Skip to content

Use tensors instead of numpy arrays in feed dicts #344

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

Merged
boomanaiden154 merged 6 commits into from
May 19, 2025
Merged

Use tensors instead of numpy arrays in feed dicts #344

Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
5b9dde5
[𝘀𝗽𝗿] changes to main this commit is based on
boomanaiden154 May 18, 2025
afc255b
[𝘀𝗽𝗿] initial version
boomanaiden154 May 18, 2025
d3c9561
[𝘀𝗽𝗿] changes introduced through rebase
boomanaiden154 May 19, 2025
af97746
rebase
boomanaiden154 May 19, 2025
f1b902f
[𝘀𝗽𝗿] changes introduced through rebase
boomanaiden154 May 19, 2025
4455e85
rebase
boomanaiden154 May 19, 2025
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
25 changes: 12 additions & 13 deletions gematria/model/python/model_base.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -60,7 +60,6 @@
INVALID_THROUGHPUT_VALUE = -1

_BASIC_BLOCK_INDEX_TF_DTYPE = tf.dtypes.int32
_BASIC_BLOCK_INDEX_NUMPY_DTYPE = _BASIC_BLOCK_INDEX_TF_DTYPE.as_numpy_dtype()

# A type variable that is either a basic block or block with throughput. The
# advantage over typing.Union is that in each context, the typevar represents
Expand Down Expand Up @@ -755,20 +754,20 @@ def _finalize_batch(self, include_expected_outputs: bool) -> FeedDict:
"""
schedule = self._make_batch_feed_dict()
if self._create_delta_block_index:
schedule['delta_block_index'] = np.array(
self._batch_delta_block_index, dtype=_BASIC_BLOCK_INDEX_NUMPY_DTYPE
schedule['delta_block_index'] = tf.constant(
self._batch_delta_block_index, dtype=_BASIC_BLOCK_INDEX_TF_DTYPE
)
if include_expected_outputs:
schedule['expected_outputs'] = np.reshape(
np.array(self._batch_expected_outputs, dtype=self.numpy_dtype),
schedule['expected_outputs'] = tf.reshape(
tf.constant(self._batch_expected_outputs, dtype=self.dtype),
[-1, self.num_tasks],
)
schedule['output_mask'] = np.array(self._batch_mask, dtype=bool)
schedule['output_mask'] = tf.constant(
self._batch_mask, dtype=tf.dtypes.bool
)
if self._use_deltas:
schedule['expected_outputs_deltas'] = np.reshape(
np.array(
self._batch_expected_outputs_deltas, dtype=self.numpy_dtype
),
schedule['expected_outputs_deltas'] = tf.reshape(
tf.constant(self._batch_expected_outputs_deltas, dtype=self.dtype),
[-1, self.num_tasks],
)

Expand Down Expand Up @@ -1304,16 +1303,16 @@ def _compute_loss(self, schedule: FeedDict) -> loss_utils.LossComputation:
output = self(schedule, train=True)
loss = loss_utils.LossComputation(
output['output'],
tf.constant(schedule['expected_outputs']),
tf.constant(schedule['output_mask']),
schedule['expected_outputs'],
schedule['output_mask'],
percentile_ranks=self._collected_percentile_ranks,
dtype=self.dtype,
)

if self._use_deltas:
self._delta_loss = loss_utils.LossComputation(
output['output_deltas'],
tf.constant(schedule['expected_outputs_deltas']),
schedule['expected_outputs_deltas'],
output['output_mask_deltas'],
percentile_ranks=self._collected_percentile_ranks,
dtype=self.dtype,
Expand Down
Loading