horse or human inceptionv3 miscatagorizing

[tameyer1]

Hello

I was working on the horse or human transfer learning section. I am using the datasets for training and validation from the urls you supply. No matter what horse image I try the model classifies it as human. I figured my local setup might have issues so I opened and ran the colab file for transfer learning in the chapter 3 folder. I received the same incorrect results using a random horse picture from the web as well as when I uploaded some of the validation horse images from the colab project.

[Rutraz]

Hey

I was having a similar issue. I realized that the images of horses and humans used to train the model on the first iteration of chapter 3 were 300 x 300 and later in the chapter when we arrive at transfer learning the inputs were 150 x 150. So I altered the code given and put it back to 300 x 300 and that seems to work for me. Maybe the reason behind it is because the input needs to match the shape of the data (in this case 300× 300).

[tameyer1]

I had been trying it the opposite way of setting the training, validation, and test images to 150,150. I went ahead and tested with 300,300 in both the model training and test and receive the same result: running the validation images against the model always returns human and 1. for the class. So still no luck but thanks for the reply.

My code:
#Training
import urllib.request
import zipfile
from tensorflow.keras.preprocessing.image import ImageDataGenerator
from tensorflow.keras import layers
from tensorflow.keras import Model
from tensorflow.keras.applications.inception_v3 import InceptionV3
from tensorflow.keras.optimizers import RMSprop

weights_url = "https://storage.googleapis.com/mledu-datasets/inception_v3_weights_tf_dim_ordering_tf_kernels_notop.h5"
weights_file = "inception_v3.h5"
urllib.request.urlretrieve(weights_url, weights_file)

pre_trained_model = InceptionV3(input_shape=(300, 300, 3),
include_top=False,
weights=None)

pre_trained_model.load_weights(weights_file)

for layer in pre_trained_model.layers:
layer.trainable = False

last_layer = pre_trained_model.get_layer('mixed7')
print('last layer output shape: ', last_layer.output_shape)
last_output = last_layer.output

x = layers.Flatten()(last_output)

x = layers.Dense(1024, activation='relu')(x)

x = layers.Dropout(0.2)(x)

x = layers.Dense(1, activation='sigmoid')(x)

model = Model(pre_trained_model.input, x)

model.compile(optimizer=RMSprop(lr=0.001),
loss='binary_crossentropy',
metrics=['acc'])

training_url = "https://storage.googleapis.com/laurencemoroney-blog.appspot.com/horse-or-human.zip"
training_file_name = "horse-or-human.zip"
training_dir = 'horse-or-human/training/'
urllib.request.urlretrieve(training_url, training_file_name)
zip_ref = zipfile.ZipFile(training_file_name, 'r')
zip_ref.extractall(training_dir)
zip_ref.close()

validation_url = "https://storage.googleapis.com/laurencemoroney-blog.appspot.com/validation-horse-or-human.zip"
validation_file_name = "validation-horse-or-human.zip"
validation_dir = 'horse-or-human/validation/'
urllib.request.urlretrieve(validation_url, validation_file_name)

zip_ref = zipfile.ZipFile(validation_file_name, 'r')
zip_ref.extractall(validation_dir)
zip_ref.close()

train_datagen = ImageDataGenerator(rescale=1./255.,
rotation_range=40,
width_shift_range=0.2,
height_shift_range=0.2,
shear_range=0.2,
zoom_range=0.2,
horizontal_flip=True)

#Note that the validation data should not be augmented!
test_datagen = ImageDataGenerator(rescale=1.0/255.)

#Flow training images in batches of 20 using train_datagen generator
train_generator = train_datagen.flow_from_directory(training_dir,
batch_size=20,
class_mode='binary',
target_size=(300, 300))

#Flow validation images in batches of 20 using test_datagen generator
validation_generator = test_datagen.flow_from_directory(validation_dir,
batch_size=20,
class_mode='binary',
target_size=(300, 300))

history = model.fit_generator(
train_generator,
validation_data=validation_generator,
epochs=15,
verbose=1)

model.save('horse_or_human.h5')

#Test script

import tensorflow as tf
from tensorflow.keras.preprocessing import image
import numpy as np
import os

model = tf.keras.models.load_model('horse_or_human.h5')

#predicting images
path = 'horse-or-human/validation/horses/'
directories = os.listdir(path)

for file in directories:
    img = image.load_img(path + file, target_size=(300, 300))
    x = image.img_to_array(img)
    x = np.expand_dims(x, axis=0)

    images = np.vstack([x])
    classes = model.predict(images)
    print(classes)
    print(classes[0])
    if classes[0] > 0.5:
        print(file + " image is a human")
    else:
        print(file + " image is a horse")

[emin]

yes, I've the same problem as well. Using 300 adjustment didn't work either. I don't know the issue, considering I've started with this book to my ML journey, I'd like to know what's wrong with this.

@lmoroney

[bartpotrykus]

@tameyer1 Hey, I know your question is one year old but I encountered same issue. The reason is we were not normalizing when using a test image.
x = x / 255.0
is the line that made the classification correct.

[semajson]

I hit the exact same issue.

It looks like this might be a bug in: https://github.com/lmoroney/tfbook/blob/master/chapter3/transfer_learning.ipynb

where rather than:

for fn in uploaded.keys():
 
  # predicting images
  path = fn
  img = image.load_img(path, target_size=(150, 150))
  x = image.img_to_array(img)
  x = np.expand_dims(x, axis=0)

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

it should be:

for fn in uploaded.keys():
 
  # predicting images
  path = fn
  img = image.load_img(path, target_size=(150, 150))
  x = image.img_to_array(img)
  x = np.expand_dims(x, axis=0)
  x = x / 255.0

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

However, difference that might explain why i needed to add x = x / 255.0 on my machine

• I'm running on my WSL env, using tensorflow-cpu (version 2.11.0)
• I had to replace img = image.load_img(path, target_size=(150, 150)) with img = tf.keras.utils.load_img(test_image, target_size=(300, 300)) to get my code to not error
• i had to replace x = image.img_to_array(img) with x = tf.keras.utils.img_to_array(img) to get my code to not error

I think the next step to figure out if this is an error in https://github.com/lmoroney/tfbook/blob/master/chapter3/transfer_learning.ipynb is to run that code on colab code to see if x = x / 255.0 is needed

[semajson]

I ran the https://github.com/lmoroney/tfbook/blob/master/chapter3/transfer_learning.ipynb on colab, and needed to make these following changes:

added:

• import tensorflow as tf
• x = x / 255.0

Changes:

• img = image.load_img(path, target_size=(150, 150)) -> img = tf.keras.utils.load_img(path, target_size=(150, 150))
• x = image.img_to_array(img) -> x = tf.keras.utils.img_to_array(img)

After these changes, things ran correctly and images I uploaded were identified correctly

[semajson]

@lmoroney looks like people can't just raise PRs against this repo.

What is the correct to get this issue fixed?

[semajson]

Ah i just realised what i was doing wrong - leave this with me and i will raise a PR