_model.training_init_op)
else:
two_tower_model.restore_model(sess, checkpoint_dir)
# sess.run(tf.global_variables_initializer())
# youtube_dnn_model.load_saved_model(sess, savedmodel_dir)
print("restore model finished!!")
two_tower_model.predict_topk_score()
if local == 0:
writer = two_tower_model.write_table()
# # 读取历史参数，保证增量更新
# if tf.gfile.Exists(checkpoint_dir):
# two_tower_model.restore_model(sess, checkpoint_dir)
while True:
# # 更新数据到 timeline 的第一种种方式
# trace = timeline.Timeline(step_stats=run_metadata.step_stats)
# trace_file = open('timeline.ctf.json', 'w')
# trace_file.write(trace.generate_chrome_trace_format())
# 更新数据到profiler
# my_profiler.add_step(step=int(train_step), run_meta=run_metadata)
if is_train:
train_step = two_tower_model.global_step.eval()
# #
# _, loss, learning_rate = sess.run([train, losses, two_tower_model.learning_rate],
# run_metadata=run_metadata, options=run_options)
_, loss, learning_rate = sess.run(
[train, losses, two_tower_model.learning_rate])
loss_sum += loss
if train_step % two_tower_model.PRINT_STEP == 0:
if train_step == 0:
print(
'time: %s\tEpoch: %d\tGlobal_Train_Step: %d\tTrain_loss: %.8f\tLearning_rate:%.8f'
% (time.strftime("%Y-%m-%d %H:%M:%S",
time.localtime()),
two_tower_model.train_epoches, train_step,
loss_sum, learning_rate))
else:
print(
'time: %s\tEpoch: %d\tGlobal_Train_Step: %d\tTrain_loss: %.8f\tLearning_rate:%.8f'
% (time.strftime("%Y-%m-%d %H:%M:%S",
time.localtime()),
two_tower_model.train_epoches, train_step,
loss_sum / two_tower_model.PRINT_STEP,
learning_rate))
if train_step % two_tower_model.SAVE_STEP == 0:
two_tower_model.save_model(sess=sess,
path=checkpoint_dir)
loss_sum = 0.0
# local test
# user_ids, user_click_hash, target_ids, target_hash = sess.run(
# [
# two_tower_model.user_item_click_avg_embed,
# two_tower_model.gender_one_hot,
# two_tower_model.target_item_list,
# two_tower_model.target_item_idx,
# ])
# print(user_ids)
# print(user_click_hash)
# print(target_ids)
# print(target_hash)
# print(label.shape)
# break
else:
if local == 0:
sess.run(writer)
else:
user_embeddings, item_embeddings, logits = sess.run([
tf.reduce_join(tf.as_string(
two_tower_model.user_embedding_final),
1,
separator=","),
tf.reduce_join(tf.as_string(
two_tower_model.item_embeding_final),
1,
separator=","),
two_tower_model.user_topk_item
])
print("users:")
print(user_embeddings)
print("items:")
print(item_embeddings)
print("logits:")
print(logits)
if pred_step % 1000 == 0:
print("%d finished" % pred_step)
pred_step += 1
except tf.errors.OutOfRangeError:
print('time: %s\t %d records copied' %
(time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()),
two_tower_model.global_step.eval()))
# if is_train == 1:
# two_tower_model.save_model(sess=sess, path=checkpoint_dir)
# profile_code_builder = option_builder.ProfileOptionBuilder()
# # profile_code_builder.with_node_names(show_name_regexes=['main.*'])
# profile_code_builder.with_min_execution_time(min_micros=15)
# profile_code_builder.select(['micros']) # 可调整为 'bytes', 'occurrence'
# profile_code_builder.order_by('micros')
# profile_code_builder.with_max_depth(6)
# my_profiler.profile_python(profile_code_builder.build())
# my_profiler.profile_operations(profile_code_builder.build())
# my_profiler.profile_name_scope(profile_code_builder.build())
# my_profiler.profile_graph(profile_code_builder.build())
if is_train == 1:
print("time: %s\tckpt model save start..." %
time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()))
two_tower_model.save_model(sess=sess, path=checkpoint_dir)
print("time: %s\tsave_model save start..." %
time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()))
two_tower_model.save_model_as_savedmodel(
sess=sess,
dir=saved_model_dir,
inputs=two_tower_model.saved_model_inputs,
outputs=two_tower_model.saved_model_outputs)
def _detokenize_tensor(self, tokens):
return tf.reduce_join(tokens, axis=0, separator=" ")
sparse_decoded, log_prob = tf.nn.ctc_greedy_decoder(logits, seqs_len)
dense_decoded = tf.sparse_to_dense(sparse_decoded[0].indices,
sparse_decoded[0].dense_shape,
sparse_decoded[0].values)
sess.run(tf.global_variables_initializer())
table = tf.contrib.lookup.index_to_string_table_from_file(params.get('char2idx'),key_column_index = 0,
value_column_index = 1, delimiter = '\t',
default_value=' ')
tf.tables_initializer().run(session = sess)
expected_chars = table.lookup(dense_decoded)
join_expected = tf.reduce_join(expected_chars, separator = '', axis = 1)
test = dense_decoded.eval(session = sess)
#test_seqs_len = length(test)
print(test.shape)
#print(test_seqs_len.eval(session = sess))
#
#
#logits = tf.transpose(features['logits'], (1,0,2))
#
#
#
def tensor_predict(words_list):
num_classes = FLAGS.num_classes
num_layers = FLAGS.num_layers
num_steps = FLAGS.num_steps
embedding_size = FLAGS.embedding_size
hidden_size = FLAGS.hidden_size
keep_prob = FLAGS.keep_prob
vocab_size = FLAGS.vocab_size
vocab_path = FLAGS.vocab_path
prop_limit = FLAGS.prop_limit
checkpoint_path = FLAGS.checkpoint_path
# split 1-D String dense Tensor to words SparseTensor
sentences = tf.placeholder(dtype=tf.string,
shape=[None],
name='input_sentences')
sparse_words = tf.string_split(sentences, delimiter=' ')
# slice SparseTensor
valid_indices = tf.less(sparse_words.indices,
tf.constant([num_steps], dtype=tf.int64))
valid_indices = tf.reshape(
tf.split(valid_indices, [1, 1], axis=1)[1], [-1])
valid_sparse_words = tf.sparse_retain(sparse_words, valid_indices)
excess_indices = tf.greater_equal(sparse_words.indices,
tf.constant([num_steps], dtype=tf.int64))
excess_indices = tf.reshape(
tf.split(excess_indices, [1, 1], axis=1)[1], [-1])
excess_sparse_words = tf.sparse_retain(sparse_words, excess_indices)
# sparse to dense
words = tf.sparse_to_dense(
sparse_indices=valid_sparse_words.indices,
output_shape=[valid_sparse_words.dense_shape[0], num_steps],
sparse_values=valid_sparse_words.values,
default_value='_PAD')
# dict words to token ids
# with open(os.path.join(vocab_path, 'words_vocab.txt'), 'r') as data_file:
# words_table_list = [line.strip() for line in data_file if line.strip()]
# words_table_tensor = tf.constant(words_table_list, dtype=tf.string)
# words_table = lookup.index_table_from_tensor(mapping=words_table_tensor, default_value=3)
words_table = lookup.index_table_from_file(os.path.join(
vocab_path, 'words_vocab.txt'),
default_value=3)
words_ids = words_table.lookup(words)
# blstm model predict
with tf.variable_scope('model', reuse=None):
logits, _ = model.inference(words_ids,
valid_sparse_words.dense_shape[0],
num_steps,
vocab_size,
embedding_size,
hidden_size,
keep_prob,
num_layers,
num_classes,
is_training=False)
# using softmax
# props = tf.nn.softmax(logits)
# max_prop_values, max_prop_indices = tf.nn.top_k(props, k=1)
# predict_scores = tf.reshape(max_prop_values, shape=[-1, num_steps])
# predict_labels_ids = tf.reshape(max_prop_indices, shape=[-1, num_steps])
# predict_labels_ids = tf.to_int64(predict_labels_ids)
# using crf
logits = tf.reshape(logits, shape=[-1, num_steps, num_classes])
transition_params = tf.get_variable("transitions",
[num_classes, num_classes])
sequence_length = tf.constant(num_steps,
shape=[logits.get_shape()[0]],
dtype=tf.int64)
predict_labels_ids, _ = crf_utils.crf_decode(logits, transition_params,
sequence_length)
predict_labels_ids = tf.to_int64(predict_labels_ids)
predict_scores = tf.constant(1.0,
shape=predict_labels_ids.get_shape(),
dtype=tf.float32)
# replace untrusted prop that less than prop_limit
trusted_prop_flag = tf.greater_equal(
predict_scores, tf.constant(prop_limit, dtype=tf.float32))
replace_prop_labels_ids = tf.to_int64(
tf.fill(tf.shape(predict_labels_ids), 4))
predict_labels_ids = tf.where(trusted_prop_flag, predict_labels_ids,
replace_prop_labels_ids)
# dict token ids to labels
# with open(os.path.join(vocab_path, 'labels_vocab.txt'), 'r') as data_file:
# labels_table_list = [line.strip() for line in data_file if line.strip()]
# labels_table_tensor = tf.constant(labels_table_list, dtype=tf.string)
# labels_table = lookup.index_to_string_table_from_tensor(mapping=labels_table_tensor, default_value='O')
labels_table = lookup.index_to_string_table_from_file(os.path.join(
vocab_path, 'labels_vocab.txt'),
default_value='O')
predict_labels = labels_table.lookup(predict_labels_ids)
# extract real blstm predict label in dense and save to sparse
valid_sparse_predict_labels = tf.SparseTensor(
indices=valid_sparse_words.indices,
values=tf.gather_nd(predict_labels, valid_sparse_words.indices),
dense_shape=valid_sparse_words.dense_shape)
# create excess label SparseTensor with 'O'
excess_sparse_predict_labels = tf.SparseTensor(
indices=excess_sparse_words.indices,
values=tf.fill(tf.shape(excess_sparse_words.values), 'O'),
dense_shape=excess_sparse_words.dense_shape)
# concat SparseTensor
sparse_predict_labels = tf.SparseTensor(
indices=tf.concat(axis=0,
values=[
valid_sparse_predict_labels.indices,
excess_sparse_predict_labels.indices
]),
values=tf.concat(axis=0,
values=[
valid_sparse_predict_labels.values,
excess_sparse_predict_labels.values
]),
dense_shape=excess_sparse_predict_labels.dense_shape)
sparse_predict_labels = tf.sparse_reorder(sparse_predict_labels)
# join SparseTensor to 1-D String dense Tensor
# remain issue, num_split should equal the real size, but here limit to 1
join_labels_list = []
slice_labels_list = tf.sparse_split(sp_input=sparse_predict_labels,
num_split=1,
axis=0)
for slice_labels in slice_labels_list:
slice_labels = slice_labels.values
join_labels = tf.reduce_join(slice_labels,
reduction_indices=0,
separator=' ')
join_labels_list.append(join_labels)
format_predict_labels = tf.stack(join_labels_list, name='predict_labels')
saver = tf.train.Saver()
tables_init_op = tf.tables_initializer()
with tf.Session() as sess:
sess.run(tables_init_op)
ckpt = tf.train.get_checkpoint_state(checkpoint_path)
if ckpt and ckpt.model_checkpoint_path:
print('read model from {}'.format(ckpt.model_checkpoint_path))
saver.restore(sess, ckpt.model_checkpoint_path)
else:
print('No checkpoint file found at %s' % checkpoint_path)
return
# crf tensor
predict_labels_list = sess.run(format_predict_labels,
feed_dict={sentences: words_list})
# save graph into .pb file
graph = tf.graph_util.convert_variables_to_constants(
sess, sess.graph_def, ["init_all_tables", "predict_labels"])
tf.train.write_graph(graph, '.', 'ner_graph.pb', as_text=False)
return predict_labels_list
