microclip-checkpoints

MicroCLIP checkpoints

Every trained checkpoint behind the results in github.com/umutonuryasar/microclip: a CLIP-style image–text model built from scratch on a single-GPU budget, used to test whether SigLIP's sigmoid loss keeps its small-batch advantage over softmax (InfoNCE) at small scale. At this scale it does not: softmax matched or beat sigmoid at every batch size from 128 to 512.

Live demo: huggingface.co/spaces/umutonuryasar/microclip

Contents

One best.pt per run: the lowest-validation-loss weights over training. The main comparison is 3 seeds (42/43/44) at batch 128 and 512; batch 256 and all ablations are single-seed. Checkpoints are byte-identical to the training outputs; the MD5 sums below let you verify that.

Also included: eval_per_run.csv and eval_summary.csv (identical to results/ in the repo), eval_results.csv (an earlier single-seed evaluation pass) and queue_log.txt (the training queue's timestamps).

Loading

Each file is a dict with a single model key holding the state dict, loadable with weights_only=True. The config defines the architecture; it must be the one the run was trained with, per the table above.

import torch
from microclip.config import load_config
from microclip.data.tokenizer import CaptionTokenizer
from microclip.models.microclip import MicroCLIP

cfg = load_config("configs/softmax_b512.yml")
tokenizer = CaptionTokenizer(cfg["tokenizer"]["path"])
model = MicroCLIP(cfg, vocab_size=tokenizer.vocab_size, max_len=cfg["data"]["max_text_len"])
state = torch.load("runs/softmax_b512_s42/best.pt", map_location="cpu", weights_only=True)
model.load_state_dict(state["model"])

Evaluate with the repo's script, for example COCO 5K retrieval:

python scripts/evaluate.py --config configs/softmax_b512.yml \
    --checkpoint runs/softmax_b512_s42/best.pt --task retrieval

Limits

Trained from scratch on COCO train2017 only, for 30 epochs (the recipe ablations for 10). Absolute retrieval quality is modest by design; see the repo README for the full results and limitations.