Predicting on a single image

Part 1 (2019)
#1

Hi guys

I am working on a Image Regression task. The model is working fine and now I would like to test this on some real life sample cases. However, I am unable to use learn.predict and getting errors:

  1. Here is my model:

    learn = Learner(data, model, loss_func = F.l1_loss, model_dir = “/temp/model/”)

I used Learner approach because I customized cnn_learner model to add few linear layers in the end for my regression task.

Now I am using following method to predict on a single image:

img = open_image('../input/pic.jpg')
learn.predict(img.data)

But this is giving me following error:
’Tensor’ object has no attribute 'apply_tfms’

Can anyone please tell me how to resolve this issue?

Best Regards
Abhik

#2

Try:
learn.predict(img)
The result of open should be and Image(ItemBase) that has apply_tfms method

#3

Thanks Stefano, but unfortunately, it didn’t work. I tried everything but apply_tfms is unwilling to go.

Here is the screen shot of error:

image
image.png724×420 53.7 KB

This is what I tried for the above error:
learn.predict(img.data)

I am marking @sgugger here if I am doing something horribly wrong

#4

You should do learn.predict(img), as Stefano said, not img.data.

2 Likes
#5

Thanks @sgugger, I did exactly that and the error was the same which I pasted above as screen-shot (Tensor object has no attribute ‘apply_tfms’)

Here is the code:

img = open_image('../input/picture-1/55632758_1035600193317551_7808214316078858240_n.jpg')
`learn.predict(img)`
#6

In the screenshot above you said you typed learn.predict(img.data) not learn.predict(img).
This code should work, so we can’t help without knowing more about how that learner was created.

#7

Here is the data creation:

    tfms = get_transforms(max_zoom=1., max_lighting=0.20,
                      max_warp=0., xtra_tfms=[crop_pad(size=600, padding_mode='zeros', row_pct=0.,col_pct=0.), symmetric_warp(magnitude=(-0.1,0.1)) ])
path_csv = '../input/appa-real-face-cropped/'
path_folder = '../input/appa-real-face-cropped/final_files/final_files/'

data = ImageList.from_csv(path, csv_name = '/kaggle/input/appa-real-face-cropped/labels.csv', folder = '.').split_by_rand_pct(0.2).label_from_df(label_cls=FloatList).transform(tfms, size=224).databunch(bs=64).normalize(imagenet_stats)

Here is the model I created:

class AgeModel(nn.Module):
    def __init__(self):
        super().__init__()
        layers = list(models.resnet50(pretrained=True).children())[:-2]
        layers += [AdaptiveConcatPool2d(),Flatten()]
        layers += [nn.Linear(4096, 1024), nn.ReLU(), nn.Linear(1024,256), nn.ReLU(), nn.Linear(256,64), nn.ReLU(), nn.Linear(64,1)]
        self.agemodel = nn.Sequential(*layers)
    def forward(self, x):
        return self.agemodel(x).squeeze()

Here is the learner:

learn = Learner(data, model, loss_func = F.l1_loss, model_dir = "/temp/model/")

Then I am splitting the model like this:

learn.split([model.agemodel[2],model.agemodel[5], model.agemodel[-7]])

Then training:

learn.freeze_to(-1)
learn.fit_one_cycle(5)
#8

Ok, and the full stack trace please. You truncated the error message up there.

#9

Sorry, I had to get the code running again to get the error :slight_smile:

Here is the full error (this time I got something different :slight_smile: ):

image
image.png943×278 25.1 KB

#10

How do you created your data bunch?

#11

This is how I created my databunch:

tfms = get_transforms(max_zoom=1., max_lighting=0.20,
                      max_warp=0., xtra_tfms=[crop_pad(size=600, padding_mode='zeros', row_pct=0.,col_pct=0.), symmetric_warp(magnitude=(-0.1,0.1)) ])
path_csv = '../input/appa-real-face-cropped/'
path_folder = '../input/appa-real-face-cropped/final_files/final_files/'

data = ImageList.from_csv(path, csv_name = '/kaggle/input/appa-real-face-cropped/labels.csv', folder = '.').split_by_rand_pct(0.2).label_from_df(label_cls=FloatList).transform(tfms, size=224).databunch(bs=64).normalize(imagenet_stats)
#12

This is the line that’s causing your problem. You should only squeeze the last axis because when you want to predict, you pass a batch of size 1 and you end up squeezing that dimension, which makes your output a rank-0 tensor instead of being a tensor of size 1.

2 Likes
#13

Hi @sgugger: thanks a lot! But can I ask you how to squeeze only the last axis? Sorry for troubling you on this…

#14

Last dimension is -1, so self.agemodel(x).squeeze(-1).

1 Like
#15

Thanks a lot!! :slight_smile:

I will make the changes and let you know if this solution worked.

Thanks again for your time and help.

#16

@sgugger : it worked like magic… thanks again!! :smiley:

1 Like