RuntimeError: CUDNN_STATUS_BAD_PARAM

vision
1

Hello everyone,
Can someone help with this error? I am running an inception_v3 model via transfer learning with Pytorch, but I receive a CUDNN_STATUS_BAD_PARAM. My input dataset consists of 224x224 patch image (size: [32,3,224,224]). Following is the model structure:

class PretrainedModel(nn.Module):
    def __init__(self, dropout):
        super().__init__()
        model = models.inception_v3(pretrained=True)
        num_ftrs = model.fc.in_features
        model.dropout = dropout
        model.fc = nn.Linear(num_ftrs, 2)
        
        self.model = model

    def forward(self, x):
        return self.model(x)

Bellow is a screenshot of the error I receive:

image
image922×694 38.3 KB

Feedback much appreciated. Thanks

2

The Inception model expects input data with the size [batch_size, 3, 299, 299].
Could you try to increase the spatial size of your input and try it again?
Also note, that models.inception_v3 will return the aux_logits with the output in the default setting.
In case you don’t need the aux_logits, you should just return the first return value:

return self.model(x)[0]
3

Thank you @ptrblck after changing the input size to 299299 instead of 224224 and including aux_logits = False as parameter to the model (an alternative to returning the first value as in self.model(x)[0]) the network runs good now