def main():
image_dir = '../../vqa_images/'
local_images = [
f for f in listdir(image_dir) if isfile(join(image_dir, f))
]
parser = argparse.ArgumentParser()
parser.add_argument(
'-model',
type=str,
default='../models/mlp_num_hidden_units_1024_num_hidden_layers_3.json')
parser.add_argument(
'-weights',
type=str,
default=
'../models/mlp_num_hidden_units_1024_num_hidden_layers_3_epoch_70.hdf5'
)
parser.add_argument('-sample_size', type=int, default=1)
args = parser.parse_args()
model = model_from_json(open(args.model).read())
model.load_weights(args.weights)
model.compile(loss='categorical_crossentropy', optimizer='rmsprop')
print 'Model loaded and compiled'
images_val = open('../data/preprocessed/images_val2014_all.txt',
'r').read().decode('utf8').splitlines()
nlp = English()
print 'Loaded word2vec features'
labelencoder = joblib.load('../models/labelencoder.pkl')
vgg_model_path = '../features/coco/vgg_feats.mat'
features_struct = scipy.io.loadmat(vgg_model_path)
VGGfeatures = features_struct['feats']
print 'Loaded vgg features'
image_ids = open('../features/coco_vgg_IDMap.txt').read().splitlines()
img_map = {}
for ids in image_ids:
id_split = ids.split()
img_map[id_split[0]] = int(id_split[1])
image_sample = random.sample(local_images, args.sample_size)
for image in image_sample:
img = Image.open(image_dir + image)
img.show()
q = unicode(raw_input("Ask a question about the image:"))
coco_id = str(int(image[-16:-4]))
timesteps = len(nlp(q))
X_q = get_questions_matrix_sum([q], nlp)
X_i = get_images_matrix([coco_id], img_map, VGGfeatures)
X_batch = np.hstack((X_q, X_i))
y_predict = model.predict_classes(X_batch, verbose=0)
print labelencoder.inverse_transform(y_predict)
#img.close()
raw_input('Press enter to continue...')
def main():
parser = argparse.ArgumentParser()
parser.add_argument('-model', type=str, required=True)
parser.add_argument('-weights', type=str, required=True)
parser.add_argument('-results', type=str, required=True)
args = parser.parse_args()
model = model_from_json(open(args.model).read())
model.load_weights(args.weights)
model.compile(loss='categorical_crossentropy', optimizer='rmsprop')
questions_val = open('../data/preprocessed/questions_val2014.txt',
'r').read().decode('utf8').splitlines()
answers_val = open('../data/preprocessed/answers_val2014.txt',
'r').read().decode('utf8').splitlines()
images_val = open('../data/preprocessed/images_val2014.txt',
'r').read().decode('utf8').splitlines()
vgg_model_path = '../features/coco/vgg_feats.mat'
print 'Model compiled, weights loaded...'
labelencoder = joblib.load('../models/labelencoder.pkl')
features_struct = scipy.io.loadmat(vgg_model_path)
VGGfeatures = features_struct['feats']
print 'loaded vgg features'
image_ids = open('../features/coco_vgg_IDMap.txt').read().splitlines()
img_map = {}
for ids in image_ids:
id_split = ids.split()
img_map[id_split[0]] = int(id_split[1])
nlp = English()
print 'loaded word2vec features'
nb_classes = 1000
y_predict_text = []
batchSize = 128
widgets = ['Evaluating ', Percentage(), ' ', Bar(marker='#',left='[',right=']'),
' ', ETA()]
pbar = ProgressBar(widgets=widgets)
for qu_batch,an_batch,im_batch in pbar(zip(grouper(questions_val, batchSize, fillvalue=questions_val[0]),
grouper(answers_val, batchSize, fillvalue=answers_val[0]),
grouper(images_val, batchSize, fillvalue=images_val[0]))):
X_q_batch = get_questions_matrix_sum(qu_batch, nlp)
if 'language_only' in args.model:
X_batch = X_q_batch
else:
X_i_batch = get_images_matrix(im_batch, img_map , VGGfeatures)
X_batch = np.hstack((X_q_batch, X_i_batch))
y_predict = model.predict_classes(X_batch, verbose=0)
y_predict_text.extend(labelencoder.inverse_transform(y_predict))
correct_val=0
incorrect_val=0
f1 = open(args.results, 'w')
for prediction, truth, question, image in zip(y_predict_text, answers_val, questions_val, images_val):
temp_count=0
for _truth in truth.split(';'):
if prediction == _truth:
temp_count+=1
if temp_count>2:
correct_val+=1
else:
incorrect_val+=1
f1.write(question.encode('utf-8'))
f1.write('\n')
f1.write(image.encode('utf-8'))
f1.write('\n')
f1.write(prediction)
f1.write('\n')
f1.write(truth.encode('utf-8'))
f1.write('\n')
f1.write('\n')
f1.write('Final Accuracy is ' + str(float(correct_val)/(incorrect_val+correct_val)))
f1.close()
f1 = open('../results/overall_results.txt', 'a')
f1.write(args.weights + '\n')
f1.write(str(float(correct_val)/(incorrect_val+correct_val)) + '\n')
f1.close()
print 'Final Accuracy on the validation set is', float(correct_val)/(incorrect_val+correct_val)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('-model', type=str, required=True)
parser.add_argument('-weights', type=str, required=True)
parser.add_argument('-results', type=str, required=True)
args = parser.parse_args()
model = model_from_json(open(args.model).read())
model.load_weights(args.weights)
model.compile(loss='categorical_crossentropy', optimizer='rmsprop')
questions_val = open('../data/preprocessed/questions_val2014.txt',
'r').read().decode('utf8').splitlines()
answers_val = open('../data/preprocessed/answers_val2014.txt',
'r').read().decode('utf8').splitlines()
images_val = open('../data/preprocessed/images_val2014.txt',
'r').read().decode('utf8').splitlines()
vgg_model_path = '../features/coco/vgg_feats.mat'
print 'Model compiled, weights loaded...'
labelencoder = joblib.load('../models/labelencoder.pkl')
features_struct = scipy.io.loadmat(vgg_model_path)
VGGfeatures = features_struct['feats']
print 'loaded vgg features'
image_ids = open('../features/coco/coco_vgg_IDMap.txt').read().splitlines()
img_map = {}
for ids in image_ids:
id_split = ids.split()
img_map[id_split[0]] = int(id_split[1])
nlp = English()
print 'loaded word2vec features'
nb_classes = 1000
y_predict_text = []
batchSize = 128
widgets = [
'Evaluating ',
Percentage(), ' ',
Bar(marker='#', left='[', right=']'), ' ',
ETA()
]
pbar = ProgressBar(widgets=widgets)
for qu_batch, an_batch, im_batch in pbar(
zip(grouper(questions_val, batchSize, fillvalue=questions_val[0]),
grouper(answers_val, batchSize, fillvalue=answers_val[0]),
grouper(images_val, batchSize, fillvalue=images_val[0]))):
X_q_batch = get_questions_matrix_sum(qu_batch, nlp)
if 'language_only' in args.model:
X_batch = X_q_batch
else:
X_i_batch = get_images_matrix(im_batch, img_map, VGGfeatures)
X_batch = np.hstack((X_q_batch, X_i_batch))
y_predict = model.predict_classes(X_batch, verbose=0)
y_predict_text.extend(labelencoder.inverse_transform(y_predict))
correct_val = 0
incorrect_val = 0
f1 = open(args.results, 'w')
for prediction, truth, question, image in zip(y_predict_text, answers_val,
questions_val, images_val):
temp_count = 0
for _truth in truth.split(';'):
if prediction == _truth:
temp_count += 1
if temp_count > 2:
correct_val += 1
else:
incorrect_val += 1
f1.write(question.encode('utf-8'))
f1.write('\n')
f1.write(image.encode('utf-8'))
f1.write('\n')
f1.write(prediction)
f1.write('\n')
f1.write(truth.encode('utf-8'))
f1.write('\n')
f1.write('\n')
f1.write('Final Accuracy is ' +
str(float(correct_val) / (incorrect_val + correct_val)))
f1.close()
f1 = open('../results/overall_results.txt', 'a')
f1.write(args.weights + '\n')
f1.write(str(float(correct_val) / (incorrect_val + correct_val)) + '\n')
f1.close()
print 'Final Accuracy on the validation set is', float(correct_val) / (
incorrect_val + correct_val)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('-num_hidden_units', type=int, default=1024)
parser.add_argument('-num_hidden_layers', type=int, default=3)
parser.add_argument('-dropout', type=float, default=0.5)
parser.add_argument('-activation', type=str, default='tanh')
parser.add_argument('-language_only', type=bool, default=False)
parser.add_argument('-num_epochs', type=int, default=10)
parser.add_argument('-model_save_interval', type=int, default=10)
parser.add_argument('-batch_size', type=int, default=128)
args = parser.parse_args()
questions_train = open('../data/preprocessed/questions_train2014.txt',
'r').read().decode('utf8').splitlines()
answers_train = open('../data/preprocessed/answers_train2014_modal.txt',
'r').read().decode('utf8').splitlines()
images_train = open('../data/preprocessed/images_train2014.txt',
'r').read().decode('utf8').splitlines()
vgg_model_path = '../features/coco/vgg_feats.mat'
maxAnswers = 1000
questions_train, answers_train, images_train = selectFrequentAnswers(
questions_train, answers_train, images_train, maxAnswers)
#encode the remaining answers
labelencoder = preprocessing.LabelEncoder()
labelencoder.fit(answers_train)
nb_classes = len(list(labelencoder.classes_))
joblib.dump(labelencoder, '../models/labelencoder.pkl')
features_struct = scipy.io.loadmat(vgg_model_path)
VGGfeatures = features_struct['feats']
print 'loaded vgg features'
image_ids = open('../features/coco_vgg_IDMap.txt').read().splitlines()
id_map = {}
for ids in image_ids:
id_split = ids.split()
id_map[id_split[0]] = int(id_split[1])
nlp = English()
print 'loaded word2vec features...'
img_dim = 4096
word_vec_dim = 300
model = Sequential()
if args.language_only:
model.add(
Dense(args.num_hidden_units,
input_dim=word_vec_dim,
init='uniform'))
else:
model.add(
Dense(args.num_hidden_units,
input_dim=img_dim + word_vec_dim,
init='uniform'))
model.add(Activation(args.activation))
if args.dropout > 0:
model.add(Dropout(args.dropout))
for i in xrange(args.num_hidden_layers - 1):
model.add(Dense(args.num_hidden_units, init='uniform'))
model.add(Activation(args.activation))
if args.dropout > 0:
model.add(Dropout(args.dropout))
model.add(Dense(nb_classes, init='uniform'))
model.add(Activation('softmax'))
json_string = model.to_json()
if args.language_only:
model_file_name = '../models/mlp_language_only_num_hidden_units_' + str(
args.num_hidden_units) + '_num_hidden_layers_' + str(
args.num_hidden_layers)
else:
model_file_name = '../models/mlp_num_hidden_units_' + str(
args.num_hidden_units) + '_num_hidden_layers_' + str(
args.num_hidden_layers)
open(model_file_name + '.json', 'w').write(json_string)
print 'Compiling model...'
model.compile(loss='categorical_crossentropy', optimizer='rmsprop')
print 'Compilation done...'
print 'Training started...'
for k in xrange(args.num_epochs):
#shuffle the data points before going through them
index_shuf = range(len(questions_train))
shuffle(index_shuf)
questions_train = [questions_train[i] for i in index_shuf]
answers_train = [answers_train[i] for i in index_shuf]
images_train = [images_train[i] for i in index_shuf]
progbar = generic_utils.Progbar(len(questions_train))
for qu_batch, an_batch, im_batch in zip(
grouper(questions_train,
args.batch_size,
fillvalue=questions_train[-1]),
grouper(answers_train,
args.batch_size,
fillvalue=answers_train[-1]),
grouper(images_train,
args.batch_size,
fillvalue=images_train[-1])):
X_q_batch = get_questions_matrix_sum(qu_batch, nlp)
if args.language_only:
X_batch = X_q_batch
else:
X_i_batch = get_images_matrix(im_batch, id_map, VGGfeatures)
X_batch = np.hstack((X_q_batch, X_i_batch))
Y_batch = get_answers_matrix(an_batch, labelencoder)
loss = model.train_on_batch(X_batch, Y_batch)
progbar.add(args.batch_size, values=[("train loss", loss)])
#print type(loss)
if k % args.model_save_interval == 0:
model.save_weights(model_file_name +
'_epoch_{:02d}.hdf5'.format(k))
model.save_weights(model_file_name + '_epoch_{:02d}.hdf5'.format(k))
def main():
parser = argparse.ArgumentParser()
parser.add_argument('-num_hidden_units', type=int, default=1024)
parser.add_argument('-num_hidden_layers', type=int, default=3)
parser.add_argument('-dropout', type=float, default=0.5)
parser.add_argument('-activation', type=str, default='tanh')
parser.add_argument('-language_only', type=bool, default= False)
parser.add_argument('-num_epochs', type=int, default=100)
parser.add_argument('-model_save_interval', type=int, default=10)
parser.add_argument('-batch_size', type=int, default=128)
parser.add_argument('-word_vector', type=str, default='')
args = parser.parse_args()
questions_train = open('../data/preprocessed/questions_train2014.txt', 'r').read().decode('utf8').splitlines()
answers_train = open('../data/preprocessed/answers_train2014_modal.txt', 'r').read().decode('utf8').splitlines()
images_train = open('../data/preprocessed/images_train2014.txt', 'r').read().decode('utf8').splitlines()
vgg_model_path = '../features/coco/vgg_feats.mat'
maxAnswers = 1000
questions_train, answers_train, images_train = selectFrequentAnswers(questions_train,answers_train,images_train, maxAnswers)
#encode the remaining answers
labelencoder = preprocessing.LabelEncoder()
labelencoder.fit(answers_train)
nb_classes = len(list(labelencoder.classes_))
joblib.dump(labelencoder,'../models/labelencoder.pkl')
features_struct = scipy.io.loadmat(vgg_model_path)
VGGfeatures = features_struct['feats']
print 'loaded vgg features'
image_ids = open('../features/coco_vgg_IDMap.txt').read().splitlines()
id_map = {}
for ids in image_ids:
id_split = ids.split()
id_map[id_split[0]] = int(id_split[1])
# Code to choose the word vectors, default is Goldberg but GLOVE is preferred
if args.word_vector == 'glove':
nlp = spacy.load('en', vectors='en_glove_cc_300_1m_vectors')
else:
nlp = English()
print 'loaded ' + args.word_vector + ' word2vec features...'
img_dim = 4096
word_vec_dim = 300
model = Sequential()
if args.language_only:
model.add(Dense(args.num_hidden_units, input_dim=word_vec_dim, init='uniform'))
else:
model.add(Dense(args.num_hidden_units, input_dim=img_dim+word_vec_dim, init='uniform'))
model.add(Activation(args.activation))
if args.dropout>0:
model.add(Dropout(args.dropout))
for i in xrange(args.num_hidden_layers-1):
model.add(Dense(args.num_hidden_units, init='uniform'))
model.add(Activation(args.activation))
if args.dropout>0:
model.add(Dropout(args.dropout))
model.add(Dense(nb_classes, init='uniform'))
model.add(Activation('softmax'))
json_string = model.to_json()
if args.language_only:
model_file_name = '../models/mlp_language_only_num_hidden_units_' + str(args.num_hidden_units) + '_num_hidden_layers_' + str(args.num_hidden_layers)
else:
model_file_name = '../models/mlp_num_hidden_units_' + str(args.num_hidden_units) + '_num_hidden_layers_' + str(args.num_hidden_layers)
open(model_file_name + '.json', 'w').write(json_string)
print 'Compiling model...'
model.compile(loss='categorical_crossentropy', optimizer='rmsprop')
print 'Compilation done...'
print 'Training started...'
for k in xrange(args.num_epochs):
#shuffle the data points before going through them
index_shuf = range(len(questions_train))
shuffle(index_shuf)
questions_train = [questions_train[i] for i in index_shuf]
answers_train = [answers_train[i] for i in index_shuf]
images_train = [images_train[i] for i in index_shuf]
progbar = generic_utils.Progbar(len(questions_train))
for qu_batch,an_batch,im_batch in zip(grouper(questions_train, args.batch_size, fillvalue=questions_train[-1]),
grouper(answers_train, args.batch_size, fillvalue=answers_train[-1]),
grouper(images_train, args.batch_size, fillvalue=images_train[-1])):
X_q_batch = get_questions_matrix_sum(qu_batch, nlp)
if args.language_only:
X_batch = X_q_batch
else:
X_i_batch = get_images_matrix(im_batch, id_map, VGGfeatures)
X_batch = np.hstack((X_q_batch, X_i_batch))
Y_batch = get_answers_matrix(an_batch, labelencoder)
loss = model.train_on_batch(X_batch, Y_batch)
# fix for the Keras v0.3 issue #9
progbar.add(args.batch_size, values=[("train loss", loss[0])])
#print type(loss)
if k%args.model_save_interval == 0:
model.save_weights(model_file_name + '_epoch_{:02d}.hdf5'.format(k))
model.save_weights(model_file_name + '_epoch_{:02d}.hdf5'.format(k))
def main():
cwd = os.getcwd()
parser = argparse.ArgumentParser()
parser.add_argument('-num_hidden_units', type=int, default=1024)
parser.add_argument('-num_hidden_layers', type=int, default=3)
parser.add_argument('-dropout', type=float, default=0.5)
parser.add_argument('-activation', type=str, default='tanh')
parser.add_argument('-language_only', type=bool, default= False)
parser.add_argument('-num_epochs', type=int, default=2)
parser.add_argument('-model_save_interval', type=int, default=10)
parser.add_argument('-model_weights_path', type=str, default=cwd+'/vgg/vgg16_weights.h5')
parser.add_argument('-batch_size', type=int, default=128)
parser.add_argument('-questions_train',type=str, default = cwd+'/data/preprocessed/questions_train2015.txt')
parser.add_argument('-answers_train',type=str, default = cwd+'/data/preprocessed/answers_train2015_modal.txt')
parser.add_argument('-im_dir',type=str, default =cwd+'/data/preprocessed/scene_img_abstract_v002_train2015/')
#parser.add_argument('-questions_train',type=str, default = cwd+'/data/preprocessed/questions_train2014.txt')
args = parser.parse_args()
questions_train = open(args.questions_train, 'r').read().decode('utf8').splitlines()
answers_train = open(args.answers_train, 'r').read().decode('utf8').splitlines()
images_train = open(cwd+'/data/preprocessed/images_train2015.txt', 'r').read().decode('utf8').splitlines()
#vgg_model_path = cwd+'/features/coco/vgg_feats.mat' #this needs to change
maxAnswers = 100
questions_train, answers_train, images_train = selectFrequentAnswers(questions_train,answers_train,images_train, maxAnswers)
#encode the remaining answers
labelencoder = preprocessing.LabelEncoder()
labelencoder.fit(answers_train)
nb_classes = len(list(labelencoder.classes_))
joblib.dump(labelencoder,cwd+'/models/labelencoder.pkl')
#features_struct = scipy.io.loadmat(vgg_model_path)
#VGGfeatures = features_struct['feats']
# print 'loaded vgg features'
# image_ids = open(cwd+'/features/coco_vgg_IDMap.txt').read().splitlines()
# id_map = {}
# for ids in image_ids:
# id_split = ids.split()
# id_map[id_split[0]] = int(id_split[1])
vgg_model = vgg16.VGG_16(args.model_weights_path)
sgd = SGD(lr=0.1, decay=1e-6, momentum=0.9, nesterov=True)
vgg_model.compile(optimizer=sgd, loss='categorical_crossentropy')
print 'loaded vgg model...'
nlp = English()
print 'loaded word2vec features...'
img_dim = 4096
word_vec_dim = 300
model = Sequential()
if args.language_only:
model.add(Dense(args.num_hidden_units, input_dim=word_vec_dim, init='uniform'))
else:
model.add(Dense(args.num_hidden_units, input_dim=img_dim+word_vec_dim, init='uniform'))
model.add(Activation(args.activation))
if args.dropout>0:
model.add(Dropout(args.dropout))
for i in xrange(args.num_hidden_layers-1):
model.add(Dense(args.num_hidden_units, init='uniform'))
model.add(Activation(args.activation))
if args.dropout>0:
model.add(Dropout(args.dropout))
model.add(Dense(nb_classes, init='uniform'))
model.add(Activation('softmax'))
json_string = model.to_json()
model_file_name = cwd+'/models/mlp_num_hidden_units_' + str(args.num_hidden_units) + '_num_hidden_layers_' + str(args.num_hidden_layers)
open(model_file_name + '.json', 'w').write(json_string)
print 'Training started...'
id_map = {}
f1 = open('abstract_image_precompute')
f2 = open('abstract_image_precompute_reverse')
VGGfeatures = np.loadtxt(f1)
VGGfeatures_reverse = np.loadtxt(f2)
f1.close()
f2.close()
for k in xrange(args.num_epochs):
#shuffle the data points before going through them
index_shuf = range(len(questions_train))
shuffle(index_shuf)
questions_train = [questions_train[i] for i in index_shuf]
answers_train = [answers_train[i] for i in index_shuf]
images_train = [images_train[i] for i in index_shuf]
progbar = generic_utils.Progbar(len(questions_train))
for qu_batch,an_batch,im_batch in zip(grouper(questions_train, args.batch_size, fillvalue=questions_train[-1]),
grouper(answers_train, args.batch_size, fillvalue=answers_train[-1]),
grouper(images_train, args.batch_size, fillvalue=images_train[-1])):
X_q_batch = get_questions_matrix_sum(qu_batch, nlp)
im_path = args.im_dir +"abstract_v002_train2015_"
print 'getting image features...'
X_i_batch = get_images_matrix(im_batch, VGGfeatures, VGGfeatures_reverse)
# X_i_batch = get_images_matrix_from_model(vgg_model, im_batch, im_path, id_map)
X_batch = np.hstack((X_q_batch, X_i_batch))
Y_batch = get_answers_matrix(an_batch, labelencoder)
print 'running training on batch...'
loss = model.train_on_batch(X_batch, Y_batch)
progbar.add(args.batch_size, values=[("train loss", loss)])
if k%args.model_save_interval == 0:
model.save_weights(model_file_name + '_epoch_{:02d}.hdf5'.format(k))
model.save_weights(model_file_name + '_epoch_{:02d}.hdf5'.format(k))
def main():
parser = argparse.ArgumentParser()
parser.add_argument('-model', type=str, required=True)
parser.add_argument('-results', type=str, required=True)
args = parser.parse_args()
model = load_model(args.model)
model.compile(loss='categorical_crossentropy', optimizer='rmsprop')
questions_val = open('./data/questions_val2014.txt',
'r').read().decode('utf8').splitlines()
answers_val = open('./data/answers_val2014_all.txt',
'r').read().decode('utf8').splitlines()
images_val = open('./data/images_val2014_all.txt',
'r').read().decode('utf8').splitlines()
print 'Model compiled, weights loaded...'
labelencoder = joblib.load('./labelencoder.pkl')
image_ids = open("./id_map.txt").read().splitlines()
id_map = {}
for ids in image_ids:
id_split = ids.split()
id_map[int(id_split[0])] = int(id_split[1]) - 1
sherlock_features = np.load('./sherlock_features_temp.npy')
nlp = English()
print 'loaded word2vec features'
nb_classes = 1000
y_predict_text = []
batchSize = 128
widgets = ['Evaluating ', Percentage(), ' ', Bar(marker='#', left='[', right=']'),
' ', ETA()]
pbar = ProgressBar(widgets=widgets)
for qu_batch, an_batch, im_batch in pbar(zip(grouper(questions_val, batchSize, fillvalue=questions_val[0]),
grouper(answers_val, batchSize, fillvalue=answers_val[0]),
grouper(images_val, batchSize, fillvalue=images_val[0]))):
X_q_batch = get_questions_matrix_sum(qu_batch, nlp)
X_i_batch = get_images_matrix(im_batch, id_map, sherlock_features)
X_batch = np.hstack((X_q_batch, X_i_batch))
y_predict = model.predict_classes(X_batch, verbose=0)
y_predict_text.extend(labelencoder.inverse_transform(y_predict))
correct_val = 0.0
total = 0
f1 = open(args.results, 'w')
for prediction, truth, question, image in zip(y_predict_text, answers_val, questions_val, images_val):
temp_count = 0
for _truth in truth.split(';'):
if prediction == _truth:
temp_count += 1
if temp_count > 2:
correct_val += 1
else:
correct_val += float(temp_count) / 3
total += 1
f1.write(question.encode('utf-8'))
f1.write('\n')
f1.write(image.encode('utf-8'))
f1.write('\n')
f1.write(prediction)
f1.write('\n')
f1.write(truth.encode('utf-8'))
f1.write('\n')
f1.write('\n')
f1.write('Final Accuracy is ' + str(correct_val / total))
f1.close()
f1 = open('../results/overall_results.txt', 'a')
f1.write(args.weights + '\n')
f1.write(str(correct_val / total) + '\n')
f1.close()
print 'Final Accuracy on the validation set is', correct_val / total
def main():
parser = argparse.ArgumentParser()
parser.add_argument("-model", type=str, required=True)
parser.add_argument("-weights", type=str, required=True)
parser.add_argument("-results", type=str, required=True)
args = parser.parse_args()
model = model_from_json(open(args.model).read())
model.load_weights(args.weights)
model.compile(loss="categorical_crossentropy", optimizer="rmsprop")
questions_val = open("../data/preprocessed/questions_val2014.txt", "r").read().decode("utf8").splitlines()
answers_val = open("../data/preprocessed/answers_val2014.txt", "r").read().decode("utf8").splitlines()
images_val = open("../data/preprocessed/images_val2014.txt", "r").read().decode("utf8").splitlines()
vgg_model_path = "../features/coco/vgg_feats.mat"
print "Model compiled, weights loaded..."
labelencoder = joblib.load("../models/labelencoder.pkl")
features_struct = scipy.io.loadmat(vgg_model_path)
VGGfeatures = features_struct["feats"]
print "loaded vgg features"
image_ids = open("../features/coco/coco_vgg_IDMap.txt").read().splitlines()
img_map = {}
for ids in image_ids:
id_split = ids.split()
img_map[id_split[0]] = int(id_split[1])
nlp = English()
print "loaded word2vec features"
nb_classes = 1000
y_predict_text = []
batchSize = 128
widgets = ["Evaluating ", Percentage(), " ", Bar(marker="#", left="[", right="]"), " ", ETA()]
pbar = ProgressBar(widgets=widgets)
for qu_batch, an_batch, im_batch in pbar(
zip(
grouper(questions_val, batchSize, fillvalue=questions_val[0]),
grouper(answers_val, batchSize, fillvalue=answers_val[0]),
grouper(images_val, batchSize, fillvalue=images_val[0]),
)
):
X_q_batch = get_questions_matrix_sum(qu_batch, nlp)
if "language_only" in args.model:
X_batch = X_q_batch
else:
X_i_batch = get_images_matrix(im_batch, img_map, VGGfeatures)
X_batch = np.hstack((X_q_batch, X_i_batch))
y_predict = model.predict_classes(X_batch, verbose=0)
y_predict_text.extend(labelencoder.inverse_transform(y_predict))
correct_val = 0
incorrect_val = 0
f1 = open(args.results, "w")
for prediction, truth, question, image in zip(y_predict_text, answers_val, questions_val, images_val):
temp_count = 0
for _truth in truth.split(";"):
if prediction == _truth:
temp_count += 1
if temp_count > 2:
correct_val += 1
else:
incorrect_val += 1
f1.write(question.encode("utf-8"))
f1.write("\n")
f1.write(image.encode("utf-8"))
f1.write("\n")
f1.write(prediction)
f1.write("\n")
f1.write(truth.encode("utf-8"))
f1.write("\n")
f1.write("\n")
f1.write("Final Accuracy is " + str(float(correct_val) / (incorrect_val + correct_val)))
f1.close()
f1 = open("../results/overall_results.txt", "a")
f1.write(args.weights + "\n")
f1.write(str(float(correct_val) / (incorrect_val + correct_val)) + "\n")
f1.close()
print "Final Accuracy on the validation set is", float(correct_val) / (incorrect_val + correct_val)
training_loss = []
for i in xrange(args.num_epochs):
progbar = generic_utils.Progbar(len(questions_train))
for qu_batch, an_batch, im_batch in zip(
grouper(questions_train,
args.batch_size,
fillvalue=questions_train[-1]),
grouper(answers_train,
args.batch_size,
fillvalue=answers_train[-1]),
grouper(images_train, args.batch_size,
fillvalue=images_train[-1])):
# logging.debug("One batch done")
x_q_batch = get_questions_matrix_sum(qu_batch, nlp)
# logging.debug("length of qu_batch is %d", len(qu_batch))
# logging.debug("Shape of x_q_batch is: %s", x_q_batch.shape)
x_i_batch = get_images_matrix(im_batch, id_map, sherlock_features)
# logging.debug("shape of x_i_batch is %s", x_i_batch.shape)
x_batch = np.hstack((x_q_batch, x_i_batch))
y_batch = get_answers_matrix(an_batch, labelencoder)
loss = model.train_on_batch(x_batch, y_batch)
training_loss.append(loss)
progbar.add(args.batch_size, values=[("train_loss", loss)])
# print "\n" # if __name__ == '__main__':
# main()
plt.plot(training_loss)
plt.title("Training loss for the network")
plt.savefig('loss.png')
def process_input(self, question):
return get_questions_matrix_sum(question, self._nlp)
