P8-P9：Tensorboard的使用（Pytorch小土堆学习笔记）

安装直接在terminal上安装就行，注意要进入pytorch环境（用conda activate pytorch来激活）

首先说包，这个包是将内容告诉给tensorboard，并让它展示的

from torch.utils.tensorboard import SummaryWriter

summarywriter部分关键注解 

Help on class SummaryWriter in module torch.utils.tensorboard.writer:
class SummaryWriter(builtins.object)
 |  SummaryWriter(log_dir=None, comment='', purge_step=None, max_queue=10, flush_secs=120, filename_suffix='')
 |  
 |  Writes entries directly to event files in the log_dir to be
 |  consumed by TensorBoard.
 |  
 |  The `SummaryWriter` class provides a high-level API to create an event file
 |  in a given directory and add summaries and events to it. The class updates the
 |  file contents asynchronously. This allows a training program to call methods
 |  to add data to the file directly from the training loop, without slowing down
 |  training.
 |  
 |  Methods defined here:
 |  
 |  __enter__(self)
 |  
 |  __exit__(self, exc_type, exc_val, exc_tb)
 |  
 |  __init__(self, log_dir=None, comment='', purge_step=None, max_queue=10, flush_secs=120, filename_suffix='')
 |      Creates a `SummaryWriter` that will write out events and summaries
 |      to the event file.
 |      
 |      Args:
 |          log_dir (string): Save directory location. Default is
 |            runs/**CURRENT_DATETIME_HOSTNAME**, which changes after each run.
 |            Use hierarchical folder structure to compare
 |            between runs easily. e.g. pass in 'runs/exp1', 'runs/exp2', etc.
 |            for each new experiment to compare across them.
 |          comment (string): Comment log_dir suffix appended to the default
 |            ``log_dir``. If ``log_dir`` is assigned, this argument has no effect.
 |          purge_step (int):
 |            When logging crashes at step :math:`T+X` and restarts at step :math:`T`,
 |            any events whose global_step larger or equal to :math:`T` will be
 |            purged and hidden from TensorBoard.
 |            Note that crashed and resumed experiments should have the same ``log_dir``.
 |          max_queue (int): Size of the queue for pending events and
 |            summaries before one of the 'add' calls forces a flush to disk.
 |            Default is ten items.
 |          flush_secs (int): How often, in seconds, to flush the
 |            pending events and summaries to disk. Default is every two minutes.
 |          filename_suffix (string): Suffix added to all event filenames in
 |            the log_dir directory. More details on filename construction in
 |            tensorboard.summary.writer.event_file_writer.EventFileWriter.
 |      
 |      Examples::
 |      
 |          from torch.utils.tensorboard import SummaryWriter
 |      
 |          # create a summary writer with automatically generated folder name.
 |          writer = SummaryWriter()
 |          # folder location: runs/May04_22-14-54_s-MacBook-Pro.local/
 |      
 |          # create a summary writer using the specified folder name.
 |          writer = SummaryWriter("my_experiment")
 |          # folder location: my_experiment
 |      
 |          # create a summary writer with comment appended.
 |          writer = SummaryWriter(comment="LR_0.1_BATCH_16")
 |          # folder location: runs/May04_22-14-54_s-MacBook-Pro.localLR_0.1_BATCH_16/
 |  
 | 

解释：它是一个向 log_dir写入的事件文件，这个文件可以被tensorboard解析。需要输入一个文件夹的名称，不输入的话默认文件夹为runs/**CURRENT_DATETIME_HOSTNAME**。其他的参数当前并不重要，需要的话可以自己看看。

使用的基本套路：

from torch.utils.tensorboard import SummaryWriter
writer=SummaryWriter(logs)

writer.add_scalar()
writer.add_image()

writer.close()

先说第一个add_scalar

这个是add_scalar的相关文档说明，主要注意一下前三个参数（标题，y轴，x轴）

  add_scalars(self, main_tag, tag_scalar_dict, global_step=None, walltime=None)
 |      Adds many scalar data to summary.
 |      
 |      Args:
 |          main_tag (string): The parent name for the tags
 |          tag_scalar_dict (dict): Key-value pair storing the tag and corresponding values
 |          global_step (int): Global step value to record
 |          walltime (float): Optional override default walltime (time.time())
 |            seconds after epoch of event
 |      
 |      Examples::
 |      
 |          from torch.utils.tensorboard import SummaryWriter
 |          writer = SummaryWriter()
 |          r = 5
 |          for i in range(100):
 |              writer.add_scalars('run_14h', {'xsinx':i*np.sin(i/r),
 |                                              'xcosx':i*np.cos(i/r),
 |                                              'tanx': np.tan(i/r)}, i)
 |          writer.close()
 |          # This call adds three values to the same scalar plot with the tag
 |          # 'run_14h' in TensorBoard's scalar section.
 |      
 |      Expected result:
 |      
 |      .. image:: _static/img/tensorboard/add_scalars.png
 |         :scale: 50 %

 练手代码：

from torch.utils.tensorboard import SummaryWriter

writer=SummaryWriter("logs")
for i in range(100):
    writer.add_scalar("y=2x",2*i,i)
writer.close()

然后点击运行，打开下面的这个并找到上面代码所在的文件夹（通过cd命令）

然后按照下图输入命令行（第一行和第三行） 

tensorboard --logdir=y=2x --port=6007 #换端口

 然后复制网址在网页中打开

如果出现这种情况应该是你的文件名没有搞对，可以检查一下文件夹是否创建以及文件夹里面有没有文件（太多文件有的时候也会影响）

再说第二个 add_image

文档说明：

add_images(self, tag, img_tensor, global_step=None, walltime=None, dataformats='NCHW')
 |      Add batched image data to summary.
 |      
 |      Note that this requires the ``pillow`` package.
 |      
 |      Args:
 |          tag (string): Data identifier
 |          img_tensor (torch.Tensor, numpy.array, or string/blobname): Image data
 |          global_step (int): Global step value to record
 |          walltime (float): Optional override default walltime (time.time())
 |            seconds after epoch of event
 |          dataformats (string): Image data format specification of the form
 |            NCHW, NHWC, CHW, HWC, HW, WH, etc.
 |      Shape:
 |          img_tensor: Default is :math:`(N, 3, H, W)`. If ``dataformats`` is specified, other shape will be
 |          accepted. e.g. NCHW or NHWC.
 |      
 |      Examples::
 |      
 |          from torch.utils.tensorboard import SummaryWriter
 |          import numpy as np
 |      
 |          img_batch = np.zeros((16, 3, 100, 100))
 |          for i in range(16):
 |              img_batch[i, 0] = np.arange(0, 10000).reshape(100, 100) / 10000 / 16 * i
 |              img_batch[i, 1] = (1 - np.arange(0, 10000).reshape(100, 100) / 10000) / 16 * i
 |      
 |          writer = SummaryWriter()
 |          writer.add_images('my_image_batch', img_batch, 0)
 |          writer.close()
 |      
 |      Expected result:
 |      
 |      .. image:: _static/img/tensorboard/add_images.png
 |         :scale: 30 %

 解释：tag：图片标题； img_tensor：图片类型（只有这三种torch.Tensor, numpy.array, or string/blobname）下面的JpegImagePlugin格式不行

 用poencv读取的是numpy类型，转换类型（主要是tensor类型变换）在后续文章深入介绍

这里说一个转换成numpy类型的方式

最后整体的代码：

from torch.utils.tensorboard import SummaryWriter
import numpy as np
from PIL import Image

writer = SummaryWriter("P9_logs_tensorboard2")

image_path1 = "data/train/bees_image/16838648_415acd9e3f.jpg"
image_path2 = "data/train/ants_image/9715481_b3cb4114ff.jpg"
#转换图片类型以符合add_imge的要求
img_PIL1 = Image.open(image_path1)
img_PIL2 = Image.open(image_path2)
img_array1 = np.array(img_PIL1)
img_array2 = np.array(img_PIL2)

# print(type(img_array1))
# print(img_array.shape)

writer.add_image("train", img_array1,1, dataformats='HWC')
writer.add_image("train", img_array2,2, dataformats='HWC')
#数字表示步长，输入不同的数字可以通过滑动进度条看不同步长的图
#dataformats表示图像传入的格式（高|宽|通道），如果不写，默认为（通道|高|宽）

writer.close()

 

 如果更改完代码之后刷新tensorboard并没有变化，需要将SummaryWriter创建的文件夹删掉，重新编译