def load_resource(self):
self.player = cv2.imread(os.path.join(self.resource_dir, 'player.png'),
0)
# network initization
self.net = JumpModel()
self.net_fine = JumpModelFine()
self.img = tf.placeholder(tf.float32, [None, 640, 720, 3], name='img')
self.img_fine = tf.placeholder(tf.float32, [None, 320, 320, 3],
name='img_fine')
self.label = tf.placeholder(tf.float32, [None, 2], name='label')
self.is_training = tf.placeholder(np.bool, name='is_training')
self.keep_prob = tf.placeholder(np.float32, name='keep_prob')
self.pred = self.net.forward(self.img, self.is_training,
self.keep_prob)
self.pred_fine = self.net_fine.forward(self.img_fine, self.is_training,
self.keep_prob)
self.sess = tf.Session()
self.sess.run(tf.global_variables_initializer())
all_vars = tf.all_variables()
var_coarse = [k for k in all_vars if k.name.startswith('coarse')]
var_fine = [k for k in all_vars if k.name.startswith('fine')]
self.saver_coarse = tf.train.Saver(var_coarse)
self.saver_fine = tf.train.Saver(var_fine)
self.saver_coarse.restore(self.sess, self.ckpt)
self.saver_fine.restore(self.sess, self.ckpt_fine)
print('==== successfully restored ====')
def inference(path):
net = JumpModel()
img = tf.placeholder(tf.float32, [None, 640, 720, 3], name='img')
label = tf.placeholder(tf.float32, [None, 2], name='label')
is_training = tf.placeholder(np.bool, name='is_training')
keep_prob = tf.placeholder(np.float32, name='keep_prob')
lr = tf.placeholder(np.float32, name='lr')
pred = net.forward(img, is_training, keep_prob)
saver = tf.train.Saver()
sess = tf.Session()
sess.run(tf.global_variables_initializer())
ckpt = tf.train.get_checkpoint_state('./train_logs')
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
print('==== successfully restored ====')
val_img = get_a_img(path)
feed_dict = {
img: val_img,
is_training: False,
keep_prob: 1.0,
}
pred_out = sess.run(pred, feed_dict=feed_dict)
return pred_out
def load_resource(self):
self.player = cv2.imread(os.path.join(self.resource_dir, 'player.png'),
0)
# network initization
self.net = JumpModel()
self.img = tf.placeholder(tf.float32, [None, 640, 720, 3], name='img')
self.label = tf.placeholder(tf.float32, [None, 2], name='label')
self.is_training = tf.placeholder(np.bool, name='is_training')
self.keep_prob = tf.placeholder(np.float32, name='keep_prob')
self.pred = self.net.forward(self.img, self.is_training,
self.keep_prob)
self.saver = tf.train.Saver()
self.sess = tf.Session()
self.sess.run(tf.global_variables_initializer())
ckpt = tf.train.get_checkpoint_state(
os.path.join(self.resource_dir, 'train_logs2'))
if ckpt and ckpt.model_checkpoint_path:
self.saver.restore(self.sess, ckpt.model_checkpoint_path)
print('==== successfully restored ====')
def load_resource(self):
self.player = cv2.imread(os.path.join(self.resource_dir, 'player.png'),
0)
# network initialization
self.net = JumpModel()
self.net_fine = JumpModelFine()
self.img = tf.placeholder(tf.float32,
shape=(None, 640, 720, 3),
name='img')
self.img_fine = tf.placeholder(tf.float32,
shape=(None, 320, 320, 3),
name='img_fine')
self.label = tf.placeholder(tf.float32, [None, 2], name='label')
self.is_training = tf.placeholder(np.bool, name='is_training')
self.keep_prob = tf.placeholder(
np.float32, name='keep_prob') # get_target_position()用到
# 这里的第三个参数self.keep_prob必须是string类型的
# self.pred = self.net.forward(self.img, self.is_training, self.keep_prob)
# self.pred_fine = self.net_fine.forward(self.img_fine, self.is_training, self.keep_prob)
# 第三个参数直接写成string类型,两个模型-分别为 coarse 与 fine
self.pred = self.net.forward(self.img, self.is_training, "coarse")
self.pred_fine = self.net_fine.forward(self.img_fine, self.is_training,
"fine")
self.sess = tf.Session()
self.sess.run(tf.global_variables_initializer())
all_vars = tf.global_variables(
) # tf.all_variables() # Please use tf.global_variables instead.
var_coarse = [k for k in all_vars if k.name.startswith('coarse')]
var_fine = [k for k in all_vars if k.name.startswith('fine')]
self.saver_coarse = tf.train.Saver(var_coarse)
self.saver_fine = tf.train.Saver(var_fine)
self.saver_coarse.restore(self.sess, self.ckpt)
self.saver_fine.restore(self.sess, self.ckpt_fine)
print('==== successfully restored ====')
6], name[name.index('_w_') + 3:name.index('_w_') + 6]
x, y = int(x), int(y)
img = cv2.imread(img_name)
label = np.array([x, y], dtype=np.float32)
return img[np.newaxis, :, :, :], label.reshape((1, label.shape[0]))
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('-g', '--gpu', default=None, type=int)
args = parser.parse_args()
if args is not None:
os.environ['CUDA_VISIBLE_DEVICES'] = str(args.gpu)
net = JumpModel()
dataset = JumpData()
img = tf.placeholder(tf.float32, [None, 640, 720, 3], name='img')
label = tf.placeholder(tf.float32, [None, 2], name='label')
is_training = tf.placeholder(np.bool, name='is_training')
keep_prob = tf.placeholder(np.float32, name='keep_prob')
lr = tf.placeholder(np.float32, name='lr')
pred = net.forward(img, is_training, keep_prob, 'coarse')
loss = tf.reduce_mean(tf.sqrt(tf.pow(pred - label, 2) + 1e-12))
tf.summary.scalar('loss', loss)
optimizer = tf.train.AdamOptimizer(lr)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
train_op = optimizer.minimize(loss)
saver = tf.train.Saver()