def build_theano_fn_resnet(self):
t0 = time.time()
print '%s build theano fn resnet' % self.rank
x = T.ftensor4('images')
y = T.ivector('label')
model = resnet50.build_model(x)
prob = lasagne.layers.get_output(model['prob'], deterministic=True)
self.params = lasagne.layers.get_all_params(model['prob'],
trainable=True)
cost = -T.log(prob[T.arange(prob.shape[0]), y] + 1e-6).mean()
grads = T.grad(cost, self.params)
grads_all_reduced = self.grad_all_reduce(grads)
updates = adam(grads_all_reduced, self.params)
self.train_fn = theano.function([x, y], [cost, y],
updates=updates,
accept_inplace=True)
print '%s finished build theano fn, used %.3f' % (self.rank,
time.time() - t0)
formatter = logging.Formatter('%(asctime)s %(message)s', '%m/%d/%Y %I:%M:%S %p')
fh = logging.FileHandler('ln_hs_mscoco_captions.log')
fh.setLevel(logging.DEBUG)
fh.setFormatter(formatter)
sh = logging.StreamHandler(sys.stdout)
sh.setLevel(logging.INFO)
sh.setFormatter(formatter)
logger.addHandler(fh)
logger.addHandler(sh)
logger.info('Loading the ResNet50 model.')
# First get the ResNet50 model
resnet_weights_path = '/home/noury/modelzoo/resnet50.pkl'
resnet = build_model()
model_params = pickle.load(open(resnet_weights_path, 'rb'))
lasagne.layers.set_all_param_values(resnet['prob'], model_params['values'])
mean_im = model_params['mean_image'].reshape((1, 3, 224, 224)).astype(np.float32)
# Load the files needed for the MS COCO Captions dataset
train_images_path = '/home/noury/datasets/mscoco/train2014'
valid_images_path = '/home/noury/datasets/mscoco/val2014'
train_annotations_filepath = '/home/noury/datasets/mscoco/annotations/captions_train2014.json'
valid_annotations_filepath = '/home/noury/datasets/mscoco/annotations/captions_val2014.json'
coco_captions = pickle.load(open('coco_captions_trainval2014.pkl', 'rb'))
train_buckets = coco_captions['train buckets']
valid_buckets = coco_captions['valid buckets']
wordset = coco_captions['raw wordset']
word2idx = {}
word2idx['<PAD>'] = 0
num_epochs = 20
hidden_layer = 512
model_name = 'RESNET50' # 'VGG19','RESNET50'
X_train, X_valid, Y_train, Y_valid = utils.split_train_valid(
utils.read_all_training_file(
train_folder='./input/train/', label_file='./input/labels.csv', image_size=img_size),
test_percentage=0.2)
num_of_clases = 120 #Y_train.shape[1]
if model_name == 'VGG19':
model = vgg19.build_model(
num_of_clases, image_size=img_size, hidden_layer=hidden_layer)
elif model_name == 'RESNET50':
model = resnet50.build_model(
num_of_clases, image_size=img_size, hidden_layer=hidden_layer)
model.compile(loss='categorical_crossentropy',
optimizer='adam', metrics=['accuracy'])
model.summary()
# model.fit(x=X_train, y=Y_train, batch_size=batch,
# epochs=num_epochs, validation_data=(X_valid, Y_valid))
# data augmentation
generator = utils.augment_image_generator()
train_generator = generator.flow(
x=X_train, y=Y_train, batch_size=batch, shuffle=True)
#fit model
model.fit_generator(generator=train_generator,
steps_per_epoch=X_train.shape[0] // batch,