DataLoader object has no attribute 'after_batch'

fastai
1

I am trying to implement KFold c.v. using both PyTorch and Fastai.

for fold, (train_idx, valid_idx) in enumerate(kfold.split(glomer_dataset)):
    print(f'FOLD {fold}')
    print('--------------------------------')
    
    train_subsampler = SubsetRandomSampler(train_idx)
    valid_subsampler = SubsetRandomSampler(valid_idx)
    
    train_loader = DataLoader(my_dataset, batch_size=BS, sampler=train_subsampler)
    valid_loader = DataLoader(my_dataset, batch_size=BS, sampler=valid_subsampler)
    
    dls = DataLoaders(train_loader, valid_loader)
    
    opt = ranger
    learn = unet_learner(dls, resnet34, loss=nn.CrossEntropyLoss() , self_attention=True, act_cls=Mish, opt_func=opt)
    lr = 3e-4
    learn.fit_flat_cos(3, slice(lr))

When I run the code above, I get this error:

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
Cell In[10], line 14
     11 dls = DataLoaders(train_loader, valid_loader)
     13 opt = ranger
---> 14 learn = unet_learner(dls, resnet34, loss=nn.CrossEntropyLoss() , self_attention=True, act_cls=Mish, opt_func=opt)
     15 lr = 3e-4
     16 learn.fit_flat_cos(3, slice(lr))

File /opt/conda/lib/python3.10/site-packages/fastai/vision/learner.py:262, in unet_learner(dls, arch, normalize, n_out, pretrained, config, loss_func, opt_func, lr, splitter, cbs, metrics, path, model_dir, wd, wd_bn_bias, train_bn, moms, **kwargs)
    260 meta = model_meta.get(arch, _default_meta)
    261 n_in = kwargs['n_in'] if 'n_in' in kwargs else 3
--> 262 if normalize: _add_norm(dls, meta, pretrained, n_in)
    264 n_out = ifnone(n_out, get_c(dls))
    265 assert n_out, "`n_out` is not defined, and could not be inferred from data, set `dls.c` or pass `n_out`"

File /opt/conda/lib/python3.10/site-packages/fastai/vision/learner.py:196, in _add_norm(dls, meta, pretrained, n_in)
    194 if stats is None: return
    195 if n_in != len(stats[0]): return
--> 196 if not dls.after_batch.fs.filter(risinstance(Normalize)):
    197     dls.add_tfms([Normalize.from_stats(*stats)],'after_batch')

File /opt/conda/lib/python3.10/site-packages/fastcore/basics.py:496, in GetAttr.__getattr__(self, k)
    494 if self._component_attr_filter(k):
    495     attr = getattr(self,self._default,None)
--> 496     if attr is not None: return getattr(attr,k)
    497 raise AttributeError(k)

AttributeError: 'DataLoader' object has no attribute 'after_batch'

I assume it is due to a compatibility issue between PyTorch DataLoader and Fastai, but I am not sure.

2

There should be a normalize or add_norm param to unet_learner, set this to False

1 Like
3

I have added normalize=False:

    learn = unet_learner(dls, resnet34, n_out=3, loss=nn.CrossEntropyLoss() , self_attention=True,
                         act_cls=Mish, opt_func=opt, normalize=False,
    )

Now I get this error:

AttributeError                            Traceback (most recent call last)
Cell In[17], line 14
     11 dls = DataLoaders(train_loader, valid_loader)
     13 opt = ranger
---> 14 learn = unet_learner(dls, resnet34, n_out=3, loss=nn.CrossEntropyLoss() , self_attention=True,
     15                      act_cls=Mish, opt_func=opt, normalize=False,
     16 ) 
     17 lr = LR
     18 learn.fit_flat_cos(3, slice(lr))

File /opt/conda/lib/python3.10/site-packages/fastai/vision/learner.py:266, in unet_learner(dls, arch, normalize, n_out, pretrained, config, loss_func, opt_func, lr, splitter, cbs, metrics, path, model_dir, wd, wd_bn_bias, train_bn, moms, **kwargs)
    264 n_out = ifnone(n_out, get_c(dls))
    265 assert n_out, "`n_out` is not defined, and could not be inferred from data, set `dls.c` or pass `n_out`"
--> 266 img_size = dls.one_batch()[0].shape[-2:]
    267 assert img_size, "image size could not be inferred from data"
    268 model = create_unet_model(arch, n_out, img_size, pretrained=pretrained, **kwargs)

File /opt/conda/lib/python3.10/site-packages/fastcore/basics.py:496, in GetAttr.__getattr__(self, k)
    494 if self._component_attr_filter(k):
    495     attr = getattr(self,self._default,None)
--> 496     if attr is not None: return getattr(attr,k)
    497 raise AttributeError(k)

AttributeError: 'DataLoader' object has no attribute 'one_batch

I assume the problem is with DataLoader again.

4

I guess the unet_learner doesn’t support custom dataloaders :person_shrugging:

Just follow unet_learner’s source yourself and pull a batch of data out to create it, rather than using unet_learner directly

1 Like