Input 0 is incompatible with layer model about keras-unet

[karolzak]

image = np.array(Image.open("test.jpg").resize((384, 384)))
print("shape: ", image.shape)
shape: (384, 384, 3)
pr_mask = model.predict(image).round()

Ok, I see your problem. When providing inputs for prediction you need to serve images in batches - if you want to fit just a single image you need to reshape it from (384, 384, 3) to (1, 384, 384, 3)

[parcodepie]

Indeed it was the problem.
Here is the new working code

image = np.array(Image.open("test.jpg").resize((384, 384)))
images_list = []
images_list.append(np.array(image))
x = np.asarray(images_list)
pr_mask = model.predict(x).round()

plt.imshow(
    pr_mask[0]
)
plt.show()

also I'm saving the image to the disk

data = Image.fromarray(np.reshape(pr_mask[0], (384, 384)) )
data = data.convert("L")
data.save('xxx.png')

the model needs more training but it's a good start

Thanks @karolzak for your time

[karolzak]

Hi @parcodepie
Seems to me you made a mistake when resizing the test image. You used 348x348 instead of 384x384

[parcodepie]

Thanks @karolzak I edited it but it throws another error
Input 0 is incompatible with layer model: expected shape=(None, 384, 384, 3), found shape=(32, 384, 3)

[karolzak]

Your input array still have a wrong shape. It expects a batch of arrays of size 384x384x3 while you're providing 32 arrays of size 384x3. Something is wrong with your code

[parcodepie]

I'm trying to test only one image what could went wrong?

here is the whole test code

import cv2
import numpy as np
from keras_net.keras_unet.utils import get_patches
from keras_unet.models import satellite_unet
from PIL import Image

model_file = 'model_satellite.h5'
input_shape = (384, 384, 3)

model = satellite_unet(
    input_shape
)

model.load_weights(model_file)
image = np.array(Image.open("test.jpg").resize((384, 384)))
print("shape: ", image.shape)
shape:  (384, 384, 3)
pr_mask = model.predict(image).round()

cv2.imshow(
    "mask: ", pr_mask
)

[gordonhu608]

Thanks @karolzak I edited it but it throws another error
Input 0 is incompatible with layer model: expected shape=(None, 384, 384, 3), found shape=(32, 384, 3)

Hi, so how did u fix the issue when you are testing many images instead of just one?

[hhhhhhhhhhhhhhhhho]

hi, i have a same problem ,too

how fix that problem ?

Input 0 is incompatible with layer model: expected shape=(None, 64, 64, 3), found shape=(32, 64, 3) .. . .

[karolzak]

Hello @hhhhhhhhhhhhhhhhho ! Great username :)

As it was mentioned begore you need to serve images as an array. The solution mentioned above will work:

image = np.array(Image.open("test.jpg").resize((384, 384)))
images_list = []
images_list.append(np.array(image))
x = np.asarray(images_list)
pr_mask = model.predict(x).round()

plt.imshow(
pr_mask[0]
)
plt.show()

[Anne-Andresen]

Had the same problem solved it by adding an additional dimension
code before:

for i in range(8):
    img = io.imread(TEST_PATH + str(i) + '.png')
    print('shape: ', img.shape)
    pred = model.predict(img)
    prediction = np.squeeze(pred) > .5
    plt.imsave(os.path.join(save_path, str(i) + '_predicted.png'), prediction)

Code after:

for i in range(8):
    img = io.imread(TEST_PATH + str(i) + '.png')
    print('shape: ', img.shape)
    img = img[np.newaxis, :,  :, :]
    pred = model.predict(img)
    prediction = np.squeeze(pred) > .5
    plt.imsave(os.path.join(save_path, str(i) + '_predicted.png'), prediction)

[tehreemnaqvi]

@karolzak I have the same error:

ValueError: Input 0 is incompatible with layer model: expected shape=(None, 9, 4), found shape=(9, 4).

This is my original data shape and I'm trying to generate synthetic data from TimeGAN. When I run the model, got the above error.

[karolzak]

@tehreemnaqvi Network is expecting batches of inputs of size (9,4). Please read through this thread. Your problem was already answered above.