WZMIAOMIAO/deep-learning-for-image-processing · error · KeyError

%s is not compatible with %s. Specify --weights '' or specif

Error message

%s is not compatible with %s. Specify --weights '' or specify a --cfg compatible with %s. See https://github.com/ultralytics/yolov3/issues/657

What it means

When resuming/loading pretrained weights, train() filters checkpoint params by matching numel with the current model and loads with strict=False; if a checkpoint key is absent from the model's state_dict, model.state_dict()[k] raises KeyError, which is re-raised as a clearer KeyError stating that --weights and --cfg are incompatible, linking to yolov3 issue #657.

Source

Thrown at pytorch_object_detection/yolov3_spp/train.py:108

    pg = [p for p in model.parameters() if p.requires_grad]
    optimizer = optim.SGD(pg, lr=hyp["lr0"], momentum=hyp["momentum"],
                          weight_decay=hyp["weight_decay"], nesterov=True)

    scaler = torch.cuda.amp.GradScaler() if opt.amp else None

    start_epoch = 0
    best_map = 0.0
    if weights.endswith(".pt") or weights.endswith(".pth"):
        ckpt = torch.load(weights, map_location=device)

        # load model
        try:
            ckpt["model"] = {k: v for k, v in ckpt["model"].items() if model.state_dict()[k].numel() == v.numel()}
            model.load_state_dict(ckpt["model"], strict=False)
        except KeyError as e:
            s = "%s is not compatible with %s. Specify --weights '' or specify a --cfg compatible with %s. " \
                "See https://github.com/ultralytics/yolov3/issues/657" % (opt.weights, opt.cfg, opt.weights)
            raise KeyError(s) from e

        # load optimizer
        if ckpt["optimizer"] is not None:
            optimizer.load_state_dict(ckpt["optimizer"])
            if "best_map" in ckpt.keys():
                best_map = ckpt["best_map"]

        # load results
        if ckpt.get("training_results") is not None:
            with open(results_file, "w") as file:
                file.write(ckpt["training_results"])  # write results.txt

        # epochs
        start_epoch = ckpt["epoch"] + 1
        if epochs < start_epoch:
            print('%s has been trained for %g epochs. Fine-tuning for %g additional epochs.' %
                  (opt.weights, ckpt['epoch'], epochs))
            epochs += ckpt['epoch']  # finetune additional epochs

View on GitHub (pinned to 1ec3fe6f37)

Solutions

1. Load with --weights '' (train from scratch with the given cfg), or
2. Make --cfg match the checkpoint's architecture (same layer sizes / classes), or
3. Strip/convert the incompatible layers from the checkpoint before loading

Example fix

// before
python train.py --cfg cfg/yolov3-custom4.cfg --weights weights/yolov3-spp.pt
// after (option A: scratch)
python train.py --cfg cfg/yolov3-custom4.cfg --weights ''
// or (option B: matching cfg)
python train.py --cfg cfg/yolov3-spp.cfg --weights weights/yolov3-spp.pt

import torch
ckpt = torch.load(opt.weights, map_location='cpu')
cs = set(ckpt['model'].keys()); ms = set(model.state_dict().keys())
missing = cs - ms
assert not missing, f'cfg incompatible with weights, e.g. {sorted(missing)[:3]}'

try:
    model.load_state_dict(ckpt['model'], strict=False)
except KeyError as e:
    raise SystemExit('weights/cfg mismatch: start with --weights \'\'' or use the matching cfg') from e

When it happens

Trigger: Running train.py with --weights pointing to a checkpoint whose architecture differs from --cfg (different number of classes, different backbone width/depth, or a cfg edited after the checkpoint was saved).

Common situations: Fine-tuning a COCO-pretrained .pt (80 classes) with a custom cfg (N classes) without adjusting; using yolov3.cfg weights with yolov3-spp.cfg; modified hyp/cfg after resuming an old run.

Related errors

• %s is not compatible with %s. Specify --weights '' or specif
• conv2d filter size must be int type.
• not found weights file: {}
• not found weights file: {}
• expected stages_repeats as list of 3 positive ints

AI-assisted analysis of WZMIAOMIAO/deep-learning-for-image-processing@1ec3fe6f37 (2026-08-30). Data as JSON: /api/errors/3c0f6c60d1bd6e6d. Report an issue: GitHub.