I’m open to making norm type configurable. I’ll look into it.
1 Like
Just found this super cool little API change! I was doing a show_results on the planets dataset and:
Our correct and incorrect predictions are highlighted in red and green! Such a subtle change but it’s fantastic! 
10 Likes
Also separate question @sgugger is there a method for utilizing a MultiCategory with using get_y’s such as RegexLabeller? Essentially what I am trying to do is use the sigmoid operation with the PETs dataset and currently (if I am reading how to do multi-label right) it looks like I need to convert everything into a csv first.(I can understand the reasoning behind keeping the csv mandatory for such a task in the sense of you don’t accidentally do it.)
You don’t need to use a csv. Multicategory expect labels to be in a list, so jsut add a transform that puts your label l in [l] then you should be fine.
2 Likes
Awesome! Easy enough to do Thanks! For those wondering what that looks like, here it is in DataBlock form:
def multi_l(l): return [l]
pets_multi = DataBlock(blocks=(ImageBlock, MultiCategoryBlock),
get_items=get_image_files,
splitter=RandomSplitter(),
get_y=[RegexLabeller(pat = r'/([^/]+)_\d+.jpg$'), multi_l])
1 Like
This is entirely custom, just want some input to how I did it and how it could be adjusted for a show function. I’m doing multiple keypoints and when I call show_results I get a tensor size mismatch (as currently show_results on a TensorPoint expects one point I believe). Here is how I’m manually doing it now:
x, y = dbunch.one_batch()
with torch.no_grad():
res = learn.model(x)
pt = res[0].view(-1, 2)
a, b = dbunch.decode_batch((x[0], pt.unsqueeze(0)))[0]
ctx = a.show(figsize=(9,9))
b.show(ctx=ctx)
How would I go about adapting this for multiple key-points? Or is there something I’m missing and regular can do the job
For reference, there is a size mismatch error when I do a regular show_results
It’s hard to know what the problem is with no information about how you built your DataBunch and no stack trace. In general, the approach to customize show_batch is to define type-dipsatched version of the show_batch function like in vision.data.
1 Like
Thanks! I’ll look into it! I’m making it like such:
def img2kpts(f): return path/f'{str(f)}.cat'
def get_points(coords:array):
kpts = []
for i in range(0, len(coords), 2):
kpts.append([coords[i], coords[i+1]])
return tensor(kpts)
def get_y(f:Path):
pts = np.genfromtxt(img2kpts(f))[1:]
return get_points(pts)
def get_ip(img:PILImage, pnts:array): return TensorPoint.create(pnts, img_size=img.size)
dblock = DataBlock(blocks=(ImageBlock, PointsBlock),
get_items=get_image_files,
splitter=RandomSplitter(),
get_y=lambda o: get_y(o.name))
And here is that runtime error log:
---------------------------------------------------------------------------
RuntimeError Traceback (most recent call last)
<ipython-input-340-c3b657dcc9ae> in <module>()
----> 1 learn.show_results()
21 frames
/usr/local/lib/python3.6/dist-packages/fastai2/learner.py in show_results(self, ds_idx, dl, max_n, **kwargs)
327 b = dl.one_batch()
328 _,_,preds = self.get_preds(dl=[b], with_decoded=True)
--> 329 self.dbunch.show_results(b, preds, max_n=max_n, **kwargs)
330
331 def show_training_loop(self):
/usr/local/lib/python3.6/dist-packages/fastai2/data/core.py in show_results(self, b, out, max_n, ctxs, show, **kwargs)
82 x,y,its = self.show_batch(b, max_n=max_n, show=False)
83 b_out = b[:self.n_inp] + (tuple(out) if is_listy(out) else (out,))
---> 84 x1,y1,outs = self.show_batch(b_out, max_n=max_n, show=False)
85 res = (x,x1,None,None) if its is None else (x, y, its, outs.itemgot(slice(self.n_inp,None)))
86 if not show: return res
/usr/local/lib/python3.6/dist-packages/fastai2/data/core.py in show_batch(self, b, max_n, ctxs, show, **kwargs)
76 def show_batch(self, b=None, max_n=9, ctxs=None, show=True, **kwargs):
77 if b is None: b = self.one_batch()
---> 78 if not show: return self._pre_show_batch(b, max_n=max_n)
79 show_batch(*self._pre_show_batch(b, max_n=max_n), ctxs=ctxs, max_n=max_n, **kwargs)
80
/usr/local/lib/python3.6/dist-packages/fastai2/data/core.py in _pre_show_batch(self, b, max_n)
70 b = self.decode(b)
71 if hasattr(b, 'show'): return b,None,None
---> 72 its = self._decode_batch(b, max_n, full=False)
73 if not is_listy(b): b,its = [b],L((o,) for o in its)
74 return detuplify(b[:self.n_inp]),detuplify(b[self.n_inp:]),its
/usr/local/lib/python3.6/dist-packages/fastai2/data/core.py in _decode_batch(self, b, max_n, full)
64 f = self.after_item.decode
65 f = compose(f, partial(getattr(self.dataset,'decode',noop), full = full))
---> 66 return L(batch_to_samples(b, max_n=max_n)).map(f)
67
68 def _pre_show_batch(self, b, max_n=9):
/usr/local/lib/python3.6/dist-packages/fastcore/foundation.py in map(self, f, *args, **kwargs)
360 else f.format if isinstance(f,str)
361 else f.__getitem__)
--> 362 return self._new(map(g, self))
363
364 def filter(self, f, negate=False, **kwargs):
/usr/local/lib/python3.6/dist-packages/fastcore/foundation.py in _new(self, items, *args, **kwargs)
313 @property
314 def _xtra(self): return None
--> 315 def _new(self, items, *args, **kwargs): return type(self)(items, *args, use_list=None, **kwargs)
316 def __getitem__(self, idx): return self._get(idx) if is_indexer(idx) else L(self._get(idx), use_list=None)
317 def copy(self): return self._new(self.items.copy())
/usr/local/lib/python3.6/dist-packages/fastcore/foundation.py in __call__(cls, x, *args, **kwargs)
39 return x
40
---> 41 res = super().__call__(*((x,) + args), **kwargs)
42 res._newchk = 0
43 return res
/usr/local/lib/python3.6/dist-packages/fastcore/foundation.py in __init__(self, items, use_list, match, *rest)
304 if items is None: items = []
305 if (use_list is not None) or not _is_array(items):
--> 306 items = list(items) if use_list else _listify(items)
307 if match is not None:
308 if is_coll(match): match = len(match)
/usr/local/lib/python3.6/dist-packages/fastcore/foundation.py in _listify(o)
240 if isinstance(o, list): return o
241 if isinstance(o, str) or _is_array(o): return [o]
--> 242 if is_iter(o): return list(o)
243 return [o]
244
/usr/local/lib/python3.6/dist-packages/fastcore/foundation.py in __call__(self, *args, **kwargs)
206 if isinstance(v,_Arg): kwargs[k] = args.pop(v.i)
207 fargs = [args[x.i] if isinstance(x, _Arg) else x for x in self.pargs] + args[self.maxi+1:]
--> 208 return self.fn(*fargs, **kwargs)
209
210 #Cell
/usr/local/lib/python3.6/dist-packages/fastcore/utils.py in _inner(x, *args, **kwargs)
351 if order is not None: funcs = funcs.sorted(order)
352 def _inner(x, *args, **kwargs):
--> 353 for f in L(funcs): x = f(x, *args, **kwargs)
354 return x
355 return _inner
/usr/local/lib/python3.6/dist-packages/fastcore/transform.py in decode(self, o, full)
181
182 def decode (self, o, full=True):
--> 183 if full: return compose_tfms(o, tfms=self.fs, is_enc=False, reverse=True, split_idx=self.split_idx)
184 #Not full means we decode up to the point the item knows how to show itself.
185 for f in reversed(self.fs):
/usr/local/lib/python3.6/dist-packages/fastcore/transform.py in compose_tfms(x, tfms, is_enc, reverse, **kwargs)
121 for f in tfms:
122 if not is_enc: f = f.decode
--> 123 x = f(x, **kwargs)
124 return x
125
/usr/local/lib/python3.6/dist-packages/fastcore/transform.py in decode(self, x, **kwargs)
60 def use_as_item(self): return ifnone(self.as_item_force, self.as_item)
61 def __call__(self, x, **kwargs): return self._call('encodes', x, **kwargs)
---> 62 def decode (self, x, **kwargs): return self._call('decodes', x, **kwargs)
63 def setup(self, items=None): return self.setups(items)
64 def __repr__(self): return f'{self.__class__.__name__}: {self.use_as_item} {self.encodes} {self.decodes}'
/usr/local/lib/python3.6/dist-packages/fastcore/transform.py in _call(self, fn, x, split_idx, **kwargs)
68 f = getattr(self, fn)
69 if self.use_as_item or not is_listy(x): return self._do_call(f, x, **kwargs)
---> 70 res = tuple(self._do_call(f, x_, **kwargs) for x_ in x)
71 return retain_type(res, x)
72
/usr/local/lib/python3.6/dist-packages/fastcore/transform.py in <genexpr>(.0)
68 f = getattr(self, fn)
69 if self.use_as_item or not is_listy(x): return self._do_call(f, x, **kwargs)
---> 70 res = tuple(self._do_call(f, x_, **kwargs) for x_ in x)
71 return retain_type(res, x)
72
/usr/local/lib/python3.6/dist-packages/fastcore/transform.py in _do_call(self, f, x, **kwargs)
72
73 def _do_call(self, f, x, **kwargs):
---> 74 return x if f is None else retain_type(f(x, **kwargs), x, f.returns_none(x))
75
76 add_docs(Transform, decode="Delegate to `decodes` to undo transform", setup="Delegate to `setups` to set up transform")
/usr/local/lib/python3.6/dist-packages/fastcore/dispatch.py in __call__(self, *args, **kwargs)
96 if not f: return args[0]
97 if self.inst is not None: f = MethodType(f, self.inst)
---> 98 return f(*args, **kwargs)
99
100 def __get__(self, inst, owner):
/usr/local/lib/python3.6/dist-packages/fastai2/vision/core.py in decodes(self, x)
211
212 def encodes(self, x:TensorPoint): return _scale_pnts(x, self._get_sz(x), self.do_scale, self.y_first)
--> 213 def decodes(self, x:TensorPoint): return _unscale_pnts(x, self._get_sz(x))
214
215 #Cell
/usr/local/lib/python3.6/dist-packages/fastai2/vision/core.py in _unscale_pnts(y, sz)
186 return TensorPoint(res, img_size=sz)
187
--> 188 def _unscale_pnts(y, sz): return TensorPoint((y+1) * tensor(sz).float()/2, img_size=sz)
189
190 #Cell
/usr/local/lib/python3.6/dist-packages/fastai2/torch_core.py in _f(self, *args, **kwargs)
256 def _f(self, *args, **kwargs):
257 cls = self.__class__
--> 258 res = getattr(super(TensorBase, self), fn)(*args, **kwargs)
259 return retain_type(res, self)
260 return _f
RuntimeError: The size of tensor a (18) must match the size of tensor b (2) at non-singleton dimension 0
Oh, this error comes from the fact your model doesn’t output predictions that are bs * n_points * 2 but are flattened.
I see. I am creating my model via cnn_learner in which it is getting that behavior.
The outputted size is a [bs, n_points*2]
How should that tensor size instead look?
Ah yes, that’s incompatible with cnn_learner. I guess we need to fix the decode in the pointscaler then, Will do tomorrow.
2 Likes
Thanks!!!
Got a bug I believe with verify_images. No matter what, it will say that it cannot identify image file. I’ve grabbed a URL in the file manually and looked at it and it was perfectly fine, however if I run verify_images it will flag it as not! @pnvijay also noticed this behavior.
Here is the relevant code:
for i, n in enumerate(classes):
print(n)
path_f = Path(files[i])
download_images(path_f, path/folders[i], max_pics=50)
for n in classes:
print(n)
verify_images(path/n, delete=True, max_size=500)
I’m looking as well now. It’s in with PIL Image.
---------------------------------------------------------------------------
OSError Traceback (most recent call last)
<ipython-input-49-9585efa78fe5> in <module>()
----> 1 img = Image.open(path_bb.ls()[0])
/usr/local/lib/python3.6/dist-packages/PIL/Image.py in open(fp, mode)
2816 for message in accept_warnings:
2817 warnings.warn(message)
-> 2818 raise IOError("cannot identify image file %r" % (filename if filename else fp))
2819
2820
OSError: cannot identify image file '/content/data/wolves/wolves/00000015.jpg'
If you tag me, please include the whole stack trace 
1 Like
Getting warmer! @sgugger got the issue finally. When I try to download a URL with show_progress as False I get the following traceback:
---------------------------------------------------------------------------
AttributeError Traceback (most recent call last)
<ipython-input-126-65db74e53af5> in <module>()
1 download_url(urls[0], Path(f"{0}{suffix[0]}"), overwrite=True,
----> 2 show_progress=False, timeout=4)
/usr/local/lib/python3.6/dist-packages/fastai2/data/external.py in download_url(url, dest, overwrite, pbar, show_progress, chunk_size, timeout, retries)
147 pbar = progress_bar(range(file_size), leave=False, parent=pbar)
148 try:
--> 149 pbar.update(0)
150 for chunk in u.iter_content(chunk_size=chunk_size):
151 nbytes += len(chunk)
AttributeError: 'NoneType' object has no attribute 'update'
Whereas True will work and use a non-corrupted image
This is where the issue is! As I then went through the rest and it worked fine
My code:
def _download_image_inner(dest, inp, timeout=4):
i,url = inp
suffix = re.findall(r'\.\w+?(?=(?:\?|$))', url)
suffix = suffix[0] if len(suffix)>0 else '.jpg'
try: download_url(url, dest/f"{i:08d}{suffix}", overwrite=True, timeout=timeout)
except Exception as e: f"Couldn't download {url}."
def download_images(url_file, dest, max_pics=1000, n_workers=8, timeout=4):
"Download images listed in text file `url_file` to path `dest`, at most `max_pics`"
urls = url_file.read().strip().split("\n")[:max_pics]
dest = Path(dest)
dest.mkdir(exist_ok=True)
parallel(partial(_download_image_inner, dest, timeout=timeout), list(enumerate(urls)), n_workers=n_workers)
1 Like
Ah yes, it misses a test. Should be fixed now.
1 Like
Thanks!
One other little tidbit, with multiple keypoints I cannot just do a batch transform, I must include a resize on item_tfms else I get a tensor-mismatch. The temporary is an item_tfm with Resize() to my size
However on the totem pole that can be pretty low
I imagine this issue actually has to deal with multiple key points being loaded in from a model perspective where they keypoints may not be where the image is. If so let me work on it a bit. If you believe otherwise:
IE:
Will not work:
dbunch = dblock.databunch(newPath, path=newPath, bs=32,batch_tfms=[*aug_transforms(size=(120, 160)), Normalize.from_stats(*imagenet_stats)])
Will work
dbunch = dblock.databunch(newPath, path=newPath, bs=32,item_tfms=Resize(120,160), batch_tfms=[*aug_transforms(size=(120, 160)), Normalize.from_stats(*imagenet_stats)])
2 Likes
This is only necessary if your images are not all of the same size (and not specific to points: you can only batch tensors of compatible shapes).
So if they’re not and we declare a size in batch_tfms, we still need one in the item_tfms? (Or a seperate resize transformation)