Building an SVM with Tensorflow

问题:

I currently have two numpy arrays:

• X - (157, 128) - 157 sets of 128 features
• Y - (157) - classifications of the feature sets

This is the code I have written to attempt to build a linear classification model of these features.

First of all I adapted the arrays to a Tensorflow dataset:

train_input_fn = tf.estimator.inputs.numpy_input_fn(
    x={"x": X},
    y=Y,
    num_epochs=None,
    shuffle=True)

I then tried to fit an SVM model:

svm = tf.contrib.learn.SVM(
    example_id_column='example_id', # not sure why this is necessary
    feature_columns=tf.contrib.learn.infer_real_valued_columns_from_input(X), # create feature columns (not sure why this is necessary) 
    l2_regularization=0.1)

svm.fit(input_fn=train_input_fn, steps=10)

But this just returns the error:

WARNING:tensorflow:float64 is not supported by many models, consider casting to float32.
WARNING:tensorflow:Using temporary folder as model directory: /tmp/tmpf1mwlR
WARNING:tensorflow:tf.variable_op_scope(values, name, default_name) is deprecated, use tf.variable_scope(name, default_name, values)
Traceback (most recent call last):
  File "/var/www/idmy.team/python/train/classifier.py", line 59, in <module>
    svm.fit(input_fn=train_input_fn, steps=10)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/util/deprecation.py", line 316, in new_func
    return func(*args, **kwargs)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/contrib/learn/python/learn/estimators/estimator.py", line 480, in fit
    loss = self._train_model(input_fn=input_fn, hooks=hooks)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/contrib/learn/python/learn/estimators/estimator.py", line 985, in _train_model
    model_fn_ops = self._get_train_ops(features, labels)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/contrib/learn/python/learn/estimators/estimator.py", line 1201, in _get_train_ops
    return self._call_model_fn(features, labels, model_fn_lib.ModeKeys.TRAIN)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/contrib/learn/python/learn/estimators/estimator.py", line 1165, in _call_model_fn
    model_fn_results = self._model_fn(features, labels, **kwargs)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/contrib/learn/python/learn/estimators/linear.py", line 244, in sdca_model_fn
    features.update(layers.transform_features(features, feature_columns))
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/contrib/layers/python/layers/feature_column_ops.py", line 656, in transform_features
    transformer.transform(column)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/contrib/layers/python/layers/feature_column_ops.py", line 847, in transform
    feature_column.insert_transformed_feature(self._columns_to_tensors)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/contrib/layers/python/layers/feature_column.py", line 1816, in insert_transformed_feature
    input_tensor = self._normalized_input_tensor(columns_to_tensors[self.name])
KeyError: ''

What am I doing wrong?

回答1:

Here's an SVM usage example which does not throw an error:

import numpy
import tensorflow as tf

X = numpy.zeros([157, 128])
Y = numpy.zeros([157], dtype=numpy.int32)
example_id = numpy.array(['%d' % i for i in range(len(Y))])

x_column_name = 'x'
example_id_column_name = 'example_id'

train_input_fn = tf.estimator.inputs.numpy_input_fn(
    x={x_column_name: X, example_id_column_name: example_id},
    y=Y,
    num_epochs=None,
    shuffle=True)

svm = tf.contrib.learn.SVM(
    example_id_column=example_id_column_name,
    feature_columns=(tf.contrib.layers.real_valued_column(
        column_name=x_column_name, dimension=128),),
    l2_regularization=0.1)

svm.fit(input_fn=train_input_fn, steps=10)

Examples passed to the SVM Estimator need string IDs. You can probably substitute back infer_real_valued_columns_from_input, but you would need to pass it a dictionary so it picks up the right name for the column. In this case it's conceptually simpler to just construct the feature column yourself.

回答2:

• The key self.name is not present in column_to_tensors dictionary that's what the error says and the value of self.name is an empty string
• I think you messed up in giving the arguments to tf.estimator.inputs.numpy_input_fn

• The solution might be changing train_input_fn line to

  train_input_fn = tf.estimator.inputs.numpy_input_fn(x=X,
                                                      y=Y,
                                                      num_epochs=None,
                                                      shuffle=True)
  

• I think the x argument has to be a numpy array and you are giving it a dictionary

• I will stick to their tutorial and do not do any fancy

  real_feature_column = real_valued_column(...)
  sparse_feature_column = sparse_column_with_hash_bucket(...)
  
  estimator = SVM(
      example_id_column='example_id',
      feature_columns=[real_feature_column, sparse_feature_column],
      l2_regularization=10.0)
  
  # Input builders
  def input_fn_train: # returns x, y
      ...
  def input_fn_eval: # returns x, y
      ...
  
  estimator.fit(input_fn=input_fn_train)
  estimator.evaluate(input_fn=input_fn_eval)
  estimator.predict(x=x)
  

===============UPDATED==============

• My updated answer will be specific to your error
• As the error says self.name is an empty string and that empty string is not present in your dictionary that you are passing to infer_real_valued_columns_from_input that creates _RealValuedColumn object
• So What I found by debugging the error is that the tf.contrib.learn.infer_real_valued_columns_from_input(X) the X that you pass has to be a dictionary so that the self.name of _RealValuedColumn object is initialized by the key of the dictionary that you pass

• So this is what I did

  import tensorflow as tf
  import numpy as np
  
  X = np.array([[1], [0], [0], [1]])
  Y = np.array([[1], [0], [0], [1]])
  
  dic = {"x": X}
  
  train_input_fn = tf.estimator.inputs.numpy_input_fn(
      x=dic,
      y=Y,
      num_epochs=None,
      shuffle=True)
  
  svm = tf.contrib.learn.SVM(example_id_column='x', feature_columns=tf.contrib.learn.infer_real_valued_columns_from_input(dic), l2_regularization=0.1)
  
  svm.fit(input_fn=train_input_fn, steps=10)
  

• Now this removes the above error an but it gives a new error TypeError: Input 'input' of 'SdcaFprint' Op has type int64 that does not match expected type of string.

• Hopefully you remove your down-vote