我正在尝试在包含一些分类变量的数据集上使用 TensorFlow。我已经用虚拟数据对它们进行了编码，但看起来它会造成麻烦，并且 TF 抱怨数据集不密集。
或者错误的原因完全不同？
我正在尝试运行一个简单的神经网络模型，其中包含 1 个具有随机梯度的隐藏层。当输入是数值变量（来自 MNIST 的数字图像）时，代码可以正常工作
--------------------------------------------------------------------------
ValueError Traceback (most recent call last) <ipython-input-473-7517101e1879> in <module>()
37 return(test_acc,round(l,5))
38
---> 39 define_batch(0.005)
40 run_batch()
<ipython-input-472-48b4e30f8e9e> in define_batch(beta)
11 shape=(batch_size, num_var))
12 tf_train_labels = tf.placeholder(tf.float32, shape=(batch_size, num_labels))
---> 13 tf_valid_dataset = tf.constant(valid_dataset)
14 tf_test_dataset = tf.constant(test_dataset)
15
/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/tensorflow/python/ops/constant_op.pyc in constant(value, dtype, shape, name)
159 tensor_value = attr_value_pb2.AttrValue()
160 tensor_value.tensor.CopyFrom(
--> 161 tensor_util.make_tensor_proto(value, dtype=dtype, shape=shape))
162 dtype_value = attr_value_pb2.AttrValue(type=tensor_value.tensor.dtype)
163 const_tensor = g.create_op(
/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/tensorflow/python/framework/tensor_util.pyc in make_tensor_proto(values, dtype, shape)
320 nparray = np.array(values, dtype=np_dt)
321 if list(nparray.shape) != _GetDenseDimensions(values):
--> 322 raise ValueError("Argument must be a dense tensor: %s" % values)
323 # python/numpy default float type is float64. We prefer float32 instead.
324 if (nparray.dtype == np.float64) and dtype is None:
ValueError: Argument must be a dense tensor: Tuesday Wednesday Thursday Friday Saturday Sunday CENTRAL \ 736114
0.0 0.0 0.0 0.0 1.0 0.0 0.0 437148 0.0 0.0 1.0 0.0 0.0 0.0 0.0 605041 0.0 0.0 0.0 0.0 0.0 0.0 0.0 444608 0.0 0.0 0.0 0.0 1.0 0.0 0.0 695549 0.0 0.0 0.0 0.0 1.0 0.0 0.0 662807 0.0 0.0 0.0 1.0 0.0 0.0 0.0 238635 0.0 0.0 0.0 0.0 0.0 1.0 0.0 549524 0.0 0.0 0.0 1.0 0.0 0.0 0.0 705478 1.0 0.0 0.0 0.0 0.0 0.0 0.0 557716 0.0 0.0 0.0 1.0 0.0 0.0 0.0 41808 0.0 0.0 0.0 0.0 0.0 1.0 0.0 227235 1.0 0.0 0.0 0.0 0.0 0.0 0.0 848719 0.0 0.0 0.0 0.0 0.0 0.0 0.0 731202 0.0 0.0 0.0 0.0 1.0 0.0 0.0 467516 1.0 0.0 0.0 0.0 0.0 0.0 1.0
这是代码摘录
# Adding regularization to the 1 hidden layer network
def define_batch(beta):
batch_size = 128
num_RELU =256
graph1 = tf.Graph()
with graph1.as_default():
# Input data. For the training data, we use a placeholder that will be fed
# at run time with a training minibatch.
tf_train_dataset = tf.placeholder(tf.float32,
shape=(batch_size, num_var))
tf_train_labels = tf.placeholder(tf.float32, shape=(batch_size, num_labels))
tf_valid_dataset = tf.constant(valid_dataset)
tf_test_dataset = tf.constant(test_dataset)
# Variables.
weights_RELU = tf.Variable(
tf.truncated_normal([num_var, num_RELU]))
biases_RELU = tf.Variable(tf.zeros([num_RELU]))
weights_layer1 = tf.Variable(
tf.truncated_normal([num_RELU, num_labels]))
biases_layer1 = tf.Variable(tf.zeros([num_labels]))
# Training computation.
logits_RELU = tf.matmul(tf_train_dataset, weights_RELU) + biases_RELU
RELU_vec = tf.nn.relu(logits_RELU)
logits_layer = tf.matmul(RELU_vec, weights_layer1) + biases_layer1
# loss = tf.reduce_mean(
# tf.nn.softmax_cross_entropy_with_logits(logits_layer, tf_train_labels))
cross_entropy = tf.nn.softmax_cross_entropy_with_logits(logits_layer, tf_train_labels,name="cross_entropy")
l2reg = tf.reduce_sum(tf.square(weights_RELU))+tf.reduce_sum(tf.square(weights_layer1))
beta = 0.005
loss = tf.reduce_mean(cross_entropy+beta*l2reg)
# Optimizer.
optimizer = tf.train.GradientDescentOptimizer(0.3).minimize(loss)
# Predictions for the training, validation, and test data.
train_prediction = tf.nn.softmax(logits_layer)
valid_prediction = tf.nn.softmax(
tf.matmul(tf.nn.relu((tf.matmul(tf_valid_dataset, weights_RELU) + biases_RELU)),weights_layer1)+biases_layer1)
test_prediction =tf.nn.softmax(
tf.matmul(tf.nn.relu((tf.matmul(tf_test_dataset, weights_RELU) + biases_RELU)),weights_layer1)+biases_layer1)
import datetime
startTime = datetime.datetime.now()
num_steps = 301 # change to 3001
def run_batch():
with tf.Session(graph=graph1) as session:
tf.initialize_all_variables().run()
print("Initialized")
for step in range(num_steps):
# Pick an offset within the training data, which has been randomized.
# Note: we could use better randomization across epochs.
offset = (step * batch_size) % (train_labels.shape[0] - batch_size)
# Generate a minibatch.
batch_data = train_dataset[offset:(offset + batch_size), :]
batch_labels = train_labels[offset:(offset + batch_size), :]
# Prepare a dictionary telling the session where to feed the minibatch.
# The key of the dictionary is the placeholder node of the graph to be fed,
# and the value is the numpy array to feed to it.
feed_dict = {tf_train_dataset : batch_data, tf_train_labels : batch_labels}
_, l, predictions, logits = session.run(
[optimizer, loss,train_prediction,logits_RELU], feed_dict=feed_dict)
if (step % 500 == 0):
print("Minibatch loss at step %d: %f" % (step, l))
print("Minibatch accuracy: %.1f%%" % accuracy(predictions, batch_labels))
print("Validation accuracy: %.1f%%" % accuracy(
valid_prediction.eval(), valid_labels))
test_acc = accuracy(test_prediction.eval(), test_labels)
print("Test accuracy: %.1f%%" % test_acc)
print('loss=%s' % l)
x = datetime.datetime.now() - startTime
print(x)
return(test_acc,round(l,5))
define_batch(0.005)
run_batch()
编辑： @gdhal 感谢您查看它
train_datasettrain_dataset.columns
Index([u'Tuesday', u'Wednesday', u'Thursday', u'Friday', u'Saturday',
u'Sunday', u'CENTRAL', u'INGLESIDE', u'MISSION', u'NORTHERN', u'PARK',
u'RICHMOND', u'SOUTHERN', u'TARAVAL', u'TENDERLOIN', u' 3H - 4H',
u' 5H - 6H', u' 7H - 8H', u' 9H - 10H', u'11H - 12H', u'13H - 14H',
u'15H - 16H', u'17H - 18H', u'19H - 20H', u'21H - 22H', u'23H - 0H',
u'Xnorm', u'Ynorm', u'Hournorm'],
dtype='object')
除了最后 3 个变量（Xnorm、Ynorm 和 Hournorm）之外，所有变量都是虚拟变量（取 0 或 1 值），它们是标准化为 [0,1] 区间的数值。
valid_datasettest_datasettrain_labelstrain_labels.describe()
count 790184
unique 39
top LARCENY/THEFT
freq 157434
Name: Category, dtype: object
valid_labelstest_labels尝试输入 numpy 数组而不是 pandas 数据框。