duoan/TorchCode

Features

• PyTorch Code Exercises - Provides a curated set of PyTorch exercises for implementing softmax, normalization, attention, and Transformer components from scratch in notebooks.
• Neural Network Operator Implementations - Has learners write softmax, normalization, and convolution operators from first principles and checks their numerical behavior and gradient flow.
• Practice Platforms - Offers a self-hostable Jupyter-based practice platform where learners implement PyTorch operators and architectures and get instant automated grading.
• Automated Grading Workflows - Imports a learner's submitted function and runs it against several test cases for numerical correctness, gradient flow and edge cases, then records pass or fail.
• Embedded Notebook Exercises - Delivers each problem as a self-contained notebook with a blank template that learners fill in and run in the same document.
• Coding Interview Preparation - Offers hands-on PyTorch coding problems mirroring machine learning interview questions, with a weekly study plan and instant feedback.
• Jupyter Notebook Curricula - Delivers each problem as a self-contained Jupyter notebook with a blank template, hints, and a reference solution.
• Multiple Choice Grading - Runs submitted functions against multiple test cases and reports a colored pass or fail for each test.
• Practice Problem Sets - Provides a broad set of notebook practice problems where learners implement operators, attention, and language models from scratch.
• Grading Systems - Runs submitted functions against multiple test cases checking correctness, gradients, and edge cases, then reports pass or fail without a human reviewer.
• From-Scratch Decoder Implementations - Has learners build attention, GPT-2 style blocks, and mixture-of-experts layers by hand instead of calling library code.
• Decoding Sampling Strategies - Covers top-k and top-p sampling, beam search and speculative decoding exercises so learners see how text generation works internally.
• Gradient Verification Tools - Verifies submitted implementations by backpropagating through them and comparing the resulting gradients, catching code that looks correct but breaks training.
• Language Model Training and Decoding Techniques - Includes exercises on optimizers, learning-rate schedules, alignment losses, beam search, and top-k and top-p sampling.
• Transformer Component Exercises - Challenges learners to build language-model blocks, expert-routing layers, low-rank adapters and vision patch embeddings from primitive operations.
• Multi-Head Attention Mechanisms - Guides learners to implement scaled dot-product, multi-head, causal, grouped-query and sliding-window attention variants from scratch.
• Training Algorithm Exercises - Provides exercises for writing optimizers, learning-rate schedules, gradient clipping, gradient accumulation and alignment losses without built-in library versions.
• Self-Contained Judge Deployments - Packages the notebook server and judge into one container image that starts with a single command.
• Exercise-Solution Pairings - Pairs every problem with a hint layer and a full reference solution notebook to study after attempting it.
• Instant Answer Feedback - Returns immediate pass or fail results after a learner submits an implementation against automated tests.
• LeetCode Solution References - Offers gentle hints first and a complete reference solution afterward for every practice problem.
• Notebook Practice Environments - Runs the full notebook practice environment in a single container with one command, needing no signup, database, or GPU.
• Practice Progress Tracking - Records solved problems, best completion times, and attempt counts locally so learners can measure improvement across sessions.
• Implementation Practice - Offers curated problems where learners implement softmax, normalization layers, and convolutions from scratch in notebooks, mirroring skills tested in interviews.
• Interview Study Paths - Lays out a weekly study plan with time estimates from foundations through attention, architectures and inference to prepare for interviews.
• Output Comparison Judges - Ships a judge that imports submitted functions and runs them against predefined test cases, checking numerical results and autograd gradients.

Star history