Skip to content

fix: prevent NaN from division by zero in TFT interpretation + fix headless test backend #2207

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
haoyu-haoyu wants to merge 1 commit into
base: main
Choose a base branch
from
+13 −7
Open

fix: prevent NaN from division by zero in TFT interpretation + fix headless test backend #2207

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
18 changes: 11 additions & 7 deletions pytorch_forecasting/models/temporal_fusion_transformer/_tft.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -849,9 +849,10 @@ def interpret_output(

attention = masked_op(attention, dim=0, op=reduction)
else:
attention = attention / masked_op(attention, dim=1, op="sum").unsqueeze(
-1
) # renormalize
denom = masked_op(
attention, dim=1, op="sum"
).clamp(min=1e-8).unsqueeze(-1)
attention = attention / denom # renormalize

interpretation = dict(
attention=attention.masked_fill(torch.isnan(attention), 0.0),
Expand Down Expand Up @@ -939,7 +940,7 @@ def plot_interpretation(self, interpretation: dict[str, torch.Tensor]):
# attention
fig, ax = plt.subplots()
attention = interpretation["attention"].detach().cpu()
attention = attention / attention.sum(-1).unsqueeze(-1)
attention = attention / attention.sum(-1).clamp(min=1e-8).unsqueeze(-1)
ax.plot(
np.arange(
-self.hparams.max_encoder_length,
Expand All @@ -956,7 +957,7 @@ def plot_interpretation(self, interpretation: dict[str, torch.Tensor]):
def make_selection_plot(title, values, labels):
fig, ax = plt.subplots(figsize=(7, len(values) * 0.25 + 2))
order = np.argsort(values)
values = values / values.sum(-1).unsqueeze(-1)
values = values / values.sum(-1).clamp(min=1e-8).unsqueeze(-1)
ax.barh(
np.arange(len(values)),
values[order] * 100,
Expand Down Expand Up @@ -1006,7 +1007,10 @@ def log_interpretation(self, outputs):
attention_occurrences = (
interpretation["encoder_length_histogram"][1:].flip(0).float().cumsum(0)
)
attention_occurrences = attention_occurrences / attention_occurrences.max()
attention_occurrences = (
attention_occurrences
/ attention_occurrences.max().clamp(min=1e-8)
)
attention_occurrences = torch.cat(
[
attention_occurrences,
Expand All @@ -1022,7 +1026,7 @@ def log_interpretation(self, outputs):
"attention"
] / attention_occurrences.pow(2).clamp(1.0)
interpretation["attention"] = (
interpretation["attention"] / interpretation["attention"].sum()
interpretation["attention"] / interpretation["attention"].sum().clamp(min=1e-8)
)

mpl_available = _check_matplotlib("log_interpretation", raise_error=False)
Expand Down
2 changes: 2 additions & 0 deletions tests/conftest.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,3 +1,5 @@
import matplotlib
matplotlib.use("Agg")
import os
import sys

Expand Down
Loading