Gfyuki

I'm brand new to machine learning (having just completed the google machine learning crash course) and thought it would be good to try my hand at a Kaggle competition as a good starter to some real problem solving. I'm using tensorflow and Python 3, all up to date (the kaggle online jupyter notebook)

The data is formatted in a dataframe like below

|Identity | Cuisine | Ingredients                |

|---------|---------|----------------------------|

|1        | italian | [beans, milk,..., tomatoes]|

|2        | indian  | [chicken, curry leaf,...]  |

I have made a vocabulary list generator to create a vocabulary set, and replace instances of those words in the ingredients array with the index of the ingredient in the vocabulary set, so my original data looks like below.

|Identity | Cuisine | Ingredients |

|---------|---------|-------------|

|1        | italian |[0, 1,..., 4]|

|2        | indian  |[5, 6,...]   |

I seperate the labels (cuisine) and the features (ingredients) into 2 seperate dataframes for ease, and I am using a tf.feature_column.categorical_column_with_vocabulary_list and subsequent tf.feature_column.indicator_column for the ingredients array.

I now however have an issue with my model not being able to read the ingredients column, and get the error

TypeError: Expected binary or unicode string, got [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]

my input function is as follows

def input_fn(features,labels,batch_size,num_epochs=None,shuffle=True):

    ds = Dataset.from_tensor_slices((features,labels))

    ds = ds.batch(batch_size).repeat(num_epochs)



    if shuffle:

        ds = ds.shuffle(10000)



    feature_batch, label_batch = ds.make_one_shot_iterator().get_next()

    return feature_batch, label_batch

which is fed into a simple function as below

training_func = lambda: input_fn(training_example,training_target,batch_size)

validati_func = lambda: input_fn(validation_example,validation_target,batch_size)



optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)

optimizer = tf.contrib.estimator.clip_gradients_by_norm(optimizer, 5.0)



classifier.train(

    input_fn=training_func,

    steps=steps_per_period

)

My urgent question is how do I fix this TypeError

In addition I also want to know if there a best practice for handling this format of data? (and if there is any built-in functionality to handle this)

I'm not completely familiar with TF API, but here's what I think is happening.

The library tells you that it can handle a binary column or a string. But you have all the ingredients listed in a single column. So the integer conversion of ingredient label is not helping.

You can instead create one column per possible list of ingredient and setting it to 1 if that ingredient is present or absent. For example, Italian cuisine will have column for tomatoes or garlic set to 1 for many records.

You can read more about get_dummies function in pandas library. If the original ingredient list comes in form of text, you can read up more about text feature extraction / bag of words APIs in scikit-learn libary.