Fastai v2 chat

rmkn85 · October 14, 2019, 9:15pm

(moved everything to previous post)

slawekbiel · October 15, 2019, 5:07pm

Where does the DataBunch.c attribute get set? It’s used in get_c() function, but I couldn’t find the code that sets it.

sgugger · October 15, 2019, 8:54pm

It may be set by one of the transforms.

ilovescience · October 15, 2019, 9:31pm

Hello,

I am unsure if these questions are for mainly v2 or both v1 and v2. I might be able to answer some of these questions:

I am unsure but I guess you could probably try this yourself and see
Don’t know
For fastai v1, you could do learn.recorder.plot(suggestion=True) but I am not sure if this is the same for v2
Don’t know
For fastai v1 we have the OverSamplingCallback that uses the weighted random sampler. I will probably write a version for fastai v2 in the next couple weeks.

fanyi · October 16, 2019, 12:18pm

I can connect to WSL and run python/jupyter, but for fastai2, it reports No module named ‘local’. So who knows what I am missing?

rmkn85 · October 16, 2019, 9:25pm

See resolution here:

rmkn85 · October 16, 2019, 9:27pm

I was referring to @jeremy’s plans for v2, exactly bacuse some was available in v1 as you mentioned.

ilovescience · October 16, 2019, 10:21pm

Ah ok I will try to answer again based on what’s in the code now:

Here is the code for the zoom. It seems it uses PyTorch .uniform_ from 1 to max_zoom, but I am not sure if it will work if max_zoom<1 and the upper bound is actually smaller than the lower bound.
Not sure
The code for the LR finder plot is here. There is no suggestion arguments right now.
Not sure
Again, I will develop this as a callback soon.

I hope this helps answer some of your questions.

jeremy · October 16, 2019, 10:48pm

There’s weighted_databunch now, although it’s not well tested.

ilovescience · October 16, 2019, 11:28pm

I was not aware of this. Does this mean there is no need for an OverSamplingCallback like there was for fastai v1?

rmkn85 · October 17, 2019, 12:01am

Why does drop_last equal shuffles?

github.com

fastai/fastai_dev/blob/c391844103e3e0bff0cfb417a8340deea74ffecd/dev/local/data/core.py#L144


    def n_subsets(self): return len(self.splits)
    def _new(self, items, **kwargs): return super()._new(items, splits=self.splits, **kwargs)
    def subset(self): raise NotImplemented


    def databunch(self, bs=16, val_bs=None, shuffle_train=True, n=None, dl_type=None, dl_kwargs=None, **kwargs):
        if dl_kwargs is None: dl_kwargs = [{}] * self.n_subsets
        ns = self.n_subsets-1
        bss = [bs] + [2*bs]*ns if val_bs is None else [bs] + [val_bs]*ns
        shuffles = [shuffle_train] + [False]*ns
        if dl_type is None: dl_type = self._dl_type
        dls = [dl_type(self.subset(i), bs=b, shuffle=s, drop_last=s, n=n if i==0 else None, **kwargs, **dk)
               for i,(b,s,dk) in enumerate(zip(bss,shuffles,dl_kwargs))]
        return DataBunch(*dls)


FilteredBase.train,FilteredBase.valid = add_props(lambda i,x: x.subset(i), 2)


#Cell
class TfmdList(FilteredBase, L):
    "A `Pipeline` of `tfms` applied to a collection of `items`"
    _default='tfms'
    def __init__(self, items, tfms, use_list=None, do_setup=True, as_item=True, split_idx=None, train_setup=True, splits=None):

jeremy · October 17, 2019, 1:06am

That’s right. I just added this thingie the other day

jeremy · October 17, 2019, 1:07am

Generally you just want your training set to have drop_last.

rmkn85 · October 17, 2019, 6:06am

So it’s implicit there that training set has shuffle=True and test set has shuffle=False?

MadeUpMasters · October 17, 2019, 2:39pm

Are there any end-to-end training examples where we can see models (especially vision) training using an example of the high-level API? I checked around the nbs but, besides a brief example in 21_tutorial_imagenette, didn’t see any.

I just got one working for audio. 97% accuracy on speaker recognition using basic fit and xresnet18, even though the rest of our audio stuff is still pretty buggy.

muellerzr · October 17, 2019, 3:38pm

IIRC Jeremy said the high level isn’t quite done yet. Just medium and low level

muellerzr · October 19, 2019, 8:23pm

Do we have an example of inference on a test set? (Or how to add that to the databunch?)

sgugger · October 19, 2019, 9:36pm

At the end of the imagenette tutorial, there is an example of adding a test set (there is a test_dl function for that). Note that this part is still under development and isn’t thoroughly tested yet.

muellerzr · October 21, 2019, 12:44pm

Do we know the equivalent for a pip install for mkl? I cannot import anything from fastai2.basics due to an import error “No module named ‘mkl’”. Which seems to be only installable via conda (but I cannot do that as I’m using Google Colab). Thoughts?

Edit: for those who do not know (like me just now!), pip is for python packages, mkl is not python

Edit x2: things I have tried:

I tried following this to install conda, and when I double checked and did a !conda install mkl-service (and did regular mkl too) it says it’s already installed.

fanyi · October 21, 2019, 1:54pm

same thing happen to me. I can only turn around by remove the “MKL” in local/core/imports.py for the time being. Hope anyone can help to solve the problem.