How to predict using learn.fit_one_cycle model

Part 1 (2020)
1

I trained a model using the MNIST_SAMPLE (3s and 7 digits).

path = untar_data(URLs.MNIST_SAMPLE)
dls = ImageDataLoaders.from_folder(path)
learn=cnn_learner(dls,resnet18,pretrained=False,loss_func=F.cross_entropy,metrics = accuracy)
learn.fit_one_cycle(1,0.1)

Now I wish to test this model on a single image.
I upload the images:

threes = (path/'train'/'3').ls()
threes_data = [Image.open(i) for i in threes]
tensor_3 =[tensor(i).float()/255 for i in threes_data]
img = tensor_3[0]

Prediction:
learn.predict(img)

Getting the following error message:
---------------------------------------------------------------------------
RuntimeError Traceback (most recent call last)
/usr/local/lib/python3.6/dist-packages/fastai/learner.py in with_events(self, f, event_type, ex, final)
154 def with_events(self, f, event_type, ex, final=noop):
–> 155 try: self(f’before{event_type}’) ;f()
156 except ex: self(f’after_cancel{event_type}’)

30 frames
RuntimeError: Given groups=1, weight of size [64, 3, 7, 7], expected input[1, 1, 28, 28] to have 3 channels, but got 1 channels instead

During handling of the above exception, another exception occurred:

TypeError                                 Traceback (most recent call last)
TypeError: expected Tensor as element 0 in argument 0, but got NoneType

During handling of the above exception, another exception occurred:

TypeError                                 Traceback (most recent call last)
/usr/local/lib/python3.6/dist-packages/fastcore/utils.py in range_of(x)
    209 def range_of(x):
    210     "All indices of collection `x` (i.e. `list(range(len(x)))`)"
--> 211     return list(range(len(x)))
    212 
    213 # Cell

TypeError: object of type 'NoneType' has no len()
2

It seems there needs to be three channels but you have given one during inference?

1 Like
3

Thank you!
Somehow this slipped by me.
I tried to solve this by reshaping the img file.

img = img.view(1,1,28,28)

type(img),img.shape

(torch.Tensor, torch.Size([1, 1, 28, 28]))

But upon predicting I get

learn.predict(img)

KeyError Traceback (most recent call last)
/usr/local/lib/python3.6/dist-packages/PIL/Image.py in fromarray(obj, mode)
2679 try:
-> 2680 mode, rawmode = _fromarray_typemap[typekey]
2681 except KeyError:

KeyError: ((1, 1, 28, 28), ‘<f4’)

During handling of the above exception, another exception occurred:

TypeError Traceback (most recent call last)
25 frames
/usr/local/lib/python3.6/dist-packages/PIL/Image.py in fromarray(obj, mode)
2680 mode, rawmode = _fromarray_typemap[typekey]
2681 except KeyError:
-> 2682 raise TypeError(“Cannot handle this data type: %s, %s” % typekey)
2683 else:
2684 rawmode = mode

TypeError: Cannot handle this data type: (1, 1, 28, 28), <f4

4

just a rookie question here (correct me if I am wrong)
When predicting, shouldn’t you pass an actual image instead of the image converted to a tensor?

1 Like
5

Can you predict by looping through the images in a folder? If so, like @jimmiemunyi says, can you try: learn.predict(path/‘test/img.jpg’), where ‘img.jpg’ is the name of the image you are trying to predict?

Two points:

  1. Are you using a ResNet Model? If not, can you tell what model you use?
  2. On the error:
    KeyError Traceback (most recent call last)
    /usr/local/lib/python3.6/dist-packages/PIL/Image.py in fromarray(obj, mode)
    2679 try:
    -> 2680 mode, rawmode = _fromarray_typemap[typekey]
    2681 except KeyError:

KeyError: ((1, 1, 28, 28), ‘<f4’)

I think the issue is with Pillow not being able to read the image. See this question: https://stackoverflow.com/questions/60138697/typeerror-cannot-handle-this-data-type-1-1-3-f4

6

Hi @jimmiemunyi the model was trained on a tensor of the image so I should probably pass it the same tensor shape.

1 Like
7

Hey,

Tried to pass an image file but it didn’t works. See attached showing first the image I used (showing it exists) and the error.

image
image1120×772 38.5 KB

8

  1. I’m using a resnet model
    learn=cnn_learner(dls,resnet18,pretrained=False,loss_func=F.cross_entropy,metrics = accuracy)

  2. The error you mentioned seems to be related to np while I’m using torch tensor. I tried multiplying by 255 but that didn’t help

So frustrating, you’d expect that predicting would be the easiest part…

9

@jimmiemunyi - this is also with regards to your suggestion.

10

thanks for clarification. hope you find a solution

1 Like
11

So, the input type needs to be in one of the formats shown in the screenshot. The error must have some simple fix, don’t be frustrated :slight_smile:

Can you tell me the source of the dataset? I will train with FastAI’s ResNet model and come back if someone has not resolved your query already!

1 Like
12

@karthikrI appreciate your support. It encourages me to keep looking for a solution.

Here’s the complete code from the lesson (ran on Google Colab):

!pip install -Uqq fastbook
import fastbook
fastbook.setup_book()
from fastai.vision.all import *
from fastbook import *
matplotlib.rc('image', cmap='Greys')

 path = untar_data(URLs.MNIST_SAMPLE)
dls = ImageDataLoaders.from_folder(path)
learn=cnn_learner(dls,resnet18,pretrained=False,loss_func=F.cross_entropy,metrics = accuracy)
learn.fit_one_cycle(1,0.1)

The only progress I made so far was to be able to get predictions for the training data using the following:

(tr,vl)= dls.one_batch()
learn.get_preds(dl=[(tr,vl)])

Of course the whole idea of a code is allow a prediction from data outside the training and test.
This I still can’t do.

Let me know if you manage to figure this out.

13

I tried debugging your code and it looks like the loss_func you pass, could be the issue.

Please change this line:
learn=cnn_learner(dls,resnet18,pretrained=False,loss_func=F.cross_entropy,metrics = accuracy)

to one of the below:

  1. When giving the loss_func: learn=cnn_learner(dls,resnet18,pretrained=False,metrics = accuracy, loss_func=CrossEntropyLossFlat())

  2. When not giving the loss_func:. learn=cnn_learner(dls,resnet18,pretrained=False,metrics = accuracy)
    #You do not have to specify a loss_func as Fastai knows by deafult. You can print it confirm
    print(learn.loss_func)

  3. Inference: Please pass an image not in your training, below is for example only

If an Individual item
#Input is an pathlib.PosixPath type
threes = (path/‘train’/‘3’).ls()
learn.predict(threes[0])

#Input is an image, in a folder:
learn.predict(’/images/three.png’)

If a batch (I got this from another thread in the forum)
dl = learn.dls.test_dl(threes)
inp, preds,_,dec_preds = learn.get_preds(dl=dl, with_input=True, with_decoded=True)
full_list = learn.dls.decode_batch((*tuplify(inp),*tuplify(dec_preds)), max_n=18) #18 inputs

print(len(full_list)) #This gives 18, the same number of images we passed as input above

#Get the Preds from the list of tuples:
numbers = [ num[1] for num in full_list]

#This is the ‘number’ predicted by the model. Since we fed all threes, we should get those
print(numbers)

3 Likes
14

It Worked :fireworks: :fireworks: :fireworks:
You rock big time!

image

If I understand correctly the change was that you replaced loss_func=CrossEntropyLosst() with loss_func=CrossEntropyLossFlat()

Can you please explain how this made a difference and more importantly how you found that was the problem and how you found the solution?

You are amazing. Thank you so much!

P.s.it’s very strange that the problem was with the learner code as it’s copy pasted from Jeremy’s lesson…

1 Like
15

Glad it worked.

Since your predict method was similar to what I use, I went backwards to see what could be the change. I looked at the loss functions in GitHub and noticed they were all ‘flat’. I felt changing your loss to something that uses ‘flat’ would work and it did (note, you can still write your custom loss functions)…

While I was debugging the code to understand why, I noticed another thread:
Image Classifier learn.predict(img) and Image Classifier learn.predict(img). It looks like FastAI v2 needs the ‘flat’ data, which your loss function would not provide.

2 Likes
16

Not in a million years would I have figured this out.
I learned so much from this interaction.

Many (Many) thanks!

1 Like