def __init__(self):
#self.q_ph = None
self.q_ph = tf.compat.v1.placeholder(dtype=tf.int32, shape=())
from gqn_inputs import DataReader
kwargs = {
'dataset': const.GQN_DATA_NAME,
'context_size': const.NUM_VIEWS,
'root': 'gqn-dataset'
}
self.train_data_reader = DataReader(mode='train', **kwargs)
self.test_data_reader = DataReader(mode='test', **kwargs)
assert const.NUM_PREDS == 1
#!/usr/bin/env python3
import sys
sys.path.append('..')
from gqn_inputs import DataReader
import tensorflow as tf
import numpy as np
import ipdb
data_reader = DataReader(dataset='shepard_metzler_5_parts',
context_size=10,
root='../gqn-dataset',
mode='test')
data = data_reader.read(batch_size=10)
with tf.train.SingularMonitoredSession() as sess:
task = sess.run(data)
pos = task.query.context.cameras[0, :, :3]
dist = np.linalg.norm(pos, axis=1)
print(dist)
# let's do some geometry here
pos /= 10.0 / 3.0
yaws = task.query.context.cameras[0, :, 3]
pitches = task.query.context.cameras[0, :, 4]
# ys = np.sin(pitches)
class GQNInput(Input):
def __init__(self):
#self.q_ph = None
self.q_ph = tf.compat.v1.placeholder(dtype=tf.int32, shape=())
from gqn_inputs import DataReader
kwargs = {
'dataset': const.GQN_DATA_NAME,
'context_size': const.NUM_VIEWS,
'root': 'gqn-dataset'
}
self.train_data_reader = DataReader(mode='train', **kwargs)
self.test_data_reader = DataReader(mode='test', **kwargs)
assert const.NUM_PREDS == 1
def data(self):
#self.q_ph = utils.tfpy.print_val(self.q_ph, 'qph')
rval = tf.case(
{
tf.equal(self.q_ph, 0):
lambda: self.train_data_reader.read(batch_size=const.BS),
tf.equal(self.q_ph, 1):
lambda: self.test_data_reader.read(batch_size=const.BS),
tf.equal(self.q_ph, 2):
lambda: self.test_data_reader.read(batch_size=const.BS)
},
exclusive=True)
rval = self.munch(rval)
if const.generate_views:
assert const.BS == 1
counter = tf.Variable(0, dtype=tf.int32)
increment_op = tf.compat.v1.assign_add(counter, 1)
counter_mod = tf.mod(counter, const.GEN_NUM_VIEWS)
# printing is a bit weird due to the async ??
counter_mod = utils.tfpy.print_val(counter_mod, 'counter is')
elev_index = (counter_mod - 1) / const.AZIMUTH_GRANULARITY
azimuth_index = tf.mod((counter_mod - 1),
const.AZIMUTH_GRANULARITY)
elev_index = tf.cast(elev_index, tf.float32)
azimuth_index = tf.cast(azimuth_index, tf.float32)
azimuth = azimuth_index * (360 / const.AZIMUTH_GRANULARITY)
azimuth = azimuth + tf.cast(azimuth > 180, tf.float32) * (-360)
azimuth /= 180 / np.pi
if const.ELEV_GRANULARITY == 1:
elev = const.MIN_ELEV
else:
elev_step = (const.MAX_ELEV -
const.MIN_ELEV) / (const.ELEV_GRANULARITY - 1)
elev = (const.MIN_ELEV + elev_index * elev_step) / (180 /
np.pi)
with tf.control_dependencies([increment_op]):
#phi, theta is the correct order here
query_cam = tf.expand_dims(tf.stack([elev, azimuth], axis=0),
axis=0)
rval.query.query_camera = query_cam
return rval
def munch(self, task):
return Munch(query=self.munchq(task.query), target=task.target)
def munchq(self, query):
return Munch(context=self.munchc(query.context),
query_camera=query.query_camera)
def munchc(self, context):
return Munch(frames=tf.unstack(context.frames, axis=1),
cameras=tf.unstack(context.cameras, axis=1))