How can we change model_dir?

Part 1 (2019)
1

Currently it is “models”, as a local directory.
Then when directories represent categories, a “new” spurious category “models” appears…
Which then appears as a row and column in confusion_matrix, etc.
I suggest renaming it “.models” as default so it will not be picked up, eg by .ls, etc.
Alternatively/additionally it could be a parameter, eg to ConvLearner.

1 Like
2

you should be able to override the directory by providing a models_name model_dir parameter, as you can see here.

Just pass in an extra parameter called model_dir='.models' to your ...Learner instantiation, if you want the weights to be saved in .models

3 
learn = ConvLearner(data, models.resnet152, metrics=accuracy, models_name='.models')

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-70-c2a85b42c6e9> in <module>()
----> 1 learn = ConvLearner(data, models.resnet152, metrics=accuracy, models_name='.models') #], pretrained=True)

~/fastai/fastai/vision/learner.py in __init__(self, data, arch, cut, pretrained, lin_ftrs, ps, custom_head, split_on, **kwargs)
     56         head = custom_head or create_head(nf, data.c, lin_ftrs, ps)
     57         model = nn.Sequential(body, head)
---> 58         super().__init__(data, model, **kwargs)
     59         self.split(ifnone(split_on,meta['split']))
     60         if pretrained: self.freeze()

TypeError: __init__() got an unexpected keyword argument 'models_name'

I did … but did not work…

4

The correct name of that argument is model_dir.

4 Likes
5

Yes, it is. Thanks!
(I will clean my eyeglasses, did not find it at first…) :wink:

1 Like
6

:slight_smile: The pointer above was to the OLD learner…

7

Oops. My bad ! Glad that you got it going. will update links and description above, so it doesn’t confuse others.