Автоматты аударма пайдаланылды
Guide On How To Set Up Text Sentiment Analysis with ML + Deployment with Streamlit
Hello there!
My name is Askhat Aubakirov, I recently completed the Data Science Course provided by OUTPEERKZ as part of the TechOrda program.
Today, I am doing data science projects with students of mine at Makerspace Petropavl to share what I know and inspire young people to join the data science world.
Here is what we have created during many of our meetings. Follow us step by step and you'll understand the crucial steps of how data scientists work.
Important: when following this guide, please make sure to install libraries via pip and refer to the files in the github repository (https://github.com/askhat-aubakirov/sentiment/tree/main)
Here is what awaits:
- On the stage of learning: in the Jupyter Notebook:
- Working with Dataset:
--- Data Loading,
--- Exploratory Data Analysis,
--- Data Cleaning,
--- Data Processing,
--- Feature Engineering.
- Preparing for the training:
--- Loading Random Forest Model,
--- Splitting data to train and test arrays.
- Training and testing
- Saving the model - On the stage of Deployment:
- Loading Streamlit
--- Creating a visual template with elements from Streamlit
--- Adapting functions from Jupyter Notebook to tokenize, lemmatize and vectorize user inputs
--- extracting only essential data for deployment
- Launching the app locally and Testing the process with users - Github repo for the project: https://github.com/askhat-aubakirov/sentiment
First, we start with a problem to solve
We will explore NLP (Natural Language Processing) and build a model to classify texts into sentiment categories. In a nutshell, it should be an app to get a text from a user and return one of the verdicts such as "positive", "negative", etc.
The data for this problem is found easily on Kaggle. Here is the link: https://www.kaggle.com/datasets/jp797498e/twitter-entity-sentiment-analysis
Loading necessary libraries:
import pandas as pd #pandas for calculations
import matplotlib.pyplot as plt #matplotlib for plotting graphs in EDA steps
import re #regular expressions library to help us clean up the dataset
import spacy #library for lemmatization function of ours
from sklearn.ensemble import RandomForestClassifier #machine Learning model we're going to use
from sklearn.metrics import accuracy_score, f1_score #metrics to measure our success
from sklearn.feature_extraction.text import TfidfVectorizer #method to convert words and text to a matrix
from sklearn.model_selection import train_test_split #splitting the data to parts
import joblib #library to save the model locally You may see what libraries we have imported so far:
- pandas: used for calculations in matrix algebra (we are going to store and access our dataset as a matrix)
- matplotlib: used to create graphs -> this way we will see if a dataset is imbalanced or too dirty via statistics and visuals
- re: super useful tool for NLP, helps create expressions to recognize complex strings and manipulate them
- spacy: library containing pre-loaded stop words and corpora of the language
- RandomForestClassifier: a model we are going to use. (What is it? Here is some visual explanation: https://youtu.be/cIbj0WuK41w
- metrics: accuracy and F1 score (more details below)
- TfidfVectorizer: great tool to convert words to numerical matrix - the form digestible by an algorithm
- train_test_split: to train a model and calculate metrics, you need a few parts, which are the training set and the test set.
- joblib: a library to save the trained model in one file -> we will use it later in deployment process
Exploratory Data Analysis (EDA)
EDA is a crucial step in data science to understand your data before analysis. It involves:
- Summarizing: Calculating statistics like mean, median, standard deviation.
- Visualizing: Creating plots like histograms, scatter plots, and box plots.
- Cleaning: Handling missing values and outliers.
For a dataset on customer sales:
- Summary: Calculate the average purchase amount, and number of unique customers.
- Visualization: Create a histogram to see the distribution of purchase amounts.
- Cleaning: Handle missing values in customer addresses.
EDA helps you identify patterns, relationships, and potential issues in your data.
First things first, we load the data (saved from Kaggle beforehand):
data = pd.read_csv('twitter_training.csv')Then we go on and take a look at the data using head() method from pandas:
data.head()And in doing so, we get this output:
2401 Borderlands Positive im getting on borderlands and i will murder you all ,
0 2401 Borderlands Positive I am coming to the borders and I will kill you...
1 2401 Borderlands Positive im getting on borderlands and i will kill you ...
2 2401 Borderlands Positive im coming on borderlands and i will murder you...
3 2401 Borderlands Positive im getting on borderlands 2 and i will murder ...
4 2401 Borderlands Positive im getting into borderlands and i can murder y...We clearly see it has some IDs (2401, etc.), Entity Names (Borderlands, etc), and Sentiment (Positive, etc. This is a class, in other words). But the thing is we don't have headers here. Let's add them:
data_valid.columns = ['ID', 'entity', 'sentiment', 'content']
data_valid.head(5) ID entity sentiment content
0 352 Amazon Neutral BBC News - Amazon boss Jeff Bezos rejects clai...
1 8312 Microsoft Negative @Microsoft Why do I pay for WORD when it funct...
2 4371 CS-GO Negative CSGO matchmaking is so full of closet hacking,...
3 4433 Google Neutral Now the President is slapping Americans in the...
4 6273 FIFA Negative Hi @EAHelp I’ve had Madeleine McCann in my cel...Now that's better, we have our headers nice and clean.
Graphs
It's high time we generated some charts and looked at the dataset statistically. We will use matplotlib. Read carefully -> code explains itself.
# Get the entity labels from the DataFrame
entity_labels = data['entity']
entity_labels = data["entity"].value_counts() #count the values we have
print(entity_labels)
plt.figure(figsize=(12, 6))
entity_labels.plot(kind='bar')
plt.title('Value Counts of entities')
plt.xlabel('entity')
plt.ylabel('Count')
plt.xticks(fontsize=10) #decrease fontsize for readability - unfortunately, readablity is close to zero
plt.show()
plt.figure(figsize=(4, 4))
entity_labels.plot(kind='box')
plt.title('Value Counts of entities')
plt.xlabel('entity')
plt.ylabel('Count')
plt.xticks(fontsize=10)
plt.show()matplotlib is referred to as plt. We simply call built-in methods to create 2 types of graphs: A bar graph and a box graph. We get this as an output:
entity
TomClancysRainbowSix 2400
MaddenNFL 2400
Microsoft 2400
LeagueOfLegends 2394
CallOfDuty 2394
Verizon 2382
CallOfDutyBlackopsColdWar 2376
ApexLegends 2376
Facebook 2370
WorldOfCraft 2364
Dota2 2364
NBA2K 2352
TomClancysGhostRecon 2346
Battlefield 2346
FIFA 2340
Xbox(Xseries) 2334
Overwatch 2334
johnson&johnson 2328
Amazon 2316
PlayStation5(PS5) 2310
HomeDepot 2310
Cyberpunk2077 2304
CS-GO 2304
GrandTheftAuto(GTA) 2304
...
Fortnite 2274
RedDeadRedemption(RDR) 2262
AssassinsCreed 2244
Name: count, dtype: int64The output gives us info on the amounts of entities in the dataset. The dataset is balanced (cause it is curated), so we have "minus one" to the number of data-related problems to solve.
Both plots show almost perfectly balanced data. This means the training process will most likely yield a pretty accurate model. Let's move forward.
Data Preprocessing
We know we have text data. But it is dirty. Tweets we are working with are filled with typos, punctuation, account handlers, URLs and numbers. We define a function to get rid of the unnecessary symbols called clean_tweet.
def clean_tweet(text):
if isinstance(text, str):
text = text.lower() # Convert to lowercase
text = re.sub(r'http\S+|www\S+|https\S+', '', text, flags=re.MULTILINE) # Remove URLs
text = re.sub(r'\@\w+|\#', '', text) # Remove mentions and hashtags
text = re.sub(r'[^\w\s]', '', text) # Remove punctuation
text = re.sub(r'\d+', '', text) # Remove numbers
else:
text = '' # Handle non-string inputs like float (NaN) by returning an empty string or handling as needed
return textThe main part here is the re library being used. It is a library of regular expressions.
The regular expression http\S+|www\S+|https\S+ is used to match URLs in text. Here's a breakdown of its components:
- http\S+: Matches the string "http" followed by one or more non-whitespace characters. This captures HTTP URLs.
- www\S+: Matches the string "www" followed by one or more non-whitespace characters. This captures URLs that start with "www".
- https\S+: Matches the string "https" followed by one or more non-whitespace characters. This captures HTTPS URLs.
- The | character is a pipe symbol that acts as an "or" operator, allowing the regular expression to match any of the three patterns.
Now we use the method apply() to use this function to every element of the dataset:
# Apply the cleaning function to the 'content' column
data['clean_content'] = data['content'].apply(clean_tweet)Well done. We have a clean text column in our dataset now:
data.head(5) ID entity sentiment content clean_content
0 2401 Borderlands Positive I am coming to the borders and I will kill you... i am coming to the borders and i will kill you...
1 2401 Borderlands Positive im getting on borderlands and i will kill you ... im getting on borderlands and i will kill you all
2 2401 Borderlands Positive im coming on borderlands and i will murder you... im coming on borderlands and i will murder you...
3 2401 Borderlands Positive im getting on borderlands 2 and i will murder ... im getting on borderlands and i will murder y...
4 2401 Borderlands Positive im getting into borderlands and i can murder y... im getting into borderlands and i can murder y...Tokenization, Lemmatization, Vectorizing
Now the best part in NLP! Take a look at explanations of these three processes:
- Breaking text into smaller units (tokens).
- Example: "The quick brown fox jumps over the lazy dog" becomes ["The", "quick", "brown", "fox", "jumps", "over", "the", "lazy", "dog"].
- Reducing words to their root form (lemma).
- Example: "running", "runs", "ran" are reduced to "run".
- Converting text into numerical representations (vectors).
- Example: Using techniques like Bag-of-Words (BoW) or TF-IDF to represent text as numerical vectors.
And now, the code:
nlp = spacy.load('en_core_web_sm')
def tokenize_and_lemmatize(text):
"""
Tokenizes and lemmatizes the input text.
Args:
text: A string representing the input text.
Returns:
a list of lemmas
"""
lemmas = []
doc = nlp(text)
for token in doc:
if token.is_alpha and not token.is_stop:
lemmas.append(token.lemma_)
new_lemma_tokens = " ".join(lemmas) #Arsen was right -> I had indentation error
return new_lemma_tokensdata['lemmas'] = data['clean_content'].apply(tokenize_and_lemmatize)And when we take a look at the dataset now, we see something like this:
ID entity sentiment content clean_content lemmas
0 352 Amazon Neutral BBC News - Amazon boss Jeff Bezos rejects clai... bbc news amazon boss jeff bezos rejects claim... bbc news amazon boss jeff bezos rejects claim ...
1 8312 Microsoft Negative @Microsoft Why do I pay for WORD when it funct... why do i pay for word when it functions so po... pay word function poorly chromebook
2 4371 CS-GO Negative CSGO matchmaking is so full of closet hacking,... csgo matchmaking is so full of closet hacking ... csgo matchmaking closet hack truly awful game
3 4433 Google Neutral Now the President is slapping Americans in the... now the president is slapping americans in the... president slap americans face commit unlawful ...
4 6273 FIFA Negative Hi @EAHelp I’ve had Madeleine McCann in my cel... hi ive had madeleine mccann in my cellar for ... hi ve madeleine mccann cellar past year little... Let's take a look at what differences to the text we've made so far:
original text "@Microsoft Why do I pay for WORD when it funct..." became "why do i pay for word when it functions so po..." and then became "pay word function poorly chromebook".
The trick is the meaning of the tweet is clear when you read just lemmas, not every single word.
And now, Vectorizing. It's conversion to numbers. We need to feed the model matrices with numbers, not words. We will use TF-IDF Vectorizer for this task:
It works like this:
1) We get tokens(words, terms) x and calculate how many times it is present in the document y, then we calculate the frequency of term x in this document y -> this is TF_x,y
2) We multiply the TF_x,y with IDF (which is logarithm of the total amount of documents divided by the amount of documents that contain the term x.
This gives us a great representation of the importance of the word. For example, we may have many words like "I", "have", "we", etc. Although there are many of them, they are not very important. On the other hand, we may have just few mentions of words like "happy", "angry" or anything like these, but they are important.
So, TF-IDF Vectorizer gives relative coefficients of frequency (any float from 0 to 1), by which we may estimate the importance of the terms.
Here is what it looks like in code:
#vectorizing the lemmatized text -> this is basically Feature Engineering, cause we have text data and define max_features
vectorizer = TfidfVectorizer(lowercase=False, max_features=10000, ngram_range=(1, 2))
vectorized_data = vectorizer.fit_transform(data['lemmas'].values.astype('U'))
vectorized_data = pd.DataFrame(vectorized_data.toarray(), columns=vectorizer.get_feature_names_out()) EXPLANATION:
vectorizer = TfidfVectorizer(lowercase=False, max_features=10000, ngram_range=(1, 2)) creates a vectorizer with parameters like lowercase=False, because applied lowercase before, max_features=10000 to determine the amount of new features' columns, and ngram_range=(1, 2) to make features consist not only of single words but combinations of two words.
That's quite handy, since in, for example, medical data cases words like "heart" and "attack" may have separate meanings, but "heart attack" has more important meaning.
Now the dataset we have looks like this and has the shape of 74681 x 10000:
aa aa aa aaa aaron ab abandon abandon sanction abc abella abella danger ... zero dawn zion zip zoe zombie zone zonestreamcx zoom zsmitty zuckerberg
0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 ... 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 ... 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
2 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 ... 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
3 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 ... 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
4 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 ... 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
5 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 ... 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
6 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 ... 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
7 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 ... 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
8 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 ... 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
9 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 ... 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0Note it has 10 000 words and combinations of words, each with their own frequency across documents. Now that IS a matrix we may feed to the model.
Model Choice and Training
We should split the dataset in order to train it and then test it to get our metrics.
For now, we are at the left part of the scheme. We have our features matrix and our classes. We need to split it to X_train, y_train (training matrix, or "questions", and target classes, or "answers"). Plus, X_test and y_test. We need that to test how good (or bad) our model performs.
X_train, X_test, y_train, y_test = train_test_split(vectorized_data, data['sentiment'], test_size=0.2, random_state=42)Model we will work with is Random Forest Classifier.
- Bootstrap Sampling: Random sampling with replacement: Multiple subsets (bootstrap samples) are created from the original dataset. Each subset may contain duplicate instances.
- Random sampling with replacement: Multiple subsets (bootstrap samples) are created from the original dataset. Each subset may contain duplicate instances.
- Decision Tree Building: Random feature selection: For each tree, a random subset of features is selected from the total number of features.Decision tree growth:Each decision tree is grown to its maximum depth without pruning.
- Random feature selection: For each tree, a random subset of features is selected from the total number of features.
- Decision tree growth:Each decision tree is grown to its maximum depth without pruning.
- Prediction: Voting: To make a prediction for a new instance, each decision tree casts a vote. The class with the most votes is chosen as the final prediction.
- Voting: To make a prediction for a new instance, each decision tree casts a vote. The class with the most votes is chosen as the final prediction.
rf = RandomForestClassifier(n_estimators=100, random_state=42)
rf.fit(X_train, y_train)n_estimators in a Random Forest classifier refers to the number of decision trees that are grown in the ensemble. random_state=42 is something of a seed -> used to ensure reproducibility. Also, it is a reference to the book "The Hitchhikers Guide to the Galaxy", where civilization asked what the meaning of life was and got "42" as an answer from the biggest computer in the galaxy. It's quite fun, isn't it?
When Random Forest Model finishes the training (took 11 minutes for me), it'll show this:
Now let's test it. We make predictions on our X_test set, getting y_pred as output and then compare results to the y_test via metrics such as Accuracy and F1 Score.
Here is the explanation of what these metrics are (along with other metrics as well):
And now, the code and output:
y_pred = rf.predict(X_test)
accuracy = accuracy_score(y_test, y_pred)
f1 = f1_score(y_test, y_pred, average='macro')
print("Accuracy:", accuracy)
print("F1 score:", f1)Accuracy: 0.8594764678315592
F1 score: 0.8567240271159241We may say our model has accuracy of 86.95% and F1 score of 85.67%
Let us now test the model on our "organic" sample.
Just give the model some text out of our heads as an input, but don't forget to preprocess it (cleaning, tokenization, lemmatization, vectorizing + compare and modify format -> columns of the input matrix has to be the same as columns of the train matrix).
Here is the code for this:
nlp = spacy.load('en_core_web_sm', disable=['parser', 'ner'])
headers = pd.read_csv("headers_only.csv")
rf_model = joblib.load('rf_model.joblib')
vectorizer = TfidfVectorizer(lowercase=False, max_features=10000, ngram_range=(1, 2))
def tokenize_and_lemmatize(text):
"""
Tokenizes and lemmatizes the input text.
Args:
text: A string representing the input text.
Returns:
a list of lemmas
"""
lemmas = []
doc = nlp(text)
for token in doc:
if token.is_alpha and not token.is_stop:
lemmas.append(token.lemma_)
new_lemma_tokens = " ".join(lemmas)
return new_lemma_tokens
def predict_condition(user_input):
"""
Preprocesses user input, predicts skin condition using the SVM model,
and returns the prediction result.
Args:
user_input: A string representing the user's description.
Returns:
A string with the predicted skin condition.
"""
try:
# Lowercase and remove punctuation
#user_input = user_input.lower()
#user_input = "".join([char for char in user_input if char.isalnum() or char.isspace()])
# Tokenize and lemmatize
# lemmatized_text = tokenize_and_lemmatize(user_input)
# Vectorize the text
vectorized_input = vectorizer.fit_transform([user_input])
vectorized_input = pd.DataFrame(vectorized_input.toarray(), columns=vectorizer.get_feature_names_out())
#vectorized_input = pd.DataFrame(vectorized_input)
# Load missing columns from training data (if any)
missing_columns = set(headers.columns) - set(vectorized_input.columns)
vectorized_input = vectorized_input.reindex(columns=headers.columns, fill_value=0)
vectorized_input = vectorized_input[headers.columns]
# Predict using the model
prediction = rf_model.predict(vectorized_input)[0]
# Return the prediction result
return prediction
except Exception as e:
# Handle potential errors during preprocessing or prediction
error_message = f"An error occurred: {str(e)}"
return None
user_input = "YAY my model works fine and quite fast! Let's test it"
user_input = user_input.lower()
prediction = predict_condition(user_input)
print(prediction) Note these lines:
- missing_columns = set(headers.columns) - set(vectorized_input.columns)
- vectorized_input = vectorized_input.reindex(columns=headers.columns, fill_value=0)
They are here to make sure the columns of the train set are the same as the columns of the input. It is a simple representation of a mechanism pretty much the same as when a person recognizes only words they know from a speech. For the words our model doesn't know, it'll simply put zeroes, which is equivalent to us ignoring something we don't know about a word or few in someone's speech.
So, the output for the "YAY my model works fine and quite fast! Let's test it" is Positive
PositiveDeployment using Streamlit
Now, let's make it an app. Streamlit is a simple library to use specially for Data Scientists, since it takes care of all UI rendering.
Remember, for the streamlit app, you need to make user input go through all of the steps of text preprocessing as the input above. But the thing is we uploaded our whole train set for the training in the Jupyter Notebook, which we don't need here, since the dataset is very big. We only need headers, so we manipulate our training set to only leave columns (details can bee seen in the train_set.ipynb).
We only need to load our dataset and use this method (it's that simple):
new_df = pd.DataFrame(columns=X_train.columns)With this in mind, we will have this code to run our streamlit app:
import spacy
import streamlit as st
import joblib
import pandas as pd
from sklearn.feature_extraction.text import TfidfVectorizer
#from sklearn.model_selection import train_test_split
#from sklearn.metrics import accuracy_score, precision_score, recall_score, f1_score
nlp = spacy.load('en_core_web_sm', disable=['parser', 'ner'])
headers = pd.read_csv("headers_only.csv")
rf_model = joblib.load('rf_model.joblib')
vectorizer = TfidfVectorizer(lowercase=False, max_features=10000, ngram_range=(1, 2))
@st.cache_resource
def load_model():
return joblib.load('rf_model.joblib')
def tokenize_and_lemmatize(text):
"""
Tokenizes and lemmatizes the input text.
Args:
text: A string representing the input text.
Returns:
a list of lemmas
"""
lemmas = []
doc = nlp(text)
for token in doc:
if token.is_alpha and not token.is_stop:
lemmas.append(token.lemma_)
new_lemma_tokens = " ".join(lemmas)
return new_lemma_tokens
def predict_condition(user_input):
"""
Preprocesses user input, predicts skin condition using the SVM model,
and returns the prediction result.
Args:
user_input: A string representing the user's description.
Returns:
A string with the predicted skin condition.
"""
try:
# Lowercase and remove punctuation
user_input = user_input.lower()
user_input = "".join([char for char in user_input if char.isalnum() or char.isspace()])
# Tokenize and lemmatize
lemmatized_text = tokenize_and_lemmatize(user_input)
# Vectorize the text
vectorized_input = vectorizer.fit_transform([lemmatized_text])
vectorized_input = pd.DataFrame(vectorized_input.toarray(), columns=vectorizer.get_feature_names_out())
vectorized_input = pd.DataFrame(vectorized_input)
# Load missing columns from training data (if any)
#missing_columns = set(headers.columns) - set(vectorized_input.columns)
vectorized_input = vectorized_input.reindex(columns=headers.columns, fill_value=0)
#vectorized_input = vectorized_input.assign(**{col: 0 for col in missing_columns})
vectorized_input = vectorized_input[headers.columns]
st.write(vectorized_input)
# Predict using the model
prediction = rf_model.predict(vectorized_input)[0]
# Return the prediction result
return prediction
except Exception as e:
# Handle potential errors during preprocessing or prediction
error_message = f"An error occurred: {str(e)}"
st.error(error_message)
return None
# Streamlit app layout and functionality
st.title(":blue[Text Sentiment Analyzer] App")
user_input_text = st.text_area("Write your text here:", height=100)
if st.button("Predict"):
prediction = predict_condition(user_input_text)
if prediction:
st.success(f"Predicted condition: {prediction}. \n")
else:
st.warning("Prediction failed. Please try again and/or consult with developers.")
# Display model information (optional)
st.header("Model Information")
st.header("This app is powered by a trained Random Forest model.", divider = 'rainbow')
st.write("- Text to be analyzed sentiment-wise")
st.write("- Data Science Club Group One")
st.write("- IG: @american_corner_petropavlovsk")
st.write("- Club host: Askhat Aubakirov")
st.write("- Publication Date: September 17th 2024")Notice there are some handlers and streamlit-specific methods. All of them can be found on their official cheat sheet: https://docs.streamlit.io/develop/quick-reference/cheat-sheet
Here is what the product looks like:
THE END
Big thanks to all of the participants of Data Club at Makerspace Petropavl for all the great work you do. Special thanks to those who added their ideas to this code: Arsen and Arthur.
I hope you found this article helpful.
Best regards,
Askhat.
My LinkedIn: https://www.linkedin.com/in/askhattio/
Whatsapp
696
Пікірлер 8
Кіру пікір қалдыру үшін
БЕХРУЗ ТОХТАМИШОВ · Қыр. 26, 2024 15:39
🔥🔥
Alizhan Nurgazy · Қыр. 25, 2024 09:03
👍👍👍👍👍👍👍👍👍👍👍👍
Ans s · Қыр. 25, 2024 00:37
🔥🔥
Ans s · Қыр. 25, 2024 00:37
🔥🔥🔥🔥
Саин Абельсеитов · Қыр. 24, 2024 17:05
Классно. Как раз поступил на курсы дата скайнс при поддержке техорда
Askhat Aubakirov · Қыр. 24, 2024 17:18
Поздравляю! Надеюсь, этот материал поможет разобраться в ML и NLP)
Lain Iwakura · Қыр. 24, 2024 15:45
супер!
Askhat Aubakirov · Қыр. 24, 2024 16:17
спасибо за отзыв)