AI/diffusers
1
0
Fork 0
mirror of https://github.com/huggingface/diffusers.git synced 2026-01-27 17:22:53 +03:00
Code Activity

Fix issue in maybe_convert_prompt (#3188)

Browse Source 
When the token used for textual inversion does not have any special symbols (e.g. it is not surrounded by <>), the tokenizer does not properly split the replacement tokens.  Adding a space for the padding tokens fixes this.
This commit is contained in:
pdoane
2023-04-25 06:17:57 -07:00
committed by Daniel Gu
parent 0431637f11
commit 9e2f445125
2 changed files with 3 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
src/diffusers/loaders.py
View File

@@ -410,7 +410,7 @@ class TextualInversionLoaderMixin:
replacement = token
i = 1
while f"{token}_{i}" in tokenizer.added_tokens_encoder:
replacement += f"{token}_{i}"
replacement += f" {token}_{i}"
i += 1
prompt = prompt.replace(token, replacement)

4
tests/pipelines/test_pipelines.py
View File

@@ -541,7 +541,7 @@ class DownloadTests(unittest.TestCase):
assert pipe.text_encoder.get_input_embeddings().weight[-3].sum().item() == 96
assert pipe.text_encoder.get_input_embeddings().weight[-2].sum().item() == 128
assert pipe.text_encoder.get_input_embeddings().weight[-1].sum().item() == 160
assert pipe._maybe_convert_prompt("<***>", pipe.tokenizer) == "<***><***>_1<***>_2"
assert pipe._maybe_convert_prompt("<***>", pipe.tokenizer) == "<***> <***>_1 <***>_2"
prompt = "hey <***>"
out = pipe(prompt, num_inference_steps=1, output_type="numpy").images
@@ -569,7 +569,7 @@ class DownloadTests(unittest.TestCase):
assert pipe.text_encoder.get_input_embeddings().weight[-3].sum().item() == 96
assert pipe.text_encoder.get_input_embeddings().weight[-2].sum().item() == 128
assert pipe.text_encoder.get_input_embeddings().weight[-1].sum().item() == 160
assert pipe._maybe_convert_prompt("<****>", pipe.tokenizer) == "<****><****>_1<****>_2"
assert pipe._maybe_convert_prompt("<****>", pipe.tokenizer) == "<****> <****>_1 <****>_2"
prompt = "hey <****>"
out = pipe(prompt, num_inference_steps=1, output_type="numpy").images