百度了一下好像这块的代码不多？找了一个能用的代码（似乎也比较旧了，可能以后用不了）
""" Random Forest. Implement Random Forest algorithm with TensorFlow, and apply it to classify handwritten digit images. This example is using the MNIST database of handwritten digits as training samples (http://yann.lecun.com/exdb/mnist/). Author: Aymeric Damien Project: https://github.com/aymericdamien/TensorFlow-Examples/ """
from __future__ import print_function
import tensorflow as tf
from tensorflow.contrib.tensor_forest.python import tensor_forest
from tensorflow.python.ops import resources
# # 忽略所有GPU的代码块，emm大概因为随机森林不能用GPU加速所以不热门？
# import os
# os.environ["CUDA_VISIBLE_DEVICES"] = ""
# Import MNIST data，下载使用mnist数据集
from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets('MNIST_data', one_hot=False)
tf.reset_default_graph() # 这个是针对类似jupyter notebook的会话运行
# 参数
num_steps = 1000 # 一共训练几轮
batch_size = 1024 # 一轮包含多少数据
num_classes = 10 # 数字化个数，即分为几类
num_features = 784 # 每张有28*28个像素
num_trees = 100 # 森林有几棵树
max_nodes = 10000 # 最大节点数
# 输入X输出Y数据占位符
X = tf.placeholder(tf.float32, shape=[None, num_features])
# 随机森林的输出必须是数字int，所以
Y = tf.placeholder(tf.int32, shape=[None])
# 随机森林参数
hparams = tensor_forest.ForestHParams(num_classes=num_classes,
num_features=num_features,
num_trees=num_trees,
max_nodes=max_nodes).fill()
# 建立森林
forest_graph = tensor_forest.RandomForestGraphs(hparams)
# Get training graph and loss
train_op = forest_graph.training_graph(X, Y) # ？？？反正定义了一个训练结点
loss_op = forest_graph.training_loss(X, Y) # ？？？反正和误差有关
# Measure the accuracy
# Returns:A tuple of (probabilities, tree_paths, variance).反正只需要第一个，是一个集合
infer_op, _, _ = forest_graph.inference_graph(X)
correct_prediction = tf.equal(
tf.argmax(infer_op, 1), tf.cast(Y, tf.int64)) # 返回布尔数组
accuracy_op = tf.reduce_mean(
tf.cast(correct_prediction, tf.float32)) # 逻辑转浮点取平均，变成正确率
# Initialize the variables (i.e. assign their default value) and forest resources
# 全部初始化
init_vars = tf.group(tf.global_variables_initializer(),
resources.initialize_resources(resources.shared_resources()))
# Start TensorFlow session
sess = tf.Session()
# Run the initializer
sess.run(init_vars)
# Training
for i in range(1, num_steps + 1):
# Prepare Data
# Get the next batch of MNIST data (only images are needed, not labels)
batch_x, batch_y = mnist.train.next_batch(batch_size)
_, l = sess.run([train_op, loss_op], feed_dict={X: batch_x, Y: batch_y})
if i % 50 == 0 or i == 1:
acc = sess.run(accuracy_op, feed_dict={X: batch_x, Y: batch_y})
print('Step %i, Loss: %f, Acc: %f' % (i, l, acc))
# Test Model测试正确率
test_x, test_y = mnist.test.images, mnist.test.labels
print("Test Accuracy:", sess.run(
accuracy_op, feed_dict={X: test_x, Y: test_y}))
打印结果
Extracting /tmp/data/train-images-idx3-ubyte.gz
Extracting /tmp/data/train-labels-idx1-ubyte.gz
Extracting /tmp/data/t10k-images-idx3-ubyte.gz
Extracting /tmp/data/t10k-labels-idx1-ubyte.gz
INFO:tensorflow:Constructing forest with params =
INFO:tensorflow:{'num_trees': 100, 'max_nodes': 10000, 'bagging_fraction': 1.0, 'feature_bagging_fraction': 1.0, 'num_splits_to_consider': 28, 'max_fertile_nodes': 0, 'split_after_samples': 250, 'valid_leaf_threshold': 1, 'dominate_method': 'bootstrap', 'dominate_fraction': 0.99, 'model_name': 'all_dense', 'split_finish_name': 'basic', 'split_pruning_name': 'none', 'collate_examples': False, 'checkpoint_stats': False, 'use_running_stats_method': False, 'initialize_average_splits': False, 'inference_tree_paths': False, 'param_file': None, 'split_name': 'less_or_equal', 'early_finish_check_every_samples': 0, 'prune_every_samples': 0, 'num_classes': 10, 'num_features': 784, 'bagged_num_features': 784, 'bagged_features': None, 'regression': False, 'num_outputs': 1, 'num_output_columns': 11, 'base_random_seed': 0, 'leaf_model_type': 0, 'stats_model_type': 0, 'finish_type': 0, 'pruning_type': 0, 'split_type': 0}
Extracting /tmp/data/train-images-idx3-ubyte.gz
Extracting /tmp/data/train-labels-idx1-ubyte.gz
Extracting /tmp/data/t10k-images-idx3-ubyte.gz
Extracting /tmp/data/t10k-labels-idx1-ubyte.gz
INFO:tensorflow:Constructing forest with params =
INFO:tensorflow:{'num_trees': 100, 'max_nodes': 10000, 'bagging_fraction': 1.0, 'feature_bagging_fraction': 1.0, 'num_splits_to_consider': 28, 'max_fertile_nodes': 0, 'split_after_samples': 250, 'valid_leaf_threshold': 1, 'dominate_method': 'bootstrap', 'dominate_fraction': 0.99, 'model_name': 'all_dense', 'split_finish_name': 'basic', 'split_pruning_name': 'none', 'collate_examples': False, 'checkpoint_stats': False, 'use_running_stats_method': False, 'initialize_average_splits': False, 'inference_tree_paths': False, 'param_file': None, 'split_name': 'less_or_equal', 'early_finish_check_every_samples': 0, 'prune_every_samples': 0, 'num_classes': 10, 'num_features': 784, 'bagged_num_features': 784, 'bagged_features': None, 'regression': False, 'num_outputs': 1, 'num_output_columns': 11, 'base_random_seed': 0, 'leaf_model_type': 0, 'stats_model_type': 0, 'finish_type': 0, 'pruning_type': 0, 'split_type': 0}
Step 1, Loss: -1.720000, Acc: 0.512695
Step 50, Loss: -256.399994, Acc: 0.917969
Step 100, Loss: -542.260010, Acc: 0.942383
Step 150, Loss: -830.960022, Acc: 0.940430
Step 200, Loss: -1118.959961, Acc: 0.955078
Step 250, Loss: -1403.420044, Acc: 0.958984
Step 300, Loss: -1684.959961, Acc: 0.961914
Step 350, Loss: -1964.099976, Acc: 0.961914
Step 400, Loss: -2236.679932, Acc: 0.971680
Step 450, Loss: -2511.560059, Acc: 0.964844
Step 500, Loss: -2782.419922, Acc: 0.969727
Step 550, Loss: -3045.379883, Acc: 0.965820
Step 600, Loss: -3310.080078, Acc: 0.974609
Step 650, Loss: -3568.800049, Acc: 0.972656
Step 700, Loss: -3827.020020, Acc: 0.977539
Step 750, Loss: -4081.060059, Acc: 0.981445
Step 800, Loss: -4331.720215, Acc: 0.971680
Step 850, Loss: -4576.040039, Acc: 0.981445
Step 900, Loss: -4818.740234, Acc: 0.982422
Step 950, Loss: -5061.339844, Acc: 0.976562
Step 1000, Loss: -5297.240234, Acc: 0.985352
Step 1050, Loss: -5529.259766, Acc: 0.973633
Step 1100, Loss: -5757.000000, Acc: 0.982422
Step 1150, Loss: -5988.939941, Acc: 0.984375
Step 1200, Loss: -6210.520020, Acc: 0.978516
Step 1250, Loss: -6429.000000, Acc: 0.986328
Step 1300, Loss: -6646.479980, Acc: 0.988281
Step 1350, Loss: -6861.580078, Acc: 0.987305
Step 1400, Loss: -7070.700195, Acc: 0.988281
Step 1450, Loss: -7282.620117, Acc: 0.985352
Step 1500, Loss: -7488.779785, Acc: 0.982422
Step 1550, Loss: -7688.120117, Acc: 0.984375
Step 1600, Loss: -7887.859863, Acc: 0.990234
Step 1650, Loss: -8085.560059, Acc: 0.991211
Step 1700, Loss: -8279.139648, Acc: 0.984375
Step 1750, Loss: -8469.519531, Acc: 0.997070
Step 1800, Loss: -8659.879883, Acc: 0.990234
Step 1850, Loss: -8843.639648, Acc: 0.985352
Step 1900, Loss: -9023.559570, Acc: 0.993164
Step 1950, Loss: -9203.799805, Acc: 0.993164
Step 2000, Loss: -9381.019531, Acc: 0.994141
Step 2050, Loss: -9556.320312, Acc: 0.994141
Step 2100, Loss: -9725.719727, Acc: 0.988281
Step 2150, Loss: -9875.740234, Acc: 0.997070
Step 2200, Loss: -9970.719727, Acc: 0.994141
Step 2250, Loss: -9998.120117, Acc: 0.995117
Step 2300, Loss: -10001.000000, Acc: 0.992188
Step 2350, Loss: -10001.000000, Acc: 0.988281
Step 2400, Loss: -10001.000000, Acc: 0.997070
Step 2450, Loss: -10001.000000, Acc: 0.995117
Step 2500, Loss: -10001.000000, Acc: 0.996094
Step 2550, Loss: -10001.000000, Acc: 0.994141
Step 2600, Loss: -10001.000000, Acc: 0.992188
Step 2650, Loss: -10001.000000, Acc: 0.993164
Step 2700, Loss: -10001.000000, Acc: 0.998047
Step 2750, Loss: -10001.000000, Acc: 0.994141
Step 2800, Loss: -10001.000000, Acc: 0.997070
Step 2850, Loss: -10001.000000, Acc: 0.993164
Step 2900, Loss: -10001.000000, Acc: 0.994141
Step 2950, Loss: -10001.000000, Acc: 0.996094
Step 3000, Loss: -10001.000000, Acc: 0.998047
Step 3050, Loss: -10001.000000, Acc: 0.991211
Step 3100, Loss: -10001.000000, Acc: 0.996094
Step 3150, Loss: -10001.000000, Acc: 0.994141
Step 3200, Loss: -10001.000000, Acc: 0.994141
Step 3250, Loss: -10001.000000, Acc: 0.995117
Step 3300, Loss: -10001.000000, Acc: 0.995117
Step 3350, Loss: -10001.000000, Acc: 0.992188
Step 3400, Loss: -10001.000000, Acc: 0.994141
Step 3450, Loss: -10001.000000, Acc: 0.994141
Step 3500, Loss: -10001.000000, Acc: 0.994141
Step 3550, Loss: -10001.000000, Acc: 0.991211
Step 3600, Loss: -10001.000000, Acc: 0.994141
Step 3650, Loss: -10001.000000, Acc: 0.994141
Step 3700, Loss: -10001.000000, Acc: 0.991211
Step 3750, Loss: -10001.000000, Acc: 0.994141
Step 3800, Loss: -10001.000000, Acc: 0.992188
Step 3850, Loss: -10001.000000, Acc: 0.995117
Step 3900, Loss: -10001.000000, Acc: 0.991211
Step 3950, Loss: -10001.000000, Acc: 0.990234
Step 4000, Loss: -10001.000000, Acc: 0.991211
Step 4050, Loss: -10001.000000, Acc: 0.988281
Step 4100, Loss: -10001.000000, Acc: 0.995117
Step 4150, Loss: -10001.000000, Acc: 0.991211
Step 4200, Loss: -10001.000000, Acc: 0.993164
Step 4250, Loss: -10001.000000, Acc: 0.992188
Step 4300, Loss: -10001.000000, Acc: 0.996094
Step 4350, Loss: -10001.000000, Acc: 0.995117
Step 4400, Loss: -10001.000000, Acc: 0.994141
Step 4450, Loss: -10001.000000, Acc: 0.992188
Step 4500, Loss: -10001.000000, Acc: 0.992188
Step 4550, Loss: -10001.000000, Acc: 0.996094
Step 4600, Loss: -10001.000000, Acc: 0.997070
Step 4650, Loss: -10001.000000, Acc: 0.993164
Step 4700, Loss: -10001.000000, Acc: 0.997070
Step 4750, Loss: -10001.000000, Acc: 0.990234
Step 4800, Loss: -10001.000000, Acc: 0.993164
Step 4850, Loss: -10001.000000, Acc: 0.992188
Step 4900, Loss: -10001.000000, Acc: 0.994141
Step 4950, Loss: -10001.000000, Acc: 0.992188
Step 5000, Loss: -10001.000000, Acc: 0.993164
Test Accuracy: 0.963emm速度和正确率比卷积差多了
对于随机森林的通俗理解 – CSDN博客 blog.csdn.net
图源网络
森林里有100颗分类的树，每棵树把X分进十个结点中的一个结点。看哪一个结点的结果多，森林的结果就是这个结点对应的标签。
至于具体怎么种树怎么剪枝，TensorFlow已经封装好了。
- 从784个像素点随机抽出一些像素，考察它们每一个像素对于结果的离散程度（即纯度）的作用，越强越好（即分类越有效），改变权值，使得树最优化。
- 建立不同的100颗树，组成森林。
