CS 6476 Project 4: PSPNet semantic segmentation weights
Final checkpoints for the six Camvid-11 experiments and the KITTI road transfer run.
Optimizer state is removed. Each file is {"epoch": int, "state_dict": ...}, the format
src/vision/test.py and part6_transfer_learning.py expect.
Code: the project-4 branch of sahil485/cs6476-projects (project-4-student/).
| File | Model | Pretrained | Aug | PPM | Aux loss | Camvid val mIoU |
|---|---|---|---|---|---|---|
| camvid_r1_simple_scratch.pth | SimpleSegmentationNet | no | no | β | β | 0.332 |
| camvid_r2_simple_pretrained.pth | SimpleSegmentationNet | yes | no | β | β | 0.255 |
| camvid_r3_simple_aug.pth | SimpleSegmentationNet | yes | yes | β | β | 0.594 |
| camvid_r4_psp_noppm.pth | PSPNet | yes | yes | no | no | 0.665 |
| camvid_r5_psp_ppm.pth | PSPNet | yes | yes | yes | no | 0.649 |
| camvid_r6_psp_aux.pth | PSPNet | yes | yes | yes | yes | 0.661 (submission) |
| kitti_pspnet.pth | PSPNet, 2 classes | from r6 | yes | yes | yes | train mIoU 0.942 |
All Camvid runs: 100 epochs, batch 32, 201x201 crops, SGD lr 0.01 (poly). KITTI: 20 epochs.
Usage
hf download sahilchat/cs6476-proj4-pspnet camvid_r6_psp_aux.pth --local-dir checkpoints
Run full-res Camvid evaluation (from project-4-student/, after uv sync):
import sys, torch; sys.path.insert(0, ".")
import train_all
from src.vision.test import test_model
args = train_all.make_args("r6_psp_aux", 100, torch.device("cuda" if torch.cuda.is_available() else "cpu"))
args.model_path = "checkpoints/camvid_r6_psp_aux.pth"
test_model(args) # writes ./camvid_r6_psp_aux/camvid-11/240/{gray,rgb_mask_predictions,results.txt}
Or load the model directly:
args.pretrained = False
from src.vision.part3_training_utils import get_model_and_optimizer
model, _ = get_model_and_optimizer(args)
model.load_state_dict(torch.load(args.model_path, map_location="cpu")["state_dict"])
model.eval()
# input: normalized (1, 3, H, W) tensor; H-1 and W-1 divisible by 8, e.g. 201x201
logits, pred, _, _ = model(x) # logits (1, 11, H, W), pred (1, H, W) class ids
For the KITTI model set args.classes = 2. For r1βr3 use make_args("r1_simple_scratch", ...) etc. so the architecture matches.
Inference Providers NEW
This model isn't deployed by any Inference Provider. π Ask for provider support