How do we use our model against a specific image?

arifoyong · February 22, 2018, 4:16pm

Hi,
I’m learning with Cats & Dogs classification.
I’m trying to predict a specific image as follow:

trn_tfms, val_tfms = tfms_from_model(resnet34, sz, aug_tfms = transforms_side_on, max_zoom = 1.1)
im = val_tfms(np.array(Image.open(fname)))

pred1 = learn.predict_array(im[None])
prob = np.argmax(np.exp(pred1))

However, the results of prob is always 0
I found out the pred1 always the following results:
array([[ 0. ,negative_number]], dtype=float32)

Example:
array([[ 0. , -106.45783]], dtype=float32)

It doesn’t matter if the image is dog or cat, the first element of pred1 is always 0.

My learn model has accuracy of 99.3% and the confusion matrix was pretty good.
I’ve checked my learn.data.classes it is ['cats', 'dogs']

Does anybody have similar issue?

quangduong · February 22, 2018, 10:11pm

I am trying to predict the label for a specific image using code sharing on this thread. I also seeing the different result between predict(isTest=true) and the predict_array() function. I try scaling img/255 before converting it to numpy and now it works. The two results are identical !

quangduong · February 22, 2018, 10:16pm

Try np.array(Image.open(fname))/255. I have a similar issue when working on my dataset. Scaling 1/255 works for me.

binarypoet · February 23, 2018, 2:34pm

Thank you @quangduong! I’ll give it a try as soon as I can.

arifoyong · February 24, 2018, 6:54am

Thank you @quangduong it works.
What is the reason behind it, why do we need to scale it ?
I don’t see that during the training, so why do we have to divide by 255 during the prediction.

ecdrid · February 24, 2018, 11:45am

Maybe it’s done automatically during training and testing when images are present there in the folders and not when we are explicitly using it against the single image?

quangduong · February 26, 2018, 4:18am

As @ecdrid pointed, I took a look into the source code when the images are present in the folders, its pixel values are first scaled into range [0, 1] before further processing (normalized, denormalized, crop, …). I am not sure when we use predict_array() img is scaled or not. But when I compared the transformed img as in this code
im = val_tfms(np.array(Image.open(fname)))
with the one when images are presented in the folder the scales of the two are very different. Plus, you can check the prob
prob = np.argmax(np.exp(pred1))
with the prob when images are presented in folders, again, they are in different magnitude scales. Then, I though something wrong with the scaling and try scale 1/255 as mentioned in previous post and it works.

davidc · March 5, 2018, 12:36am

Yes dividing by 255 fixes this issue of the same class always being returned - thanks.

Edit: Using open_image instead of Image.open will do the divide by 255 as well as converting it to an array so always use that utility function:

im = val_tfms(open_image("/my/image.png"))

Check out source code:

??open_image

JeremieCohen · March 31, 2018, 11:48am

Hello ! I am having the same error as @Vikasj (but with AttributeError: ‘PngImageFile’ object has no attribute ‘shape’)

Do you have any idea how to solve it ?

AttributeError Traceback (most recent call last)
in ()
3
4 trn_tfms, val_tfrms = tfms_from_model(arch, sz)
----> 5 im = val_tfrms(Image.open(f’{PATH}valid/dogs/arnaud.jpg’))
6 log_probs = to_np(learn.models.model(V(T(im[None]).cuda())))
7 preds = np.argmax(log_probs, axis=1)

~/fastai/courses/dl1/fastai/transforms.py in call(self, im, y)
501 crop_tfm = crop_fn_lu[crop_type](sz, tfm_y, sz_y)
502 self.tfms = tfms + [crop_tfm, normalizer, ChannelOrder(tfm_y)]
–> 503 def call(self, im, y=None): return compose(im, y, self.tfms)
504 def repr(self): return str(self.tfms)
505

~/fastai/courses/dl1/fastai/transforms.py in compose(im, y, fns)
482 for fn in fns:
483 #pdb.set_trace()
–> 484 im, y =fn(im, y)
485 return im if y is None else (im, y)
486

~/fastai/courses/dl1/fastai/transforms.py in call(self, x, y)
172 def call(self, x, y):
173 self.set_state()
–> 174 x,y = ((self.transform(x),y) if self.tfm_y==TfmType.NO
175 else self.transform(x,y) if self.tfm_y in (TfmType.PIXEL, TfmType.CLASS)
176 else self.transform_coord(x,y))

~/fastai/courses/dl1/fastai/transforms.py in transform(self, x, y)
180
181 def transform(self, x, y=None):
–> 182 x = self.do_transform(x,False)
183 return (x, self.do_transform(y,True)) if y is not None else x
184

~/fastai/courses/dl1/fastai/transforms.py in do_transform(self, x, is_y)
306 def do_transform(self, x, is_y):
307 if is_y: return scale_min(x, self.sz_y, cv2.INTER_NEAREST)
–> 308 else : return scale_min(x, self.sz, cv2.INTER_AREA )
309
310

~/fastai/courses/dl1/fastai/transforms.py in scale_min(im, targ, interpolation)
10 targ (int): target size
11 “”"
—> 12 r,c,*_ = im.shape
13 ratio = targ/min(r,c)
14 sz = (scale_to(c, ratio, targ), scale_to(r, ratio, targ))

AttributeError: ‘PngImageFile’ object has no attribute ‘shape’

jeremy · March 31, 2018, 6:01pm

This is the recommended approach and will solve most of the problems in this thread.

Bigicecream · April 5, 2018, 9:23pm

Hello, I have trained my model and got 90% accuracy,
but somehow when I predict single image it seems to be wrong, I tested more than 20 and all wrong

motor_city_cobra · April 7, 2018, 8:38pm

Actually my code isn’t quite right

I’ve been trying to get a single image classified for a few weeks now.
I have yet to figure out what code will make this happen.

Not clear to me what ‘the 2nd return val from tfms_from_model’ is.

In this iteration of my failed attempt I am running…
trn_tfms,val_tfms = tfms_from_model(arch, sz)

im = (open_image(PATH+fn))

im_val = val_tfms(im)

preds = learn.predict_array(im[None])
np.argmax(preds)

and getting the error…
RuntimeError: running_mean should contain 1722 elements not 1024

Bigicecream · April 7, 2018, 9:45pm

As far as I understand(from end of lessons 2 I believe) this is the code:
Am I right? @jeremy

But as you see, it didn’t work for me, try run it tell me if it worked for you

cedric · April 23, 2018, 4:04pm

im = val_tfrms(Image.open(f’{PATH}valid/dogs/arnaud.jpg’))

Use open_image instead of Image.open. This is the recommended approach and will avoid most problems related to AttributeError: ‘PngImageFile’ object has no attribute ‘shape’.

fn = f’{PATH}valid/dogs/arnaud.jpg’
im = val_tfrms(open_image(fn))

cedric · April 23, 2018, 4:30pm

Setting learn.precompute = False before you run that prediction will resolved the issue.

trn_tfms, val_tfms = tfms_from_model(arch, sz) # get image transformations
im = val_tfrms(open_image(PATH + fn))
learn.precompute = False # we need to set this since we are passing in an image, not a precomputed activation
preds = learn.predict_array(im[None]) # do prediction
... truncated ...
... truncated ...

motor_city_cobra · April 26, 2018, 2:09am

Does this classify a totally new image the model hasn’t seen before or do you have to show it something from the validation folder?

motor_city_cobra · April 26, 2018, 6:09am

I see that this …

learn.precompute = False
i = '/path/to/photo/'
im = val_tfms(open_image(i))
preds = learn.predict_array(im[None])
print (np.argmax(preds))

…returns 1 or 0.
But what can I do to get it to return a value BETWEEN 1 and 0?

cedric · April 26, 2018, 12:43pm

Yes, it can do both. I have tested both cases:

Image from the validation set.

Totally new image the model hasn’t seen before.
Example: https://www.k9rl.com/dog-breeds/toy-dogs/affenpinscher/

cedric · April 26, 2018, 12:45pm

It shouldn’t returns 1 or 0 or a value between 1 and 0. Your codes are incorrect. Check my previous reply for the correct codes and how the expected results should look like.

motor_city_cobra · April 26, 2018, 8:15pm

With notebook 1 it classifies dog or cat with 0 or 1. It shows probabilities between 0 and 1.
Not really relevant to my question but I’m trying with my own images with categories of want vs pass. (Stylish images I want to use for a website vs images I don’t want to use).
When I show it a new image it hasn’t seen before it returns a 1 or a 0. I want probilities like it gave on the validation images.
Here it’s giving probabilities (between 1 and 0) that it validated on. It “shouldn’t” you say, maybe because it’s not what the model is supposed to do? But it’s what I want.

(Below) These are new images after the model has been trained. They only return 1 or 0. I want probabilities instead because I know the model isn’t very good yet overall but I know anything below a 0.01 or above a 0.8 is correctly classified. The training it does on my few thousand images has an accuracy of 0.81. Pretty good but I need more images.

So I’m trying to use the model to help me build a bigger, better dataset for it to train on.

I know this is probably too much irrelevant information but, again, I want probilities if that’s possible.