conv_lstm_models.py (18 lines):
	- line 19: # TODO: add visibility feature
	- line 29: # TODO: check that the gating is learned to be 0 on the last layer,
	- line 145: # TODO Make the sizes line up with resnets...
	- line 208: self.conv1x1 = nn.Conv2d(self.nchannel, self.inp_embsize, 1)  # TODO do that before trunk?
	- line 296: if self.with_z:  # TODO replace by class decorator
	- line 316: self.conv1x1 = nn.Conv2d(self.nchannel, self.inp_embsize, 1)  # TODO do that before trunk?
	- line 334: if self.with_z:  # TODO replace by class decorator
	- line 344: self.zs = {}  # TODO replace by LookUpTable
	- line 359: # TODO decoder that starts from input embedding (after first 1x1 Conv2d)
	- line 360: # TODO check input/output size in features/channels
	- line 361: # TODO try to remove border artifacts (borders are important!)
	- line 362: # TODO hierarchical deconv
	- line 595: self.encoder = None  # TODO I hope this garbage collects.
	- line 596: # TODO actually do it with the nn.Module correspondong to midnets+midrnn
	- line 597: # TODO skip connections from midnets to corresponding decoder level.
	- line 640: # Each midnet runs a convolution + LSTM; TODO make it do multiple conv
	- line 697: # TODO Automaticall calculate padding
	- line 698: # TODO maybe do some pooling in the temporal convolution


compare_models.py (18 lines):
	- line 55: # 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
	- line 134: # TODO a structured prediction (local and or global) type of loss?
	- line 135: # TODO a mass conservation (except where we have factories?) loss?
	- line 136: # TODO the loss/model should make you pay less for small (dx, dy) in prediction
	- line 188: "KL": th.nn.KLDivLoss  # TODO (the target in training)
	- line 192: # TODO becareful about size_averaging and BPTT
	- line 223: # TODO Consistency loss between regression and classification heads?
	- line 304: with th.cuda.device(self.args.gpu):  # TODO wrap everything?
	- line 363: # TODO: move other_metrics to self.state / init, but only once we stabilize what's in it
	- line 559: v = [v[0], v[1], 2 * v[0] * v[1] / (v[0] + v[1] + 1E-9)]  # TODO F1 score after all averaging ok?
	- line 587: if args.save:  # TODO save on other metrics?
	- line 597: if recent_mean_loss > 0.99 * older_mean_loss or self.args.small:  # TODO with lr_decay only
	- line 598: with th.cuda.device(self.args.gpu):  # TODO wrap everything?
	- line 607: # TODO plot debug outputs on the valid set
	- line 626: elif state.delta_full_info: # TODO add this option, TODO change test time!
	- line 827: # TODO remove this multiple optimizer stuff since we just create a new one here
	- line 929: # TODO bptt sampled with cut ~ Bernouilli(hyperparam)
	- line 941: #if args.loss == 'SoftMargin': // TODO KL


featurizers/forward_conv2d_state_action_featurizer.cpp (3 lines):
	- line 85: //TODO do targets too
	- line 153: // Never true for forward models TODO Change for policy models
	- line 184: // Never true for forward models TODO Change for policy models


data.py (3 lines):
	- line 19: # TODO Make the data loaders all generators
	- line 51: self._feeder_queue = queue.Queue()  # TODO check
	- line 146: self._feeder_queue = queue.Queue()  # TODO check


conv_lstm_utils.py (1 line):
	- line 21: # TODO:


featurizers/coarse_conv_featurizer_units.h (1 line):
	- line 17: // TODO: The ffbs data below is used for loading frames from buffers, and


baseline.py (1 line):
	- line 36: "TODO: make the array an hdf5 instead if we want to reuse.")


utils.py (1 line):
	- line 112: # TODO skip_frames >= combine_frames


_ext.cpp (1 line):
	- line 123: // TODO: assert that features have the same size


setup.py (1 line):
	- line 51: # TODO Dynamically search for this somehow???


reduce_data.py (1 line):
	- line 91: # TODO: Both reduced data and overall data may be compressed