pytorch shape_error ai_generated true

RuntimeError: 1D target tensor expected, multi-target not supported

ID: pytorch/invalid-argument-2d-index-target

Also available as: JSON · Markdown · 中文
90%Fix Rate
88%Confidence
1Evidence
2023-08-20First Seen

Version Compatibility

VersionStatusIntroducedDeprecatedNotes
pytorch>=1.13.0 active — — —

Root Cause

A loss function like CrossEntropyLoss received a target tensor with more than one dimension (e.g., one-hot encoded labels) instead of a 1D tensor of class indices.

generic

中文

损失函数（如 CrossEntropyLoss）接收到的目标张量维度超过一维（例如独热编码标签），而非一维类别索引张量。

Official Documentation

https://pytorch.org/docs/stable/generated/torch.nn.CrossEntropyLoss.html

Workarounds

  1. 95% success Convert one-hot encoded labels to class indices using `torch.argmax(target, dim=1)` before passing to the loss function. Example: `loss = criterion(output, target.argmax(dim=1))` 
    Convert one-hot encoded labels to class indices using `torch.argmax(target, dim=1)` before passing to the loss function. Example: `loss = criterion(output, target.argmax(dim=1))`
  2. 90% success If the target is already 1D but has an extra dimension from batching, squeeze it: `target = target.squeeze()`. 
    If the target is already 1D but has an extra dimension from batching, squeeze it: `target = target.squeeze()`.
  3. 85% success Use `torch.nn.BCEWithLogitsLoss` with proper target shape (same as output) if the task is multi-label classification. 
    Use `torch.nn.BCEWithLogitsLoss` with proper target shape (same as output) if the task is multi-label classification.

中文步骤

  1. 在传递给损失函数之前，使用 `torch.argmax(target, dim=1)` 将独热编码标签转换为类别索引。示例：`loss = criterion(output, target.argmax(dim=1))`
  2. 如果目标已经是 1D 但由于批处理有额外维度，则挤压它：`target = target.squeeze()`。
  3. 如果任务是多标签分类，则使用 `torch.nn.BCEWithLogitsLoss` 并具有正确的目标形状（与输出相同）。

Dead Ends

Common approaches that don't work:

  1. 95% fail

    This still produces a 2D tensor; the loss expects 1D indices.

  2. 90% fail

    The error is about dimensionality, not dtype; CrossEntropyLoss expects Long type indices.

  3. 80% fail

    BCEWithLogitsLoss expects a different shape and value range; it will not fix the dimensionality issue.