Use Text as feature column in Tensorflows existing Estimator - tensorflow

Question

I try to build an Classification with the existing estimator to predict if an article will be sold or not.
I tried to use a linearClassifier, because I'm a beginner in Tensorflow and Pyhton.
I have a dataset with price, category and size, which is perfect for numeric or category feature columns. But I also have a description of the article, only 3-6 words per article and around 6500 different words as per my analysis.
I tried to use shared embed, with one category column per word, but this not work. And when I add all 6500 columns directly to the model it is very slow.
What is the best way and easiest way to handle the description? At best with code example. The word order doesn't matter, but for example if it's from a brand it will sell better than noname.
Many thanks for your answers
Edit: I tried with this post Tensorflow pad sequence feature column
But I now have the problem that tf.data.Dataset.from_tensor_slices((dict(dataframe), labels)) don't work
import os
import sys
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from IPython.display import clear_output
from six.moves import urllib
import tensorflow.compat.v2.feature_column as fc
from sklearn.feature_extraction.text import CountVectorizer
import tensorflow_hub as hub
from sklearn.model_selection import train_test_split
from tensorflow.python.framework.ops import disable_eager_execution
import itertools
import tensorflow as tf
from tensorflow.keras.preprocessing.text import Tokenizer
from tensorflow.keras.preprocessing.sequence import pad_sequences
from tensorflow.keras.preprocessing.text import text_to_word_sequence
dfall = pd.read_csv('./articles.csv')
# Build vacabulary
vocab_size = 6203
oov_tok = '<OOV>'
sentences = dfall['description'].to_list()
tokenizer = Tokenizer(num_words = vocab_size, oov_token="<OOV>")
tokenizer.fit_on_texts(sentences)
word_index = tokenizer.word_index
# if word_index shorter then default value of vocab_size we'll save actual size
vocab_size=len(word_index)
print("vocab_size = word_index = ",len(word_index))
# Split sentensec on tokens. here token = word
# text_to_word_sequence() has good default filter for
# charachters include basic punctuation, tabs, and newlines
dfall['description'] = dfall['description'].apply(text_to_word_sequence)
max_length = 9
# paddind and trancating setnences
# do that directly with strings without using tokenizer.texts_to_sequences()
# the feature_colunm will convert strings into numbers
dfall['description']=dfall['description'].apply(lambda x, N=max_length: (x + N * [''])[:N])
dfall['description']=dfall['description'].apply(lambda x, N=max_length: x[:N])
#dfall['description']=dfall['description'].apply(np.asarray)
dfall.head()
# Define method to create tf.data dataset from Pandas Dataframe
def df_to_dataset(dataframe, label_column, shuffle=True, batch_size=32):
dataframe = dataframe.copy()
#labels = dataframe.pop(label_column)
labels = dataframe[label_column]
ds = tf.data.Dataset.from_tensor_slices((dict(dataframe), labels))
if shuffle:
ds = ds.shuffle(buffer_size=len(dataframe))
ds = ds.batch(batch_size)
return ds
# Split dataframe into train and validation sets
train_df, val_df = train_test_split(dfall, test_size=0.2)
print(len(train_df), 'train examples')
print(len(val_df), 'validation examples')
batch_size = 32
ds = df_to_dataset(dfall, 'sold',shuffle=False,batch_size=batch_size)
train_ds = df_to_dataset(train_df, 'sold', shuffle=False, batch_size=batch_size)
val_ds = df_to_dataset(val_df, 'sold', shuffle=False, batch_size=batch_size)
# and small batch for demo
example_batch = next(iter(ds))[0]
example_batch
# Helper methods to print exxample outputs of for defined feature_column
def demo(feature_column):
feature_layer = tf.keras.layers.DenseFeatures(feature_column)
print(feature_layer(example_batch).numpy())
def seqdemo(feature_column):
sequence_feature_layer = tf.keras.experimental.SequenceFeatures(feature_column)
print(sequence_feature_layer(example_batch))
dfall.head() is
sold description category_id size_id gender price host_id lat long year month
0 1 [dünne, jacke, gepunktet, , , , , , ] 9 25 f 3.5 1 48.21534 11.29949 2019 3
1 1 [kleid, pudel, dunkelblau, gepunktet, , , , , ] 9 25 f 4.0 1 48.21534 11.29949 2019 3
2 0 [kleid, rosa, hum, hund, katze, , , , ] 9 24 f 4.0 1 48.21534 11.29949 2019 3
3 1 [kleid, hum, blau, elsa, und, anna, , , ] 9 24 f 4.0 1 48.21534 11.29949 2019 3
4 0 [kleid, blue, seven, lachsfarben, , , , , ] 9 23 f 4.5 1 48.21534 11.29949 2019 3
The result is
vocab_size = word_index = 6203
12482 train examples
3121 validation examples
---------------------------------------------------------------------------
TypeError Traceback (most recent call last)
c:\users\nibur\appdata\local\programs\python\python38\lib\site-packages\tensorflow\python\data\util\structure.py in normalize_element(element)
92 try:
---> 93 spec = type_spec_from_value(t, use_fallback=False)
94 except TypeError:
c:\users\nibur\appdata\local\programs\python\python38\lib\site-packages\tensorflow\python\data\util\structure.py in type_spec_from_value(element, use_fallback)
464
--> 465 raise TypeError("Could not build a TypeSpec for %r with type %s" %
466 (element, type(element).__name__))
TypeError: Could not build a TypeSpec for 0 [dünne, jacke, gepunktet, , , , , , ]
1 [kleid, pudel, dunkelblau, gepunktet, , , , , ]
2 [kleid, rosa, hum, hund, katze, , , , ]
3 [kleid, hum, blau, elsa, und, anna, , , ]
4 [kleid, blue, seven, lachsfarben, , , , , ]
...
15598 [gartenschuhe, pink, , , , , , , ]
15599 [sandalen, grau, blume, superfit, , , , , ]
15600 [turnschuhe, converse, grau, , , , , , ]
15601 [strickjacke, rosa, , , , , , , ]
15602 [bikinihose, schmetterling, , , , , , , ]
Name: description, Length: 15603, dtype: object with type Series
During handling of the above exception, another exception occurred:
ValueError Traceback (most recent call last)
<ipython-input-1-420304a651bd> in <module>
71
72 batch_size = 32
---> 73 ds = df_to_dataset(dfall, 'sold',shuffle=False,batch_size=batch_size)
74
75 train_ds = df_to_dataset(train_df, 'sold', shuffle=False, batch_size=batch_size)
<ipython-input-1-420304a651bd> in df_to_dataset(dataframe, label_column, shuffle, batch_size)
58 labels = dataframe[label_column]
59
---> 60 ds = tf.data.Dataset.from_tensor_slices((dict(dataframe), labels))
61 if shuffle:
62 ds = ds.shuffle(buffer_size=len(dataframe))
c:\users\nibur\appdata\local\programs\python\python38\lib\site-packages\tensorflow\python\data\ops\dataset_ops.py in from_tensor_slices(tensors)
638 Dataset: A `Dataset`.
639 """
--> 640 return TensorSliceDataset(tensors)
641
642 class _GeneratorState(object):
c:\users\nibur\appdata\local\programs\python\python38\lib\site-packages\tensorflow\python\data\ops\dataset_ops.py in __init__(self, element)
2856 def __init__(self, element):
2857 """See `Dataset.from_tensor_slices()` for details."""
-> 2858 element = structure.normalize_element(element)
2859 batched_spec = structure.type_spec_from_value(element)
2860 self._tensors = structure.to_batched_tensor_list(batched_spec, element)
c:\users\nibur\appdata\local\programs\python\python38\lib\site-packages\tensorflow\python\data\util\structure.py in normalize_element(element)
96 # the value. As a fallback try converting the value to a tensor.
97 normalized_components.append(
---> 98 ops.convert_to_tensor(t, name="component_%d" % i))
99 else:
100 if isinstance(spec, sparse_tensor.SparseTensorSpec):
c:\users\nibur\appdata\local\programs\python\python38\lib\site-packages\tensorflow\python\framework\ops.py in convert_to_tensor(value, dtype, name, as_ref, preferred_dtype, dtype_hint, ctx, accepted_result_types)
1339
1340 if ret is None:
-> 1341 ret = conversion_func(value, dtype=dtype, name=name, as_ref=as_ref)
1342
1343 if ret is NotImplemented:
c:\users\nibur\appdata\local\programs\python\python38\lib\site-packages\tensorflow\python\framework\constant_op.py in _constant_tensor_conversion_function(v, dtype, name, as_ref)
319 as_ref=False):
320 _ = as_ref
--> 321 return constant(v, dtype=dtype, name=name)
322
323
c:\users\nibur\appdata\local\programs\python\python38\lib\site-packages\tensorflow\python\framework\constant_op.py in constant(value, dtype, shape, name)
259 ValueError: if called on a symbolic tensor.
260 """
--> 261 return _constant_impl(value, dtype, shape, name, verify_shape=False,
262 allow_broadcast=True)
263
c:\users\nibur\appdata\local\programs\python\python38\lib\site-packages\tensorflow\python\framework\constant_op.py in _constant_impl(value, dtype, shape, name, verify_shape, allow_broadcast)
268 ctx = context.context()
269 if ctx.executing_eagerly():
--> 270 t = convert_to_eager_tensor(value, ctx, dtype)
271 if shape is None:
272 return t
c:\users\nibur\appdata\local\programs\python\python38\lib\site-packages\tensorflow\python\framework\constant_op.py in convert_to_eager_tensor(value, ctx, dtype)
94 dtype = dtypes.as_dtype(dtype).as_datatype_enum
95 ctx.ensure_initialized()
---> 96 return ops.EagerTensor(value, ctx.device_name, dtype)
97
98
ValueError: Failed to convert a NumPy array to a Tensor (Unsupported object type list).
I already tried to use
dfall['description']=dfall['description'].apply(np.asarray)
but then I got
ValueError: Failed to convert a NumPy array to a Tensor (Unsupported object type numpy.ndarray).
For all have same problem the solution is
tf.data.Dataset.from_tensor_slices((dataframe .to_dict(orient='list'), labels))

Answer 1

Unless there is a good reason to use Tensorflow, I would advise to start with a simple model first. Use scikit-learn and follow their tutorial on working with text data. This will show you techniques like Bag of words (BoW) embeddings or TF-IDF embeddings.
For your particular problem, one thing really interesting to try is the following: you embed your article description using BoW or TF-IDF, and you embed the rest of your features as you would for regular tabular data. And then you concatenate the embeddings and feed that to a linear classifier in scikit-learn.