Basic Dictionary commands (2):

# Looping
for key in d:
    print(key)
for value in d.values():
    print(value)
for key, value in d.items():
    print(key, value)
    
# Copying
d_copy = d.copy()
d_copy = dict(d)

# Merging
d = {'name': 'Max', 'age': 28}
d1 = {'name': 'Anna', 'age': 27}
d.update(d1)
d = {**d, **d1}

# Dictionary Comprehension
d = {x: x**2 for x in range(10)}
d = {k: v**2 for k, v in zip(['a', 'b'], range(4))}
d = {k: v for k, v in d.items() if v % 2 == 0}

Copy and deep copy in dictionaries:

# copy and deepcopy
import copy
d = {'x': [1, 2, 3]}
d1 = d.copy()
d2 = copy.deepcopy(d)

d['x'].append(4)
print(d1) # {'x': [1, 2, 3, 4]}
print(d2) # {'x': [1, 2, 3]}

activate "sticky scroll" in vscode through ctrl+shirt+p panel.

#Datashader simplifies creating meaningful visuals from large datasets by breaking down the process into clear steps. It automatically generates accurate visualizations without the need for manual parameter tweaking. Computations are optimized using Python, Numba, Dask, and CUDA, making it efficient even with huge datasets on standard hardware.

https://datashader.org/
Gallery

The walrus operator, introduced in Python 3.8, is represented by ":=". It allows the assignment of a value to a variable within an expression.

## f(x) is called 3 times
foo = [f(x), f(x)**2, f(x)**3]

## two lines of code
y = f(x)
foo = [y, y**2, y**3]

## walrus operator
foo = [y := f(x), y**2, y**3]

Walrus operator [1]

# Avoiding inefficient comprehensions

results = []
for x in data:
    res = f(x)
    if res:
        results.append(res)
        
# f(x) is called twice
results = [f(x) for x in data if f(x)]        
# walrus operator
results = [res for x in data if (res := f(x))]

# Unnecessary variables in scope
match = pattern.search(data)
if match:
    do_sth(math)

if (match := pattern.search(data)):
    do_sth(match)
    
# Processing streams in chunks
chunk = file.read(8192)
while chunk:
    process(chunk)
    chunk = file.read(8192)

# walrus operator
while chunk := file.read(8192):
    process(chunk)

Datasets for machine learning typically contain a large number of features, but such high-dimensional feature spaces are not always helpful.

In general, all the features are not equally important and there are certain features that account for a large percentage of variance in the dataset. Dimensionality reduction algorithms aim to reduce the dimension of the feature space to a fraction of the original number of dimensions. In doing so, the features with high variance are still retained—but are in the transformed feature space. And principal component analysis (PCA) is one of the most popular dimensionality reduction algorithms.

Here's a simple example in Python demonstrating PCA for dimensionality reduction before training a scikit-learn classifier.
Github

You may also need to read more about PCA here.

Applications for Students & Teaching Assistants are Open!

1️⃣ 3-week Courses (July 8 - 26, 2024):
- Computational Neuroscience: Explore the intricacies of the brain's computational processes and join an engaging community of learners.
- Deep Learning: Delve into the world of machine learning, uncovering the principles and applications of deep learning.

2️⃣ 2-week Courses (July 15 - 26, 2024):
- Computational Tools for Climate Science: Uncover the tools and techniques driving climate science research in this dynamic two-week course.
- NeuroAI - Inaugural Year!: Be part of history as we launch our first-ever NeuroAI course, designed to explore the intersection of neuroscience and artificial intelligence.

https://neuromatch.io/courses/

Incremental principal component analysis (IPCA) is typically used as a replacement for principal component analysis (PCA) when the dataset to be decomposed is too large to fit in memory.

IPCA builds a low-rank approximation for the input data using an amount of memory which is independent of the number of input data samples. It is still dependent on the input data features, but changing the batch size allows for control of memory usage.

I have made some changes on the example from sklearn documentation so one does not need to load the whole dataset in memory.

python
X_ipca = np.zeros((X.shape[0], n_components))
for i in range(3):
    ipca.partial_fit(X[i*50:(i+1)*50])
    
for i in range(3):
    X_ipca[i*50:(i+1)*50] = ipca.transform(X[i*50:(i+1)*50])

GitHub

How can you create an audiobook with a natural human voice and a customized accent? Let's say you have an EPUB file and you're tired of the robotic voice generated by common text-to-speech (TTS) systems. One of the most advanced TTS technologies available today is provided by Openvoice. You can find more information about it here.

It performs optimally with a GPU, but it's also compatible with CPU. To use it on your own machine, simply set up a virtual environment and install the package. You'll also need to download a few additional files. I'm currently using the basic setup with the default voice, but the ability to clone any voice is an incredibly exciting feature.

follow the notebook demo1, extract text from epub and replace the sample test with your favourite book.

You may need to split the book into several chapters to fit into the gpu memory and avoid killing the job.

It took me about 10 min to make audio book from Shogun, a novel which is about 500 pages.

How to Use ZSH Auto-suggestions?

ZSH is a popular Unix shell that extends the Bourne Again Shell. It comes packed with features and improvements over Bash.
If you already zsh as default terminal just use:

# Linux
git clone https://github.com/zsh-users/zsh-autosuggestions ~/.zsh/zsh-autosuggestions
# add to .zshrc
source ~/.zsh/zsh-autosuggestions/zsh-autosuggestions.zsh
# Mac
brew install zsh-autosuggestions
# add to .zshrc
source $(brew --prefix)/share/zsh-autosuggestions/zsh-autosuggestions.zsh

Read more here.

Also for #JAX 😢
Credit: Geek_code

JAX is an open-source Python library developed by Google for high-performance numerical computing, especially suited for machine learning and scientific computing. It provides a combination of automatic differentiation, just-in-time compilation, and support for GPU/TPU acceleration, making it particularly well-suited for scalable and efficient computation on large datasets. JAX is built on top of the XLA (Accelerated Linear Algebra) compiler and is heavily inspired by NumPy, making it easy 🤨 for users familiar with NumPy to transition to JAX.

Let's practice some JAX:
I recommend start with this repo and the series of videos for start.

Videos
GitHub

Then you can move to the Deep learning book.

Deep Learning with JAX
Workshop JAX

I found this to be quite useful, and it might be beneficial for you as well. There's a YouTube video course available, along with a GitHub page focusing on R and Python.
PDF can be found here: @reza_jafari_ai

PyGWalker: A Python Library for Exploratory Data Analysis with Visualization
docs
colab

Some of them are not free.

Vectorizing in JAX

def dot(v1, v2):
    return jax.numpy.dot(v1, v2)

1️⃣ Naively vectorizing

dot_naive =[dot(v1, v2) for v1, v2 in zip(v1s, v2s)]

2️⃣ Manual vectorizing

def dot_vectorized(v1s, v2s):
    return jnp.einsum("ij,ij->i", v1s, v2s)

3️⃣ Automatic vectorizing

dot_vmapped = jax.vmap(dot)

⏰ Timing

%timeit [dot(v1, v2) for v1, v2 in zip(v1s, v2s)]
%timeit dot_vectorized(v1s, v2s).block_until_ready()
%timeit dot_vmapped(v1s, v2s).block_until_ready()

# 5.15 ms ± 54.3 µs per loop
# 135 µs ± 171 ns per loop
# 543 µs ± 1.38 µs per loop

Adding JIT

dot_vectorized_jitted = jax.jit(dot_vectorized)
dot_vmapped_jitted = jax.jit(dot_vmapped)

Timing

bash
6.5 µs ± 12.9 ns per loop
6.39 µs ± 13.4 ns per loop

Notebook
ویرگول

Statistical hypothesis tests
Github

Here are some of the most important and frequently used commands in scikit-learn (sklearn):

1. Model Selection:
- train_test_split(): Split arrays or matrices into random train and test subsets.
- cross_val_score(): Evaluate a score by cross-validation.
- GridSearchCV(): Exhaustive search over specified parameter values for an estimator.
- StratifiedKFold(): Provides train/test indices to split data into train/test sets while maintaining class distribution.

2. Preprocessing:
- StandardScaler(): Standardize features by removing the mean and scaling to unit variance.
- MinMaxScaler(): Transform features by scaling each feature to a given range.
- OneHotEncoder(): Encode categorical integer features as one-hot numeric arrays.

3. Model Building:
- LinearRegression(): Ordinary least squares Linear Regression.
- LogisticRegression(): Logistic Regression (for classification tasks).
- RandomForestClassifier(): Random Forest Classifier.
- RandomForestRegressor(): Random Forest Regressor.
- GradientBoostingClassifier(): Gradient Boosting Classifier.
- GradientBoostingRegressor(): Gradient Boosting Regressor.
- DecisionTreeClassifier(): Decision Tree Classifier.

4. Model Evaluation:
- accuracy_score(): Accuracy classification score.
- precision_score(), recall_score(), f1_score(): Compute precision, recall, F-measure, and support for classification.
- mean_squared_error(): Mean squared error regression loss.
- r2_score(): R^2 (coefficient of determination) regression score function.

5. Pipeline and Feature Union:
- Pipeline(): Chain multiple estimators into one.
- FeatureUnion(): Combine several transformer objects into a new transformer.

6. Dimensionality Reduction:
- PCA(): Principal Component Analysis.
- TruncatedSVD(): Dimensionality reduction using truncated singular value decomposition.

7. Clustering:
- KMeans(): K-Means clustering.
- AgglomerativeClustering(): Agglomerative hierarchical clustering.

These are just a few of the many functionalities provided by scikit-learn for machine learning tasks.