Machine learning is simply a computer learning from data instead of following a recipe. It's meant to mimic how people (and perhaps other animals) learn while still being grounded in mathematics.
This post is meant to get you started with a basic machine learning model.
A chatbot.
Now, we're not re-creating Alexa, Siri, Cortana, or Google Assistant but we are going to create a brand new machine learning program from scratch.
This tutorial is meant to be easy assuming you know a bit of Python Programming.
Watch the entire series that corresponds to this post.
Machine learning needs data to actually, well, learn. Machines don't yet learn like you and I do but they do learn by finding patterns in things that may seem non-obvious to you and I. We'll see a lot of that in this entire post.
Before we define our data, let's talk about the goal of this ML (machine learning) project:
To answer somewhat "random" questions with pre-defined responses.
Here's what we'll try and solve:
Scenario 1
Bill: Hi there, what time do you open tomorrow for lunch?
Bot: Our hours are 9am-10pm everyday.
Scenario 2
Karen: Can I speak to your manager?
Bot: You can contact our customer support at 555-555-555.5
Scenario 3
Wade: What type of products do you have?
Bot: We carry various food items including tacos, nachos, burritos, and salads.
Let's put this into a python format:
conversations = [
{
"customer": "Hi there, what time do you open tomorrow for lunch?",
"response": "Our hours are 9am-10pm everday."
},
{
"customer": "Can I speak to your manager?",
"response": "You can contact our customer support at 555-555-5555."
},
{
"customer": "What type of products do you have?",
"response": "We carry various food items including tacos, nachos, burritos, and salads."
}
]Without machine learning our bot would look like this (uncomment next cell to run):
# while True:
# my_input = input("What is your question?\n")
# response = None
# for convo in conversations:
# if convo['customer'] == my_input:
# response = convo['response']
# if response != None:
# print(response)
# break
# print("I don't know")
# continueRight away, you should see the huge flaws in this recipe; if a customer doesn't ask a question in a specific pre-defined way, the bot fails and ultimately really sucks.
A few examples: - What if a customer says, when do you open? What do you already know the response to be? - What if a customer says, Do you sell burgers? - What if a customer says, How do I reach you on the phone?
I'm sure you could come up with many many more examples of where this really falls apart.
So let's clean up our conversations data a bit more by adding tags that describe the initial question.
convos_one = [
{
"customer": "Hi there, what time do you open tomorrow for lunch?",
"tags": ["opening", "closing", "hours"],
},
{
"customer": "Can I speak to your manager?",
"tags": ["customer_support"],
},
{
"customer": "The food was amazing thank you!",
"tags": ["customer_support", "feedback"],
},
{
"customer": "What type of products do you have?",
"tags": ["products", "menu", "inventory", "food"],
}
]convos_two = [
{
"customer": "How late is your kitchen open?",
"tags": ["opening", "hours", "closing"],
},
{
"customer": "My order was prepared incorrectly, how can I get this fixed?",
"tags": ["customer_support"],
},
{
"customer": "What kind of meats do you have?",
"tags": ["menu", "products", "inventory", "food"],
}
]convos_three = [
{
"customer": "When does your dining room open?",
"tags": ['opening', 'hours'],
},
{
"customer": "When do you open for dinner?",
"tags": ['opening', 'hours', "closing"],
},
{
"customer": "How do I contact you?",
"tags": ["contact", "customer_support"]
}
]Do you see a trend happening here? It's really easy to come up with all kinds of questions for a restaurant bot. It's also easy to see how challenging this would be to try and hard-code conditions to handle all the kinds of queries/questions customers could have.
I'm sure you've heard you need a LOT of data for machine learning. I'll just add one thing to that, you need a lot of data to have awe-inspiring machine learning projects. A simple bot for a mom-and-pop store down the street doesn't need awe-inspiring just yet. They need simple, approachable, easy to explain. That's exactly what this is. It's not a black box of millions of lines of data points. It's like 20 questions with made up on the spot tags.
In so many ways, machine learning today (in the 2020s) is like the internet of the 1990s. People have heard about it and "sort of get it" and feel like it's just this magical gimmick that only super nerds know how to do. Ha. Super nerds.
Now that we have our starting data, let's prepare for machine learning.
First, let's combine all conversations:
dataset = convos_one + convos_two + convos_three
dataset[{'customer': 'Hi there, what time do you open tomorrow for lunch?',
'tags': ['opening', 'closing', 'hours']},
{'customer': 'Can I speak to your manager?', 'tags': ['customer_support']},
{'customer': 'The food was amazing thank you!',
'tags': ['customer_support', 'feedback']},
{'customer': 'What type of products do you have?',
'tags': ['products', 'menu', 'inventory', 'food']},
{'customer': 'How late is your kitchen open?',
'tags': ['opening', 'hours', 'closing']},
{'customer': 'My order was prepared incorrectly, how can I get this fixed?',
'tags': ['customer_support']},
{'customer': 'What kind of meats do you have?',
'tags': ['menu', 'products', 'inventory', 'food']},
{'customer': 'When does your dining room open?',
'tags': ['opening', 'hours']},
{'customer': 'When do you open for dinner?',
'tags': ['opening', 'hours', 'closing']},
{'customer': 'How do I contact you?',
'tags': ['contact', 'customer_support']}]
Our conversations have the keys customer and tags. These are arbitrary names for this project and you change change them at-will. Just remember that customer equals input and tags equals output. This makes sense because in the future, we want a random customer input such as What's the menu specials today and a predicted tags output like menu or something similar.
Machine learning has all kinds of terms and acronyms that often make it a bit confusing. In general, just remember that you have some inputs and some target outputs. Here's what I mean by that:
customer: These values are really theinputvalues for our ML project. Input values are sometimes calledsource,feature,training,X,X_train/X_test/X_valid, and a few others.tags: These values are really theoutputvalues for our ML project. Output values are sometimes calledtarget,labels,y,y_train/y_test/y_valid,classes/class, and a few others.
We're using a machine learning technique known as
supervised learningwhich means we provide both theinputsandoutputsto the model. Both data points are known data that we came up with. As you know, thetags(orlabels/outputs) have been decided by a human (ie you and me) but can, eventually, be decided by a ML model itself and then verified by a human. Doing so would make the model better and better. There are many other techniques butsupervised learningis by far the most approachable for beginners.
Now that we have our data, it's time to put it into a format that works well for computers. As you may know, computers are great at numbers and not so great at text. In this case, we have to convert our text into numbers.
This is made simple by using the scikit-learn library. So let's install it below by uncommenting the cell.
# !pip install scikit-learnFirst up, let's turn our customer and tag data into 2 separate lists where the index of each item corresponds to the index of the other.
X = [customer_convo_1, customer_convo_2, ...]
y = [convo_1_tags, convo_2_tags, ...]
This is very standard practice so that X[0] is the input that corresponds to the y[0] output, X[1] is the input that corresponds to the y[1] output and so on.
inputs = [x['customer'] for x in dataset]
print(inputs)['Hi there, what time do you open tomorrow for lunch?', 'Can I speak to your manager?', 'The food was amazing thank you!', 'What type of products do you have?', 'How late is your kitchen open?', 'My order was prepared incorrectly, how can I get this fixed?', 'What kind of meats do you have?', 'When does your dining room open?', 'When do you open for dinner?', 'How do I contact you?']
outputs = [x['tags'] for x in dataset]
print(outputs)[['opening', 'closing', 'hours'], ['customer_support'], ['customer_support', 'feedback'], ['products', 'menu', 'inventory', 'food'], ['opening', 'hours', 'closing'], ['customer_support'], ['menu', 'products', 'inventory', 'food'], ['opening', 'hours'], ['opening', 'hours', 'closing'], ['contact', 'customer_support']]
assert(len(inputs) == len(outputs))If you have an
AssertionErrorabove, that means yourinputsandoutputsare not balanced. Check your data source(s) to ensure everyinputhas a correspondingoutputvalue.
Let's verify the positions of this data to show how little we actually changed the data:
idx = 4
print(inputs[idx], outputs[idx])
print(dataset[idx])How late is your kitchen open? ['opening', 'hours', 'closing']
{'customer': 'How late is your kitchen open?', 'tags': ['opening', 'hours', 'closing']}
The goal of machine learning is to produce a function that takes inputs and produces outputs (predictions). Below is, at a conceptual level, a representation of that:
def my_pred_function(inputs):
# pred
outputs = inputs * 0.39013 # this decimal represents what our model will essentially do.
return outputsNow we need to turn each inputs list and outputs list into matrices so our machine learning can do machine learning.
scikit-learn has a simple way to do this. First, let's focus on the inputs (aka customer conversations) as they are the most simple.
from sklearn.feature_extraction.text import CountVectorizer
vectorizer = CountVectorizer()
X = vectorizer.fit_transform(inputs)Technical note:
scikit-learnconverted our data into a collection of 1 dimension matrices. We need to use matrices so we can do matrix multiplication (that's how machine learning works under the hood). Innumpyspeak,Xis anarrayofarrays. If you want to see the actual vectors created, check outX.toarray()and you'll see it.
X.shape(10, 43)
X.shape is useful to describe our data.
X.shape[0] refers to the number of conversations from our final_convos list. So, X.shape[0] == len(final_convos) and X.shape[0] == len(inputs)
X.shape[1] refers to the number of words our data has. The CountVectorizer did this for us. The Machine Learning term is features related to what our data has. You can see all of the features (words minus punctuation) with:
words = vectorizer.get_feature_names()
print(words)['amazing', 'can', 'contact', 'dining', 'dinner', 'do', 'does', 'fixed', 'food', 'for', 'get', 'have', 'hi', 'how', 'incorrectly', 'is', 'kind', 'kitchen', 'late', 'lunch', 'manager', 'meats', 'my', 'of', 'open', 'order', 'prepared', 'products', 'room', 'speak', 'thank', 'the', 'there', 'this', 'time', 'to', 'tomorrow', 'type', 'was', 'what', 'when', 'you', 'your']
The vectorizer has a very limited vocabulary as you can see. Naturally, this means our ML project will always misunderstand some key conversations and that's okay. The goal for our project is to get it working first, get customers (or ourselves) using it so we can improve it with new data right away (and thus re-improve it).
len(words)43
Every one of our inputs has a list of tags, not just one tag. Let's look at what I mean:
print(inputs[0], outputs[0], isinstance(outputs[0], list))Hi there, what time do you open tomorrow for lunch? ['opening', 'closing', 'hours'] True
In machine learning, this means multi-label classification because there are multiple possible output values for each input value. This is a more challenging problem than a single label but definitely necessary for a chatbot project.
A single label dataset would look like the following:
Input: Hi there, how are you doing today?
Output: not_spam
Input: Free CELL phones just text 3ED#2
Output: spam
Notice that the output is a single str and not a list of str values. If we continued down this path, our data would always fall into 2 categories: spam or not_spam. This type of classification is called binary classification because there are only 2 possible classes for the prediction to be.
In our project, our input values can fall into multiple class items, 1 class, or no class item at all. (Remember, class = output tag = label)
from sklearn.preprocessing import MultiLabelBinarizer
mlb = MultiLabelBinarizer()
y = mlb.fit_transform(outputs)You might consider running fit_transform on a CountVectorizer like we did with our training data (aka inputs) but that doesn't work on multi-label classification. For that, we need MultiLabelBinarizer.
mlb.classes_array(['closing', 'contact', 'customer_support', 'feedback', 'food',
'hours', 'inventory', 'menu', 'opening', 'products'], dtype=object)
Calling mlb.classes_ gives us the exact order of how our classes are defined in y. So y[0] corresponds to outputs[0] but in numbers instead of words. It's pretty cool. To see this technically, run the following code:
print(y[0])
# map to classes with `zip`
y0_mapped_to_classes = dict(zip(mlb.classes_, y[0]))
print(y0_mapped_to_classes)
Then compare:
sorted(outputs[0]) == sorted([k for k,v in y0_mapped_to_classes.items() if v == 1])
yarray([[1, 0, 0, 0, 0, 1, 0, 0, 1, 0],
[0, 0, 1, 0, 0, 0, 0, 0, 0, 0],
[0, 0, 1, 1, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 1, 0, 1, 1, 0, 1],
[1, 0, 0, 0, 0, 1, 0, 0, 1, 0],
[0, 0, 1, 0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 1, 0, 1, 1, 0, 1],
[0, 0, 0, 0, 0, 1, 0, 0, 1, 0],
[1, 0, 0, 0, 0, 1, 0, 0, 1, 0],
[0, 1, 1, 0, 0, 0, 0, 0, 0, 0]])
Here we can see the matrix that is generated for us with sklearn's MultiLabelBinarizer. It's an array of one-hot arrays.
one-hot is a term that refers to the type of encoding we're using for this particular model. It's a very common practice in machine learning. Essentially turning data into
1s and0s instead of strings or any other data type.
y.shape(10, 10)
y.shape is useful to describe our data in a similar way to X.shape
y.shape[0] refers to the number of conversations from our final_convos list. So, y.shape[0] == len(final_convos) and y.shape[0] == len(outputs) and y.shape[0] == X.shape[0]
y.shape[1] refers to the unique values of all of the possible tags each conversation has; it will never repeat using the MultiLabelBinarizer.
assert y.shape[0] == X.shape[0]
assert y.shape[0] == len(inputs)If you see an AssertionError here, it's the same exact error as assert len(inputs) == len(outputs) from above. Your data is not balanced.
from sklearn.multioutput import MultiOutputClassifier
from sklearn.ensemble import RandomForestClassifier
forest = RandomForestClassifier(random_state=1)
model = MultiOutputClassifier(forest, n_jobs=-1)model.fit(X, y)MultiOutputClassifier(estimator=RandomForestClassifier(random_state=1),
n_jobs=-1)
txt = "Hi, when do you close?"
input_vector = vectorizer.transform([txt])
input_vector<1x43 sparse matrix of type '<class 'numpy.int64'>'
with 4 stored elements in Compressed Sparse Row format>
output_vector = model.predict(input_vector)
print(output_vector)[[0 0 0 0 0 0 0 0 0 0]]
preds = {}
classes = mlb.classes_
for i, val in enumerate(output_vector[0]):
preds[classes[i]] = valpreds{'closing': 0,
'contact': 0,
'customer_support': 0,
'feedback': 0,
'food': 0,
'hours': 0,
'inventory': 0,
'menu': 0,
'opening': 0,
'products': 0}
def label_predictor(txt='Hello world'):
# pred
input_vector = vectorizer.transform([txt])
output_vector = model.predict(input_vector)
preds = {}
classes = mlb.classes_
for i, val in enumerate(output_vector[0]):
preds[classes[i]] = val
return predslabel_predictor(){'closing': 0,
'contact': 0,
'customer_support': 0,
'feedback': 0,
'food': 0,
'hours': 0,
'inventory': 0,
'menu': 0,
'opening': 0,
'products': 0}
label_predictor("When do you open?"){'closing': 0,
'contact': 0,
'customer_support': 0,
'feedback': 0,
'food': 0,
'hours': 1,
'inventory': 0,
'menu': 0,
'opening': 1,
'products': 0}
label_predictor("When are you opening tomorrow?"){'closing': 0,
'contact': 0,
'customer_support': 0,
'feedback': 0,
'food': 0,
'hours': 0,
'inventory': 0,
'menu': 0,
'opening': 0,
'products': 0}
import pickle
# classes
# model
# vectorizer
model_data = {
"classes": list(mlb.classes_),
"model": model,
"vectorizer": vectorizer
}
with open("model.pkl", 'wb') as f:
pickle.dump(model_data, f)model_loaded_data = {}
with open("model.pkl", 'rb') as f:
model_loaded_data = pickle.loads(f.read())
def label_predictor_from_export(txt='Hello world',
vectorizer=None,
model=None,
classes=[],
*args,
**kwargs):
# pred
assert(vectorizer!=None)
assert(model != None)
input_vector = vectorizer.transform([txt])
output_vector = model.predict(input_vector)
assert(len(output_vector[0]) == len(classes))
preds = {}
classes = mlb.classes_
for i, val in enumerate(output_vector[0]):
preds[classes[i]] = val
return preds
label_predictor_from_export("When does your kitchen close?", **model_loaded_data){'closing': 0,
'contact': 0,
'customer_support': 0,
'feedback': 0,
'food': 0,
'hours': 1,
'inventory': 0,
'menu': 0,
'opening': 1,
'products': 0}
from sklearn.feature_extraction.text import CountVectorizer
from sklearn.preprocessing import MultiLabelBinarizer
from sklearn.multioutput import MultiOutputClassifier
from sklearn.ensemble import RandomForestClassifier
def train(dataset, train_col='customer', label_col='tags', export_path='model.pkl'):
inputs = [x[train_col] for x in dataset]
outputs = [x[label_col] for x in dataset]
assert(len(inputs) == len(outputs))
vectorizer = CountVectorizer()
X = vectorizer.fit_transform(inputs)
mlb = MultiLabelBinarizer()
y = mlb.fit_transform(outputs)
classes = list(mlb.classes_)
forest = RandomForestClassifier(random_state=1)
model = MultiOutputClassifier(forest, n_jobs=-1)
model.fit(X, y)
model_data = {
"classes": list(mlb.classes_),
"model": model,
"vectorizer": vectorizer
}
with open(export_path, 'wb') as f:
pickle.dump(model_data, f)
return export_path# datasettrain(dataset, export_path='model2.pkl')'model2.pkl'
model_loaded_data = {}
with open("model2.pkl", 'rb') as f:
model_loaded_data = pickle.loads(f.read())
label_predictor_from_export("What is your favorite menu item?", **model_loaded_data){'closing': 0,
'contact': 0,
'customer_support': 0,
'feedback': 0,
'food': 0,
'hours': 0,
'inventory': 0,
'menu': 0,
'opening': 0,
'products': 0}
!pip install pandasRequirement already satisfied: pandas in /Users/cfe/.local/share/virtualenvs/ml-hello-world-rt2goiiz/lib/python3.8/site-packages (1.0.5)
Requirement already satisfied: numpy>=1.13.3 in /Users/cfe/.local/share/virtualenvs/ml-hello-world-rt2goiiz/lib/python3.8/site-packages (from pandas) (1.19.1)
Requirement already satisfied: pytz>=2017.2 in /Users/cfe/.local/share/virtualenvs/ml-hello-world-rt2goiiz/lib/python3.8/site-packages (from pandas) (2020.1)
Requirement already satisfied: python-dateutil>=2.6.1 in /Users/cfe/.local/share/virtualenvs/ml-hello-world-rt2goiiz/lib/python3.8/site-packages (from pandas) (2.8.1)
Requirement already satisfied: six>=1.5 in /Users/cfe/.local/share/virtualenvs/ml-hello-world-rt2goiiz/lib/python3.8/site-packages (from python-dateutil>=2.6.1->pandas) (1.15.0)
import pandas as pddf = pd.DataFrame(dataset)
df.head(n=100).dataframe tbody tr th {
vertical-align: top;
}
.dataframe thead th {
text-align: right;
}
| customer | tags | |
|---|---|---|
| 0 | Hi there, what time do you open tomorrow for l... | [opening, closing, hours] |
| 1 | Can I speak to your manager? | [customer_support] |
| 2 | The food was amazing thank you! | [customer_support, feedback] |
| 3 | What type of products do you have? | [products, menu, inventory, food] |
| 4 | How late is your kitchen open? | [opening, hours, closing] |
| 5 | My order was prepared incorrectly, how can I g... | [customer_support] |
| 6 | What kind of meats do you have? | [menu, products, inventory, food] |
| 7 | When does your dining room open? | [opening, hours] |
| 8 | When do you open for dinner? | [opening, hours, closing] |
| 9 | How do I contact you? | [contact, customer_support] |
df.to_pickle("dataset.pkl")# df = pd.read_pickle("dataset.pkl")
# og_df.head()# og_df.iloc[0]['tags'][0]new_dataset = df.to_dict("records")
# print(new_dataset)# df = df.append({"customer": "Who is the manager?", "tags": ["customer_support"]}, ignore_index=True)# df.head(n=100)def append_to_df(df):
df = df.copy()
while True:
customer_input = input("What is the question?\n")
tags_input = input("Tags? Use commas to separate\n")
if tags_input != None:
tags_input = tags_input.split(",")
if not isinstance(tags_input, list):
tags_input = [tags_input]
if customer_input != None and tags_input != None:
df = df.append({"customer": customer_input, "tags": tags_input}, ignore_index=True)
tag_another = input("Tag another? Type (y) to continue or any other key to exit.")
if tag_another.lower() == "y":
continue
break
return dfnew_df = append_to_df(df)What is the question?
Who is the manager?
Tags? Use commas to separate
customer_support
Tag another? Type (y) to continue or any other key to exit.d
new_df.head(n=100).dataframe tbody tr th {
vertical-align: top;
}
.dataframe thead th {
text-align: right;
}
| customer | tags | |
|---|---|---|
| 0 | Hi there, what time do you open tomorrow for l... | [opening, closing, hours] |
| 1 | Can I speak to your manager? | [customer_support] |
| 2 | The food was amazing thank you! | [customer_support, feedback] |
| 3 | What type of products do you have? | [products, menu, inventory, food] |
| 4 | How late is your kitchen open? | [opening, hours, closing] |
| 5 | My order was prepared incorrectly, how can I g... | [customer_support] |
| 6 | What kind of meats do you have? | [menu, products, inventory, food] |
| 7 | When does your dining room open? | [opening, hours] |
| 8 | When do you open for dinner? | [opening, hours, closing] |
| 9 | How do I contact you? | [contact, customer_support] |
| 10 | Who is the manager? | [customer_support] |
new_df.to_pickle("dataset.pkl")Using fastapi
!pip install fastapi uvicorn requestsRequirement already satisfied: fastapi in /Users/cfe/.local/share/virtualenvs/ml-hello-world-rt2goiiz/lib/python3.8/site-packages (0.60.0)
Requirement already satisfied: uvicorn in /Users/cfe/.local/share/virtualenvs/ml-hello-world-rt2goiiz/lib/python3.8/site-packages (0.11.6)
Requirement already satisfied: requests in /Users/cfe/.local/share/virtualenvs/ml-hello-world-rt2goiiz/lib/python3.8/site-packages (2.24.0)
Requirement already satisfied: pydantic<2.0.0,>=0.32.2 in /Users/cfe/.local/share/virtualenvs/ml-hello-world-rt2goiiz/lib/python3.8/site-packages (from fastapi) (1.6.1)
Requirement already satisfied: starlette==0.13.4 in /Users/cfe/.local/share/virtualenvs/ml-hello-world-rt2goiiz/lib/python3.8/site-packages (from fastapi) (0.13.4)
Requirement already satisfied: websockets==8.* in /Users/cfe/.local/share/virtualenvs/ml-hello-world-rt2goiiz/lib/python3.8/site-packages (from uvicorn) (8.1)
Requirement already satisfied: uvloop>=0.14.0; sys_platform != "win32" and sys_platform != "cygwin" and platform_python_implementation != "PyPy" in /Users/cfe/.local/share/virtualenvs/ml-hello-world-rt2goiiz/lib/python3.8/site-packages (from uvicorn) (0.14.0)
Requirement already satisfied: h11<0.10,>=0.8 in /Users/cfe/.local/share/virtualenvs/ml-hello-world-rt2goiiz/lib/python3.8/site-packages (from uvicorn) (0.9.0)
Requirement already satisfied: httptools==0.1.*; sys_platform != "win32" and sys_platform != "cygwin" and platform_python_implementation != "PyPy" in /Users/cfe/.local/share/virtualenvs/ml-hello-world-rt2goiiz/lib/python3.8/site-packages (from uvicorn) (0.1.1)
Requirement already satisfied: click==7.* in /Users/cfe/.local/share/virtualenvs/ml-hello-world-rt2goiiz/lib/python3.8/site-packages (from uvicorn) (7.1.2)
Requirement already satisfied: chardet<4,>=3.0.2 in /Users/cfe/.local/share/virtualenvs/ml-hello-world-rt2goiiz/lib/python3.8/site-packages (from requests) (3.0.4)
Requirement already satisfied: urllib3!=1.25.0,!=1.25.1,<1.26,>=1.21.1 in /Users/cfe/.local/share/virtualenvs/ml-hello-world-rt2goiiz/lib/python3.8/site-packages (from requests) (1.25.9)
Requirement already satisfied: idna<3,>=2.5 in /Users/cfe/.local/share/virtualenvs/ml-hello-world-rt2goiiz/lib/python3.8/site-packages (from requests) (2.10)
Requirement already satisfied: certifi>=2017.4.17 in /Users/cfe/.local/share/virtualenvs/ml-hello-world-rt2goiiz/lib/python3.8/site-packages (from requests) (2020.6.20)
API_APP_PATH = 'app.py' # pathlib, os.path# from fastapi import FastAPI
# app = FastAPI()
# @app.get("/")
# def homepage_view():
# return {"Hello": "World"}%%writefile $API_APP_PATH
import pickle
from fastapi import FastAPI
from pydantic import BaseModel
app = FastAPI()
model_data = {}
with open("model.pkl", 'rb') as f:
model_data = pickle.loads(f.read())
class CustomerInput(BaseModel):
query:str
def predict(txt='Hello world',
vectorizer=None,
model=None,
classes=[],
*args,
**kwargs):
# pred
assert(vectorizer!=None)
assert(model != None)
input_vector = vectorizer.transform([txt])
output_vector = model.predict(input_vector)
assert(len(output_vector[0]) == len(classes))
preds = {}
for i, val in enumerate(output_vector[0]):
preds[classes[i]] = int(val)
return preds
@app.post("/predict")
def predict_view(customer_input:CustomerInput):
# storing this query data -> SQL database
my_pred = predict(customer_input.query, **model_data)
return {"query": customer_input.query, "predictions": my_pred}
# @app.post('/train')Overwriting app.py
import requests
json = {
"query": "When do you open?"
}
r = requests.post("http://127.0.0.1:8000/predict", json=json)
print(r.json()){'query': 'When do you open?', 'predictions': {'closing': 0, 'contact': 0, 'customer_support': 0, 'feedback': 0, 'food': 0, 'hours': 1, 'inventory': 0, 'menu': 0, 'opening': 1, 'products': 0}}
# label_predictor_from_export("When does your kitchen close?", **model_loaded_data)bot_responses = [
{
"responses": [
"We open at 8am everyday",
"We are open from 8am to 10pm everyday",
"8am to 10pm everyday"
],
"tags": ["hours", 'opening']
},
{
"responses": [
"Tacos & Burgers",
"Pizza"
],
"tags": ["menu", 'food']
}
]
bot_df = pd.DataFrame(bot_responses)
bot_df.head(n=100).dataframe tbody tr th {
vertical-align: top;
}
.dataframe thead th {
text-align: right;
}
| responses | tags | |
|---|---|---|
| 0 | [We open at 8am everyday, We are open from 8am... | [hours, opening] |
| 1 | [Tacos & Burgers, Pizza] | [menu, food] |
pred_response = {'query': 'When do you open?', 'predictions': {'closing': 0, 'contact': 0, 'customer_support': 0, 'feedback': 0, 'food': 0, 'hours': 1, 'inventory': 0, 'menu': 0, 'opening': 1, 'products': 0}}pred_tags = [k for k,v in pred_response['predictions'].items() if v != 0]
pred_tags['hours', 'opening']
mask = bot_df.tags.apply(lambda x: set(pred_tags) == set(x))
print(mask)0 True
1 False
Name: tags, dtype: bool
response_df = bot_df[mask] # bot_df[bot_df.tags.isin("abcs")]
response_df.head().dataframe tbody tr th {
vertical-align: top;
}
.dataframe thead th {
text-align: right;
}
| responses | tags | |
|---|---|---|
| 0 | [We open at 8am everyday, We are open from 8am... | [hours, opening] |
all_responses = list(response_df['responses'].values)
print(all_responses)[['We open at 8am everyday', 'We are open from 8am to 10pm everyday', '8am to 10pm everyday']]
responses = []
for row in all_responses:
for r in row:
responses.append(r)
responses = list(set(responses))
responses['We are open from 8am to 10pm everyday',
'We open at 8am everyday',
'8am to 10pm everyday']
import random
def predict_and_respond(txt=None, bot_df=None):
if txt == None and bot_df is None:
return "Sorry, I don't know what that means. Please contact us."
json = {
"query": txt
}
r = requests.post("http://127.0.0.1:8000/predict", json=json)
if r.status_code not in range(200, 299):
# send a signal, logging
return "Sorry, I am having trouble right now. Please try again later."
pred_response = r.json()
pred_tags = [k for k,v in pred_response['predictions'].items() if v != 0]
mask = bot_df.tags.apply(lambda x: set(pred_tags) == set(x))
response_df = bot_df[mask]
all_responses = list(response_df['responses'].values)
responses = []
for row in all_responses:
for r in row:
responses.append(r)
responses = list(set(responses))
if len(responses) == 0:
return "Sorry, I am still learning. I don't understand what you said."
return random.choice(responses)predict_and_respond("Are you open at 9:30am tomorrow?", bot_df=bot_df)"Sorry, I am still learning. I don't understand what you said."
- Get more data, a lot more
- Add more data right now. Keep adding data.
- Refine the data, move tags, remove tags, upgrade.
- Deploy to a production server:
- Ideally using this project for deploying a serverless application using FastAPI (like we did).
- Use internally, a lot.
- If this tool becomes the go-to for finding answers about your business internally, it can become the same tool for external (customer facing) as well. When in doubt, give it to customers,