preflight-ml 0.1.0

Pre-flight checks for PyTorch pipelines. Catch silent failures before they waste your GPU.

preflight

Pre-flight checks for PyTorch pipelines. Catch silent failures before they waste your GPU.

---

Most deep learning bugs don't crash your training loop — they silently produce a garbage model. NaNs in your data, labels leaking between train and val, wrong channel ordering, dead gradients. You won't know until hours later, after the GPU bill has landed.

preflight is a pre-training validation tool you run in 30 seconds before starting any training job. It's not a linter. It's a pre-flight check — the kind pilots run before the expensive thing takes off.

---

Install

pip install preflight-ml

Quickstart

Create a small Python file that exposes your dataloader:

# my_dataloader.py
import torch
from torch.utils.data import DataLoader, TensorDataset

x = torch.randn(200, 3, 224, 224)
y = torch.randint(0, 10, (200,))
dataloader = DataLoader(TensorDataset(x, y), batch_size=32)

Run preflight:

preflight run --dataloader my_dataloader.py

Output:

preflight — pre-training check report
╭────────────────────────┬──────────┬────────┬──────────────────────────────────────────────────╮
│ Check                  │ Severity │ Status │ Message                                          │
├────────────────────────┼──────────┼────────┼──────────────────────────────────────────────────┤
│ nan_inf_detection      │ FATAL    │ PASS   │ No NaN or Inf values found in 10 sampled batches │
│ normalisation_sanity   │ WARN     │ PASS   │ Normalisation looks reasonable (mean=0.001)      │
│ channel_ordering       │ WARN     │ PASS   │ Channel ordering looks correct (NCHW)            │
│ label_leakage          │ FATAL    │ PASS   │ No val_dataloader provided — skipped             │
│ split_sizes            │ INFO     │ PASS   │ train=200 samples                                │
│ vram_estimation        │ WARN     │ INFO   │ No CUDA GPU detected — skipped                   │
│ class_imbalance        │ WARN     │ PASS   │ Class distribution looks balanced                │
│ shape_mismatch         │ FATAL    │ PASS   │ No model provided — skipped                      │
│ gradient_check         │ FATAL    │ PASS   │ No model+loss provided — skipped                 │
╰────────────────────────┴──────────┴────────┴──────────────────────────────────────────────────╯

  0 fatal  0 warnings  9 passed

Pre-flight passed. Safe to start training.

Checks

preflight runs 10 checks across three severity tiers. A FATAL failure exits with code 1 and blocks CI.

Check	Severity	What it catches
nan_inf_detection	FATAL	NaN or Inf values anywhere in sampled batches
label_leakage	FATAL	Samples appearing in both train and val sets
shape_mismatch	FATAL	Dataset output shape incompatible with model input
gradient_check	FATAL	Zero gradients, dead layers, exploding gradients
normalisation_sanity	WARN	Data that looks unnormalised (raw pixel values etc.)
channel_ordering	WARN	NHWC tensors when PyTorch expects NCHW
vram_estimation	WARN	Estimated peak VRAM exceeds 90% of GPU memory
class_imbalance	WARN	Severe class imbalance beyond configurable threshold
split_sizes	INFO	Empty or degenerate train/val splits
duplicate_samples	INFO	Identical samples within a split

With a model

Pass a model file to enable shape, gradient, and VRAM checks:

# my_model.py
import torch.nn as nn
model = nn.Sequential(nn.Flatten(), nn.Linear(3 * 224 * 224, 10))

# my_loss.py
import torch.nn as nn
loss_fn = nn.CrossEntropyLoss()

preflight run \
  --dataloader my_dataloader.py \
  --model my_model.py \
  --loss my_loss.py \
  --val-dataloader my_val_dataloader.py

Configuration

Add a .preflight.toml to your repo root to configure thresholds and disable checks:

[thresholds]
imbalance_threshold = 0.05
nan_sample_batches = 20

[checks]
vram_estimation = false

[ignore]
# check = "class_imbalance"
# reason = "intentional: rare event dataset"

CI integration

Add to your GitHub Actions workflow:

- name: Install preflight
  run: pip install preflight-ml

- name: Run pre-flight checks
  run: preflight run --dataloader scripts/dataloader.py --format json

The --format json flag outputs machine-readable results. Exit code is 1 if any FATAL check fails, 0 otherwise.

List all checks

preflight checks

What preflight does NOT do

• It does not replace unit tests. Use pytest for code logic.
• It does not guarantee a correct model. Passing preflight is a minimum safety bar, not a certification.
• It does not run your full training loop. Use it as a gate before training starts.
• It does not modify your code unless you pass --fix.

Roadmap

• --fix flag — auto-patch common issues (channel ordering, normalisation)
• Dataset snapshot + drift detection (preflight diff baseline.json new_data.pt)
• Full dry-run mode (one batch through model + loss + backward)
• Jupyter magic command (%load_ext preflight)
• preflight-monai plugin for medical imaging checks
• preflight-sktime plugin for time series checks

Contributing

See CONTRIBUTING.md. New checks are welcome — each one needs a passing test, a failing test, and a fix hint.

License

MIT — see LICENSE.