机器学习中几个关键操作

固定参数不进行训练

pytorch

1. 模型中设置

# Fix the layers before conv3:
for layer in range(10):
   for p in self.vgg.features[layer].parameters():
      p.requires_grad = False

2. 变量声明时设置

# 方法1
from torch.autograd import Variable
x = Variable(torch.randn(5, 5), requires_grad=False)
print(x.numpy())
    
# 方法2
y = torch.ones(10, requires_grad=False)
print(y.numpy())

tensorflow

1. 模型中设置

import tensorflow.contrib.slim as slim
# 卷积的
net = slim.repeat(self._image, 2, slim.conv2d, 64, [3, 3], trainable=False, scope='conv1')
# 全连接的
fc6 = slim.fully_connected(pool5_flat, 4096, scope='fc6', trainable=False)
# dropout
fc6 = slim.dropout(fc6, keep_prob=0.5, is_training=True, scope='dropout6')

2. 变量声明时设置

w = tf.Variable(1.0, name="w", trainable=False)

paddlepaddle

1. 模型中设置

import paddle.fluid as fluid

# fluid.ParamAttr 方法中的 initializer 可以设为None的，这样在fc层的时候，会自动生成新的超参数值的。

w_param_attrs = fluid.ParamAttr(name="fc_weight",
                                learning_rate=1.0, # 这个学习率是要乘以 global learning_rate 和 scheduler_factor的
                                regularizer=fluid.regularizer.L2Decay(1.0),
                                trainable=True) # 代表是否可训练
x = fluid.layers.data(name='X', shape=[1], dtype='float32')
y_predict = fluid.layers.fc(input=x, size=10, param_attr=w_param_attrs)

2. 变量声明中设置

class paddle.fluid.ParamAttr(name=None, initializer=None, learning_rate=1.0, regularizer=None, trainable=True, do_model_average=False)

 

训练过程中修改学习率

pytorch

    for param_group in optimizer.param_groups:
        param_group['lr'] *= scale

tensorflow

lr = tf.Variable(cfg.TRAIN.LEARNING_RATE, trainable=False)
self.optimizer = tf.train.MomentumOptimizer(lr, cfg.TRAIN.MOMENTUM)

 rate *= cfg.TRAIN.GAMMA
 sess.run(tf.assign(lr, rate))

paddlepaddle

paddle.fluid.layers里面有学习率

https://www.paddlepaddle.org.cn/documentation/docs/zh/1.5/api_cn/layers_cn/learning_rate_scheduler_cn.html

 

Tensorboard 使用

命令

tensorboard --logdir=./ --port=8008

 

from tensorboardX import SummaryWriter

writer = SummaryWriter(logdir="./tensorboard")
for epoch_id in range(train_epoch):
   for batch_id in range(total_size):
       # 省略了好多计算loss tensor的步骤........
       avg_loss = fluid.layers.mean(loss)
       writer.add_scalar("Loss", avg_loss.numpy(), epoch_id * batch_id)
       writer.add_scalar("Loss__1.1", avg_loss.numpy() + 1.1, epoch_id * batch_id)

writer.close()

结果：

当前目录下只有一个logdir。

如果当前目录下，有多个logdir,则每个目录对于一种颜色的曲线

后面还有graph, histogram等图  https://www.jianshu.com/p/46eb3004beca

tensorflow

import tensorflow as tf

writer = tf.summary.FileWriter(self.tbdir, sess.graph)
valwriter = tf.summary.FileWriter(self.tbvaldir)

A = tf.summary.histogram('A_' + tensor.op.name, tensor)
B = tf.summary.scalar('B_' + tensor.op.name, tf.nn.zero_fraction(tensor))

val_summaries = []
val_summaries.append(B)

summary_op = tf.summary.merge_all() # A,B 两个图都有
 
summary_op_val = tf.summary.merge(val_summaries)  # 只有B

writer.add_summary(session.run(summary_op, feed_dict={}), float(iter))
valwriter.add_summary(session.run(summary_op_val, feed_dict={}), float(iter))


writer.close()
valwriter.close()

paddlepaddle

使用VisualDL， 里面有 安装 和 使用 代码

路径：https://blog.csdn.net/qq_33200967/article/details/79127175