Plot top losses has always worked great for me, but as of this morning I’ve noticed a problem:
The first picture shows correctly and then the rest of my grid just shows empty graphs. Any idea what is going on here or when it might be fixed?
3 Likes
do you have it in a colab you can share?
It’s like this in all of my notebooks now. Someone else actually tipped me off to the problem. Here’s an example I just built off of fast.ai’s Google Colab NB5:
Other interpreter functions, most_confused and confusion_matrix work great! Top losses is the only one broken.
Thanks,
Alex
Thanks guys, made Jeremy aware of this issue. Hopefully gets resolved soon 
2 Likes
Here’s the previous working version if you want to temporarily use it.
def plot_top_losses(self, k, largest=True, **kwargs):
losses,idx = self.top_losses(k, largest)
if not isinstance(self.inputs, tuple): self.inputs = (self.inputs,)
if isinstance(self.inputs[0], Tensor): inps = tuple(o[idx] for o in self.inputs)
else: inps = self.dl.create_batch(self.dl.before_batch([tuple(o[i] for o in self.inputs) for i in idx]))
b = inps + tuple(o[idx] for o in (self.targs if is_listy(self.targs) else (self.targs,)))
x,y,its = self.dl._pre_show_batch(b, max_n=k)
b_out = inps + tuple(o[idx] for o in (self.decoded if is_listy(self.decoded) else (self.decoded,)))
x1,y1,outs = self.dl._pre_show_batch(b_out, max_n=k)
if its is not None:
#plot_top_losses(x, y, its, outs.itemgot(slice(len(inps), None)), L(self.preds).itemgot(idx), losses, **kwargs)
plot_top_losses(x, y, its, outs.itemgot(slice(len(inps), None)), self.preds[idx], losses, **kwargs)
#TODO: figure out if this is needed
#its None means that a batch knows how to show itself as a whole, so we pass x, x1
#else: show_results(x, x1, its, ctxs=ctxs, max_n=max_n, **kwargs)
3 Likes
#can use by passing in your interp obj like so…
plot_top_losses(interp, 9, nrows=3)
1 Like
This is fixed in https://github.com/fastai/fastai/pull/3510 so it should work again in the next release. Alternatively you could also install the dev version of fastai for the fix.
2 Likes
Please give an example of how the fix is to be used. I tried to use plot_top_losses I received the following:
TypeError: plot_top_losses() takes from 2 to 3 positional arguments but 6 were given.
my input to interp.plot_top_losses which received the error: interp.plot_top_losses(21, nrows=7, figsize=(18,20))
should be changed to what in plot_top_losses?
I still have the same issue.
I fixed it with this:
I redefined the function:
def plot_top_losses_fix(interp, k, largest=True, **kwargs):
losses,idx = interp.top_losses(k, largest)
if not isinstance(interp.inputs, tuple): interp.inputs = (interp.inputs,)
if isinstance(interp.inputs[0], Tensor): inps = tuple(o[idx] for o in interp.inputs)
else: inps = interp.dl.create_batch(interp.dl.before_batch([tuple(o[i] for o in interp.inputs) for i in idx]))
b = inps + tuple(o[idx] for o in (interp.targs if is_listy(interp.targs) else (interp.targs,)))
x,y,its = interp.dl._pre_show_batch(b, max_n=k)
b_out = inps + tuple(o[idx] for o in (interp.decoded if is_listy(interp.decoded) else (interp.decoded,)))
x1,y1,outs = interp.dl._pre_show_batch(b_out, max_n=k)
if its is not None:
#plot_top_losses(x, y, its, outs.itemgot(slice(len(inps), None)), L(self.preds).itemgot(idx), losses, **kwargs)
plot_top_losses(x, y, its, outs.itemgot(slice(len(inps), None)), interp.preds[idx], losses, **kwargs)
#TODO: figure out if this is needed
#its None means that a batch knows how to show itself as a whole, so we pass x, x1
#else: show_results(x, x1, its, ctxs=ctxs, max_n=max_n, **kwargs)
And using the command below I can now plot the top losses:
plot_top_losses_fix(interp, 10, nrows=2)
Hope this helps for the people who have still the same problem.
1 Like
thanks.
Also as it is fixed in the current github repo can install dev version in colab from there.
# in colab.
!pip install --upgrade git+https://github.com/fastai/fastai.git
1 Like
Thank You All!
Just scratched my head with this problem. Being a newbee everything is really scary. No I might get through page one.
1 Like
Did you try from github?
I installed the upgrade and now it doesn’t seem to work at all. Before, it just plotted the first sample, as described above.
Edit: There are more than 5 samples in the test set and yes, it’s the same error for k=1, as well.
self.items is a len()=1 list that contains an array with all the file_paths from which data is read from.
It should try to get self.items[0][i_], but it doesn’t. Not sure how to solve this. Will revert to an older version for now.
Edit:
I just wanted to tell you that the solution by Valentino et al. works flawlessly for me. I’m wondering, why the updated fastai doesn’t work.
