用tensorboard可视化模型，数据和训练
Leonhardt 2022-03-26 About 2 min
# 准备
SummaryWriter:向tensorboard写入的关键对象。创建该对象时，需要指定持久化文件保存的文件夹路径。如果路径不存在，会自动创建。
运行tensorboard: tensorboard --logdir logpath
from torch.utils.tensorboard import SummaryWriter
writer = SummaryWriter('log/FMNIST')
1
2
2
# 写入图片
添加global_step参数后，可以向同一个ImageName写入多次，会有一个类似进度条的东西，通过调节可以查看不同step写入的图像。正适合用于查看GAN生成的图像变化。
writer.add_image('ImageName', img, global_step=step)
1
# 查看模型
writer.add_graph(model, model_input)
1
# Projector--可视化高维数据
通过PCA,T-SNE等方法将数据降维到三维。
- features: 需要降维显示的高维特征
- metadata: 每个feature对应的类别
- label_img: 数据点对应的图像，可在可视化中展示。可选。
writer.add_embedding(
features,
metadata=class_labels,
label_img=data,
global_step=batch_idx,
)
1
2
3
4
5
6
2
3
4
5
6
# 记录loss等标量和plt.figure
writer.add_scalar('Name', loss, global_step=step)
writer.add_figure('Name', fig, global_step=step)
1
2
2
# 绘制PR-曲线
classes[class_index]：类别标签。只是一个Name。- tensorboard_truth: ground_truth。
- tensorboard_probs: specific_class_predicted_probability。
PR-CURVE的绘制过程中specific_class_predicted_probability将用np.histogram处理，默认分成128个区间。其中，ground_truth作为weights参数计算预测为正的样本数。1-ground_truth作为weights参数计算预测为负的样本数。
def add_pr_curve_tensorboard(class_index, test_probs, test_label, global_step=0):
'''
Takes in a "class_index" from 0 to 9 and plots the corresponding
precision-recall curve
'''
tensorboard_truth = test_label == class_index
tensorboard_probs = test_probs[:, class_index]
writer.add_pr_curve(classes[class_index],
tensorboard_truth,
tensorboard_probs,
global_step=global_step)
1
2
3
4
5
6
7
8
9
10
11
12
2
3
4
5
6
7
8
9
10
11
12
# 查看权重分布
实际就是将传入的weights拉成一维，然后画出直方图。
writer.add_histogram("fc1", weights, global_step=global_step)
1
# 记录超参数
add_hparams(hpyrams, metrics)
会以列表，平行坐标或散点图的方式展示超参数对度量的影响。
writer.add_hparams(
{"lr": learning_rate, "bsize": batch_size},
{
"accuracy": sum(accuracies) / len(accuracies),
"loss": sum(losses) / len(losses),
},
)
1
2
3
4
5
6
7
2
3
4
5
6
7
# 从tfevents中读取数据
from tensorflow.python.summary.summary_iterator import summary_iterator
tfevent_filename = "path/to/events.out.tfevents.xxxx"
for event in summary_iterator(tfevent_filename):
# print(event.wall_time)
# print(event.step)
for v in event.summary.value:
print(v.tag)
print(v.HasField('image'))
print(v.HasField('simple_value'))
"""
tfevents迭代器打印出的内容
wall_time: 1648622763.1866968
file_version: "brain.Event:2"
wall_time: 1648622764.5857239
summary {
value {
tag: "Loss"
simple_value: -9831.6953125
}
}
wall_time: 1648622764.7647285
summary {
value {
tag: "Your Image"
image {
height: 266
width: 530
colorspace: 3
encoded_image_string: "\211PNG\r\n\032\n\000\000\000\rIHDR\000\000......"
}
}
}
"""
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40