Module 'SeparableConvBlock' has no attribute 'swish'

[BigSteve94]

Hey,

I am trying to export an EfficientDet D1 model with torch.jit.script to a TorchScript model, but I get the following error:

Module 'SeparableConvBlock' has no attribute 'swish' : File "../../efficientdet/model.py", line 50

    if self.activation:
        x = self.swish(x)
            ~~~~~~~~~~ <--- HERE

    return x

I am using the code to export this: state_dict_location = 'initial-model.pth' ck_point = 'last.ckpt' model = EfficientDetBackbone(in_channels=3, num_classes=6, compound_coef=1, ratios=eval(str([(1.0, 1.0), (1.4, 0.7), (0.7, 1.4)])), scales=eval(str([2 0, 2 (1.0 / 3.0), 2 ** (2.0 / 3.0)])), fpn_num_filters=None, input_size=None)

sd = torch.load(ck_point)

model.load_state_dict(sd["state_dict"])

model.backbone_net.model.set_swish(memory_efficient=False) tried this already but this is related to onnx exports

device = torch.device("cpu") model.to(device) model.eval()

scripted = torch.jit.script(model) scripted.save(f"test_saving_Effdet_model.pt")

Any idea how to solve this issue? (Using fpn_num_filters=88 and input_size=512 doesnt help)

[phungpx]

Hi @BigSteve94,

You can fix this issue by replace (

__init__:

self.activation = activation
if self.activation:
     self.swish = MemoryEfficientSwish() if not onnx_export else Swish()

__forward__:
if self.activation:
    x = self.swish(x)
return x

by:

__init__:

if activation:
     self.swish = MemoryEfficientSwish() if not onnx_export else Swish()
else:
    self.swish = nn.Identity()

__forward__:
x = self.swish(x)
return x

It is explainated in

https://discuss.pytorch.org/t/jit-scripting-nn-module-with-function-attributes/60957