4 |
5 | [Slides](https://www.slideshare.net/AlexeyGrigorev/ml-zoomcamp-11-introduction-to-machine-learning)
6 |
7 |
8 | ## Notes
9 |
10 | The concept of ML is depicted with an example of predicting the price of a car. The ML model
11 | learns from data, represented as some **features** such as year, mileage, among others, and the **target** variable, in this
12 | case, the car's price, by extracting patterns from the data.
13 |
14 | Then, the model is given new data (**without** the target) about cars and predicts their price (target).
15 |
16 | In summary, ML is a process of **extracting patterns from data**, which is of two types:
17 |
18 | * features (information about the object) and
19 | * target (property to predict for unseen objects).
20 |
21 | Therefore, new feature values are presented to the model, and it makes **predictions** from the learned patterns.
22 |
23 | | ⚠️ | 26 |
27 | The notes are written by the community. 28 | If you see an error here, please create a PR with a fix. 29 | |
30 |
4 |
5 |
6 | ## Notes
7 |
8 |
9 | Add notes here (PRs are welcome).
10 |
11 | | ⚠️ | 14 |
15 | The notes are written by the community. 16 | If you see an error here, please create a PR with a fix. 17 | |
18 |
5 |
6 | [Slides](https://www.slideshare.net/AlexeyGrigorev/ml-zoomcamp-21-car-price-prediction-project)
7 |
8 |
9 | ## Notes
10 |
11 | This project is about the creation of a model for helping users to predict car prices. The dataset was obtained from [this
12 | kaggle competition](https://www.kaggle.com/CooperUnion/cardataset).
13 |
14 | **Project plan:**
15 |
16 | * Prepare data and Exploratory data analysis (EDA)
17 | * Use linear regression for predicting price
18 | * Understanding the internals of linear regression
19 | * Evaluating the model with RMSE
20 | * Feature engineering
21 | * Regularization
22 | * Using the model
23 |
24 | The code and dataset are available at this [link](https://github.com/alexeygrigorev/mlbookcamp-code/tree/master/chapter-02-car-price).
25 |
26 | | ⚠️ | 29 |
30 | The notes are written by the community. 31 | If you see an error here, please create a PR with a fix. 32 | |
33 |
5 |
6 | [Slides](https://www.slideshare.net/AlexeyGrigorev/ml-zoomcamp-2-slides)
7 |
8 |
9 | ## Notes
10 |
11 | **Pandas attributes and methods:**
12 |
13 | * `pd.read_csv(| ⚠️ | 26 |
27 | The notes are written by the community. 28 | If you see an error here, please create a PR with a fix. 29 | |
30 |
5 |
6 | [Slides](https://www.slideshare.net/AlexeyGrigorev/ml-zoomcamp-2-slides)
7 |
8 |
9 | ## Notes
10 |
11 | **Pandas attributes and methods:**
12 |
13 | * `df[col].unique()` -> return a list of unique values in the series
14 | * `df[col].nunique()` -> return the number of unique values in the series
15 | * `df.isnull().sum()` -> return the number of null values in the dataframe
16 |
17 | **Matplotlib and seaborn methods:**
18 |
19 | * `%matplotlib inline` -> assure that plots are displayed in jupyter notebook's cells
20 | * `sns.histplot()` -> show the histogram of a series
21 |
22 | **Numpy methods:**
23 | * `np.log1p()` -> apply log transformation to a variable, after adding one to each input value.
24 |
25 | Long-tail distributions usually confuse the ML models, so the recommendation is to transform the target variable distribution to a normal one whenever possible.
26 |
27 | The entire code of this project is available in [this jupyter notebook](https://github.com/alexeygrigorev/mlbookcamp-code/blob/master/chapter-02-car-price/02-carprice.ipynb).
28 |
29 | | ⚠️ | 32 |
33 | The notes are written by the community. 34 | If you see an error here, please create a PR with a fix. 35 | |
36 |
5 |
6 | [Slides](https://www.slideshare.net/AlexeyGrigorev/ml-zoomcamp-2-slides)
7 |
8 |
9 | ## Notes
10 |
11 | The formula of linear regression can be synthesized with the dot product between features and weights. The feature vector includes the *bias* term with an *x* value of one, such as $w_{0}^{x_{i0}},\ where\ x_{i0} = 1\ for\ w_0$.
12 |
13 | When all the records are included, the linear regression can be calculated with the dot product between ***feature matrix*** and ***vector of weights***, obtaining the `y` vector of predictions.
14 |
15 | The entire code of this project is available in [this jupyter notebook](https://github.com/alexeygrigorev/mlbookcamp-code/blob/master/chapter-02-car-price/02-carprice.ipynb).
16 |
17 | | ⚠️ | 20 |
21 | The notes are written by the community. 22 | If you see an error here, please create a PR with a fix. 23 | |
24 |
5 |
6 | [Slides](https://www.slideshare.net/AlexeyGrigorev/ml-zoomcamp-2-slides)
7 |
8 |
9 | ## Notes
10 |
11 | Obtaining predictions as close as possible to $y$ target values requires the calculation of weights from the general
12 | LR equation. The feature matrix does not
13 | have an inverse because it is not square, so it is required to obtain an approximate solution, which can be
14 | obtained using the **Gram matrix**
15 | (multiplication of feature matrix ($X$) and its transpose ($X^T$)). The vector of weights or coefficients $w$ obtained with this
16 | formula is the closest possible solution to the LR system.
17 |
18 | Normal Equation:
19 |
20 | $w$ = $(X^TX)^{-1}X^Ty$
21 |
22 | Where:
23 |
24 | $X^TX$ is the Gram Matrix
25 |
26 |
27 |
28 |
29 | The entire code of this project is available in [this jupyter notebook](https://github.com/alexeygrigorev/mlbookcamp-code/blob/master/chapter-02-car-price/02-carprice.ipynb).
30 |
31 | | ⚠️ | 34 |
35 | The notes are written by the community. 36 | If you see an error here, please create a PR with a fix. 37 | |
38 |
5 |
6 | [Slides](https://www.slideshare.net/AlexeyGrigorev/ml-zoomcamp-2-slides)
7 |
8 |
9 | ## Notes
10 |
11 | * In the previous lesson we found out our predictions were a bit off from the actual target values in the training dataset. We need a way to quantify how good or bad the model is. This is where RMSE can be of help.
12 | * Root Mean Squared Error (RMSE) is a way to evaluate regression models. It measures the error associated with the model being evaluated. This numerical figure can then be used to compare models, enabling us to choose the one that gives the best predictions.
13 |
14 | $$RMSE = \sqrt{ \frac{1}{m} \sum_{i=1}^{m} {(g(x_i) - y_i)^2}}$$
15 |
16 | - $g(x_i)$ is the prediction
17 | - $y_i$ is the actual value
18 | - $m$ is the number of observations in the dataset (i.e. cars)
19 |
20 |
21 | The entire code of this project is available in [this jupyter notebook](https://github.com/alexeygrigorev/mlbookcamp-code/blob/master/chapter-02-car-price/02-carprice.ipynb).
22 |
23 | | ⚠️ | 26 |
27 | The notes are written by the community. 28 | If you see an error here, please create a PR with a fix. 29 | |
30 |
5 |
6 | [Slides](https://www.slideshare.net/AlexeyGrigorev/ml-zoomcamp-2-slides)
7 |
8 |
9 | ## Notes
10 |
11 | Calculation of the RMSE on validation partition of the dataset of car price prediction. In this way, we have a metric to evaluate the model's
12 | performance.
13 |
14 | The entire code of this project is available in [this jupyter notebook](https://github.com/alexeygrigorev/mlbookcamp-code/blob/master/chapter-02-car-price/02-carprice.ipynb).
15 |
16 | | ⚠️ | 19 |
20 | The notes are written by the community. 21 | If you see an error here, please create a PR with a fix. 22 | |
23 |
6 |
7 | [Slides](https://www.slideshare.net/AlexeyGrigorev/ml-zoomcamp-2-slides)
8 |
9 |
10 | ## Notes
11 |
12 | The feature age of the car was included in the dataset, obtained with the subtraction of the maximum year of cars and each of the years of cars.
13 | This new feature improved the model performance, measured with the RMSE and comparing the distributions of y target variable and predictions.
14 |
15 | The entire code of this project is available in [this jupyter notebook](https://github.com/alexeygrigorev/mlbookcamp-code/blob/master/chapter-02-car-price/02-carprice.ipynb).
16 |
17 | | ⚠️ | 20 |
21 | The notes are written by the community. 22 | If you see an error here, please create a PR with a fix. 23 | |
24 |
5 |
6 | [Slides](https://www.slideshare.net/AlexeyGrigorev/ml-zoomcamp-2-slides)
7 |
8 |
9 | ## Notes
10 |
11 | Categorical variables are typically represented as strings, and pandas identifies them as object types. However, some variables that appear to be numerical may actually be categorical (e.g., the number of doors a car has). All these categorical variables need to be converted to a numerical form because ML
12 | models can interpret only numerical features. It is possible to incorporate certain categories from a feature, not necessarily all of them.
13 | This transformation from categorical to numerical variables is known as One-Hot encoding.
14 |
15 | The entire code of this project is available in [this jupyter notebook](https://github.com/alexeygrigorev/mlbookcamp-code/blob/master/chapter-02-car-price/02-carprice.ipynb).
16 |
17 | | ⚠️ | 20 |
21 | The notes are written by the community. 22 | If you see an error here, please create a PR with a fix. 23 | |
24 |
5 |
6 | [Slides](https://www.slideshare.net/AlexeyGrigorev/ml-zoomcamp-2-slides)
7 |
8 |
9 | ## Notes
10 |
11 | Tuning the model consisted of finding the best regularization hyperparameter value, using the validation partition of the dataset. The model was then trained with this regularization value.
12 |
13 | The entire code of this project is available in [this jupyter notebook](https://github.com/alexeygrigorev/mlbookcamp-code/blob/master/chapter-02-car-price/02-carprice.ipynb).
14 |
15 | | ⚠️ | 18 |
19 | The notes are written by the community. 20 | If you see an error here, please create a PR with a fix. 21 | |
22 |
5 |
6 | [Slides](https://www.slideshare.net/AlexeyGrigorev/ml-zoomcamp-2-slides)
7 |
8 | ## Notes
9 |
10 | After finding the best model and its parameters, it was trained with training and validation partitions and the final RMSE was calculated on the test partition.
11 |
12 | Finally, the final model was used to predict the price of new cars.
13 |
14 | The entire code of this project is available in [this jupyter notebook](https://github.com/alexeygrigorev/mlbookcamp-code/blob/master/chapter-02-car-price/02-carprice.ipynb).
15 |
16 | | ⚠️ | 19 |
20 | The notes are written by the community. 21 | If you see an error here, please create a PR with a fix. 22 | |
23 |
5 |
6 |
7 |
8 | ## Notes
9 |
10 | In summary, this session covered some topics, including data preparation, exploratory data analysis, the validation framework, linear regression model, LR vector and
11 | normal forms, the baseline model, root mean squared error, feature engineering, regularization, tuning the model, and using the best model with new data. All these concepts
12 | were explained using the problem to predict the price of cars.
13 |
14 | | ⚠️ | 17 |
18 | The notes are written by the community. 19 | If you see an error here, please create a PR with a fix. 20 | |
21 |
4 |
5 | [Slides](https://www.slideshare.net/AlexeyGrigorev/ml-zoomcamp-3-machine-learning-for-classification) empty?
6 |
7 | ## Notes
8 |
9 | The project aims to identify customers that are likely to churn or stop to using a service. Each customer has a score associated with the probability of churning. Considering this data, the company would send an email with discounts or other promotions to avoid churning.
10 |
11 | The ML strategy applied to approach this problem is binary classification, which for one instance ($i^{th}$ customer), can be expressed as:
12 |
13 | $$\large g\left(x_{i}\right) = y_{i}$$
14 |
15 | In the formula, $y_i$ is the model's prediction and belongs to {0,1}, with 0 being the negative value or no churning, and 1 the positive value or churning. The output corresponds to the likelihood of churning.
16 |
17 | In brief, the main idea behind this project is to build a model with historical data from customers and assign a score of the likelihood of churning.
18 |
19 | For this project, we used a [Kaggle dataset](https://www.kaggle.com/blastchar/telco-customer-churn).
20 |
21 | |⚠️|The notes are written by the community.
5 |
6 | [Slides](https://www.slideshare.net/AlexeyGrigorev/ml-zoomcamp-3-machine-learning-for-classification)
7 |
8 |
9 | ## Notes
10 |
11 | Splitting the dataset with **Scikit-Learn**.
12 |
13 | **Classes, functions, and methods:**
14 |
15 | * `train_test_split` - Scikit-Learn class for splitting a dataset into two parts. The `test_size` argument states how large the test set should be. The `random_state` argument sets a random seed for reproducibility purposes.
16 | * `df.reset_index(drop=True)` - reset the indices of a dataframe and delete the previous ones.
17 | * `df.x.values` - extract the values from x series
18 | * `del df['x']` - delete x series from a dataframe
19 |
20 | The entire code of this project is available in [this jupyter notebook](https://github.com/alexeygrigorev/mlbookcamp-code/blob/master/chapter-03-churn-prediction/03-churn.ipynb).
21 |
22 | | ⚠️ | 25 |
26 | The notes are written by the community. 27 | If you see an error here, please create a PR with a fix. 28 | |
29 |
5 |
6 | [Slides](https://www.slideshare.net/AlexeyGrigorev/ml-zoomcamp-3-machine-learning-for-classification)
7 |
8 |
9 | ## Notes
10 |
11 | The EDA for this project consisted of:
12 | * Checking missing values
13 | * Looking at the distribution of the target variable (churn)
14 | * Looking at numerical and categorical variables
15 |
16 | **Functions and methods:**
17 |
18 | * `df.isnull().sum()` - returns the number of null values in the dataframe.
19 | * `df.x.value_counts()` returns the number of values for each category in x series. The `normalize=True` argument retrieves the percentage of each category. In this project, the mean of churn is equal to the churn rate obtained with the value_counts method.
20 | * `round(x, y)` - round an x number with y decimal places
21 | * `df[x].nunique()` - returns the number of unique values in x series
22 |
23 | The entire code of this project is available in [this jupyter notebook](https://github.com/alexeygrigorev/mlbookcamp-code/blob/master/chapter-03-churn-prediction/03-churn.ipynb).
24 |
25 | | ⚠️ | 28 |
29 | The notes are written by the community. 30 | If you see an error here, please create a PR with a fix. 31 | |
32 |
5 |
6 | [Slides](https://www.slideshare.net/AlexeyGrigorev/ml-zoomcamp-3-machine-learning-for-classification)
7 |
8 |
9 | ## Notes
10 |
11 | Mutual information is a concept from information theory, which measures how much we can learn about one variable if we know the value of another. In this project, we can think of this as how much do we learn about churn if we have the information from a particular feature. So, it is a measure of the importance of a categorical variable.
12 |
13 | **Classes, functions, and methods:**
14 |
15 | * `mutual_info_score(x, y)` - Scikit-Learn class for calculating the mutual information between the x target variable and y feature.
16 | * `df[x].apply(y)` - apply a y function to the x series of the df dataframe.
17 | * ` df.sort_values(ascending=False).to_frame(name='x')` - sort values in an ascending order and called the column as x.
18 |
19 | The entire code of this project is available in [this jupyter notebook](https://github.com/alexeygrigorev/mlbookcamp-code/blob/master/chapter-03-churn-prediction/03-churn.ipynb).
20 |
21 | | ⚠️ | 24 |
25 | The notes are written by the community. 26 | If you see an error here, please create a PR with a fix. 27 | |
28 |
5 |
6 | [Slides](https://www.slideshare.net/AlexeyGrigorev/ml-zoomcamp-3-machine-learning-for-classification)
7 |
8 |
9 | ## Notes
10 |
11 | One-Hot Encoding allows encoding categorical variables in numerical ones. This method represents each category of a variable as one column, and a 1 is assigned if the value belongs to the category or 0 otherwise.
12 |
13 | **Classes, functions, and methods:**
14 |
15 | * `df[x].to_dict(orient='records')` - convert x series to dictionaries, oriented by rows.
16 | * `DictVectorizer().fit_transform(x)` - Scikit-Learn class for one-hot encoding by converting x dictionaries into a sparse matrix. It does not affect the numerical variables.
17 | * `DictVectorizer().get_feature_names()` - return the names of the columns in the sparse matrix.
18 |
19 | The entire code of this project is available in [this jupyter notebook](https://github.com/alexeygrigorev/mlbookcamp-code/blob/master/chapter-03-churn-prediction/03-churn.ipynb).
20 |
21 | | ⚠️ | 24 |
25 | The notes are written by the community. 26 | If you see an error here, please create a PR with a fix. 27 | |
28 |
5 |
6 | [Slides](https://www.slideshare.net/AlexeyGrigorev/ml-zoomcamp-3-machine-learning-for-classification)
7 |
8 |
9 | ## Notes
10 |
11 | This video was about training a logistic regression model with Scikit-Learn, applying it to the validation dataset, and calculating its accuracy.
12 |
13 | **Classes, functions, and methods:**
14 |
15 | * `LogisticRegression().fit(x)` - Scikit-Learn class for training the logistic regression model.
16 | * `LogisticRegression().coef_[0]` - return the coefficients or weights of the LR model
17 | * `LogisticRegression().intercept_[0]` - return the bias or intercept of the LR model
18 | * `LogisticRegression().predict[x]` - make predictions on the x dataset
19 | * `LogisticRegression().predict_proba[x]` - make predictions on the x dataset by returning two columns with their probabilities for the two categories - soft predictions
20 |
21 | The entire code of this project is available in [this jupyter notebook](https://github.com/alexeygrigorev/mlbookcamp-code/blob/master/chapter-03-churn-prediction/03-churn.ipynb).
22 |
23 | | ⚠️ | 26 |
27 | The notes are written by the community. 28 | If you see an error here, please create a PR with a fix. 29 | |
30 |
5 |
6 | [Slides](https://www.slideshare.net/AlexeyGrigorev/ml-zoomcamp-3-machine-learning-for-classification)
7 |
8 |
9 | ## Notes
10 |
11 | This video was about the interpretation of coefficients, and training a model with fewer features.
12 |
13 | In the formula of the logistic regression model, only one of the one-hot encoded categories is multiplied by 1, and the other by 0. In this way, we only consider the appropriate category for each categorical feature.
14 |
15 | **Classes, functions, and methods:**
16 |
17 | * `zip(x,y)` - returns a new list with elements from x joined with their corresponding elements on y
18 |
19 | The entire code of this project is available in [this jupyter notebook](https://github.com/alexeygrigorev/mlbookcamp-code/blob/master/chapter-03-churn-prediction/03-churn.ipynb).
20 |
21 | | ⚠️ | 24 |
25 | The notes are written by the community. 26 | If you see an error here, please create a PR with a fix. 27 | |
28 |
5 |
6 | [Slides](https://www.slideshare.net/AlexeyGrigorev/ml-zoomcamp-3-machine-learning-for-classification)
7 |
8 |
9 | ## Notes
10 |
11 | We trained the logistic regression model with the full training dataset (training + validation), considering numerical and categorical features. Thus, predictions were made on the test dataset, and we evaluated the model using the accuracy metric.
12 |
13 | In this case, the predictions of validation and test were similar, which means that the model is working well.
14 |
15 | | ⚠️ | 18 |
19 | The notes are written by the community. 20 | If you see an error here, please create a PR with a fix. 21 | |
22 |
5 |
6 | [Slides](https://www.slideshare.net/AlexeyGrigorev/ml-zoomcamp-3-machine-learning-for-classification)
7 |
8 |
9 | ## Notes
10 |
11 | In this session, we worked on a project to predict churning in customers from a company. We learned the feature importance of numerical and categorical variables, including risk ratio, mutual information, and correlation coefficient. Also, we understood one-hot encoding and implemented logistic regression with Scikit-Learn.
12 |
13 | | ⚠️ | 16 |
17 | The notes are written by the community. 18 | If you see an error here, please create a PR with a fix. 19 | |
20 |
5 |
6 |
7 | [Slides](https://www.slideshare.net/AlexeyGrigorev/ml-zoomcamp-4-evaluation-metrics-for-classification)
8 |
9 |
10 | ## Notes
11 |
12 | The fourth week of Machine Learning Zoomcamp is about different metrics to evaluate a binary classifier. These measures include accuracy, confusion table, precision, recall, ROC curves(TPR, FRP, random model, and ideal model), AUROC, and cross-validation.
13 |
14 | For this project, we used a [Kaggle dataset](https://www.kaggle.com/blastchar/telco-customer-churn) about churn prediction.
15 |
16 | Add notes from the video (PRs are welcome)
17 |
18 | | ⚠️ | 21 |
22 | The notes are written by the community. 23 | If you see an error here, please create a PR with a fix. 24 | |
25 |
4 |
5 | [Slides](https://www.slideshare.net/AlexeyGrigorev/ml-zoomcamp-4-evaluation-metrics-for-classification)
6 |
7 |
8 | ## Notes
9 |
10 | The Area under the ROC curves can tell us how good is our model with a single value. The AUROC of a random model is 0.5, while for an ideal one is 1.
11 |
12 | In other words, AUC can be interpreted as the probability that a randomly selected positive example has a greater score than a randomly selected negative example.
13 |
14 | **Classes and methods:**
15 |
16 | * `auc(x, y)` - sklearn.metrics class for calculating area under the curve of the x and y datasets. For ROC curves x would be false positive rate, and y true positive rate.
17 | * `roc_auc_score(x, y)` - sklearn.metrics class for calculating area under the ROC curves of the x false positive rate and y true positive rate datasets.
18 | * `randint(x, y, size=z)` - np.random class for generating random integers from the “discrete uniform” distribution; from `x` (inclusive) to `y` (exclusive) of size `z`.
19 |
20 | The entire code of this project is available in [this jupyter notebook](notebook.ipynb).
21 |
22 | Add notes from the video (PRs are welcome)
23 |
24 | | ⚠️ | 27 |
28 | The notes are written by the community. 29 | If you see an error here, please create a PR with a fix. 30 | |
31 |
4 |
5 | [Slides](https://www.slideshare.net/AlexeyGrigorev/ml-zoomcamp-4-evaluation-metrics-for-classification)
6 |
7 |
8 | ## Notes
9 |
10 | General definitions:
11 |
12 | * **Metric:** A single number that describes the performance of a model
13 | * **Accuracy:** Fraction of correct answers; sometimes misleading
14 | * Precision and recall are less misleading when we have class imbalance
15 | * **ROC Curve:** A way to evaluate the performance at all thresholds; okay to use with imbalance
16 | * **K-Fold CV:** More reliable estimate for performance (mean + std)
17 |
18 | In brief, this weeks was about different metrics to evaluate a binary classifier. These measures included accuracy, confusion table, precision, recall, ROC curves(TPR, FPR, random model, and ideal model), and AUROC. Also, we talked about a different way to estimate the performance of the model and make the parameter tuning with cross-validation.
19 |
20 | The code of this project is available in [this jupyter notebook](https://github.com/alexeygrigorev/mlbookcamp-code/blob/master/course-zoomcamp/04-evaluation/notebook.ipynb).
21 |
22 | Add notes from the video (PRs are welcome)
23 |
24 | | ⚠️ | 27 |
28 | The notes are written by the community. 29 | If you see an error here, please create a PR with a fix. 30 | |
31 |
4 |
5 | [Slides](https://www.slideshare.net/AlexeyGrigorev/ml-zoomcamp-5-model-deployment)
6 |
7 |
8 | ## Notes
9 | In this chapter we learned these topics:
10 | - We learned how to save the model and load it to re-use it without running the previous code.
11 | - How to deploy the model in a web service.
12 | - How to create a virtual environment.
13 | - How to create a container and run our code in any operating systems.
14 | - How to deploy our code in a public web service and access it externally from outside a local computer.
15 |
16 | In the next chapter we will learn the algorithms such as Decision trees, Random forests and Gradient boosting as an alternative way of combining decision tress.
17 |
18 | Add notes from the video (PRs are welcome)
19 |
20 |
21 | | ⚠️ | 24 |
25 | The notes are written by the community. 26 | If you see an error here, please create a PR with a fix. 27 | |
28 |
| ⚠️ | 15 |
16 | The notes are written by the community. 17 | If you see an error here, please create a PR with a fix. 18 | |
19 |
5 |
6 |
7 | [Slides](https://www.slideshare.net/AlexeyGrigorev/ml-zoomcamp-6-decision-trees-and-ensemble-learning)
8 |
9 |
10 | ## Notes
11 |
12 | In this session we'll learn about decision trees and ensemble learning algorithms. The questions that we try to address this week are, "What are decision trees? How are they different from ensemble algorithms? How can we implement and fine-tune these models to make binary classification predictions?"
13 |
14 | To be specific, we'll use [credit scoring data](https://github.com/gastonstat/CreditScoring) to build a model that predicts whether a bank should lend loan to a client or not. The bank takes these decisions based on the historical record.
15 |
16 | In the credit scoring classification problem,
17 | - if the model returns 0, this means, the client is very likely to payback the loan and the bank will approve the loan.
18 | - if the model returns 1, then the client is considered as a `defaulter` and the bank may not approve the loan.
19 |
20 | Add notes from the video (PRs are welcome)
21 |
22 |
23 | | ⚠️ | 26 |
27 | The notes are written by the community. 28 | If you see an error here, please create a PR with a fix. 29 | |
30 |
4 |
5 | [Slides](https://www.slideshare.net/AlexeyGrigorev/ml-zoomcamp-6-decision-trees-and-ensemble-learning)
6 |
7 |
8 | ## Notes
9 |
10 | We select the final model from decision tree, random forest, or xgboost based on the best auc scores. After that we prepare the `df_full_train` and `df_test` to train and evaluate the final model. If there is not much difference between model auc scores on the train as well as test data then the model has generalized the patterns well enough.
11 |
12 | Generally, XGBoost models perform better on tabular data than other machine learning models but the downside is that these model are easy to overfit cause of the high number of hyperparameter. Therefore, XGBoost models require a lot more attention for parameters tuning to optimize them.
13 |
14 | Add notes from the video (PRs are welcome)
15 |
16 |
17 | | ⚠️ | 20 |
21 | The notes are written by the community. 22 | If you see an error here, please create a PR with a fix. 23 | |
24 |
4 |
5 | [Slides](https://www.slideshare.net/AlexeyGrigorev/ml-zoomcamp-6-decision-trees-and-ensemble-learning)
6 |
7 |
8 | ## Notes
9 |
10 | - Decision trees learn if-then-else rules from data.
11 | - Finding the best split: select the least impure split. This algorithm can overfit, that's why we control it by limiting the max depth and the size of the group.
12 | - Random forest is a way of combining multiple decision trees. It should have a diverse set of models to make good predictions.
13 | - Gradient boosting trains model sequentially: each model tries to fix errors of the previous model. XGBoost is an implementation of gradient boosting.
14 |
15 | Add notes from the video (PRs are welcome)
16 |
17 |
18 | | ⚠️ | 21 |
22 | The notes are written by the community. 23 | If you see an error here, please create a PR with a fix. 24 | |
25 |
| ⚠️ | 20 |
21 | The notes are written by the community. 22 | If you see an error here, please create a PR with a fix. 23 | |
24 |
4 |
5 | [Slides](https://www.slideshare.net/AlexeyGrigorev/ml-zoomcamp-8-neural-networks-and-deep-learning-250592316)
6 |
7 |
8 | `ModelCheckpoint` callback is used with training the model to save a model or weights in a checkpoint file at some interval, so the model or weights can be loaded later to continue the training from the state saved or to use for deployment.
9 |
10 | **Classes, function, and attributes**:
11 |
12 | - `keras.callbacks.ModelCheckpoint`: ModelCheckpoint class from keras callbacks api
13 | - `filepath`: path to save the model file
14 | - `monitor`: the metric name to monitor
15 | - `save_best_only`: only save when the model is considered the best according to the metric provided in `monitor`
16 | - `mode`: overwrite the save file based on either maximum or the minimum scores according the metric provided in `monitor`
17 |
18 | ## Notes
19 |
20 | Add notes from the video (PRs are welcome)
21 |
22 | * checkpointing saves the model after each training iteration
23 | * checkpoint conditions may include reaching the best performance
24 | * keras callbacks
25 |
26 | | ⚠️ | 29 |
30 | The notes are written by the community. 31 | If you see an error here, please create a PR with a fix. 32 | |
33 |
4 |
5 | [Slides](https://www.slideshare.net/AlexeyGrigorev/ml-zoomcamp-8-neural-networks-and-deep-learning-250592316)
6 |
7 |
8 | In this section we increase the image input size from `150` to `299`, reduce the amount of data augmentation parameters and lower the learning rate. This gives us the best results than any previous experiments.
9 |
10 | ## Notes
11 |
12 | Add notes from the video (PRs are welcome)
13 |
14 |
15 | | ⚠️ | 18 |
19 | The notes are written by the community. 20 | If you see an error here, please create a PR with a fix. 21 | |
22 |
4 |
5 | - We can use pre-trained models for general image classification
6 | - Convolutional layers let us turn an image into a vector
7 | - Dense layers use the vector to make the predictions
8 | - Instead of training a model from scratch, we can use transfer learning and re-use already trained convolutional layers
9 | - First, train a small model (150x150) before training a big one (299x299)
10 | - Learning rate - how fast the model trains. Fast learner aren't always best ones
11 | - We can save the best model using callbacks and checkpointing
12 | - To avoid overfitting, use dropout and augmentation
13 |
14 | ## Notes
15 |
16 | Add notes from the video (PRs are welcome)
17 |
18 |
19 | | ⚠️ | 22 |
23 | The notes are written by the community. 24 | If you see an error here, please create a PR with a fix. 25 | |
26 |
5 |
6 | In the last session, we built and trained a clothes classification deep learning model using `Keras` and `TensorFlow`. This session focuses on deploying it. The model categorizes images of clothing items (e.g., 👕 t-shirts, 👖 pants, etc.) uploaded by users on a website. Deployment will be done using **AWS Lambda**, a serverless solution to execute code without managing servers, and instead of `TensorFlow`, we will use `TensorFlow-lite`.
7 |
8 | Refer to [updates.md](updates.md) for info on running TF lite
9 | in 2024.
10 |
11 |
12 | ## Notes
13 |
14 | Add notes from the video (PRs are welcome)
15 |
16 | * introduction to the topic of the week: deploying a deep learning model to the cloud, aws lambda and tensorflow lite
17 |
18 | | ⚠️ | 21 |
22 | The notes are written by the community. 23 | If you see an error here, please create a PR with a fix. 24 | |
25 |
5 |
6 |
7 |
8 |
9 |
10 | ## Notes
11 |
12 | Add notes from the video (PRs are welcome)
13 |
14 |
15 | | ⚠️ | 18 |
19 | The notes are written by the community. 20 | If you see an error here, please create a PR with a fix. 21 | |
22 |
5 |
6 |
7 |
8 |
9 |
10 | ## Notes
11 |
12 | Add notes from the video (PRs are welcome)
13 |
14 |
15 | | ⚠️ | 18 |
19 | The notes are written by the community. 20 | If you see an error here, please create a PR with a fix. 21 | |
22 |
5 |
6 |
7 |
8 |
9 |
10 | ## Notes
11 |
12 | Add notes from the video (PRs are welcome)
13 |
14 |
15 | | ⚠️ | 18 |
19 | The notes are written by the community. 20 | If you see an error here, please create a PR with a fix. 21 | |
22 |
4 |
5 |
6 | ## Notes
7 |
8 | Add notes from the video (PRs are welcome)
9 |
10 |
11 | | ⚠️ | 14 |
15 | The notes are written by the community. 16 | If you see an error here, please create a PR with a fix. 17 | |
18 |
5 |
6 |
7 | [Slides](https://www.slideshare.net/AlexeyGrigorev/ml-zoomcamp-10-kubernetes)
8 |
9 |
10 | ## Notes
11 |
12 | Add notes from the video (PRs are welcome)
13 |
14 | * same use case as in the session before: classifying images of clothes
15 | * using tensorflow serving, written in C++, with focus on inference
16 | * gRPC binary protocol
17 | * deploying to kubernetes
18 | * 1st component: gateway (download image, resize, turn into numpy array - computationally not expensive - can be done with CPU)
19 | * 2nd component: model (matrix multiplications - computationally expensive - thus use GPU)
20 | * scaling the two components independently: i.e. 5 gateways handing images to 1 model
21 | * two components in two different docker container (lesson four)
22 | * kubernetes main concepts (lesson five)
23 | * running kubernetes on your local machine (lesson six)
24 | * deploy the two services to kubernetes (lesson seven)
25 | * move from local to cloud (lesson eight)
26 |
27 | | ⚠️ | 30 |
31 | The notes are written by the community. 32 | If you see an error here, please create a PR with a fix. 33 | |
34 |
4 |
5 | In the session 10 we covered the following concepts:
6 |
7 | - TF-Serving is a system for deploying TensorFlow models
8 | - When using TF-Serving, we need a component for pre-processing
9 | - Kubernetes is a container orchestration platform
10 | - To deploy something on Kubernetes, we need to specify a deployment and a service
11 | - You can use Docker compose and Kind for local experiments
12 |
13 | ## Notes
14 |
15 | Add notes from the video (PRs are welcome)
16 |
17 |
18 | | ⚠️ | 21 |
22 | The notes are written by the community. 23 | If you see an error here, please create a PR with a fix. 24 | |
25 |
| ⚠️ | 20 |
21 | The notes are written by the community. 22 | If you see an error here, please create a PR with a fix. 23 | |
24 |
5 |
6 |
7 |
8 |
9 |
10 | ## Notes
11 |
12 | Add notes from the video (PRs are welcome)
13 |
14 |
15 | | ⚠️ | 18 |
19 | The notes are written by the community. 20 | If you see an error here, please create a PR with a fix. 21 | |
22 |
5 |
6 |
7 |
8 |
9 |
10 | ## Notes
11 |
12 | Add notes from the video (PRs are welcome)
13 |
14 |
15 | | ⚠️ | 18 |
19 | The notes are written by the community. 20 | If you see an error here, please create a PR with a fix. 21 | |
22 |
5 |
6 |
7 |
8 |
9 |
10 | ## Notes
11 |
12 | Add notes from the video (PRs are welcome)
13 |
14 |
15 | | ⚠️ | 18 |
19 | The notes are written by the community. 20 | If you see an error here, please create a PR with a fix. 21 | |
22 |
5 |
6 |
7 |
8 |
9 |
10 | ## Notes
11 |
12 | Add notes from the video (PRs are welcome)
13 |
14 |
15 | | ⚠️ | 18 |
19 | The notes are written by the community. 20 | If you see an error here, please create a PR with a fix. 21 | |
22 |
5 |
6 |
7 |
8 |
9 |
10 | ## Notes
11 |
12 | Add notes from the video (PRs are welcome)
13 |
14 |
15 | | ⚠️ | 18 |
19 | The notes are written by the community. 20 | If you see an error here, please create a PR with a fix. 21 | |
22 |
5 |
6 |
7 |
8 |
9 |
10 | ## Notes
11 |
12 | Add notes from the video (PRs are welcome)
13 |
14 |
15 | | ⚠️ | 18 |
19 | The notes are written by the community. 20 | If you see an error here, please create a PR with a fix. 21 | |
22 |
5 |
6 |
7 |
8 |
9 |
10 | ## Notes
11 |
12 | Add notes from the video (PRs are welcome)
13 |
14 | | ⚠️ | 17 |
18 | The notes are written by the community. 19 | If you see an error here, please create a PR with a fix. 20 | |
21 |
| ⚠️ | 15 |
16 | The notes are written by the community. 17 | If you see an error here, please create a PR with a fix. 18 | |
19 |
| ⚠️ | 18 |
19 | The notes are written by the community. 20 | If you see an error here, please create a PR with a fix. 21 | |
22 |
20 |
21 | 
22 |
23 |
25 |
26 | You can watch this video to see how your learning in public posts may look like:
27 |
28 |
29 | 
30 |
31 |
32 |
33 | Send a PR if you want to suggest improvements for this document
34 |
--------------------------------------------------------------------------------