How to train machine learning with strings?

Question

I have made input data for machine learning as csv file . The data are 2D arrays input and label Example

[[55:32:1:23:41:243:255:11:182:192:231:201],"play"]

[[23:222:225],"talk"]

[[23:132:215:111:29:192],"talk"]

[55:32:1:23:41:243:255:11:182:192:231:201]	play
[23:222:225]	talk

I tried to train using the follwing code

import tensorflow as tf
import numpy as np
np.set_printoptions(precision=3, suppress=True)
import pandas as pd
from tensorflow.keras import layers
from tensorflow.keras.layers.experimental import preprocessing
import io

data = pd.read_csv('./newTest4.csv',  header=None)
data_features=data.copy()
data_labels=data_features.pop(0)
data_features=np.array(data_features)
data_labels=np.array(data_labels)
data_labels 


data_model=tf.keras.Sequential ([
layers.Dense(64),
layers.Dense(1)
])
data_model.compile(loss=tf.losses.MeanSquaredError(),optimizer=tf.optimizers.Adam())


data_model.fit(data_features,data_labels,epochs=100)

But the output was

UnimplementedError:  Cast string to float is not supported
     [[node mean_squared_error/Cast (defined at <ipython-input-18-ce25e735eaa4>:1) ]] [Op:__inference_train_function_1561]
Function call stack:
train_function

Answer 1

You need a way that the model can predict the output. If you have a fixed amount of strings that you want to predict, you have to map each unique string to a binary variable.

An example is a 2-dimensional vector where the first dimension represents "play" and the second dimension represents "talk".

Your data then looks like this:

[[55:32:1:23:41:243:255:11:182:192:231:201],[1,0]] # "play", no "talk"

[[23:222:225], [0,1]] # no "play", "talk"

Now, the model can learn to predict whether the output is [1,0] (play) or [0,1] (talk).

This representation is called one-hot encoding, you can read about it in this blogpost!