遵循虹膜分类演示： https://www.tensorflow.org/get_started/tflearn#construct_a_deep_neural_network_classifier https://www.tensorflow.org/get_started/monitors
我期望使用eeg数据作为行为分类的输入数据获得相同的良好准确性。但是，使用以下代码获得的准确度低于70％：
import pandas as pd
import numpy as np
import tensorflow as tf
import shutil
IRIS_TRAINING = "eeg_training2.csv"
IRIS_TEST ="eeg_test.csv"
# Define the training inputs
def get_train_inputs():
x = tf.constant(training_set.data)
y = tf.constant(training_set.target)
return x, y
# Define the test inputs
def get_test_inputs():
x = tf.constant(test_set.data)
y = tf.constant(test_set.target)
return x, y
# Needed to make sure the logging output is visible.
tf.logging.set_verbosity(tf.logging.INFO)
shutil.rmtree('./tmp',ignore_errors=True)
# Load datasets
training_set = tf.contrib.learn.datasets.base.load_csv_with_header(
filename=IRIS_TRAINING, target_dtype=np.int, features_dtype=np.float32)
test_set = tf.contrib.learn.datasets.base.load_csv_with_header(
filename=IRIS_TEST, target_dtype=np.int, features_dtype=np.float32)
print('Training set lenght:' + str(len(training_set)))
print('Testing set lenght:' + str(len(test_set)))
print('database loaded')
validation_metrics = {
"accuracy":
tf.contrib.learn.MetricSpec(
metric_fn=tf.contrib.metrics.streaming_accuracy,
prediction_key=tf.contrib.learn.PredictionKey.CLASSES),
"precision":
tf.contrib.learn.MetricSpec(
metric_fn=tf.contrib.metrics.streaming_precision,
prediction_key=tf.contrib.learn.PredictionKey.CLASSES),
"recall":
tf.contrib.learn.MetricSpec(
metric_fn=tf.contrib.metrics.streaming_recall,
prediction_key=tf.contrib.learn.PredictionKey.CLASSES)
}
validation_monitor = tf.contrib.learn.monitors.ValidationMonitor(
test_set.data,
test_set.target,
every_n_steps=50,
metrics=validation_metrics,
early_stopping_metric="loss",
early_stopping_metric_minimize=True,
early_stopping_rounds=200)
# Define the features used in the DNN model: 14 features with real value
feature_columns = [tf.contrib.layers.real_valued_column("", dimension=14)]
# Build 3 layer DNN with 25, 50, 25 units respectively.
classifier = tf.contrib.learn.DNNClassifier(feature_columns=feature_columns,
hidden_units=[25, 50, 25],
n_classes=2,
model_dir="./tmp/eeg-model",
#dropout=0.1,
config=tf.contrib.learn.RunConfig(save_checkpoints_secs=10))
# Fit model.
#classifier.fit(input_fn=get_train_inputs, steps=2000)
classifier.fit(x=training_set.data,
y=training_set.target,
steps=5000,
monitors=[validation_monitor])
# Evaluate accuracy.
accuracy_score = classifier.evaluate(input_fn=get_test_inputs,
steps=1)["accuracy"]
print("\nTest Accuracy: {0:f}".format(accuracy_score*100) + '%\n')
# Classify two new flower samples. first: 0 second: 1
def new_samples():
return np.array(
[[4289.23,4006.15,4264.62,4121.03,4327.18,4614.87,4067.69,4614.87,4196.92,4225.64,4192.31,4269.23,4595.9,4350.77],
[4291.79,4009.23,4265.64,4125.13,4330.26,4621.03,4071.79,4619.49,4193.33,4229.74,4195.38,4273.85,4601.03,4349.23],
[4293.85,4008.72,4264.62,4126.67,4333.33,4623.08,4075.9,4617.95,4193.85,4233.85,4202.05,4275.38,4607.18,4355.9],
[4295.38,4010.77,4266.67,4126.67,4333.33,4618.97,4075.9,4610.26,4193.85,4234.36,4205.13,4274.87,4605.64,4361.03],
[4294.36,4016.92,4273.33,4132.82,4335.9,4618.46,4077.95,4610.26,4196.92,4232.31,4205.64,4279.49,4609.23,4358.46],
[4289.74,4012.82,4273.33,4130.77,4333.85,4617.95,4081.03,4613.85,4195.9,4223.59,4202.05,4282.05,4609.23,4352.31],
[4286.15,4000,4265.13,4121.03,4325.64,4615.38,4078.97,4612.31,4193.85,4220.51,4194.87,4275.9,4598.97,4347.18],
[4286.15,3995.38,4263.59,4121.03,4328.21,4617.44,4074.87,4611.79,4200.51,4222.56,4191.28,4271.28,4594.36,4348.21],
[4287.18,3996.41,4271.79,4123.59,4333.33,4618.97,4074.36,4614.87,4205.13,4226.15,4194.36,4271.79,4596.41,4350.77],
[4286.67,3994.87,4273.85,4121.54,4330.77,4616.41,4075.9,4620.51,4203.08,4230.77,4197.44,4271.79,4598.46,4350.77],
[4308.21,3993.33,4260.51,4096.92,4328.72,4607.69,4060.51,4612.31,4202.56,4228.21,4191.79,4271.28,4606.67,4365.64],
[4296.92,3986.15,4249.23,4092.82,4327.69,4609.74,4064.62,4613.33,4206.67,4229.23,4192.31,4265.13,4596.41,4357.95],
[4294.36,3985.64,4245.13,4092.82,4327.69,4616.92,4066.15,4620.51,4210.77,4229.74,4196.41,4264.62,4596.41,4355.38],
[4295.38,3990.26,4247.69,4095.9,4328.21,4614.36,4059.49,4617.44,4204.1,4229.23,4197.44,4265.64,4604.1,4360],
[4300,3992.82,4254.87,4103.59,4329.74,4611.28,4056.92,4612.31,4205.13,4230.26,4198.46,4265.64,4603.08,4365.13],
[4306.67,3994.36,4263.08,4107.18,4331.28,4614.36,4061.03,4614.87,4214.87,4232.82,4203.08,4268.72,4602.05,4366.67],
[4311.79,4000,4266.67,4108.72,4334.36,4613.85,4063.08,4615.38,4213.85,4234.36,4208.21,4276.41,4610.26,4373.85],
[4313.33,4006.67,4265.64,4116.41,4335.9,4609.74,4060.51,4612.82,4205.64,4235.9,4212.31,4282.56,4617.95,4380.51],
[4314.36,4008.21,4265.64,4118.97,4336.41,4611.79,4061.03,4616.92,4212.82,4242.05,4217.95,4284.62,4619.49,4378.97],
[4317.95,4008.21,4266.15,4114.87,4337.44,4620,4066.67,4622.05,4220.51,4247.69,4218.97,4286.67,4620,4378.97]], dtype=np.float32)
predictions = list(classifier.predict(input_fn=new_samples))
print("New Samples, Class Predictions: {}\n".format(predictions))#Ground truth: {1,2}
有人尝试使用Tensorflow的DNN进行eeg分类吗？我在stackoverflow看到了一个帖子，但是他们使用的cnn不是eeg数据分类的最佳选择。我的数据来自UCI的存储库：https://archive.ics.uci.edu/ml/datasets/EEG+Eye+State#。我使用前6000个实例进行训练，最后1000个用于测试，但结果非常差。我改变了隐藏层的结构并增加了运行次数，但结果仍然很糟糕。任何有关代码的建议，以获得更高的准确性将是值得赞赏的。关于eens数据分类的Tensorflow中更好的方法的任何建议也将受到欢迎。