Hi everyone!
It just stumbled over the fastai-v2 notebook that shows how to do image segmentation with a Dynamic Unet and a resnet34 encoder on teh Camvid dataset.
Out of curiosity I tried to use different encoders than the pre-defined resnet34 ,like squeezenet, densnet and vgg that are built into fastai-v2 but all resulted in the same error:
I know it is possible to do this in v1 with e.g. VGG (see kaggle kernel). Can anybody help?
I’ll be showing an example with this in my study group but take a look at what unet_learner is doing and try following how to build the model from its code manually. You’ll see that we create an encoder from our model and wrap a head based on Unet to it.
For a custom encoder and generating the body, I describe the process in my notebook here:
@muellerzr: thanks for the custom encoder, could you please provide a link to wrap it in UNET/DynamicUnet.
I wound up going with this: https://github.com/muellerzr/Practical-Deep-Learning-for-Coders-2.0/blob/master/Computer%20Vision/06_Hybridizing_Models.ipynb as I didn’t get time to do that implementation. This adapts the Unet for classification and regression.
@muellerzr i get below error when i build custom Unet
RuntimeError: running_mean should contain 582 elements not 1164
Any help
171 def _do_one_batch(self):
--> 172 self.pred = self.model(*self.xb)
173 self('after_pred')
174 if len(self.yb):
/opt/conda/lib/python3.7/site-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
1049 if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
1050 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1051 return forward_call(*input, **kwargs)
1052 # Do not call functions when jit is used
1053 full_backward_hooks, non_full_backward_hooks = [], []
/opt/conda/lib/python3.7/site-packages/fastai/layers.py in forward(self, x)
405 for l in self.layers:
406 res.orig = x
--> 407 nres = l(res)
408 # We have to remove res.orig to avoid hanging refs and therefore memory leaks
409 res.orig, nres.orig = None, None
/opt/conda/lib/python3.7/site-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
1069 input = bw_hook.setup_input_hook(input)
1070
-> 1071 result = forward_call(*input, **kwargs)
1072 if _global_forward_hooks or self._forward_hooks:
1073 for hook in itertools.chain(
/opt/conda/lib/python3.7/site-packages/torch/nn/modules/batchnorm.py in forward(self, input)
176 bn_training,
177 exponential_average_factor,
--> 178 self.eps,
179 )
180
/opt/conda/lib/python3.7/site-packages/torch/nn/functional.py in batch_norm(input, running_mean, running_var, weight, bias, training, momentum, eps)
2274 training=training,
2275 momentum=momentum,
-> 2276 eps=eps,
2277 )
2278 if training:
/opt/conda/lib/python3.7/site-packages/torch/overrides.py in handle_torch_function(public_api, relevant_args, *args, **kwargs)
1250 # Use `public_api` instead of `implementation` so __torch_function__
1251 # implementations can do equality/identity comparisons.
-> 1252 result = overloaded_arg.__torch_function__(public_api, types, args, kwargs)
1253
1254 if result is not NotImplemented:
/opt/conda/lib/python3.7/site-packages/fastai/torch_core.py in __torch_function__(self, func, types, args, kwargs)
338 convert=False
339 if _torch_handled(args, self._opt, func): convert,types = type(self),(torch.Tensor,)
--> 340 res = super().__torch_function__(func, types, args=args, kwargs=kwargs)
341 if convert: res = convert(res)
342 if isinstance(res, TensorBase): res.set_meta(self, as_copy=True)
/opt/conda/lib/python3.7/site-packages/torch/_tensor.py in __torch_function__(cls, func, types, args, kwargs)
1021
1022 with _C.DisableTorchFunction():
-> 1023 ret = func(*args, **kwargs)
1024 return _convert(ret, cls)
1025
/opt/conda/lib/python3.7/site-packages/torch/nn/functional.py in batch_norm(input, running_mean, running_var, weight, bias, training, momentum, eps)
2280
2281 return torch.batch_norm(
-> 2282 input, weight, bias, running_mean, running_var, training, momentum, eps, torch.backends.cudnn.enabled
2283 )
2284
RuntimeError: running_mean should contain 582 elements not 1164
11): ResBlock(
(convpath): Sequential(
(0): ConvLayer(
(0): Conv2d(291, 291, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(1): ReLU()
)
(1): ConvLayer(
(0): Conv2d(291, 291, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
)
)
(idpath): Sequential()
(act): ReLU(inplace=True)
)
(12): AdaptiveConcatPool2d(
(ap): AdaptiveAvgPool2d(output_size=1)
(mp): AdaptiveMaxPool2d(output_size=1)
)
(13): Flatten(full=False)
(14): BatchNorm1d(1164, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(15): Dropout(p=0.25, inplace=False)
(16): Linear(in_features=1164, out_features=512, bias=False)
(17): ReLU(inplace=True)
(18): BatchNorm1d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(19): Dropout(p=0.5, inplace=False)
(20): Linear(in_features=512, out_features=4, bias=False)
i think output of layer prior to adaptive pool is not correct, 291*2 =582 *2=1164
I user resnext 50 here…
removing *2 from NF dsnt gives the error
But then there is no upsampling layer added