babalae/better-genshin-impact · error · InvalidOperationException

角色头像模型元素输出数量异常：logits={logits.Length}, elements={elementType

Error message

角色头像模型元素输出数量异常：logits={logits.Length}, elements={elementTypes.Length}

What it means

Thrown as InvalidOperationException by AvatarGridIconRecognizer when the ONNX model's 'element_logits' output tensor length does not match the count of distinct element types derived from avatar.csv prototypes. The recognizer maps each logit index to a sorted distinct element type; a mismatch means the model and the CSV are out of sync (different training-time class count vs. prototype data).

Source

Thrown at BetterGenshinImpact/GameTask/Common/Job/AvatarGridIconRecognizer.cs:127

            .Select(group => group.OrderByDescending(candidate => candidate.Score).First())
            .OrderByDescending(candidate => candidate.Score)
            .FirstOrDefault() ?? AvatarGridIconCandidate.Empty;

        if (!recognizeElementType || candidate == AvatarGridIconCandidate.Empty)
        {
            return candidate;
        }

        var logits = results.First(result => result.Name == "element_logits").AsEnumerable<float>().ToArray();
        var elementTypes = _prototypes
            .Select(prototype => prototype.ElementType)
            .Where(elementType => !string.IsNullOrWhiteSpace(elementType))
            .Distinct(StringComparer.Ordinal)
            .OrderBy(elementType => elementType, StringComparer.Ordinal)
            .ToArray();
        if (logits.Length != elementTypes.Length || logits.Length == 0)
        {
            throw new InvalidOperationException(
                $"角色头像模型元素输出数量异常：logits={logits.Length}, elements={elementTypes.Length}");
        }

        var predictedElementType = elementTypes[Array.IndexOf(logits, logits.Max())];
        return candidate with { ElementType = predictedElementType };
    }

    /// <summary>
    /// 按模型训练协议生成头像和元素图标两个输入张量。
    /// </summary>
    /// <remarks>
    /// 完整头像缩放为 115x115；元素输入必须从该图左上角裁剪 48x48 后再缩放为 64x64。
    /// 两个输入均执行 BGR→RGB 及 [-1,1] 归一化，不能用完整头像代替元素输入。
    /// </remarks>
    internal static (DenseTensor<float> Image, DenseTensor<float> ElementImage) CreateInputTensors(Mat mat)
    {
        using Mat resized = mat.Resize(new Size(InputSize, InputSize));
        using Mat elementRoi = resized.SubMat(0, ElementRoiSize, 0, ElementRoiSize);

View on GitHub (pinned to a7cb36712d)

Solutions

  1. Ensure avatar.onnx and avatar.csv are from the same release/training run so element class counts align.
  2. Regenerate avatar.csv from the same training data used to export the ONNX model.
  3. If the model is intentionally updated, update the CSV and re-test element recognition.

Example fix

null
Defensive patterns

Strategy: validation

Validate before calling

// After loading prototypes, verify element count matches model output metadata
var distinctElements = prototypes.Select(p => p.ElementType).Where(e => !string.IsNullOrWhiteSpace(e)).Distinct().Count();
// Compare against model metadata: _session.ModelMetadata or output node shape

Type guard

null

Try / catch

try { recognizer.Recognize(avatar, element, true); }
catch (InvalidOperationException ex) when (ex.Message.Contains("元素输出数量异常")) { /* model/CSV mismatch — update assets */ }

Prevention

When it happens

Trigger: Calling Recognize(avatarMat, elementMat, recognizeElementType: true) when the avatar.onnx model was updated/retrained with a different number of element classes than the distinct element_type values present in avatar.csv.

Common situations: The ONNX model file (avatar.onnx) was replaced with a newer version that has more/fewer element output nodes, but avatar.csv was not regenerated to match. The prototype CSV has blank element_type values that change the distinct count. A partial update of the AvatarGridIcon asset bundle.

Related errors

AI-assisted analysis of babalae/better-genshin-impact@a7cb36712d (2026-08-13). Data as JSON: /api/errors/aeae9f6b315253e9. Report an issue: GitHub.