Passenger list page model and sorter
Two small MobileNetV3-small models (PyTorch state dicts) from passenger-list-reader. Both run on CPU. The model classes and input preparation are in that repo.
page_model.pt(5 MB,layout_train.Net): looks at a whole page, shrunk to 320 x 320 grayscale, and says whether it lists passengers, plus where the name column and the first and last rows are. On 49 pages from reels nothing was trained on, it kept 20 of 20 name pages and wrongly kept 1 of 29 others.sort_model.pt(4 MB,sort_train.Net): sorts a name strip, scaled to 64 pixels high and padded to 448 wide, as typed, handwritten, or neither. 96.0% right on 1,187 test strips; it confused typed with handwritten once.
Trained on New York arrival lists, 1897 to 1957 (National Archives microfilm T715). Labels came from Claude reading pages and strips. Other layouts would likely need retraining.
Use
import torch
from huggingface_hub import hf_hub_download
from sort_train import Net # from the GitHub repo
net = Net()
net.load_state_dict(torch.load(hf_hub_download("horatio12/passenger-list-page-models", "sort_model.pt")))
net.eval()
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐ Ask for provider support