(YJ Park) #21

Hello @farlion,

I will try to reproduce your notebook later and we will see how it goes. Thank you for the information.


(Amrit ) #22

Hey @farlion
I was able to get the same error as you using your code. Not sure if this is what you want but this code worked.

def nasnet(pre): return nasnetalarge(pretrained = 'imagenet' if pre else None)
arch = nasnetalarge
batch_size = 64
img_size = 80
stats = ([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
transforms = tfms_from_stats(stats, img_size, aug_tfms=transforms_side_on, max_zoom=1.1)
data = ImageClassifierData.from_csv(PATH, 'train', f'{PATH}labels.csv', test_name='test', val_idxs=validation_set_idxs, tfms=transforms, bs=batch_size)
def get_data(img_size, batch_size):
    transforms = tfms_from_stats(stats, img_size, aug_tfms=transforms_side_on, max_zoom=1.1)
    data = ImageClassifierData.from_csv(PATH, 'train', f'{PATH}labels.csv', test_name='test',
                                        num_workers=4, val_idxs=validation_set_idxs,
                                        tfms=transforms, bs=batch_size)
    return data if img_size > 300 else data.resize(340, 'tmp')

data = get_data(img_size, batch_size)
learn = ConvLearner.pretrained(nasnet, data)

(YJ Park) #23


I was able to reproduce the error you are getting below:

The difference between your version and nasnet notebook by Jeremy (i.e. nasnet.ipynb) is mainly as follow:

Your version


It seems that

  1. nasnet.ipynb uses means and sds to transform data;
  2. ‘model_features[nasnet]=4032*2’ is missing in your notebook; and
  3. You want to make sure you call image because when I used the ‘arch’ variable defined at the beginning, it still had the same error message coming up.

It may be good to have a look at nasnet.ipynb and replicate the steps taken there to see whether things are working. Good luck!

(Florian Peter) #24

@YJP @amritv Thank you so much for the detailed steps, I wasn’t even aware there is a nasnet.ipynb notebook! Will take a look and adapt accordingly.

Thanks for your help, you two are rockstars!


Update: Aaaaaaaaand it’s traaaaaining! (:

(Amrit ) #25

@farlion it’s a good feeling when it trains ain’t it. Lol

(YJ Park) #26

@amritv, @farlion
Have you tried nasnet on the Dog Breed data? Has it worked OK for you?
I am getting an overfitting issue quite quickly with dropout 0.5 (or underfitting with more dropout rates but not improving its log loss) and it simply cannot get to a log loss figure anywhere near 0.20 that we can get from other models. Any advice would be truly appreciated. Thank you.

(Florian Peter) #27

@YJP absolute same here!

My validation loss with nasnet seems to flatten out around a very high 0.75.


I’ve tried data augmentation, higher dropout, unfreezing, lower learning rates, cycle_mult, larger and smaller images…nothing quite seems to help.

(Amrit ) #28

I have not tried it on the dog breed data but did on the plant seedling data. I too was getting overfitting fairly quickly with lower accuracy compared to other models.

(YJ Park) #29

Thank you both for the information. Will try further and let you know if there is any improvement.

(RobG) #30

Hi all, wonderful resources here, thanks! Just 2 weeks into DL, and python, and 3.5 lessons into fastai. It’s been a fun learning experience.

For NASNet and Dog Breed the significant change I made was changing dropout from 1/2 to 2/3, I think that made the difference, managing < 0.152 loss on kaggle and 0.959 accuracy without even unfreezing layers. 0.142 when ensembled. Happy to stop there. Now onto Plant Seedlings.

(Amrit ) #31

That’s awesome, can you give me a guide on how long it took to train? That was another issue for me, the training was sloooooow!

(YJ Park) #32


Thank you for your helpful post. I tried with the dropout rate you suggested.

I realised the main difference between my notebook and Jeremy’s nasnet.ipynb was an allocation of a validation set. It looks like nasnet.ipynb did not allocate any validation set and the default is


So I tried nasnet on dog breed images with no validation set (I think the default is one when there is no specific val_idxs) and the accuracy went up to 95.6% with a training log loss around 0.16.

(RobG) #33

I’m very new to this so forgive me if these are naive speculations rather than true insight.

My guess is a higher dropout might work because from my reading NASNet looks like a very wide architecture at each layer.

Also, I wouldn’t say one should drop the val_idx’s to zero at the outset. I don’t. My guess is that the reason this works for you at the outset is that NASNet already works well ‘out the box’. It doesn’t need to attenuate against a validation set. So including the val_set at the start, which means not having one, doesn’t have a detrimental effect. This isn’t surprising because the dog images are very very similar to imagenet. EDIT: Looks like if you don’t include a validation set from a csv load then fastai just calculates one for you.

For this reason, I don’t think the model needs long to train at all. It’s about getting those hyper-parameters right. When we do run an action, it doesn’t improve much from training much at all. NASNet is slow, but we don’t need to get it to do much. I did run 45 min training after precompute = false, but it didn’t do very much and 10 min training might have done just as well.

Final guess is that I think the Jeremy’s nasnet notebook doesn’t contain any validation set because its meant as a demonstrator of a new architecture not to solve a problem.

One thing I wasn’t able to get to work correctly is unfreezing. unfreeze() wasn’t successful, nor the freeze_to() already in fastai, nor the new freeze_to() in Jeremy’s nasnet.ipynb. I see he uses ‘17’ in his notebook, from my reading it looks like it has 17 layers with a final fc layer. I’m not sure what if any new effect freezing to layer 17 would have as that’s what it already was? Any info about how to unfreeze and to what layer would be useful for further use cases. Maybe this is covered in future lessons I’ve not watched.

Glad the hint helped!

(RobG) #34

Re no val_idx. Actually, looking at - I’ve only just started with python - but it looks like for a csv load if you don’t include a value it calculates a default set for you, which early lessons did outside the function, and for a PATH load, like jeremy’s nasnet.iypnb, well the validation images are in a path anyway. So in your case with dog breed it just did it for you. Try passing in [0] for val_idx and see if it improves on kaggle (I don’t think a local logloss calculation is any longer meaningful with an empty validation set).

(YJ Park) #35


You are correct and apologies for the incorrect comment about the default in Thank you for looking into this.

Looking at ‘from_csv’ below from, it does pull through 20% if it has not been specified. So the default will be 20% of the images, not 1 image.

    def from_csv(cls, path, folder, csv_fname, bs=64, tfms=(None,None),
               val_idxs=None, suffix='', test_name=None, continuous=False, skip_header=True, num_workers=8):
        """ Read in images and their labels given as a CSV file.

        This method should be used when training image labels are given in an CSV file as opposed to
        sub-directories with label names.

            path: a root path of the data (used for storing trained models, precomputed values, etc)
            folder: a name of the folder in which training images are contained.
            csv_fname: a name of the CSV file which contains target labels.
            bs: batch size
            tfms: transformations (for data augmentations). e.g. output of `tfms_from_model`
            val_idxs: index of images to be used for validation. e.g. output of `get_cv_idxs`.
                If None, default arguments to get_cv_idxs are used.
            suffix: suffix to add to image names in CSV file (sometimes CSV only contains the file name without file
                    extension e.g. '.jpg' - in which case, you can set suffix as '.jpg')
            test_name: a name of the folder which contains test images.
            continuous: TODO
            skip_header: skip the first row of the CSV file.
            num_workers: number of workers

        fnames,y,classes = csv_source(folder, csv_fname, skip_header, suffix, continuous=continuous)

        **val_idxs = get_cv_idxs(len(fnames)) if val_idxs is None else val_idxs**
        ((val_fnames,trn_fnames),(val_y,trn_y)) = split_by_idx(val_idxs, np.array(fnames), y)

        test_fnames = read_dir(path, test_name) if test_name else None
        if continuous:
            f = FilesIndexArrayRegressionDataset
            f = FilesIndexArrayDataset if len(trn_y.shape)==1 else FilesNhotArrayDataset
        datasets = cls.get_ds(f, (trn_fnames,trn_y), (val_fnames,val_y), tfms,
                               path=path, test=test_fnames)
        return cls(path, datasets, bs, num_workers, classes=classes)

In my case, I found a log loss figure improved in Kaggle when I trained on a whole data set -1 (val_idxs = [0]).

About unfreezing for the Dog Breed classification data, apparently it is best to use the pre-trained model for this particular case:

(Jeremy Howard) #36

Thanks for the PR to! I’ve merged that now. If there are any improvements suggested to the notebook, I’ll happily take PRs there too :slight_smile:

(Greg Stoddard) #37

Hi all,

I just spun up my instance on Paperspace for lesson 1 and I’m getting an import error when running the imports in lesson1.ipynb. The error message gives a “Syntax error: “Return” outside of function” and it traces it back to line 620 in My guess is that it was broken by the recent merge but I’m not sure. I’m attaching the error below.

(YJ Park) #38


Hi Jeremy, ‘return model’ was outside of def nasnetalarge. Sorry but could you please change this. Thank you in advance.

(Surya Mohan) #39

Hi all, I just started coding up lesson 1 on AWS and when I was importing the fastai libraries, it gave the same error ‘return outside function’.

Please let me know if you figure out a way to fix this.

(Francesco) #40

Hi Sumo, is this your error?