Skip to content

knowledge.md

Latest commit

 

History

History
52 lines (37 loc) · 2.59 KB

knowledge.md

File metadata and controls

52 lines (37 loc) · 2.59 KB
Raw

TensorflowAI Working Context

Repository shape

  • Root: /home/shuvrajeet/Documents/TensorflowAI
  • Main areas: ComputerVision, NaturalLanguageProcessing, ReinforcementLearning.
  • TensorFlow-first research code organized as numbered learning/project families.
  • Many model families use standalone sibling folders.

Usual project structure

Most standalone implementations contain some combination of:

  • config.py — paths, hyperparameters, runtime settings
  • dataset.py — local loading and preprocessing
  • model.py — architecture and model-specific learning logic
  • train_and_test.py or train.py — execution entrypoint
  • run.sh or train.sh — repeatable command-line workflow
  • logs/, checkpoints/, artifacts/, samples/ — experiment outputs

Working conventions

  • Inspect the target folder and neighboring sibling before editing.
  • Preserve the existing folder-local style and public CLI/run behavior.
  • Do not introduce shared helpers or broad refactors unless explicitly requested.
  • Prefer resumable workflows and reuse valid artifacts; use explicit force/reset options for rebuilds.
  • Keep README, config, CLI, and path documentation synchronized with interface changes.
  • Put progress reporting at the slowest meaningful operation without flooding logs.
  • Use syntax/static checks as limited evidence; verify runtime, dependency, device, GPU, and multi-GPU behavior separately.
  • Keep TensorFlow as the default stack where practical.
  • For experiments, use the requested fixed environment/algorithm/support pairing and report aggregate results honestly.

Repository cautions

  • This workspace contains nested Git repositories and substantial pre-existing generated/cache changes.
  • Preserve unrelated user changes; inspect status and diffs narrowly around the requested target.
  • Generated logs, images, checkpoints, caches, and datasets should not be treated as source intent.
  • Exact paths, entrypoints, tracebacks, and current checkout contents take precedence over assumptions or older notes.

Project families observed

  • Computer vision: diffusion, GANs, image classification, and object detection.
  • NLP: attention, sequence models, transformers, pretrained models, and RedditStory.
  • Reinforcement learning: standalone environments, policy-gradient/REINFORCE variants, vectorized agents, and game projects.

How to use this file

Add project-specific decisions, conventions, commands, paths, known failures, and validation boundaries below. Keep entries short and dated so future prompts can provide this file as context.

Project trail