PyTorch Patterns
Idiomatic PyTorch: device-agnostic code, reproducible training loops, efficient data loading, and speed.
- What
- Idiomatic PyTorch: device-agnostic code, reproducible training loops, efficient data loading, and speed.
- Cost
- Free
- Needs
- Use "PyTorch Patterns" with your Muse.
- Install
- Copy the installer prompt below into your Muse — your agent does the rest.
Curated by Skill Harbor: idiomatic PyTorch patterns for building robust, efficient, and reproducible deep learning applications. Three core principles: device-agnostic code (never hardcode .cuda(), always route through a device variable), reproducibility first (seed torch, CUDA, numpy, and random together, deterministic cuDNN), and explicit shape management with annotated tensor shapes in every forward pass. Covers clean nn.Module structure with explicit weight initialization (kaiming for Linear/Conv, ones/zeros for BatchNorm), complete training loops (mixed precision with GradScaler, gradient clipping, zero_grad(set_to_none=True)), proper validation (always model.eval() with torch.no_grad(), never leave dropout active), efficient DataLoader configuration (num_workers, pin_memory, persistent_workers, drop_last), custom datasets and collate functions for variable-length data, full checkpointing (model plus optimizer plus epoch state, weights_only=True for secure loading), and performance (AMP, gradient checkpointing trading compute for memory, torch.compile in PyTorch 2.0+). Includes an anti-pattern table: forgetting eval mode during validation, in-place ops breaking autograd, calling .item() before backward, moving the model to GPU inside the loop. By @affaan-m, listed here with credit to its creator. From the affaan-m/ECC repository (MIT). Honest caveats: pure guidance, nothing to install; you need Python with PyTorch installed, and a GPU to benefit from the performance sections. Skill Harbor never reviews the code, review it yourself before use.
Version:
Install
Copy the install package below, then paste it into MuseWe automatically scan every listing and flag anything suspicious, but we can't review every line of community code. Here's how to check a build yourself before you install it.
Use "PyTorch Patterns" with your Muse. Prerequisites: the Muse app (mobile or web). Pure guidance skill, nothing to install. Python with PyTorch installed (a GPU is recommended for the performance sections). 1. Open the skill: https://github.com/affaan-m/ECC/blob/main/skills/pytorch-patterns/SKILL.md and copy the full SKILL.md text. 2. Paste it into a chat with Muse and add: "Review [my training loop / my model / my data loading] against these PyTorch patterns: [paste your code]." 3. Ask Muse to flag violations (hardcoded .cuda(), missing eval mode in validation, no seed control, slow default DataLoader) and propose the corrected code. Tip: if a run will not reproduce, start with the reproducibility checklist (seeds, deterministic cuDNN, pinned data order) before touching the model. Safety: a skill is plain-text instructions; it runs nothing by itself. Review any code before running it, never paste secrets into a chat, and review anything Muse proposes before it acts.
Saved to your recent installs. Find it anytime on /connect.
Questions
How do I install a build?
Every product page includes a copy-paste install prompt. Paste it into your Muse and it sets the build up for you — no manual configuration.
Where does my money go?
Straight to the seller. Skill Harbor never processes payments: checkout happens on the seller’s own page, usually Stripe.
What does the ✓ next to a creator’s name mean?
It means we confirmed the identity of the person behind the listing. It says nothing about the code itself — always check a build before installing it.