Input shape of image for the classifiers is wrong

[federicoromeo]

I downloaded the 40 classifiers here, but when i try to run it on an image it gives strange exceptions.

If i run the classifier with:

classifier.run( image , None )

where image is:

image = np.asarray(Image.open('drive/MyDrive/sample_img.jpg').resize((256,256) ) #(256, 256, 3)
image = np.expand_dims(image, axis=0) #(1, 256, 256, 3) = (NWHC)
image = convert_images_from_uint8(images=image,nhwc_to_nchw=True) #(1, 3, 256, 256) = (NCHW)

( NB: convert_images_from_uint8 defined here )

this error comes:
_UnimplementedError (see above for traceback): Generic conv implementation only supports NHWC tensor format for now.
[[node celebahq-classifier-20-goatee/Run/celebahq-classifier-20-goatee/FromRGB_lod0/Conv2D (defined at :159) ]]

But if i comment the function convert_images_from_uint8, thus giving an NWHC image, this arises:

_InvalidArgumentError: input depth must be evenly divisible by filter depth: 256 vs 3
[[{{node celebahq-classifier-20-goatee/Run/celebahq-classifier-20-goatee/FromRGB_lod0/Conv2D}}]]

It seems to me that there are some problems/confusion among the dimensions, since the 256 appearing in the error is not a random number.
How can i fix it?

[federicoromeo]

I even tried to use

logits = classifier.get_output_for( image , None )
>>> print(logits)
>>> <tf.Tensor 'celebahq-classifier-00-male_1/scores_out:0' shape=(1, 1) dtype=float32>

predictions = tf.nn.softmax(tf.concat([logits, -logits], axis=1))
>>> print(predictions)
>>> <tf.Tensor 'Softmax:0' shape=(1, 2) dtype=float32>

as the authors did in the original StyleGAN2 paper repo here but i end up with a tf.Tensor from which i am not able to retrieve the classifier's value

[betterze]

We provide codes for running the classifiers in GetAttribute.py. Please refer to DCI '3 annotate the images using pretrained classifers' for how to use the GetAttribute.py.

Specifically, in line 67-69

            tmp_imgs=imgs[(i*batch_size):((i+1)*batch_size)]  #imgs is a numpy array of shape (n, h, w, 3)
            tmp_imgs=convert_images_from_uint8(tmp_imgs, drange=[-1,1], nhwc_to_nchw=True)
            tmp = classifier.run(tmp_imgs, None)

[federicoromeo]

Using GetAttribute.py it just gives me zeros for all the classes, even if i provide a clear example of a male and a female. Isn't it strange? (but at least it works, don't know how)

I already tried the exact code you just posted in my first code snippet of this issue here; i repost it to be more similar to yours:

i = 0
batch_size = 2

image1 = np.asarray( Image.open('drive/MyDrive/sample_img.jpg').resize((256,256)) )
image2 = np.asarray( Image.open('drive/MyDrive/sample_img2.jpg').resize((256,256)) )
imgs = np.array([image1,image2]) #(2,256,256,3)
np.save('imgs.npy', imgs)
imgs = np.load('imgs.npy') #(2,256,256,3)

for classifier in classifiers:
    tmp_imgs = imgs[(i*batch_size):((i+1)*batch_size)] #(2,256,256,3)
    tmp_imgs = convert_images_from_uint8(tmp_imgs, drange=[-1,1], nhwc_to_nchw=True) #(2,3,256,256)
    tmp = classifier.run( tmp_imgs , None)

and the result remains the same:
_UnimplementedError (see above for traceback): Generic conv implementation only supports NHWC tensor format for now.
or
_InvalidArgumentError: input depth must be evenly divisible by filter depth: 256 vs 3 if i comment line 68

[federicoromeo]

I replied, since the issue remains. Thanks for your time, i attend your response, FR. Da: Zongze ***@***.***> Inviato: domenica 27 novembre 2022 16:11 A: ***@***.***> Cc: Federico ***@***.***>; ***@***.***> Oggetto: Re: [betterze/StyleSpace] Input shape of image for the classifiers is wrong (Issue #32) We provide codes for running the classifiers in GetAttribute.py<https://github.com/betterze/StyleSpace/blob/main/GetAttribute.py>. Please refer to DCI<https://github.com/betterze/StyleSpace#dci-metric> '3 annotate the images using pretrained classifers' for how to use the GetAttribute.py. Specifically, in line 67-69<https://github.com/betterze/StyleSpace/blob/main/GetAttribute.py#L67> tmp_imgs=imgs[(i*batch_size):((i+1)*batch_size)] #imgs is a numpy array of shape (n, h, w, 3) tmp_imgs=convert_images_from_uint8(tmp_imgs, drange=[-1,1], nhwc_to_nchw=True) tmp = classifier.run(tmp_imgs, None) — Reply to this email directly, view it on GitHub<#32 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AOWYJQQSOIWHQUICKVRYCHLWKN2ZNANCNFSM6AAAAAASLRB3JA>. You are receiving this because you authored the thread.Message ID: ***@***.***>

[betterze]

Are you using tensorflow gpu version? I believe the error is caused by tensorflow cpu version.

[federicoromeo]

I'm using tensorflow cpu version since i don't have a physical GPU but i rely on Google Colab's.
In a couple of days i will be provided one, so i will try with that!

Which version of tf-gpu should i use? I remember some of your code needed tensorflow 1.x, so in colab i installed manually
pip install tensorflow 1.13.2

[betterze]

I use tf 1.14.

[federicoromeo]

Even doing ! pip install tensorflow-gpu==1.14 the result remains the same... whereas by running the script GetAttribute.py ii works but i obtain only zeros. Any suggestions?

In the meanwhile can you at least share the attribute.csv file in the complete version? I mean with 40 columns, one for classifier, since the one uploaded only has 15.

And which are the header rows (images)? I followed the generation procedure from GetCode.py, they are generated from W.npy. Are there any ways i can trace them back?

[federicoromeo]

Any hints?