What causes "dataset have {} classes, but input {}" in WZMIAOMIAO/deep-learning-for-image-processing?

Set --num_classes to the dataset's actual class count (len(train_dataset.labels)).

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

dataset have {} classes, but input {}

Error message

dataset have {} classes, but input {}

What it means

The multi-GPU training script cross-checks args.num_classes (parsed from CLI/config, used to size the model head) against len(train_dataset.labels) (the class count in the dataset JSON). A mismatch means the model output layer would be sized wrongly for the data, so a ValueError reporting both counts is raised before training starts.

Source

Thrown at pytorch_classification/mini_imagenet/train_multi_gpu_using_launch.py:58

                                     transforms.RandomHorizontalFlip(),
                                     transforms.ToTensor(),
                                     transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])]),
        "val": transforms.Compose([transforms.Resize(256),
                                   transforms.CenterCrop(224),
                                   transforms.ToTensor(),
                                   transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])])}

    data_root = args.data_path
    json_path = "./classes_name.json"
    # 实例化训练数据集
    train_dataset = MyDataSet(root_dir=data_root,
                              csv_name="new_train.csv",
                              json_path=json_path,
                              transform=data_transform["train"])

    # check num_classes
    if args.num_classes != len(train_dataset.labels):
        raise ValueError("dataset have {} classes, but input {}".format(len(train_dataset.labels),
                                                                        args.num_classes))

    # 实例化验证数据集
    val_dataset = MyDataSet(root_dir=data_root,
                            csv_name="new_val.csv",
                            json_path=json_path,
                            transform=data_transform["val"])

    # 给每个rank对应的进程分配训练的样本索引
    train_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset)
    val_sampler = torch.utils.data.distributed.DistributedSampler(val_dataset)

    # 将样本索引每batch_size个元素组成一个list
    train_batch_sampler = torch.utils.data.BatchSampler(
        train_sampler, batch_size, drop_last=True)

    nw = min([os.cpu_count(), batch_size if batch_size > 1 else 0, 8])  # number of workers
    if rank == 0:

View on GitHub (pinned to 1ec3fe6f37)

Solutions

Set --num_classes to the dataset's actual class count (len(train_dataset.labels)).
Regenerate class_indices.json / new_train.csv from the current dataset so the label map matches.
Re-run parse_data.py (or equivalent dataset-split script) after changing the training data.

Example fix

// before
python -m torch.distributed.launch --nproc_per_node=2 train_multi_gpu_using_launch.py --num_classes=1000
// after
python -m torch.distributed.launch --nproc_per_node=2 train_multi_gpu_using_launch.py --num_classes=64

Defensive patterns

Strategy: validation

Validate before calling

import json
with open(json_path) as f:
    class_indices = json.load(f)
if args.num_classes != len(class_indices):
    raise ValueError(f"--num_classes={args.num_classes} but dataset has {len(class_indices)} classes")

Type guard

def num_classes_matches(args, dataset_labels) -> bool:
    return getattr(args, "num_classes", None) == len(dataset_labels)

Try / catch

try:
    train_loop(args)
except ValueError as e:
    if "classes" in str(e):
        print("Regenerate class_indices.json / new_train.csv, then pass the matching --num_classes")
    raise

Prevention

Derive num_classes from class_indices.json instead of typing it
Regenerate class_indices.json and CSVs after any dataset change
Print len(train_dataset.labels) when the training script starts
Don't reuse launch commands from previous datasets

When it happens

Trigger: Running train_multi_gpu_using_launch.py with --num_classes N where N differs from the number of classes recorded in class_indices.json generated from new_train.csv.

Common situations: Reusing a launch command from a previous dataset, forgetting to regenerate class_indices.json after adding/removing classes, copy-pasting num_classes=1000 (ImageNet) for Mini-ImageNet.

Related errors

AI-assisted analysis of WZMIAOMIAO/deep-learning-for-image-processing@1ec3fe6f37 (2026-08-30). Data as JSON: /api/errors/3daafb723d97c6a4. Report an issue: GitHub.