demo

mini-batch 实现

data_size = 8
batch_size = 2

dataset = tf.data.Dataset.range(data_size).shuffle(1).batch(batch_size)

iterator = tf.data.Iterator.from_structure(tf.int64, tf.TensorShape([batch_size, ]))
sess.run(iterator.make_initializer(dataset))

for _ in range(data_size // batch_size):
    i = sess.run(iterator.get_next())
    print(i)

# Output:
# [0 1]
# [2 3]
# [4 5]
# [6 7]

# Shuffle的参数代表采样区buffer大小，1代表buffer大小为1，即不随机

flops and params

run_meta = tf.RunMetadata()

opts = tf.profiler.ProfileOptionBuilder.float_operation()    
flops = tf.profiler.profile(sess.graph, run_meta=run_meta, cmd='op', options=opts)

opts = tf.profiler.ProfileOptionBuilder.trainable_variables_parameter()    
params = tf.profiler.profile(sess.graph, run_meta=run_meta, cmd='op', options=opts)

flops.total_float_ops
params.total_parameters

show cpu or gpu

import tensorflow as tf; sess = tf.Session(config=tf.ConfigProto(log_device_placement=True))

如果显示CPU的信息，就是CPU版本

custom metrics

tf.metrics.mean(tf.losses.softmax_cross_entropy(multi_class_labels, logits, loss_collection=None))

multi-one-hot

# some sparse features, from raw maker_id to index
maker_id_list = [1, 3, 9, 14, 2]
one_hot_enc = tf.one_hot(indices=maker_id_list, depth=16)
with tf.Session() as sess:
    feature_list = tf.reduce_sum(one_hot_enc, reduction_indices=0)
    print(sess.run(feature_list))

坑

BatchNorm 的坑

https://blog.csdn.net/qq_25737169/article/details/79616671