ultralytics/yolov5 · error · TypeError

no matching TensorFlow activation found for PyTorch activati

Error message

no matching TensorFlow activation found for PyTorch activation {act}

What it means

models/tf.py activations() raises TypeError when asked to convert a PyTorch activation module that has no mapped TensorFlow equivalent. Only nn.LeakyReLU, nn.Hardswish, and nn.SiLU/SiLU are handled; any other activation (nn.ReLU, nn.Mish, nn.ELU, ...) reaches the else branch. This function is used while building the Keras/graph copy of a model for TF exports, so the error surfaces during export, not training.

Source

Thrown at models/tf.py:704

            selected_classes,
            paddings=[[0, topk_all - tf.shape(selected_boxes)[0]]],
            mode="CONSTANT",
            constant_values=-1.0,
        )
        valid_detections = tf.shape(selected_inds)[0]
        return padded_boxes, padded_scores, padded_classes, valid_detections


def activations(act=nn.SiLU):
    """Converts PyTorch activations to TensorFlow equivalents, supporting LeakyReLU, Hardswish, and SiLU/Swish."""
    if isinstance(act, nn.LeakyReLU):
        return lambda x: keras.activations.relu(x, alpha=0.1)
    elif isinstance(act, nn.Hardswish):
        return lambda x: x * tf.nn.relu6(x + 3) * 0.166666667
    elif isinstance(act, (nn.SiLU, SiLU)):
        return lambda x: keras.activations.swish(x)
    else:
        raise TypeError(f"no matching TensorFlow activation found for PyTorch activation {act}")


def representative_dataset_gen(dataset, ncalib=100):
    """Generate representative dataset for calibration by yielding transformed numpy arrays from the input dataset."""
    for n, (path, img, im0s, vid_cap, string) in enumerate(dataset):
        im = np.transpose(img, [1, 2, 0])
        im = np.expand_dims(im, axis=0).astype(np.float32)
        im /= 255
        yield [im]
        if n >= ncalib:
            break


def run(
    weights=ROOT / "yolov5s.pt",  # weights path
    imgsz=(640, 640),  # inference size h,w
    batch_size=1,  # batch size
    dynamic=False,  # dynamic batch size

View on GitHub (pinned to 20d1d78a08)

Solutions

1. Retrain/convert with a supported activation (SiLU is the YOLOv5 default) if you do not control the export code.
2. Extend the mapping in models/tf.py activations() with an equivalent TF op for your activation (see exampleFix).
3. Export to a backend that does not need the TF graph copy (onnx, engine) as a workaround.

Example fix

# before
else:
    raise TypeError(f"no matching TensorFlow activation found for PyTorch activation {act}")

# after (add a branch before the else)
elif isinstance(act, nn.ReLU):
    return lambda x: tf.nn.relu(x)
else:
    raise TypeError(f"no matching TensorFlow activation found for PyTorch activation {act}")

import torch.nn as nn

UNSUPPORTED = (nn.ReLU, nn.ELU, nn.PReLU, nn.GELU, nn.SELU, nn.CELU, nn.Tanh, nn.Softplus, nn.Softsign)

def model_acts_exportable(model) -> bool:
    """True if every activation module has a TF mapping in models/tf.py."""
    return not any(isinstance(m, UNSUPPORTED) for m in model.modules())

import torch.nn as nn

UNSUPPORTED = (nn.ReLU, nn.ELU, nn.PReLU, nn.GELU, nn.SELU, nn.CELU, nn.Tanh, nn.Softplus, nn.Softsign)

def activations_convertible(model: nn.Module) -> bool:
    """True if no activation module lacks a TF mapping in models/tf.py activations()."""
    return not any(isinstance(m, UNSUPPORTED) for m in model.modules())

try:
    keras_act = activations(type(m))
except TypeError:
    keras_act = keras.activations.relu  # conservative fallback for exports

When it happens

Trigger: Exporting a custom YOLOv5 variant whose YAML/blocks use nn.ReLU or nn.Mish activations via export.py --include saved_model/tflite/pb; loading a third-party checkpoint with unusual activation modules and converting it to TF.

Common situations: Custom architectures (mish-variant YOLO forks); ablation experiments swapping activation functions; trying to export classification heads with nn.ELU.

Related errors

• failed to load ONNX file: {onnx}
• ERROR: YOLOv5 TF.js inference is not supported

AI-assisted analysis of ultralytics/yolov5@20d1d78a08 (2026-08-15). Data as JSON: /api/errors/458af9e712750d1e. Report an issue: GitHub.