Summary: 49 instances, 45 unique

Text	Count
# TODO Automaticall calculate padding	1
// TODO: assert that features have the same size	1
// TODO: The ffbs data below is used for loading frames from buffers, and	1
# TODO: check that the gating is learned to be 0 on the last layer,	1
# TODO maybe do some pooling in the temporal convolution	1
# TODO Hungarian (scipy.optimize.linear_sum_assignment)? How do we deal with dead units? Right now hack to have an approx of the avg bias	1
if recent_mean_loss > 0.99 * older_mean_loss or self.args.small:  # TODO with lr_decay only	1
# TODO a mass conservation (except where we have factories?) loss?	1
self.zs = {}  # TODO replace by LookUpTable	1
"KL": th.nn.KLDivLoss  # TODO (the target in training)	1
//TODO do targets too	1
# Each midnet runs a convolution + LSTM; TODO make it do multiple conv	1
# TODO hierarchical deconv	1
// Never true for forward models TODO Change for policy models	2
# TODO bptt sampled with cut ~ Bernouilli(hyperparam)	1
# TODO becareful about size_averaging and BPTT	1
# TODO: add visibility feature	1
# TODO skip_frames >= combine_frames	1
elif state.delta_full_info: # TODO add this option, TODO change test time!	1
self.encoder = None  # TODO I hope this garbage collects.	1
# TODO check input/output size in features/channels	1
with th.cuda.device(self.args.gpu):  # TODO wrap everything?	1
# TODO:	1
self.conv1x1 = nn.Conv2d(self.nchannel, self.inp_embsize, 1)  # TODO do that before trunk?	2
# TODO decoder that starts from input embedding (after first 1x1 Conv2d)	1
# TODO Dynamically search for this somehow???	1
v = [v[0], v[1], 2 * v[0] * v[1] / (v[0] + v[1] + 1E-9)]  # TODO F1 score after all averaging ok?	1
# TODO skip connections from midnets to corresponding decoder level.	1
# TODO Make the data loaders all generators	1
self._feeder_queue = queue.Queue()  # TODO check	2
with th.cuda.device(self.args.gpu):  # TODO wrap everything?	1
if self.with_z:  # TODO replace by class decorator	2
# TODO actually do it with the nn.Module correspondong to midnets+midrnn	1
#if args.loss == 'SoftMargin': // TODO KL	1
# TODO Consistency loss between regression and classification heads?	1
# TODO a structured prediction (local and or global) type of loss?	1
# TODO Make the sizes line up with resnets...	1
# TODO try to remove border artifacts (borders are important!)	1
# TODO: move other_metrics to self.state / init, but only once we stabilize what's in it	1
# TODO plot debug outputs on the valid set	1
if args.save:  # TODO save on other metrics?	1
# TODO: Both reduced data and overall data may be compressed	1
"TODO: make the array an hdf5 instead if we want to reuse.")	1
# TODO the loss/model should make you pay less for small (dx, dy) in prediction	1
# TODO remove this multiple optimizer stuff since we just create a new one here	1