skills.rest
jasrajtulsi
jasrajtulsiCOMMUNITY
Explore all skills by this author →
GitHub

diffdock

Detect gradient-flow degradation in deep convolutional training and link per-layer anomalies to validation-performance drops.

Updated Jul 1, 2026
Search Tags:#gradient monitoring#cnn training#training instability#vanishing gradients#exploding gradients#lead time analysis#research debugging
One-click install
npx skills add https://github.com/jasrajtulsi/GRAD-SCOPE --skill diffdock-jasrajtulsi
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: diffdock
Source: https://github.com/jasrajtulsi/GRAD-SCOPE/tree/main/.claude/skills/diffdock
Command: npx skills add https://github.com/jasrajtulsi/GRAD-SCOPE --skill diffdock-jasrajtulsi

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill requires pandas, rdkit, torch, torch-geometric, torch-scatter, torch-sparse, torch-cluster, numpy, scipy, biopython, pytorch-lightning, pyyaml, esm, and includes scripts (resource) and references (resource) and assets (resource) components.

What problem does it solve?

GRAD-SCOPE helps researchers detect when deep CNN training is degrading by tracking per-layer gradient-flow signals and connecting gradient anomalies to later validation-performance drops.

Core Features & Use Cases

  • Logs lightweight per-layer gradient statistics at every training step using hooks.
  • Classifies layers by architecture type and normalized depth for comparable analysis across models.
  • Measures lead time between vanishing, exploding, or collapsed gradients and visible training degradation.
  • Use it to monitor CIFAR-style ResNet and VGG experiments, diagnose instability, and summarize results for research analysis.

Quick Start

Use the GRAD-SCOPE skill to analyze a training run and summarize per-layer gradient anomalies with their lead time to validation performance drops.

Frequently Asked Questions about diffdock

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I detect vanishing gradients in my CNN training before validation performance drops?▼

Detect vanishing gradients in CNN training by logging lightweight per-layer gradient statistics at every step. This approach measures lead time between gradient anomalies and visible training degradation, diagnosing instability early.

Why does my ResNet training fail and how can I analyze gradient flow across layers?▼

Analyze ResNet training failure by tracking per-layer gradient-flow signals and classifying layers by architecture type and normalized depth. This connects gradient anomalies to validation performance drops for comparable analysis across models.

Can I use PyTorch hooks to monitor exploding gradients in VGG experiments?▼

Yes, you can monitor exploding gradients in VGG experiments using PyTorch hooks to log per-step gradient statistics. This identifies training instability by comparing gradient dynamics across different network depths.

What is gradient collapse and how do I spot it during deep convolutional network training?▼

Gradient collapse is the degradation of gradient flow in deep convolutional training. Spot it by tracking per-layer gradient statistics with hooks, which links gradient anomalies to later validation-performance drops.

Do I need PyTorch Geometric and RDKit to run gradient monitoring scripts for research debugging?▼

You need PyTorch, PyTorch Geometric, and pandas for scripted validation workflows. RDKit and other dependencies support optional functionality, while the core focus remains on analyzing training instability and gradient dynamics.

What is the best way to compare gradient dynamics across different model depths?▼

Compare gradient dynamics across model depths by classifying layers according to architecture type and normalized depth. This method logs per-step gradient statistics to summarize per-layer anomalies for research analysis.

Related Skills

Explore More Education & Research Skills→
FeRhodiumFeRhodium
7

training-migration

Migrate PyTorch training code to Ascend NPU with AMP and HCCL.

Community
Advanced
dotlab-hqdotlab-hq

cudf-analytics

Perform GPU-accelerated groupby aggregations and statistical summaries on large CSV datasets.

Official
Intermediate
ChimeraFoundationaChimeraFoundationa

pytorch-lightning

Abstract PyTorch training boilerplate into LightningModule and Trainer workflows.

Community
Advanced
jasrajtulsijasrajtulsi

stable-baselines3

Instrument per-layer gradient dynamics in deep CNN training runs.

Community
Advanced