[TOC]
# Python 3.4+ only.
from importlib import reload
# Condition 1: foo is a folder, ./foo/utils.py, reload utils.py
reload(foo.utils)
from foo.utils import *
# Condition 2: foo.py is a script, reload foo.py
reload(foo)
# Condition 3: foo.py is a script, Foo is a class in foo.py, reload Foo
relaod(foo)
from foo import FooPython中的对象之间赋值时是按引用传递的，如果需要拷贝对象，需要使用标准库中的 copy 模块。
-
copy.copy浅拷贝 只拷贝父对象，不会拷贝对象的内部的子对象。 -
copy.deepcopy深拷贝 拷贝对象及其子对象>>> import copy >>> a = [1,2,3,4,['a','b']] #原始对象 >>> b = a #赋值，传对象的引用 >>> c = copy.copy(a) >>> d = copy.deepcopy(a) >>> a.append(5) >>> a[4].append('c') >>> print 'a=',a a= [1, 2, 3, 4, ['a', 'b', 'c'], 5] >>> print 'b=',b b= [1, 2, 3, 4, ['a', 'b', 'c'], 5] >>> print 'c=',c c= [1, 2, 3, 4, ['a', 'b', 'c']] >>> print 'd=',d d= [1, 2, 3, 4, ['a', 'b']]
-
for和if同一行时，if放在for后面mask_ = [[1 for j in range(6) if i<j] for i in range(6)] >>> mask_ [[1, 1, 1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1], [1], []]
-
for和if else同一行时，if else放在for前面mask_ = [[1 if i<j else 0 for j in range(6)] for i in range(6)] >>> mask_ [[0, 1, 1, 1, 1, 1], [0, 0, 1, 1, 1, 1], [0, 0, 0, 1, 1, 1], [0, 0, 0, 0, 1, 1], [0, 0, 0, 0, 0, 1], [0, 0, 0, 0, 0, 0]]
torch.multiprocessing 和 from IPython import embed 不要同时用，会有 bug
import operator
def num_params():
""" Return total num of params of the model """
total_num = 0
for var in tf.trainable_variables():
shape = var.get_shape()
total_num += functools.reduce(operator.mul, [dim.value for dim in shape], 1)
return total_numclass PlateauLRDecay:
""" Adjust learning rate.
Args:
init_lr: float, initial learning rate.
epoch_patience: int, the epoch interval to reduce lr.
period_patience: int, the upper-bound number of epoch_patience.
min_lr: float, the minimum learning rate.
rate: float, reduce rate.
verbose: bool, if display learning rate on shell or not.
"""
def __init__(self, init_lr, epoch_patience, period_patience, min_lr=0.00001, rate=0.4, verbose=False):
self.lr = init_lr
self.epoch_patience = epoch_patience
self.period_patience = period_patience
self.min_lr = min_lr
self.rate = rate
self.verbose = verbose
self.prev_best_epoch_num = 0
self.prev_best_loss = float('inf')
if self.lr <= self.min_lr:
self.lr = self.min_lr
self.is_min_lr = True
else:
self.is_min_lr = False
def update_lr(self, loss, epoch_num):
""" Update learning rate.
Args:
loss: float, loss of every epoch.
epoch_num: int, epoch.
Return:
bool, if update or not.
"""
if loss < self.prev_best_loss:
self.prev_best_loss = loss
self.prev_best_epoch_num = epoch_num
else:
epochs = epoch_num - self.prev_best_epoch_num
if self.is_min_lr is True or epochs >= self.epoch_patience * self.period_patience:
self.lr = 0.0
elif epochs % self.epoch_patience == 0:
# reduce lr
self.lr = min(self.lr * self.rate, self.min_lr)
if self.is_min_lr is False and self.lr == self.min_lr:
self.is_min_lr = True
self.prev_best_epoch_num = epoch_num
if self.verbose:
print('Reduce lr to ', self.lr)
return True
return Falsedef init_summary_writer(self, root_dir):
""" Init tensorboard writer """
tf_board_dir = 'tfb_dir'
folder = os.path.join(root_dir, tf_board_dir)
self.train_summary_writer = tf.summary.FileWriter(os.path.join(folder, 'train'), self.sess.graph)
self.valid_summary_writer = tf.summary.FileWriter(os.path.join(folder, 'valid'))
self.test_summary_writer = tf.summary.FileWriter(os.path.join(folder, 'test'))
def write_summary(self, epoch_num, kv_pairs, phase):
""" Write summary into tensorboard """
if phase == RunnerPhase.TRAIN:
summary_writer = self.train_summary_writer
elif phase == RunnerPhase.VALIDATE:
summary_writer = self.valid_summary_writer
elif phase == RunnerPhase.PREDICT:
summary_writer = self.test_summary_writer
else:
raise RuntimeError('Unknow phase: ' + phase)
if summary_writer is None:
return
for key, value in kv_pairs.items():
metrics = tf.Summary()
metrics.value.add(tag=key, simple_value=value)
summary_writer.add_summary(metrics, epoch_num)
summary_writer.flush()http://wuxiaoqian.blogspot.com/2017/07/tfscan.html
def dynamic_run(self, seq_type_emb, dtime):
def move_forward_fn(accumulator, item):
pass
return h_t, init_state
initial_state = list()
initial_h_t = list()
h_ts, cell_states = tf.scan(move_forward_fn,
elements,
initializer=(initial_h_t, initial_state))
return h_ts, cell_stateshttps://zhuanlan.zhihu.com/p/45673869
-
tf.split(input, num_split, dimension)
dimension 的意思就是输入张量的哪一个维度，如果是 0 就表示对第 0 维度进行切割。num_split 就是切割的数量，如果是 2 就表示输入张量被切成 2 份，每一份是一个列表。
import tensorflow as tf; A = [[1,2,3],[4,5,6]] x = tf.split(A, 3, 1) with tf.Session() as sess: c = sess.run(x) for ele in c: print( ele ) # Out: # [[1] # [4]] # [[2] # [5]] # [[3] # [6]]
-
torch.tensor.split(tensor,split_size_or_sections,dim=0)
- 第一个参数是待分割张量
- 第二个参数有两种形式。
- 第一种是分割份数；
- 第二种这是分割方案，这是一个list，待分割张量将会分割为len（list）份，每一份的大小取决于list中的元素
- 第三个参数为分割维度
section=[1,2,1,2,2] d=torch.randint(0, 10, (8,4)) print(torch.split(d,section,dim=0)) #输出结果为： (tensor([[5, 8, 7, 9]]), tensor([[1, 4, 9, 3], [1, 3, 0, 4]]), tensor([[2, 4, 4, 2]]), tensor([[0, 3, 4, 8], [5, 7, 6, 3]]), tensor([[9, 2, 7, 1], [7, 5, 8, 8]]))
https://blog.csdn.net/YiRanNingJing/article/details/79451786
tf.cumsum(
x,
axis=0,
exclusive=False,
reverse=False,
name=None
)函数 tf.cumsum 是 cumulative sum缩写，计算累积和，即沿着tensor（张量）x的某一个维度axis，计算累积和。
参数解释：
-
x, 即我们要计算累积和的tensor。
-
axis=0, 默认是沿着x的第0维计算累积和。
-
exclusive=False, 表示输出结果的第一元素是否与输入的第一个元素一致。默认exclusive=False，表示输出的第一个元素与输入的第一个元素一致（By default, this op performs an inclusive cumsum, which means that the first element of the input is identical to the first element of the output）。这是官方文档的解释。当我们对一个数组arr（或其他什么东东）进行累积求和时，我们要对累积和sum进行初始化，初始化的方式有两种，一种是将累积和初始化为0，即sum=0，一种是使用数组arr的第一个元素对累积和进行初始化，即sum=arr[0]。所以参数exclusive描述的是如何对累积和进行初始化。
-
reverse=False, 表示是否逆向累积求和。默认reverse=False，即正向累积求和。
a = [[1 ,2, 3],
[4, 5, 6],
[7, 8, 9]]
# axis=0
sum1 = tf.cumsum(a, axis=0)
# sum1 = [[ 1, 2, 3],
# [ 5, 7, 9],
# [12, 15, 18]]
# exclusive=True
sum4 = tf.cumsum(a, exclusive=True)
# sum4= [[0, 0, 0],
# [1, 2, 3],
# [5, 7, 9]]#
# reverse=True
sum5 = tf.cumsum(a, reverse=True)
# sum5 = [[12, 15, 18],
# [11, 13, 15],
# [ 7, 8, 9]]# PyTorch 中的 ModuleList 和 Sequential: 区别和使用场景
# The model is defined before, the codes below counts the number of parameters in training model
num_parameters_train = sum(p.numel() for p in model.parameters() if p.requires_grad)import torch.multiprocessing as mp
processes = []
params_list = [['mts_archive', 'ArabicDigits', 'fcn', 'adam'],
['mts_archive', 'AUSLAN', 'fcn', 'adam'],
['mts_archive', 'CharacterTrajectories', 'fcn', 'adam']]
num_processes = len(params_list)
for i in range(num_processes):
p = mp.Process(target=main, args=(params_list[i]))
p.start()
processes.append(p)
for p in processes:
p.join()# Pytorch will only use one GPU by default. You can easily run your operations on multiple GPUs
# by making your model run parallelly using `DataParallel`
model = nn.DataParallel(model)# import module
from tensorboardX import SummaryWriter
# define logger
logger = SummaryWriter(dir_logs)
# write variable to tensorboard
logger.add_scalars('{}/loss'.format(log_model_name), \
{'loss': loss_epoch.item()}, epoch)# view logs across browser
tensorboard --logdir ./ [--port 6007]pip install --upgrade git+https://github.com/sovrasov/flops-counter.pytorch.gitimport torch
import torchvision.models as models
from ptflops import get_model_complexity_info
with torch.cuda.device(0):
net = models.densenet161()
flops, params = get_model_complexity_info(net, (3, 224, 224), as_strings=True, print_per_layer_stat=True)
print('Flops: ' + flops)
print('Params: ' + params)# A collapsible section with markdown (work in github)
<details>
<summary>Click to expand!</summary>
<p>
## Heading
1. A numbered
2. list
* With some
* Sub bullets
</p>
</details>- 效果展示
| 空格类型 | 写法 | 效果演示 | 效果描述 |
|---|---|---|---|
| 两个quad空格 | a \qquad b | 两个m的宽度 | |
| quad空格 | a \quad b | 一个m的宽度 | |
| 大空格 | a\ b | 1/3m宽度 | |
| 中等空格 | a;b | 2/7m宽度 | |
| 小空格 | a,b | 1/6m宽度 | |
| 没有空格 | ab | ||
| 紧贴 | a!b | 缩进1/6m宽度 |
-
Key words:
\vphantom{\frac11}
-
Equal vertical space
\begin{align}
f_1(x) &= \frac{15x}{3} \\
f_2(x) &= \vphantom{\frac11} 3x + 5 \\
f_3(x) &= \vphantom{\frac11} 4x + 13
\end{align}- Unequal vertical space
\begin{align}
f_1(x) &= \frac{15x}{3} \\
f_2(x) &= 3x + 5 \\
f_3(x) &= 4x + 13
\end{align}# data:
# [['306', '3.27', 'Fake Pruning(60%)'],
# ['422', '3.11', 'Fake Pruning(40%)']]
x = [float(i) for i in data[:,0]]
y = [float(i) for i in data[:,1]]
label = data[:,2]
fig, ax = plt.subplots(figsize=(5,5))
# different color of two part of points
ax.scatter(x[:9], y[:9], c='black', marker='o')
ax.scatter(x[9:], y[9:], c='red', marker='x')
ax.set_xlim(300, 450)
ax.set_ylim(3, 3.4)
# set interval of coordinate axis
ax.xaxis.set_major_locator(MultipleLocator(1))
ax.yaxis.set_major_locator(MultipleLocator(100))
ax.set_xlabel('FLOPs(M)', fontsize=15)
ax.set_ylabel('CIFAR-10 Error(%)', fontsize=15)
# set grid's shape
plt.grid(ls='--')
for i, txt in enumerate(label):
if i < 1:
ax.annotate(txt, (x[i]+5,y[i]), fontsize=10)
else:
ax.annotate(txt, (x[i]+5,y[i]), color='r', fontsize=10)
plt.show()
plt.savefig('./result.png')
# save the figure to pdf file
plt.savefig('result.pdf', format='pdf')-
效果展示
线条风格 实线 虚线 破折线 点画线 无线条 代码表示 - : – -. None 或 , 线条颜色 红 洋红 黄 绿 青 蓝 黑 白 代码表示 r m y g c b k w 标记 描述 标记 描述 标记 描述 o 圆圈 . 点 * 星号 + 加号 v 朝下三角 ^ 朝上三角 < 朝左三角 > 朝右三角 D 大菱形 d 小菱形 s 正方形 p 五边形 H 大六边形 h 小六边形 p 八边形 x ✘号 None或, 无标记 - https://blog.csdn.net/guoziqing506/article/details/78975150
- 博客包含以下内容：
- 绘制一个最简单的折线图
- 绘制不同风格的线条
- 坐标轴的控制
- 坐标范围
- 坐标标题
- 坐标间隔设定
- 多图叠加
- 多曲线
- 多图与多子图
- 标题和图例
- 图像标题
- 图例
- 网格，背景色以及文字注释
- 添加网格
- 背景色
- 文字注释
