fastai efficientdet fails on learn.validate() with AttributeError: 'NoneType' object has no attribute 'shape'

[potipot]

🐛 Bug

when trying to simply validate metrics for an efficientdet model with fastai

KeyError: 'image_id'

AttributeError: 'NoneType' object has no attribute 'shape'

it fails when trying to read the batch size automatically: in accumulate, find_bs

class AvgLoss(Metric):
    "Average the losses taking into account potential different batch sizes"
    def reset(self):           self.total,self.count = 0.,0
    def accumulate(self, learn):
        bs = find_bs(learn.yb)
        self.total += learn.to_detach(learn.loss.mean())*bs
        self.count += bs
    @property
    def value(self): return self.total/self.count if self.count != 0 else None
    @property
    def name(self):  return "loss"

To Reproduce
Steps to reproduce the behavior:

1. Go to https://colab.research.google.com/drive/1i4aXYu4wIKA7eLUK86GwTm7lq7zku_oF?usp=sharing

[potipot]

I found that applying a following patch:

from fastai.learner import AvgLoss
from fastai.torch_core import find_bs

@patch
def accumulate(self:AvgLoss, learn):
    #bs = find_bs(learn.yb)
    bs = find_bs(learn.xb)
    self.total += learn.to_detach(learn.loss.mean())*bs
    self.count += bs

fixes the issue. Im not sure though how should we store that patch in icevision. Anny suggestions?

[potipot]

a linked issue: https://forums.fast.ai/t/solved-fastai2-how-to-change-the-flow-for-metrics-calculation/70851/2

[ganesh3]

I too had this issue which got solved with this patch and online help from @potipot . Thanks a lot. Hopefully, this is pushed as a PR.

[potipot]

@lgvaz this is still an issue. Just wanted to ask for your opinion on where to put this fastai patch?

[lgvaz]

@potipot i get access denied on the colab notebook, can you change the permissions?

[lgvaz]

Wow, this error is really subtle, specially the why it was not happening before.

The error was introduced by #630 but the bug was there since the beginning. To first understand it we have to take a look at how find_bs works:

def find_bs(b):
    "Recursively search the batch size of `b`."
    return item_find(b).shape[0]
    
def item_find(x, idx=0):
    "Recursively takes the `idx`-th element of `x`"
    if is_listy(x): return item_find(x[idx])
    if isinstance(x,dict):
        key = list(x.keys())[idx] if isinstance(idx, int) else idx
        return item_find(x[key])
    return x

In our case x that is passed to item_find will be a dictionary (the effdet target), note that find_bs calls item_find with idx=0, so key = list(x.keys())[**idx**] will take the first key from the dict.

And this is what #630 changed, before the first key in the dict was bbox:

And calling find_bs gives returns .shape[0] of the first prediction, which in this case is 3. Which is incorrect, the batch size is actually 16. So this is a silent bug.

After it started being img_size:

Now find_bs is called on None, hence the error.

---

The solution is to do what @potipot proposed here but without patching, instead we should do the same thing that we do for torchvision models, when creating the learner (models/torchvision/fastai/learner.py):

    class RCNNAvgLoss(fastai.AvgLoss):
        def accumulate(self, learn):
            bs = len(learn.yb)
            self.total += fastai.to_detach(learn.loss.mean()) * bs
            self.count += bs

    recorder = [cb for cb in learn.cbs if isinstance(cb, fastai.Recorder)][0]
    recorder.loss = RCNNAvgLoss()

We also have to check if the same happens for the mmdet models.

@potipot would you like to do a PR for this one?

[potipot]

@lgvaz im not sure whats the idea behind find_bs cause in the test cases it always is 1. Seems like accumulate operates on the current batch element, while the actual batch_size is 2. Copying the solution from RCNNAvgLoss works but now I'm wondering if this isn't some other issue in fact.

[potipot]

In the case of rcnn_learner the results are actually different. Don't know why this is necessary in case of rcnn_but this is what I get:

In[18]: len(learn.yb), find_bs(learn.yb), learn.dls.bs
Out[18]: (1, 2, 2)

[lgvaz]

Wait, I got a bit confused, what solution actually gives the correct batch size?

For rcnn yb is a list of dictonaries, so len(yb) will give the number of samples in the current batch. This is different for effdet where yb is a dictionary. The implementation would be a bit different for both cases

[potipot]

Problem is, I'm not sure what this is trying to do. I wasn't able to get values other than 1 there. So I guess the idea is to take an average of the loss and weigh it with the batch size. We could actually ignore this and just put a blank 1, so no weighing would happen.

When training via tests with bs=2 using RCNN and EfficientDet both I was always getting a single element (not a batch) from learn.yb.

[potipot]

The problem is that the test dataset (fridge_ds) has only 1 element in the validation set. Therefore I cannot test which list corresponds to the actual batch size, as it is always 1. :D I will increase the size of validation set to say 3 and then test again.