First off: I'm aware of this post, but it doesn't provide an answer.
I am building my model like this:
from keras.models import Sequential
from keras.layers import Conv2D, MaxPooling2D, Dense, Flatten, Dropout
from keras.preprocessing.image import ImageDataGenerator # for data augmentation
import pandas as pd # to save .csv files
from time import perf_counter # to track runtime
from keras.metrics import TrueNegatives, TruePositives, FalseNegatives, FalsePositives
def build_model(dimension):
model = Sequential()
model.add(Conv2D(32, (11,11), activation='relu',
input_shape=(dimension, dimension, 3)))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Conv2D(64, (5, 5), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Conv2D(128, (3, 3), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Conv2D(128, (3, 3), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Flatten()) # to prepare for dropout
model.add(Dropout(0.2)) # to prevent overfitting
model.add(Dense(256, activation='relu'))
model.add(Dense(1, activation='sigmoid'))
model.compile(optimizer='rmsprop',
loss='binary_crossentropy',
metrics=['accuracy',
TruePositives(),
TrueNegatives(),
FalsePositives(),
FalseNegatives()
]
)
return model
def train_model(epoch, batch_size, run, subrun):
dimension = 200
model = build_model(dimension)
train_datagen = ImageDataGenerator(validation_split=0.2,
# samplewise_std_normalization=True,
rotation_range=40,
rescale=1./255,
shear_range=0.2,
zoom_range=0.2,
horizontal_flip=True,
width_shift_range=0.2,
height_shift_range=0.2,
fill_mode='nearest'
)
training_set = train_datagen.flow_from_directory('train6',
target_size=(dimension, dimension),
color_mode='rgb', # default
class_mode='binary',
batch_size=batch_size,
save_to_dir=None,
interpolation='nearest',
subset='training')
validation_set = train_datagen.flow_from_directory('train6',
target_size=(dimension, dimension),
color_mode='rgb', # default
class_mode='binary',
batch_size=batch_size,
save_to_dir=None,
# if 'str', saves augmented images for visualisation
interpolation='nearest',
subset='validation')
start_time = perf_counter() # start counting
history = model.fit_generator(training_set,
epochs=epoch,
steps_per_epoch=training_set.samples // batch_size,
validation_data=validation_set,
validation_steps=validation_set.samples // batch_size,
verbose=2)
stop_time = perf_counter() # stop counting
# saving trained model & history file
model.save_weights('models/cat_dog_classifier_{0}_{1}.h5'.format(run, subrun)) # save model weights
hist_pd = pd.DataFrame(history.history) # making panda file of history.history
hist_csv_file = 'histories/history_{0}_{1}.csv'.format(run, subrun) # defining name for csv file
with open(hist_csv_file, mode='w') as f: # saving the pd file as csv
hist_pd.to_csv(f)
return stop_time - start_time
And I use the following code to get the probabilities:
from build_model import build_model
from keras.preprocessing import image
import numpy as np
run = 'A28'
subrun = 1
dimension = 200
# build model
model = build_model(dimension)
model.load_weights('models/cat_dog_classifier_{0}_{1}.h5'.format(run, subrun))
# Get test image ready
amount_of_images = 10
predictions = np.zeros((amount_of_images, 2))
labels = np.zeros(amount_of_images)
for i in range(amount_of_images):
image_name = 1 + i # choose what image to start from
test_image = image.load_img('test1/{}.jpg'.format(image_name), target_size=(dimension, dimension))
test_image = image.img_to_array(test_image)
test_image = np.expand_dims(test_image, axis=0)
label = model.predict_classes(test_image, batch_size=1)
labels[i] = label
prediction = model.predict(test_image, batch_size=1)
print(prediction)
print(labels)
When I train my model without using the rescaling or the normalisation, the predictions are probabilities as expected. However, when I use either, it returns only 0s and 1s (the same labels as predict_classes). I have tried to run the dummy code provided in the link above, and it worked as expected; which makes sense I suppose, as the script also runs fine when I haven't used rescaling. However, I would really like to be able to use rescaling. Does anyone have any idea what is going wrong?
