Skip to content
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

Fix Marian model conversion #30173

Merged
merged 4 commits into from
May 1, 2024
Merged

Fix Marian model conversion #30173

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
8d951e3
fix marian model coversion
zucchini-nlp Apr 10, 2024
8108eea
uncomment that line
zucchini-nlp Apr 11, 2024
4158798
remove unnecessary code
zucchini-nlp Apr 17, 2024
6535883
revert tie_weights, doesn't hurt
zucchini-nlp Apr 29, 2024
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
7 changes: 5 additions & 2 deletions src/transformers/models/marian/convert_marian_tatoeba_to_pytorch.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -34,7 +34,6 @@

DEFAULT_REPO = "Tatoeba-Challenge"
DEFAULT_MODEL_DIR = os.path.join(DEFAULT_REPO, "models")
LANG_CODE_URL = "https://datahub.io/core/language-codes/r/language-codes-3b2.csv"
Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

this one gives 404, so I just loaded dataset to the hub

ISO_URL = "https://cdn-datasets.huggingface.co/language_codes/iso-639-3.csv"
ISO_PATH = "lang_code_data/iso-639-3.csv"
LANG_CODE_PATH = "lang_code_data/language-codes-3b2.csv"
Expand Down Expand Up @@ -277,13 +276,17 @@ def write_model_card(self, model_dict, dry_run=False) -> str:
json.dump(metadata, writeobj)

def download_lang_info(self):
global LANG_CODE_PATH
Path(LANG_CODE_PATH).parent.mkdir(exist_ok=True)
import wget
from huggingface_hub import hf_hub_download

if not os.path.exists(ISO_PATH):
wget.download(ISO_URL, ISO_PATH)
if not os.path.exists(LANG_CODE_PATH):
wget.download(LANG_CODE_URL, LANG_CODE_PATH)
LANG_CODE_PATH = hf_hub_download(
repo_id="huggingface/language_codes_marianMT", filename="language-codes-3b2.csv", repo_type="dataset"
)

def parse_metadata(self, model_name, repo_path=DEFAULT_MODEL_DIR, method="best"):
p = Path(repo_path) / model_name
Expand Down
4 changes: 4 additions & 0 deletions src/transformers/models/marian/convert_marian_to_pytorch.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -622,6 +622,10 @@ def load_marian_model(self) -> MarianMTModel:
bias_tensor = nn.Parameter(torch.FloatTensor(self.final_bias))
model.model.decoder.embed_tokens.weight = decoder_wemb_tensor

# handle tied embeddings, otherwise "from_pretrained" loads them incorrectly
if self.cfg["tied-embeddings"]:
Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

and this is the actual fix, which gives equal weights for tied parameters

model.lm_head.weight.data = model.model.decoder.embed_tokens.weight.data.clone()

model.final_logits_bias = bias_tensor

if "Wpos" in state_dict:
Expand Down
Loading