Charlie Hewitt
91 linhas
3.0 KiB
Python
91 linhas
3.0 KiB
Python
from keras.preprocessing import image
|
|
from keras.models import Sequential
|
|
from keras.layers import Conv2D, MaxPooling2D, Activation, Dropout, Flatten, Dense
|
|
from keras.utils.np_utils import to_categorical
|
|
from math import floor
|
|
import csv
|
|
import numpy as np
|
|
|
|
# TODO classifier for emotions too (use to_categorical on labels)
|
|
CLASSIFY_OR_REGRESS = 1
|
|
|
|
def load_images(paths, labels, batch_size=32):
|
|
while True:
|
|
batch_n = 0
|
|
batch_d = []
|
|
batch_l = []
|
|
for i in range(batch_size):
|
|
if batch_n*batch_size + i > len(paths):
|
|
batch_n = 0
|
|
break
|
|
path = paths[batch_n*batch_size + i]
|
|
img = image.load_img(path, target_size=(256, 256))
|
|
x = image.img_to_array(img)
|
|
y = labels[batch_n*batch_size + i]
|
|
batch_d.append(x)
|
|
batch_l.append(y)
|
|
batch_d = np.array(batch_d).reshape(batch_size, 256, 256, 3)
|
|
batch_l = np.array(batch_l).reshape(batch_size, 2)
|
|
yield (batch_d, batch_l)
|
|
batch_n += 1
|
|
|
|
def load_paths(p):
|
|
labels = []
|
|
paths = []
|
|
with open(p, 'r') as csvfile:
|
|
r = csv.reader(csvfile, delimiter=',')
|
|
count = 0
|
|
for row in r:
|
|
if 'sub' in row[0]:
|
|
continue
|
|
valence = float(row[-2])
|
|
arousal = float(row[-1])
|
|
if arousal == -2 or valence == -2:
|
|
continue
|
|
# emotion = int(row[-3])
|
|
path = 'data_p/' + row[0]
|
|
paths.append(path)
|
|
labels.append([valence, arousal])
|
|
return paths, labels
|
|
|
|
model = Sequential()
|
|
|
|
# TODO define model here
|
|
model.add(Conv2D(16, (9, 9), activation='relu', input_shape=(256,256,3)))
|
|
model.add(Conv2D(16, (9, 9), activation='relu', input_shape=(256,256,3)))
|
|
model.add(MaxPooling2D(pool_size=(2, 2)))
|
|
model.add(Conv2D(32, (5, 5), activation='relu'))
|
|
model.add(Conv2D(32, (5, 5), activation='relu'))
|
|
model.add(MaxPooling2D(pool_size=(2, 2)))
|
|
model.add(Conv2D(64, (3, 3), activation='relu'))
|
|
model.add(Conv2D(64, (3, 3), activation='relu'))
|
|
model.add(MaxPooling2D(pool_size=(2, 2)))
|
|
model.add(Flatten())
|
|
model.add(Dense(1024, activation='relu'))
|
|
model.add(Dropout(0.4))
|
|
model.add(Dense(1024, activation='relu'))
|
|
model.add(Dropout(0.4))
|
|
if CLASSIFY_OR_REGRESS == 0:
|
|
model.add(Dense(7, activation='softmax'))
|
|
model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
|
|
else:
|
|
model.add(Dense(2, activation='linear'))
|
|
model.compile(loss='mean_squared_error', optimizer='adam')
|
|
|
|
t_paths, t_labels = load_paths('training.csv')
|
|
v_paths, v_labels = load_paths('validation.csv')
|
|
batch_size = 32
|
|
|
|
model.fit_generator(
|
|
load_images(t_paths,t_labels, batch_size),
|
|
steps_per_epoch=2000 // batch_size,
|
|
epochs=50,
|
|
validation_data=load_images(v_paths,v_labels, batch_size),
|
|
validation_steps=800 // batch_size)
|
|
|
|
for k in model.layers:
|
|
if type(k) is keras.layers.Dropout:
|
|
model.layers.remove(k)
|
|
|
|
model.save_weights('aff_net.h5')
|