You can deploy this model and start earning money today!

Shiym/
anyflow-biomegacorr-stage1-flow-50k

Name: anyflow-biomegacorr-stage1-flow-50k
Availability: InStock

Views

No views yet

AnyFlow Bi-OmegaCorr — Stage 1 Flow (50k)

This repository contains the exact Stage 1 checkpoint from the AnyFlow / Bi-OmegaCorr curriculum after 50,000 optical-flow training steps.

Artifact

Field	Value
File	`anyflow_biomegacorr_stage1_flow_50k.pt`
Size	4,731,466,401 bytes
SHA256	`9db208657c7f18ad8392a0e7db963661c9bf08030728452db6744610b1fbe045`
Step	50,000
Epoch	3
Batch in epoch	6,224
Final logged minibatch loss	0.05562
Final logged minibatch `epe_i2j`	10.66567

The final loss and EPE values are the last logged training minibatch, not a held-out validation result.

Lineage

Training code commit: 87f1e3d2b2727eab061d086233d283e383a31bb7
Current compatible/fixed main: 1e6e356cfce908c6f82b65e2451749be362f1d5c
Training manifest: 116,731 pure optical-flow pairs from FlyingChairs, FlyingThings3D, HD1K, KITTI 2015, Sintel, and Spring.
Manifest SHA256: 935140861e7df819568c72b15e53897db1590b92020bb02c2bbd13110c66e880.
The Stage 1 manifest is byte-identical between the audited geometry-v2 and camera-v3 releases. The track/camera repair therefore does not require retraining Stage 1.
W&B training-log backfill: https://wandb.ai/xiugo/vggt-omega-bicorr/runs/anyflow-gv2-main87f1e3d-stage01-flow-backfill

Loading

This is a native PyTorch training checkpoint, not a Transformers from_pretrained() directory.

python

from huggingface_hub import hf_hub_download
import torch

path = hf_hub_download(
    repo_id="Shiym/anyflow-biomegacorr-stage1-flow-50k",
    filename="anyflow_biomegacorr_stage1_flow_50k.pt",
)
checkpoint = torch.load(path, map_location="cpu")
print(checkpoint["step"])

The checkpoint includes model and optimizer/scaler training state. For the clean Stage 2 initialization, load its model weights as the predecessor checkpoint while starting a fresh Stage 2 optimizer.

Scope

The deliberately stopped Stage 2 diagnostic checkpoints are not published here. They contain runs affected by recurring non-finite sparse-label losses and must not be resumed.