def test_mathops_ten():
mathops_10 = MathOpsTen(seed=19)
in_node_1 = mathops_10.get_placeholder("input_1")
#in_node_2 = mathops_8.get_placeholder("input_2")
n0 = tf.is_finite(in_node_1)
n1 = tf.reduce_all(n0)
n2 = tf.cast(n0, dtype=tf.float64)
n3 = tf.cast(n1, dtype=tf.float64)
n4 = tf.add(n2, n3)
n5 = tf.cast(tf.truncatediv(tf.cast(n4, dtype=tf.int32), 3), dtype=tf.float64)
n6 = tf.reciprocal(n5) # should be inf now
n7 = tf.cast(tf.is_inf(n6), dtype=tf.float64)
n8 = tf.cast(tf.is_nan(n6), dtype=tf.float64)
n9 = tf.squared_difference(n7, n8)
w = tf.Variable(tf.random_normal([4, 3], dtype=tf.float64), name="w")
n10 = tf.reverse(w, axis=[-1])
n11 = tf.add(n10, n9)
n12 = tf.reciprocal(tf.multiply(n11, [[0, 1, 1], [1, 1, 1], [0, 1, 0], [1, 0, 0]]))
#n13 = tf.reduce_any(tf.is_inf(n12))
#n14 = tf.cast(n13, dtype=tf.float64)
out_node = tf.identity(n12, name="output")
placeholders = [in_node_1]
predictions = [out_node]
# Run and persist
tfp = TensorFlowPersistor(save_dir="g_10")
predictions_after_freeze = tfp.set_placeholders(placeholders) \
.set_output_tensors(predictions) \
.set_test_data(mathops_10.get_test_data()) \
.build_save_frozen_graph()
print(predictions_after_freeze[0].shape)
def preprocess(self, text: str, seq_length: int = 100) -> list:
"""Split the lyrics of a song into multiple substrings.
Preprocess a string containing lyrics, extracting substrings that
have a fixed, specified size from it. Return the substrings as
sequences of integers rather than characters.
:param text: a string containing lyrics
:param seq_length: fixed size of each output sequence
:return: a list of substrings extracted from the song, as lists of ints
"""
# convert the string to a list of integers:
text_as_chars = tf.strings.bytes_split(text)
text_as_int = tf.map_fn(fn=lambda c: self.char2idx.lookup(c),
elems=text_as_chars,
dtype=tf.int32)
text_as_int = tf.boolean_mask(text_as_int, text_as_int > -1)
# compute the number of characters in the text:
text_size = tf.size(text_as_int)
# increase the sequence length by 1, for character-level prediction:
seq_length += 1
# create subsequences from the original sequence:
trail = tf.truncatemod(text_size, seq_length)
n_seqs = tf.truncatediv(text_size, seq_length)
to_keep = text_size - trail
sequences = tf.reshape(text_as_int[:to_keep], [n_seqs, seq_length])
# shuffle the substrings:
sequences = tf.random.shuffle(sequences)
return sequences
def _body(jl, ju, value, array):
jm = tf.truncatediv(ju + jl, 2) # compute a midpoint,
#jm = tf.Print(jm,[jl,jm,ju])
value_ = tf.gather(array, jm)
#array[jl] <= value < array[ju]
#value_ = tf.Print(value_,[value_,value])
jl = tf.where(value >= value_, jm, jl)
ju = tf.where(value < value_, jm, ju)
return (jl, ju, value, array)
def length_penalty(sequence_lengths, penalty_factor):
"""Calculates the length penalty according to
https://arxiv.org/abs/1609.08144
Args:
sequence_lengths: The sequence length of all hypotheses, a tensor
of shape [beam_size, vocab_size].
penalty_factor: A scalar that weights the length penalty.
Returns:
The length penalty factor, a tensor fo shape [beam_size].
"""
return tf.truncatediv(
(5. + tf.cast(sequence_lengths, dtype=tf.float32))**penalty_factor,
(5. + 1.)**penalty_factor)
def wav_to_spectrogram(wav_data):
# [samples_n, channels_n]
tf.assert_equal(
tf.shape(wav_data)[1], model.channels_n[-1],
f"Audio data has {tf.shape(wav_data)[1]} channels, but needs to have {model.channels_n[-1]} for the model"
)
wav_data.set_shape([None, model.channels_n[-1]])
# truncate audio (only a tiny bit)
samples_n = tf.shape(wav_data)[0]
last_block = tf.truncatediv(samples_n, model.freq_n[-1])
wav_data = wav_data[:last_block * model.freq_n[-1], :]
wav_data = tf.expand_dims(
wav_data, axis=0) # add batch dimension (t_to_repr expects it)
mdct_norm = audio_representation.t_to_repr(wav_data) # -1..1
# [batches_n=1, blocks_n, freqs_n, channels_n]
return mdct_norm
def get_key_points_from_score_maps(score_maps):
"""
Arguments:
@score_maps: of shape (N, H, W, C)
Return:
@key_points_xy: of shape (N, C, 2)
@key_points_scores: of shape (N, C)
"""
N, _, W, C = score_maps.shape.as_list()
assert C == NUM_KEY_POINTS
max_flat = tf.argmax(tf.reshape(score_maps, [N, -1, C]), axis=1)
max_y = tf.truncatediv(max_flat, W)
max_x = max_flat - max_y * W
key_points_xy = tf.stack([max_x, max_y], axis=2)
key_points_scores = tf.squeeze(tf.reduce_max(score_maps, axis=[1, 2]))
return key_points_xy, key_points_scores
x = tf.add(a, b, name='add')
writer = tf.summary.FileWriter('./graphs/simple', tf.get_default_graph())
with tf.Session() as sess:
print(sess.run(x))
writer.close()
# 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.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)
y = tf.ones_like(t_0)
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()
# 例子2：div的奇思妙用
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))) # 对应元素相除， 取商数
print(sess.run(tf.divide(b, a))) # 对应元素相除
print(sess.run(tf.truediv(b, a))) # 对应元素 相除
# print(sess.run(tf.realdiv(b, a)))
print(sess.run(tf.floordiv(b, a))) # 结果向下取整, 但结果dtype与输入保持一致
print(sess.run(tf.truncatediv(b, a))) # 对应元素 截断除 取余
print(sess.run(tf.floor_div(b, a)))
# 例子3：乘法
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)))
# 例子4：Python 基础数据类型
t_0 = 19
x = tf.zeros_like(t_0)
y = tf.ones_like(t_0)
print(x)
print(y)
tf.minimum(x, y, name=None) # 返回两tensor中的最小值 (x < y ? x : y) # 三角函数和反三角函数 tf.cos(x, name=None) tf.sin(x, name=None) tf.tan(x, name=None) tf.acos(x, name=None) tf.asin(x, name=None) tf.atan(x, name=None) # 其它 tf.div(x, y, name=None) # python 2.7 除法, x/y-->int or x/float(y)-->float tf.truediv(x, y, name=None) # python 3 除法, x/y-->float tf.floordiv(x, y, name=None) # python 3 除法, x//y-->int tf.realdiv(x, y, name=None) # 返回一个tensor与x具有相同类型 tf.truncatediv(x, y, name=None) # tf.floor_div(x, y, name=None) tf.truncatemod(x, y, name=None) tf.floormod(x, y, name=None) tf.cross(x, y, name=None) tf.add_n(inputs, name=None) # inputs: A list of Tensor objects, each with same shape and type tf.squared_difference(x, y, name=None) # 对应元素差平方 # 矩阵数学函数 # 矩阵乘法(tensors of rank >= 2) tf.matmul(a, b, transpose_a=False, transpose_b=False, adjoint_a=False, adjoint_b=False, a_is_sparse=False, b_is_sparse=False, name=None) # 转置，可以通过指定 perm=[1, 0] 来进行轴变换 tf.transpose(a, perm=None, name='transpose') # 在张量 a 的最后两个维度上进行转置 tf.matrix_transpose(a, name='matrix_transpose') # Matrix with two batch dimensions, x.shape is [1, 2, 3, 4]
* tf.substract: 减法 * tf.multiply: 乘法 * tf.div: 除法，服从python2的除法规则,x/y，如果x,y都是整数结果为int类型，否则结果为float * tf.divide:除法 * tf.truediv: 使用python3的除法操作语义，x/y, 结果为float，如果要使用整除，使用x//y或tf.floordiv * tf.floordiv: 除法，取整 * tf.truncatediv * tf.mod: 取模（取余） ''' x=tf.constant(7) y=tf.constant(4) mul=tf.multiply(x,y) div=tf.div(x,y) divide=tf.divide(x,y) truediv=tf.truediv(x,y) truncatediv=tf.truncatediv(x,y) sess=tf.Session() out=sess.run([mul,div,divide,truediv,truncatediv]) print(out) ''' 2. pairwise 交叉相乘 * tf.cross(a,b,name): 计算a,b中元素的两两之间的乘积 ''' a=tf.constant([1,2,3]) b=tf.constant([0,2,3]) cross=tf.cross(a,b) out=sess.run(cross) print(out)
'''
TensorFlow’s zillion operations for division.
tf.div - TF style division
tf.division - Python's style division
'''
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))) # ==> [[0 0] [1 1]]
print(sess.run(tf.divide(b, a))) # ==> [[0. 0.5] [1. 1.5]]
print(sess.run(tf.truediv(b, a))) # ==> [[0. 0.5] [1. 1.5]]
print(sess.run(tf.floordiv(b, a))) # ==> [[0 0] [1 1]]
# print(sess.run(tf.realdiv(b, a))) # ErrorL only works for real values
print(sess.run(tf.truncatediv(b, a))) # ==> [[0 0] [1 1]]
print(sess.run(tf.floor_div(b, a))) # ==> [[0 0] [1 1]]
'''
tf.add_n: Allow to add multiple tensors.
tf.add_n([1, b, c] => a + b + c
'''
a = tf.constant(1, name="a")
b = tf.constant(2, name="b")
c = tf.constant(3, name="c")
add_n = tf.add_n([a, b, c])
with tf.Session() as sess:
print(sess.run(add_n))
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
y = tf.ones_like(t_0) # ==> 1
def beam_search_step(time_, logits, beam_state, config):
"""Performs a single step of Beam Search Decoding.
Args:
time_: Beam search time step, should start at 0. At time 0 we assume
that all beams are equal and consider only the first beam for
continuations.
logits: Logits at the current time step. A tensor of shape `[B, vocab_size]`
beam_state: Current state of the beam search. An instance of `BeamState`
config: An instance of `BeamSearchConfig`
Returns:
A new beam state.
"""
# Calculate the current lengths of the predictions
prediction_lengths = beam_state.lengths
previously_finished = beam_state.finished
# Calculate the total log probs for the new hypotheses
# Final Shape: [beam_width, vocab_size]
probs = tf.nn.log_softmax(logits)
probs = mask_probs(probs, config.eos_token, previously_finished)
total_probs = tf.expand_dims(beam_state.log_probs, 1) + probs
# Calculate the continuation lengths
# We add 1 to all continuations that are not EOS and were not
# finished previously
lengths_to_add = tf.one_hot([config.eos_token] * config.beam_width,
config.vocab_size, 0, 1)
add_mask = (1 - tf.cast(previously_finished, dtype=tf.int32))
lengths_to_add = tf.expand_dims(add_mask, 1) * lengths_to_add
new_prediction_lengths = tf.expand_dims(prediction_lengths,
1) + lengths_to_add
# Calculate the scores for each beam
scores = hyp_score(log_probs=total_probs,
sequence_lengths=new_prediction_lengths,
config=config)
scores_flat = tf.reshape(scores, [-1])
# During the first time step we only consider the initial beam
scores_flat = tf.cond(pred=tf.convert_to_tensor(value=time_) > 0,
true_fn=lambda: scores_flat,
false_fn=lambda: scores[0])
# Pick the next beams according to the specified successors function
next_beam_scores, word_indices = config.choose_successors_fn(
scores_flat, config)
next_beam_scores.set_shape([config.beam_width])
word_indices.set_shape([config.beam_width])
# Pick out the probs, beam_ids, and states according to the chosen predictions
total_probs_flat = tf.reshape(total_probs, [-1], name="total_probs_flat")
next_beam_probs = tf.gather(total_probs_flat, word_indices)
next_beam_probs.set_shape([config.beam_width])
next_word_ids = tf.math.mod(word_indices, config.vocab_size)
next_beam_ids = tf.truncatediv(word_indices, config.vocab_size)
# Append new ids to current predictions
next_finished = tf.logical_or(
tf.gather(beam_state.finished, next_beam_ids),
tf.equal(next_word_ids, config.eos_token))
# Calculate the length of the next predictions.
# 1. Finished beams remain unchanged
# 2. Beams that are now finished (EOS predicted) remain unchanged
# 3. Beams that are not yet finished have their length increased by 1
lengths_to_add = tf.cast(tf.not_equal(next_word_ids, config.eos_token),
dtype=tf.int32)
lengths_to_add = (1 -
tf.cast(next_finished, dtype=tf.int32)) * lengths_to_add
next_prediction_len = tf.gather(beam_state.lengths, next_beam_ids)
next_prediction_len += lengths_to_add
next_state = BeamSearchState(log_probs=next_beam_probs,
lengths=next_prediction_len,
finished=next_finished)
output = BeamSearchStepOutput(scores=next_beam_scores,
predicted_ids=next_word_ids,
beam_parent_ids=next_beam_ids)
return output, next_state
t2_1 = tf.ones_like(t2, name="t2_1") # 3x3 tensor, all elem = True
#
a = tf.constant([2, 2], name='a')
b = tf.constant([[0, 1], [2, 3]], name='b')
c = tf.constant([10, 20], name='c')
d = tf.constant([2, 3], name='d')
ops = [
# division
tf.div(b, a),
tf.divide(b, a),
tf.truediv(b, a),
tf.floordiv(b, a),
# tf.realdiv(b, a) # nly for real value
tf.truncatediv(b, a),
# moltiplication
tf.multiply(c, d), # element-wise
tf.tensordot(c, d, 1), # output 1 value
tf.add(t1_0, t1_1)
]
writer = tf.summary.FileWriter(utils.logdir, tf.get_default_graph())
with tf.Session() as sess:
# run() output is a numpy array
[print(sess.run(op)) for op in ops]
writer.close()