Skip to content

Resolve TODOs. #2

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
mishajw wants to merge 6 commits into
base: main
Choose a base branch
from
Open

Resolve TODOs. #2

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
4193a47
Remove unused latent masking code.
mishajw Oct 28, 2022
af7105c
Remove unused if.
mishajw Oct 28, 2022
b653d9a
Wrap diagonal representation fn.
mishajw Oct 28, 2022
2d92588
Remove redundant TODO.
mishajw Oct 28, 2022
e0c3254
Remove unused print statement.
mishajw Oct 28, 2022
83c66dc
Formatting fixes.
mishajw Oct 29, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 0 additions & 2 deletions experiments.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -40,8 +40,6 @@ class Experiment:
quadrant_threshold: int = 0
sparsity: int = 1 # repr_loss scaled up by sparsity, applied every 1/sparsity
latent_noise_std: float = 0
latent_masking: bool = False
latent_masking_incentive: float = 0.1


baseline = Experiment(
Expand Down
34 changes: 11 additions & 23 deletions main.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -346,12 +346,16 @@ def _hyperparameter_search(*experiments_iterable: Experiment) -> None:
df = []
for train_result in train_results:
last_step = max(step_result.step for step_result in train_result.step_results)
reconstruction_loss_p2 = np.mean([
step_result.reconstruction_loss_p2
for step_result in train_result.step_results
if step_result.step >= last_step * 0.9
])
df.append(dict(tag=train_result.tag, reconstruction_loss_p2=reconstruction_loss_p2))
reconstruction_loss_p2 = np.mean(
[
step_result.reconstruction_loss_p2
for step_result in train_result.step_results
if step_result.step >= last_step * 0.9
]
)
df.append(
dict(tag=train_result.tag, reconstruction_loss_p2=reconstruction_loss_p2)
)
df = pd.DataFrame(df)
fig, ax = plt.subplots()
sns.barplot(data=df, x="tag", y="reconstruction_loss_p2", ax=ax)
Expand Down Expand Up @@ -438,8 +442,6 @@ def _train(experiment: Experiment) -> TrainResult:
target_latent_fn: Callable
repr_loss_mask_fn: Callable

latent_use_mask_fn = torch.nn.Sigmoid()

if experiment.loss_quadrants == "all":
repr_loss_mask_fn = lambda x: torch.ones(x.shape[0])
repr_loss_scale = 1.0
Expand All @@ -455,7 +457,6 @@ def _train(experiment: Experiment) -> TrainResult:
raise ValueError(
f"Loss quadrant must be 'all', 'bin_sum' or 'bin_val', got {experiment.loss_quadrants}."
)
print(f"repr_loss_scale = {repr_loss_scale}")

if experiment.use_class:
reconstruction_loss_fn = torch.nn.CrossEntropyLoss()
Expand Down Expand Up @@ -492,9 +493,6 @@ def _train(experiment: Experiment) -> TrainResult:
step_results = []
encoder_to_decoder_idx = list(range(len(models)))
for step in range(experiment.num_batches):
# TODO: Delete this if?
if experiment.dropout_prob is not None and step == 9000:
pass
if experiment.shuffle_decoders:
random.shuffle(encoder_to_decoder_idx)

Expand Down Expand Up @@ -527,14 +525,6 @@ def _train(experiment: Experiment) -> TrainResult:
vector_input = vector

latent_repr = encoder(vector_input)
if experiment.latent_masking:
latent_repr = latent_repr[: experiment.preferred_rep_size]
repr_mask = latent_use_mask_fn(latent_repr)
# TODO: Unused variable?
repr_use_loss = torch.mean(repr_mask)
else:
repr_use_loss = torch.Tensor(0)

noise = torch.normal(
mean=0, std=experiment.latent_noise_std, size=latent_repr.shape
)
Expand All @@ -555,7 +545,6 @@ def _train(experiment: Experiment) -> TrainResult:
target_latent_fn(latent_repr),
)
# Scaling here to compensate for quadrant sparsity
# TODO: Should we roll this into `experiment.representation_loss`?
representation_loss *= repr_loss_scale
loss = reconstruction_loss
if experiment.representation_loss is not None:
Expand Down Expand Up @@ -690,8 +679,7 @@ def _make_diagonal_repr_fn(rep_size: int) -> Callable:
def diagonal_repr_target(_input: torch.Tensor) -> torch.Tensor:
assert rep_size + 1 <= _input.shape[1]
dir_1 = _input[:, :rep_size]
# TODO: Wrap this so that we don't use N+1 variables?
dir_2 = _input[:, 1 : rep_size + 1]
dir_2 = torch.concat([_input[:, 1:rep_size], _input[:, :1]], dim=1)
repr_target = (dir_1 + dir_2) / np.sqrt(2)
return repr_target

Expand Down