https://blog.youkuaiyun.com/kaixinjiuxing666/article/details/81004010
https://www.aiuai.cn/aifarm646.html#2.TensorBoardX-Graph%E5%8F%AF%E8%A7%86%E5%8C%96
https://blog.youkuaiyun.com/bigbennyguo/article/details/87956434#_add__50
看COCO文件的
2.画激活情况
用于检查深层网络里的层激活与权值分布情况，避免梯度消失等
for name, param in model.named_parameters():
writer.add_histogram(
name, param.clone().data.numpy(), epoch_index)Some comments on Speed
-
Matplotlib operations can be very slow as Matplotlib runs in python rather than native code, so please watch out for runtime speed. There is still a room for improvement, which will be addressed in the near future.
-
Moreover, it might be also a good idea to draw plots from the main code (rather than having a TF op) and add them as image summaries. Please use this library at your best discernment.
Thread-safety issue
Please use object-oriented matplotlib APIs (e.g. Figure, AxesSubplot) instead of pyplot APIs (i.e. matplotlib.pyplot or plt.XXX()) when creating and drawing plots. This is because pyplot APIs are not thread-safe, while the TensorFlow plot operations are usually executed in multi-threaded manners.
For example, avoid any use of pyplot (or plt):
# DON'T DO LIKE THIS !!!
def figure_heatmap(heatmap):
fig = plt.figure() # <--- NO!
plt.imshow(heatmap)
return figand do it like:
def figure_heatmap(heatmap):
fig = matplotlib.figure.Figure() # or just `fig = tfplot.Figure()`
ax = fig.add_subplot(1, 1, 1) # ax: AxesSubplot
# or, just `fig, ax = tfplot.subplots()`
ax.imshow(heatmap)
return fig # fig: FigureFor example, tfplot.subplots() is a good replacement for plt.subplots() to use inside plot functions. Alternatively, you can just take advantage of automatic injection of fig and/or ax.
# def add_heatmap(feature_maps, name):
# '''
#
# :param feature_maps:[B, H, W, C]
# :return:
# '''
#
# def figure_attention(activation):
# fig, ax = tfp.subplots()
# im = ax.imshow(activation, cmap='jet')
# fig.colorbar(im)
# return fig
#
# heatmap = tf.reduce_sum(feature_maps, axis=-1)
# heatmap = tf.squeeze(heatmap, axis=0)
# tfp.summary.plot(name, figure_attention, [heatmap])
# def make_r_gt_mask(fet_h, fet_w, img_h, img_w, gtboxes):
# assert gtboxes.shape[1] == 8, "gtboxes shape should be (-1, 8)"
# gtboxes = back_forward_convert(gtboxes)
# gtboxes = np.reshape(gtboxes, [-1, 6]) # [x, y, w, h, theta, label]
#
# areas = gtboxes[:, 2] * gtboxes[:, 3]
# arg_areas = np.argsort(-1 * areas) # sort from large to small
# gtboxes = gtboxes[arg_areas]
#
# fet_h, fet_w = int(fet_h), int(fet_w)
# mask = np.zeros(shape=[fet_h, fet_w], dtype=np.int32)
# for a_box in gtboxes:
# # print(a_box)
# box = cv2.boxPoints(((a_box[0], a_box[1]), (a_box[2], a_box[3]), a_box[4]))
# box = np.reshape(box, [-1, ])
# label = a_box[-1]
# new_box = []
# for i in range(8):
# if i % 2 == 0:
# x = box[i]
# new_x = int(x * fet_w / float(img_w))
# new_box.append(new_x)
# else:
# y = box[i]
# new_y = int(y*fet_h/float(img_h))
# new_box.append(new_y)
#
# new_box = np.int0(new_box).reshape([4, 2])
# color = int(label)
# # print(type(color), color)
# cv2.fillConvexPoly(mask, new_box, color=color)
# # print (mask.dtype)
# return mask
# def vis_mask_tfsmry(mask, name):
# '''
#
# :param mask:[H, W]. It's a tensor, not array
# :return:
# '''
#
# def figure_attention(activation):
# fig, ax = tfp.subplots()
# im = ax.imshow(activation, cmap='jet')
# fig.colorbar(im)
# return fig
#
# heatmap = mask*10
#
# tfp.summary.plot(name, figure_attention, [heatmap])| cv2.imshow("test", img) |
| cv2.waitKey(0) |
|
|
def test_plt():
from PIL import Image
import matplotlib.pyplot as plt
import numpy as np
a = np.random.rand(20, 30)
print (a.shape)
# plt.subplot()
b = plt.imshow(a)
plt.show()
tf.summary.image('positive_anchor', pos_in_img)
tf.summary.image('negative_anchors', neg_in_img)
tf.summary.image('Compare/gtboxes_gpu:%d' % i, gtboxes_in_img)
if cfgs.ADD_BOX_IN_TENSORBOARD:
detections_in_img = show_box_in_tensor.draw_boxes_with_categories_and_scores(
img_batch=img,
boxes=outputs[0],
scores=outputs[1],
labels=outputs[2])
tf.summary.image('Compare/final_detection_gpu:%d' % i, detections_in_img)7、tf.summary.merge_all
merge_all 可以将所有summary全部保存到磁盘，以便tensorboard显示。如果没有特殊要求，一般用这一句就可一显示训练时的各种信息了。格式：tf.summaries.merge_all(key='summaries')
8、tf.summary.FileWriter
指定一个文件用来保存图。格式：tf.summary.FileWritter(path,sess.graph)
可以调用其add_summary（）方法将训练过程数据保存在filewriter指定的文件中
Tensorflow Summary 用法示例:
-
tf.summary.scalar('accuracy',acc) #生成准确率标量图 -
merge_summary = tf.summary.merge_all() -
train_writer = tf.summary.FileWriter(dir,sess.graph)#定义一个写入summary的目标文件，dir为写入文件地址 -
......(交叉熵、优化器等定义) -
for step in xrange(training_step): #训练循环 -
train_summary = sess.run(merge_summary,feed_dict = {...})#调用sess.run运行图，生成一步的训练过程数据 -
train_writer.add_summary(train_summary,step)#调用train_writer的add_summary方法将训练过程以及训练步数保存
此时开启tensorborad：
- tensorboard --logdir=/summary_dir
便能看见accuracy曲线了。
另外，如果我不想保存所有定义的summary信息，也可以用tf.summary.merge方法有选择性地保存信息：
9、tf.summary.merge
格式：tf.summary.merge(inputs, collections=None, name=None)
一般选择要保存的信息还需要用到tf.get_collection()函数
示例：
-
tf.summary.scalar('accuracy',acc) #生成准确率标量图 -
merge_summary = tf.summary.merge([tf.get_collection(tf.GraphKeys.SUMMARIES,'accuracy'),...(其他要显示的信息)]) -
train_writer = tf.summary.FileWriter(dir,sess.graph)#定义一个写入summary的目标文件，dir为写入文件地址 -
......(交叉熵、优化器等定义) -
for step in xrange(training_step): #训练循环 -
train_summary = sess.run(merge_summary,feed_dict = {...})#调用sess.run运行图，生成一步的训练过程数据 -
train_writer.add_summary(train_summary,step)#调用train_writer的add_summary方法将训练过程以及训练步数保存
使用tf.get_collection函数筛选图中summary信息中的accuracy信息，这里的 tf.GraphKeys.SUMMARIES 是summary在collection中的标志。当然，也可以直接：
-
acc_summary = tf.summary.scalar('accuracy',acc) #生成准确率标量图 -
merge_summary = tf.summary.merge([acc_summary ,...(其他要显示的信息)]) #这里的[]不可省
扫一扫
1万+
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
