ALTI attribution method #217
Open
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
gsarti
commented
Aug 14, 2023
inseq/attr/feat/ops/alti.py
Outdated
Comment on lines
92
to
106
| # TODO: Implement decoder-only and encoder-decoder variants | ||
| # Resulting tensors (pre-rollout): | ||
| # `decoder_contributions` is a tensor of ALTI contributions for each token in the target sequence | ||
| # with shape (batch_size, n_layers, target_seq_len, target_seq_len) | ||
| decoder_contributions = torch.zeros_like(decoder_self_attentions[:, :, 0, ...]) | ||
| if self.forward_func.is_encoder_decoder: | ||
| # `encoder_contributions` is a tensor of ALTI contributions for each token in the source sequence | ||
| # with shape (batch_size, n_layers, source_seq_len, source_seq_len) | ||
| encoder_contributions = torch.zeros_like(encoder_self_attentions[:, :, 0, ...]) | ||
| # `cross_contributions` is a tensor of ALTI contributions of shape | ||
| # (batch_size, n_layers, target_seq_len, source_seq_len) | ||
| cross_contributions = torch.zeros_like(cross_attentions[:, :, 0, ...]) | ||
| else: | ||
| encoder_contributions = None | ||
| cross_contributions = None |
There was a problem hiding this comment.
Currently the ALTI method is not implemented, but tensors matching the desired output shape are provided in the method where they should be produced (i.e. running the method works, but produces 0s instead of actual scores).
The method can be ran as
import inseq
model = inseq.load_model("gpt2", "alti")
out = model.attribute("This is a test.")
out* origin/main: Attributed behavior for contrastive step functions (#228) Fix command for installing pre-commit hooks. (#229) Remove `max_input_length` from `model.encode` (#227) Migrate to `ruff format` (#225) Remove contrast_target_prefixes from contrastive step functions (#224) Step functions fixes, add `in_context_pvi` (#223) Format fixes, add Attanasio et al. (2023) to readme Add Sequential IG method (#222) Fix LIME and Occlusion outputs (#220) Update citation information Bump dependencies Add end_pos for contrast_targets_alignments Fix dummy output viz in console Minor fixes
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Description
This PR introduces the ALTI attribution method, following the notes detailed in #200. The implementation available at mt-upc/logit-explanations was used to set up a functioning template providing access to all internals needed to perform the ALTI computation (in/outs of every module, attentions and hidden states, plus a
configassociated to theAttributionModelfor accessing specific modules in the Transformer block). The core logic of the method is ininseq/attr/feat/ops/alti.py.TODOs:
get_logit_contributionsspecific to the production of ALTI contributions (pre-rollout, as this will be handled separately post-attribution via ad-hoc aggregator classes)cc @gegallego @javiferran