How can I predict? about two-branch-plant-disease

[joaopauloschuler]

@fjmsouza, thank you for sharing the solution!

[joaopauloschuler]

Hi @fjmsouza,
I think that you might have the same problem as reported here:
joaopauloschuler/two-path-noise-lab-plant-disease#1

The two-path images are loaded with:

train_x, val_x, test_x, train_y, val_y, test_y, classweight, classes = cai.datasets.load_images_from_folders(seed=7, root_dir=data_dir, lab=True, 
  verbose=Verbose, bipolar=False, base_model_name='plant_leaf',
  training_size=0.6, validation_size=0.2, test_size=0.2,
  target_size=(input_shape[0],input_shape[1]), 
  has_training=True, has_validation=True, has_testing=True, 
  smart_resize=True)

So, for your prediction to work, you'll need to use the same image transformations:

• RGB to LAB.
• Bipolar = False (numbers are positive).
• smart_resize = True.

I would expect values on each LAB channel to be similar to:

Channel  0  min: 0.0  max: 1.0
Channel  1  min: 0.20254138  max: 0.899006
Channel  2  min: 0.24579802  max: 0.94360983

You may look at the source code of cai.datasets.load_images_from_folders and check what transformations you would need.

[joaopauloschuler]

Calling the function cai.datasets.load_images_from_files will probably solve the problem:

def load_images_from_files(file_names, target_size=(224,224), dtype='float32', smart_resize=False, lab=False, rescale=False, bipolar=False):
    """Creates an array with images from an array with file names.
    # Arguments
        file_names: array with file names.
        target_size: output image size.
        dtype: output type.
        smart_resize: indicates if aspec ration should be kept adding padding.
        lab: indicates if LAB color encoding should be used.
    """

[joaopauloschuler]

Just fixed the comments on the source code: joaopauloschuler/k-neural-api@162f83a.

This is the updated version:

def add_padding_to_make_img_array_squared(img):
  """ Adds padding to make the image squared.
  # Arguments
      img: an image as an array.
  """

def load_images_from_files(file_names, target_size=(224,224), dtype='float32', smart_resize=False, lab=False, rescale=False, bipolar=False):
    """Creates an array with images from an array with file names.
    # Arguments
        file_names: array with file names.
        target_size: output image size.
        dtype: output type.
        smart_resize: indicates if aspect ratio should be kept via padding.
        lab: indicates if LAB color encoding should be used.

[fjmsouza]

Ok, I'm applying your tips, for while...
Would be something like this:

import numpy as np
from google.colab import files
import keras.utils as image

uploaded = files.upload()

for fn in uploaded.keys():
 
  path = '/content/' + fn
  img = cai.datasets.load_images_from_files(file_names=path)
  x = image.img_to_array(img)
  x = np.expand_dims(x, axis=0)

  images = np.vstack([x])
  classes = model.predict(images, batch_size=10)

?

and, for while:

searching for a solution, but your help would be welcome

[fjmsouza]

I think that we'll work around with the similar tip in #issue

Running this:

# Here's a codeblock just for fun. You should be able to upload an image here 
# and have it classified without crashing
import numpy as np
from google.colab import files
import keras.utils as image
import cv2

model = tensorflow.keras.models.load_model('/content/two-path-inception-v2.8-False-0.2-best_result.hdf5',custom_objects={'CopyChannels': cai.layers.CopyChannels})
uploaded = files.upload()

for fn in uploaded.keys():
 
  # predicting images
  path = '/content/' + fn
  imgs = cv2.imread(path)
  imgs = cv2.cvtColor(imgs, cv2.COLOR_BGR2RGB)
  imgs = cv2.cvtColor(imgs, cv2.COLOR_RGB2LAB)
  imgs=cv2.resize(imgs, (128,128))
  imm_array=np.array(imgs)
  imm_array=imm_array/255
  imm_array=imm_array.reshape(1, 128, 128, 3)

  predictions = model.predict(imm_array)
  prediction_score = tf.nn.softmax(predictions)
  predicated_class = np.argmax(prediction_score)
  print(
      "This image most likely belongs to {} with a {:.2f} percent confidence."
      .format(classes[predicated_class], 100 * np.max(prediction_score)))

Thank you!