From 6c123b7231d8bdd5c3b88a95b1eeb7b047425623 Mon Sep 17 00:00:00 2001 From: Admin Date: Fri, 17 Jul 2026 11:38:24 +0000 Subject: [PATCH] fix: correct Math model output transpose from [C,B,T] to [B,T,C] --- internal/captcha/handler.go | 153 +++--------------------------------- 1 file changed, 13 insertions(+), 140 deletions(-) diff --git a/internal/captcha/handler.go b/internal/captcha/handler.go index 011f9b6..f955418 100644 --- a/internal/captcha/handler.go +++ b/internal/captcha/handler.go @@ -392,156 +392,29 @@ func preprocessMath(img image.Image) ([]float32, error) { } func decodeMathFromOutput(output []float32) string { - // Math 模型输出固定: numChars=51, batch=1, timesteps=19, total=969 - // 可能的格式: [51, 1, 19] 或 [19, 1, 51] + // Math 模型输出: [C, B, T] = [51, 1, 19] + // 需要转置为 [B, T, C] = [1, 19, 51] numChars := 51 batch := 1 timesteps := 19 - totalElems := numChars * batch * timesteps // 969 - if len(output) != totalElems { - // 尝试推断 timesteps + if len(output) != numChars*batch*timesteps { timesteps = len(output) / numChars if timesteps <= 0 { return "" } } - // 尝试两种格式 - // 格式1: [T, B, C] = [19, 1, 51] - result1 := decodeMathFormat(output, timesteps, batch, numChars, true) - // 格式2: [C, B, T] = [51, 1, 19] - result2 := decodeMathFormat(output, timesteps, batch, numChars, false) - - // 返回更长的结果（更可能是正确的） - if len(result1) >= len(result2) { - return result1 - } - return result2 -} - -func decodeMathFormat(output []float32, timesteps, batch, numChars int, isTBC bool) string { - var data []float32 - - if isTBC { - // 格式: [T, B, C] -> [B, T, C] - data = make([]float32, batch*timesteps*numChars) - for t := 0; t < timesteps; t++ { - for b := 0; b < batch; b++ { - for c := 0; c < numChars; c++ { - srcIdx := t*batch*numChars + b*numChars + c - dstIdx := b*timesteps*numChars + t*numChars + c - if srcIdx < len(output) && dstIdx < len(data) { - data[dstIdx] = output[srcIdx] - } - } - } - } - } else { - // 格式: [C, B, T] -> [B, T, C] - data = make([]float32, batch*timesteps*numChars) - for c := 0; c < numChars; c++ { - for b := 0; b < batch; b++ { - for t := 0; t < timesteps; t++ { - srcIdx := c*batch*timesteps + b*timesteps + t - dstIdx := b*timesteps*numChars + t*numChars + c - if srcIdx < len(output) && dstIdx < len(data) { - data[dstIdx] = output[srcIdx] - } - } - } - } - } - - // CTC 解码 - result := "" - lastIdx := -1 - - for t := 0; t < timesteps; t++ { - maxIdx := 0 - maxProb := float32(-math.MaxFloat32) - - for c := 0; c < numChars; c++ { - idx := t*numChars + c - if idx < len(data) && data[idx] > maxProb { - maxProb = data[idx] - maxIdx = c - } - } - - if maxIdx != 0 && maxIdx != lastIdx { - if maxIdx-1 < len(mathChars) { - result += string(mathChars[maxIdx-1]) - } - } - lastIdx = maxIdx - } - - return result -} - -func decodeMath(output []float32, outputShape []int64) string { - // 原版逻辑： - // if preds.shape[1] == 1: # Batch size is 1 at dim 1 implies [T, B, C] - // preds = np.transpose(preds, (1, 0, 2)) # [B, T, C] - // preds_idx = np.argmax(preds, axis=2) # [B, T] - - if len(outputShape) < 3 || len(output) == 0 { - return "" - } - - // 输出可能是 [T, B, C] 或 [C, B, T] - // 判断依据：如果 shape[1] == 1，说明 batch 在中间，格式是 [T, B, C] - // 否则格式是 [C, B, T] 或 [num_chars, batch, timesteps] - - dim0 := int(outputShape[0]) - dim1 := int(outputShape[1]) - dim2 := int(outputShape[2]) - - // 处理动态维度 - if dim0 <= 0 { dim0 = 51 } - if dim1 <= 0 { dim1 = 1 } - if dim2 <= 0 { dim2 = len(output) / (dim0 * dim1) } - - var timesteps, batch, numChars int - var data []float32 - - // 判断输出格式 - if dim1 == 1 { - // 格式: [T, B, C] = [19, 1, 51] - // 需要转置为 [B, T, C] = [1, 19, 51] - timesteps = dim0 - batch = dim1 - numChars = dim2 - - data = make([]float32, batch*timesteps*numChars) - for t := 0; t < timesteps; t++ { - for b := 0; b < batch; b++ { - for c := 0; c < numChars; c++ { - srcIdx := t*batch*numChars + b*numChars + c - dstIdx := b*timesteps*numChars + t*numChars + c - if srcIdx < len(output) && dstIdx < len(data) { - data[dstIdx] = output[srcIdx] - } - } - } - } - } else { - // 格式: [C, B, T] = [51, 1, 19] - // 需要转置为 [B, T, C] = [1, 19, 51] - numChars = dim0 - batch = dim1 - timesteps = dim2 - - data = make([]float32, batch*timesteps*numChars) - for c := 0; c < numChars; c++ { - for b := 0; b < batch; b++ { - for t := 0; t < timesteps; t++ { - srcIdx := c*batch*timesteps + b*timesteps + t - dstIdx := b*timesteps*numChars + t*numChars + c - if srcIdx < len(output) && dstIdx < len(data) { - data[dstIdx] = output[srcIdx] - } + // 转置: [C, B, T] -> [B, T, C] + // output[c*batch*timesteps + b*timesteps + t] -> data[b*timesteps*numChars + t*numChars + c] + data := make([]float32, batch*timesteps*numChars) + for c := 0; c < numChars; c++ { + for b := 0; b < batch; b++ { + for t := 0; t < timesteps; t++ { + srcIdx := c*batch*timesteps + b*timesteps + t + dstIdx := b*timesteps*numChars + t*numChars + c + if srcIdx < len(output) && dstIdx < len(data) { + data[dstIdx] = output[srcIdx] } } }