def get_test_losses(opts, sample, output, name='loss'):
"""Get testing loss funcion
Input:
- opts (options) - object with all relevant options stored
- sample (dict) - sample from training dataset
- output (tf.Tensor) - output from network on sample
- name (string, optional) - name prefix for tensorflow scoping (default loss)
Output:
- gt_l1_loss (tf.Tensor) - L1 loss against ground truth
- gt_l2_loss (tf.Tensor) - L2 loss against ground truth
- gt_bce_loss (tf.Tensor) - BCE loss against ground truth
- ssame_m (tf.Tensor) - Mean similarity of corresponding points
- ssame_var (tf.Tensor) - Standard dev. of similarity of corresponding points
- sdiff_m (tf.Tensor) - Mean similarity of non-corresponding points
- sdiff_var (tf.Tensor) - Standard dev. of similarity of non-corresponding
points
"""
emb = sample['TrueEmbedding']
output_sim = tfutils.get_sim(output)
sim_true = tfutils.get_sim(emb)
if opts.loss_type == 'bce':
osim = tf.sigmoid(output_sim)
osim_log = output_sim
else:
osim = output_sim
osim_log = tf.log(tf.abs(output_sim) + 1e-9)
gt_l1_loss = loss_fns['l1'](sim_true, osim, add_loss=False)
gt_l2_loss = loss_fns['l2'](sim_true, osim, add_loss=False)
gt_bce_loss = loss_fns['bce'](sim_true, osim, add_loss=False)
num_same = tf.reduce_sum(sim_true)
num_diff = tf.reduce_sum(1 - sim_true)
ssame_m, ssame_var = tf.nn.weighted_moments(osim, None, sim_true)
sdiff_m, sdiff_var = tf.nn.weighted_moments(osim, None, 1 - sim_true)
return gt_l1_loss, gt_l2_loss, gt_bce_loss, ssame_m, ssame_var, sdiff_m, sdiff_var
def get_loss(opts, sample, output, return_gt=False, name='train'):
"""Get total loss funcion with main loss functions and regularizers
Input:
- opts (options) - object with all relevant options stored
- sample (dict) - sample from training dataset
- output (tf.Tensor) - output from network on sample
- return_gt (boolean, optional) - return ground truth losses (default False)
- name (string, optional) - name prefix for tensorflow scoping (default train)
Output:
- loss (tf.Tensor) - total summed loss value
- gt_l1_loss (tf.Tensor) - L1 loss against ground truth (only returned
if return_gt=True)
- gt_l2_loss (tf.Tensor) - L2 loss against ground truth (only returned
if return_gt=True)
"""
emb = sample['TrueEmbedding']
output_sim = tfutils.get_sim(output)
if opts.use_abs_value:
output_sim = tf.abs(output_sim)
sim_true = tfutils.get_sim(emb)
# Figure out if we are using unsupervised to know if we can use GT Adjcency
if opts.use_unsupervised_loss:
v = opts.dataset_params.views[-1]
p = opts.dataset_params.points[-1]
b = opts.batch_size
sim = sample['AdjMat'] + tf.eye(num_rows=v*p, batch_shape=[b])
else:
sim = sim_true
if opts.full_tensorboard: # Add to tensorboard
tf.summary.image('Output Similarity {}'.format(name), tf.expand_dims(output_sim, -1))
tf.summary.image('Embedding Similarity {}'.format(name), tf.expand_dims(sim, -1))
# Our main loss, the reconstruction loss
reconstr_loss = loss_fns[opts.loss_type](sim, output_sim)
if opts.full_tensorboard: # Add to tensorboard
tf.summary.scalar('Reconstruction Loss {}'.format(name), reconstr_loss)
# Geometric loss terms and tensorboard
if opts.geometric_loss > 0:
geo_loss = get_geometric_loss(opts, sample, output_sim, name='geom_loss_{}'.format(name))
if opts.full_tensorboard: # Add to tensorboard
tf.summary.scalar('Geometric Loss {}'.format(name), geo_loss)
geo_loss_gt = get_geometric_loss(opts, sample, sim_true)
tf.summary.scalar('Geometric Loss GT {}'.format(name), geo_loss_gt)
loss = opts.reconstruction_loss * reconstr_loss + opts.geometric_loss * geo_loss
else:
loss = reconstr_loss
tf.summary.scalar('Total Loss {}'.format(name), loss)
# Compare to loss vs ground truth (almost always do that)
if return_gt:
output_sim_gt = output_sim
if opts.loss_type == 'bce':
output_sim_gt = tf.sigmoid(output_sim)
gt_l1_loss = loss_fns['l1'](sim_true, output_sim_gt, add_loss=False)
gt_l2_loss = loss_fns['l2'](sim_true, output_sim_gt, add_loss=False)
if opts.full_tensorboard and opts.use_unsupervised_loss: # Add to tensorboard
tf.summary.scalar('GT L1 Loss {}'.format(name), gt_l1_loss)
tf.summary.scalar('GT L2 Loss {}'.format(name), gt_l2_loss)
return loss, gt_l1_loss, gt_l2_loss
else:
return loss
def get_output_sim(self, out_graph):
out_shape = [self.batch_size, -1, self.final_embedding_dim]
output = tf.reshape(out_graph.nodes, out_shape)
output_sim = tfutils.get_sim(output)
if self.use_abs_value:
output_sim = tf.abs(output_sim)
return output_sim
def get_loss(opts, sample, output, return_gt=False, name='train'):
emb = sample['TrueEmbedding']
output_sim = tfutils.get_sim(output)
if opts.use_end_bias:
end_bias = get_end_bias()
output_sim = output_sim + end_bias
if opts.use_abs_value:
output_sim = tf.abs(output_sim)
sim_true = tfutils.get_sim(emb)
if opts.use_unsupervised_loss:
v = opts.dataset_params.views[-1]
p = opts.dataset_params.points[-1]
b = opts.batch_size
sim = sample['AdjMat'] + tf.eye(num_rows=v*p, batch_shape=[b])
else:
sim = sim_true
if opts.full_tensorboard:
tf.summary.image('Output Similarity {}'.format(name), tf.expand_dims(output_sim, -1))
tf.summary.image('Embedding Similarity {}'.format(name), tf.expand_dims(sim, -1))
reconstr_loss = loss_fns[opts.loss_type](sim, output_sim)
if opts.full_tensorboard:
tf.summary.scalar('Reconstruction Loss {}'.format(name), reconstr_loss)
if opts.geometric_loss > 0:
geo_loss = get_geometric_loss(opts, sample, output_sim, name='geom_loss_{}'.format(name))
if opts.full_tensorboard:
tf.summary.scalar('Geometric Loss {}'.format(name), geo_loss)
geo_loss_gt = get_geometric_loss(opts, sample, sim_true)
tf.summary.scalar('Geometric Loss GT {}'.format(name), geo_loss_gt)
loss = opts.reconstruction_loss * reconstr_loss + opts.geometric_loss * geo_loss
else:
loss = reconstr_loss
tf.summary.scalar('Total Loss {}'.format(name), loss)
if return_gt:
output_sim_gt = output_sim
if opts.loss_type == 'bce':
output_sim_gt = tf.sigmoid(output_sim)
gt_l1_loss = loss_fns['l1'](sim_true, output_sim_gt, add_loss=False)
gt_l2_loss = loss_fns['l2'](sim_true, output_sim_gt, add_loss=False)
if opts.full_tensorboard and opts.use_unsupervised_loss:
tf.summary.scalar('GT L1 Loss {}'.format(name), gt_l1_loss)
tf.summary.scalar('GT L2 Loss {}'.format(name), gt_l2_loss)
return loss, gt_l1_loss, gt_l2_loss
else:
return loss
def get_test_losses(opts, sample, output, return_gt=False, name='loss'):
emb = sample['TrueEmbedding']
output_sim = tfutils.get_sim(output)
sim_true = tfutils.get_sim(emb)
if opts.use_end_bias:
end_bias = get_end_bias()
output_sim = output_sim + end_bias
if opts.loss_type == 'bce':
osim = tf.sigmoid(output_sim)
osim_log = output_sim
else:
osim = output_sim
osim_log = tf.log(tf.abs(output_sim) + 1e-9)
gt_l1_loss = loss_fns['l1'](sim_true, osim, add_loss=False)
gt_l2_loss = loss_fns['l2'](sim_true, osim, add_loss=False)
gt_bce_loss = loss_fns['bce'](sim_true, osim, add_loss=False)
num_same = tf.reduce_sum(sim_true)
num_diff = tf.reduce_sum(1 - sim_true)
ssame_m, ssame_var = tf.nn.weighted_moments(osim, None, sim_true)
sdiff_m, sdiff_var = tf.nn.weighted_moments(osim, None, 1 - sim_true)
return gt_l1_loss, gt_l2_loss, gt_bce_loss, ssame_m, ssame_var, sdiff_m, sdiff_var
print("network")
network = MLPInteractiveNetwork()
print(network)
print("data")
dataset = data_util.datasets.get_dataset(opts)
sample = dataset.load_batch('train')
print(sample.keys())
print(sample['true_match'])
print("output")
output = network(sample['graph'])
print(output.nodes)
output_batch = tf.reshape(output.nodes, [ opts.batch_size, -1, 25 ])
print(output_batch)
output_sim = tfutils.get_sim(output_batch)
print(output_sim)
diff = tf.sparse_add(-output_sim, sample['adj_mat'])
loss = tf.reduce_mean(tf.abs(diff))
b = opts.batch_size
v = opts.dataset_params.views[-1]
p = opts.dataset_params.points[-1]
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
pass
init=tf.global_variables_initializer()
sess.run(init)
# for i in range(opts.dataset_params.sizes['train'] // opts.batch_size):