def predict(self,
coord,
fits_info,
learning_rate=0.01,
max_steps=2000,
hidden1=36,
hidden2=144,
hidden3=576,
batch_size=100):
network_model_dir = 'assets/log/{}_{}/l3_lr{}_ms{}_h1.{}_h2.{}_h3.{}_bs{}/model.ckpt-{}'.format(
self.region, self.exp_num, learning_rate, max_steps, hidden1, hidden2,
hidden3, batch_size, max_steps - 1)
with tf.Graph().as_default():
num_coord = len(coord)
coord_placeholder, psf_labels_placeholder = placeholder_inputs(
num_coord)
psf_pred = inference(coord_placeholder, hidden1, hidden2, hidden3)
new_saver = tf.train.Saver()
sess = tf.Session()
new_saver.restore(sess, network_model_dir)
feed_dict = {coord_placeholder: coord}
psf_predictions = sess.run(psf_pred, feed_dict=feed_dict)
if do_preprocess:
psf_predictions += self.chip_avg_train_data.ravel()
result_dir = 'assets/predictions/{}_{}/tf_psfwise/l3_lr{}_ms{}_h1.{}_h2.{}_h3.{}_bs{}/'.format(
self.region, self.exp_num, learning_rate, max_steps, hidden1,
hidden2, hidden3, batch_size)
utils.write_predictions(result_dir,
psf_predictions,
fits_info,
method='tf_psfwise')
def predict(self,
learning_rate=0.01, max_steps=2000, hidden1=36, hidden2=144, hidden3=576, batch_size=100, tag='validate'):
fits_info = []
for chip_num in range(36):
chip_data_name = 'chip_{}_data'.format(tag)
chip_psf_data = self.psf_data[chip_num]
fits_info += [data[0:4] for data in chip_psf_data[chip_data_name]]
coord = [data[0:2] for data in chip_psf_data[chip_data_name]]
coord = np.array(coord)
network_model_dir = 'assets/log/tf_chipwise/{}_{}/l3_lr{}_ms{}_h1.{}_h2.{}_h3.{}_bs{}/chn{}/model.ckpt-{}'.format(self.region, self.exp_num, learning_rate, max_steps, hidden1, hidden2, hidden3, batch_size, chip_num, max_steps-1)
with tf.Graph().as_default():
num_coord = len(coord)
coord_placeholder, psf_labels_placeholder = placeholder_inputs(num_coord)
psf_pred = inference(coord_placeholder, hidden1, hidden2, hidden3)
new_saver = tf.train.Saver()
sess = tf.Session()
new_saver.restore(sess, network_model_dir)
feed_dict = {
coord_placeholder: coord
}
if chip_num == 0:
psf_predictions = sess.run(psf_pred, feed_dict=feed_dict)
else:
chip_predictions = sess.run(psf_pred, feed_dict=feed_dict)
psf_predictions = np.concatenate((psf_predictions, chip_predictions), axis=0)
if do_preprocess:
psf_predictions += self.chip_avg_train_data.ravel()
result_dir = 'assets/predictions/{}_{}/tf_psfwise_chip/l3_lr{}_ms{}_h1.{}_h2.{}_h3.{}_bs{}/'.format(self.region, self.exp_num, learning_rate, max_steps, hidden1, hidden2, hidden3, batch_size)
utils.write_predictions(result_dir, psf_predictions, fits_info, method='tf_psfwise')
def predict(self,
coord,
fits_info,
learning_rate=0.01,
max_steps=1000,
hidden1=3,
hidden2=6,
batch_size=100):
pixel_predictions = []
try:
with open(
'assets/cache/{}_{}/l2_lr{}_ms{}_h1.{}_h2.{}_bs{}_pixel_predictions.p'
.format(self.region, self.exp_num, learning_rate, max_steps,
hidden1, hidden2, batch_size), 'rb') as pickle_file:
pixel_predictions = pickle.load(pickle_file)['pixel_predictions']
except FileNotFoundError:
exit()
for pixel_num in range(2304):
network_model_dir = 'assets/log/pixel_wise/{}_{}/l2_lr{}_ms{}_h1.{}_h2.{}_bs{}/pn{}/model.ckpt-{}'.format(
self.region, self.exp_num, learning_rate, max_steps, hidden1,
hidden2, batch_size, pixel_num, max_steps - 1)
with tf.Graph().as_default():
num_coord = len(coord)
coord_placeholder, pixel_labels_placeholder = placeholder_inputs(
num_coord)
pixel_pred = inference(coord_placeholder, hidden1, hidden2)
new_saver = tf.train.Saver()
sess = tf.Session()
new_saver.restore(sess, network_model_dir)
feed_dict = {coord_placeholder: coord}
sub_pixel_predictions = sess.run(pixel_pred,
feed_dict=feed_dict)
pixel_predictions.append(
sub_pixel_predictions.reshape(
sub_pixel_predictions.shape[0]))
# TODO: Design pixelwise saving directory
# TODO: Or just assemble the information together
pixel_predictions = np.array(pixel_predictions)
pickle.dump(
{'pixel_predictions': pixel_predictions},
open(
'assets/cache/{}_{}/l2_lr{}_ms{}_h1.{}_h2.{}_bs{}_pixel_predictions.p'
.format(self.region, self.exp_num, learning_rate, max_steps,
hidden1, hidden2, batch_size), 'wb'))
pixel_predictions = pixel_predictions.T.copy()
if do_preprocess:
pixel_predictions += self.chip_avg_train_data.ravel()
result_dir = 'assets/predictions/{}_{}/tf_pixelwise/l2_lr{}_ms{}_h1.{}_h2.{}_bs{}/'.format(
self.region, self.exp_num, learning_rate, max_steps, hidden1, hidden2,
batch_size)
utils.write_predictions(result_dir,
pixel_predictions,
fits_info,
method='tf_pixelwise')
def predict(self, coord, fits_info, order):
poly_name = 'poly_{}'.format(str(order))
if not (poly_name in self.cal_info):
poly_interpolation(self, order)
coeffs = self.cal_info[poly_name]
psf_predictions = np.array([utils.poly_val_all(the_coord[0], the_coord[1], coeffs, order) for the_coord in coord])
if do_preprocess:
psf_predictions += self.chip_avg_train_data.ravel()
result_dir = 'assets/predictions/{}_{}/poly/{}/'.format(self.region, self.exp_num, poly_name)
utils.write_predictions(result_dir, psf_predictions, fits_info, method=poly_name)
def predict(self, coord, fits_info):
if not ("poly1" in self.cal_info):
poly1_interpolation(self)
opt_A, opt_B, opt_C = self.cal_info['poly1']
psf_predictions = np.array([
the_coord[0] * opt_A + the_coord[1] * opt_B + opt_C
for the_coord in coord
])
if do_preprocess:
psf_predictions += self.chip_avg_train_data.ravel()
result_dir = 'assets/predictions/{}_{}/poly/poly1/'.format(
self.region, self.exp_num)
utils.write_predictions(result_dir, psf_predictions, fits_info, 'poly1')
def collect_origin_data(self, tag='train'):
origin_psf = []
fits_info = []
part_name = 'chip_{}_data'.format(tag)
for chip_psf_data in self.psf_data:
origin_psf += [
data[4].ravel() for data in chip_psf_data[part_name]
] if not do_preprocess else [
data[4].ravel() + self.chip_avg_train_data.ravel()
for data in chip_psf_data[part_name]
]
fits_info += [data[0:4] for data in chip_psf_data[part_name]]
origin_psf = np.array(origin_psf)
result_dir = 'assets/predictions/{}_{}/origin/{}/'.format(
self.region, self.exp_num, tag)
utils.write_predictions(result_dir, origin_psf, fits_info, method=tag)