One of the variables needed for gradient computation has been modified by an inplace operation

[pgsrv]

Hi! I'm not sure is this a bug.

I've used tf_efficientnetv2_m_in21ft1k in the modified training script. And I get:
RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.cuda.HalfTensor [32, 24, 150, 150]], which is output 0 of ReluBackward1, is at version 2; expected version 1 instead. Hint: enable anomaly detection to find the operation that failed to compute its gradient, with torch.autograd.set_detect_anomaly(True).

Then I've added torch.autograd.set_detect_anomaly(True):
File ".../lib/python3.8/site-packages/torch/nn/modules/activation.py", line 102, in forward return F.relu(input, inplace=self.inplace) File ".../lib/python3.8/site-packages/torch/nn/functional.py", line 1296, in relu result = torch.relu_(input)

And I've found in ConvBnAct (https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/efficientnet_blocks.py#L61):
self.act1 = act_layer(inplace=True)

I've replaced inplace=True to inplace=False and the error is gone.
Is it ok to replace inplace=True to inplace=False?

Thank you!

[rwightman]

@pgsrv yes, it's perfectly okay to change inplace to False there... it doesn't appear to be a problem in normal classification use with no mods to the network, so I assume you've either modified the network from defaults, or a using it in a different task with additional stems, heads, or taps... ?

[pgsrv]

Yes, I use tf_efficientnetv2_m_in21ft1k as a backbone net with custom classification head:

class Net(torch.nn.Module):
    def __init__(self, backbone, nb_classes): 
        super(Net, self).__init__()
        self.nb_classes = nb_classes
        self.backbone = backbone
        self.backbone.classifier = torch.nn.Identity()
        self.clsf_head = torch.nn.Linear(self.backbone.num_features * self.backbone.global_pool.feat_mult(), self.nb_classes, bias=True)

    def forward(self, x):
        x = self.backbone(x)
        output = self.clsf_head(x)
        return output

By the way I've tried tf_efficientnet_b3/4/5_ns with default inplace=True and it was ok