def pcl_loss(pclA, tMatP, tMatT, **kwargs): # batchSize=Sne
"""
Generate a ground truth point cloud using ground truth transformation
Generate a prediction point cloud using predicted transformation
L2 difference between ground truth and predicted point cloud is the loss value
"""
# pclA, tMatP, tMatT are in batches
# tMatP, tMatT should get a 0,0,0,1 row and be reshaped to 4x4
tMatP = tf.concat([
tMatP,
tf.constant(
np.repeat(np.array([[0, 0, 0, 1]], dtype=np.float32),
kwargs.get('activeBatchSize'),
axis=0))
], 1)
tMatT = tf.concat([
tMatT,
tf.constant(
np.repeat(np.array([[0, 0, 0, 1]], dtype=np.float32),
kwargs.get('activeBatchSize'),
axis=0))
], 1)
tMatP = tf.reshape(tMatP, [kwargs.get('activeBatchSize'), 4, 4])
tMatT = tf.reshape(tMatT, [kwargs.get('activeBatchSize'), 4, 4])
# pclA should get a row of ones
pclA = tf.concat([
pclA,
tf.constant(
np.ones([kwargs.get('activeBatchSize'), 1,
kwargs.get('pclCols')],
dtype=np.float32))
], 1)
pclP = tf.matmul(tMatP, pclA)
pclT = tf.matmul(tMatT, pclA)
return model_base.loss(pclP, pclT, **kwargs)
def weighted_loss(tMatP, tMatT, **kwargs):
mask = np.array([[100, 100, 100, 1, 100, 100, 100, 1, 100, 100, 100, 1]],
dtype=np.float32)
mask = np.repeat(mask, kwargs.get('activeBatchSize'), axis=0)
tMatP = tf.multiply(mask, tMatP)
tMatT = tf.multiply(mask, tMatT)
return model_base.loss(tMatP, tMatT, **kwargs)
def loss(pred, target, **kwargs): # batchSize=Sne
"""Add L2Loss to all the trainable variables.
Add summary for "Loss" and "Loss/avg".
Args:
logits: Logits from inference().
labels: Labels from distorted_inputs or inputs(). 1-D tensor
of shape [batch_size, heatmap_size ]
Returns:
Loss tensor of type float.
"""
return model_base.loss(pred, target, **kwargs)
def weighted_params_loss(targetP, targetT, **kwargs):
# Alpha, Beta, Gamma are -Pi to Pi periodic radians - mod over pi to remove periodicity
mask = np.array([[np.pi, np.pi, np.pi, 1, 1, 1]], dtype=np.float32)
mask = np.repeat(mask, kwargs.get('activeBatchSize'), axis=0)
targetP = tf_mod(targetP, mask)
targetT = tf_mod(targetT, mask)
# Importance weigting on angles as they have smaller values
mask = np.array([[1000, 1000, 1000, 1, 1, 1]], dtype=np.float32)
mask = np.repeat(mask, kwargs.get('activeBatchSize'), axis=0)
targetP = tf.multiply(targetP, mask)
targetT = tf.multiply(targetT, mask)
return model_base.loss(targetP, targetT, **kwargs)
def loss(pHAB, tHAB, **kwargs): # batchSize=Sne
return model_base.loss(pHAB, tHAB, **kwargs)
