SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

set eos_token_id to None to generate until max length (#16989)

Browse Source 
Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>
This commit is contained in:
Yih-Dar
2022-04-28 19:47:38 +02:00
committed by GitHub
parent 01562dac7e
commit 5af5735f62
5 changed files with 24 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
tests/encoder_decoder/test_modeling_encoder_decoder.py
View File

@@ -413,7 +413,10 @@ class EncoderDecoderMixin:
enc_dec_model = EncoderDecoderModel(encoder=encoder_model, decoder=decoder_model) enc_dec_model = EncoderDecoderModel(encoder=encoder_model, decoder=decoder_model)
# Generate until max length # Generate until max length
enc_dec_model.config.decoder.eos_token_id = None if hasattr(enc_dec_model.config, "eos_token_id"):
enc_dec_model.config.eos_token_id = None
if hasattr(enc_dec_model.config, "decoder") and hasattr(enc_dec_model.config.decoder, "eos_token_id"):
enc_dec_model.config.decoder.eos_token_id = None
enc_dec_model.to(torch_device) enc_dec_model.to(torch_device)
# Bert does not have a bos token id, so use pad_token_id instead # Bert does not have a bos token id, so use pad_token_id instead

6
tests/encoder_decoder/test_modeling_tf_encoder_decoder.py
View File

@@ -314,6 +314,12 @@ class TFEncoderDecoderMixin:
encoder_model, decoder_model = self.get_encoder_decoder_model(config, decoder_config) encoder_model, decoder_model = self.get_encoder_decoder_model(config, decoder_config)
enc_dec_model = TFEncoderDecoderModel(encoder=encoder_model, decoder=decoder_model) enc_dec_model = TFEncoderDecoderModel(encoder=encoder_model, decoder=decoder_model)
# Generate until max length
if hasattr(enc_dec_model.config, "eos_token_id"):
enc_dec_model.config.eos_token_id = None
if hasattr(enc_dec_model.config, "decoder") and hasattr(enc_dec_model.config.decoder, "eos_token_id"):
enc_dec_model.config.decoder.eos_token_id = None
# Bert does not have a bos token id, so use pad_token_id instead # Bert does not have a bos token id, so use pad_token_id instead
generated_output = enc_dec_model.generate( generated_output = enc_dec_model.generate(
input_ids, decoder_start_token_id=enc_dec_model.config.decoder.pad_token_id input_ids, decoder_start_token_id=enc_dec_model.config.decoder.pad_token_id

3
tests/speech_encoder_decoder/test_modeling_speech_encoder_decoder.py
View File

@@ -347,7 +347,8 @@ class EncoderDecoderMixin:
enc_dec_model.to(torch_device) enc_dec_model.to(torch_device)
# make sure EOS token is set to None to prevent early stopping of generation # make sure EOS token is set to None to prevent early stopping of generation
enc_dec_model.config.eos_token_id = None if hasattr(enc_dec_model.config, "eos_token_id"):
enc_dec_model.config.eos_token_id = None
if hasattr(enc_dec_model.config, "decoder") and hasattr(enc_dec_model.config.decoder, "eos_token_id"): if hasattr(enc_dec_model.config, "decoder") and hasattr(enc_dec_model.config.decoder, "eos_token_id"):
enc_dec_model.config.decoder.eos_token_id = None enc_dec_model.config.decoder.eos_token_id = None

6
tests/vision_encoder_decoder/test_modeling_tf_vision_encoder_decoder.py
View File

@@ -300,6 +300,12 @@ class TFVisionEncoderDecoderMixin:
encoder_model, decoder_model = self.get_encoder_decoder_model(config, decoder_config) encoder_model, decoder_model = self.get_encoder_decoder_model(config, decoder_config)
enc_dec_model = TFVisionEncoderDecoderModel(encoder=encoder_model, decoder=decoder_model) enc_dec_model = TFVisionEncoderDecoderModel(encoder=encoder_model, decoder=decoder_model)
# Generate until max length
if hasattr(enc_dec_model.config, "eos_token_id"):
enc_dec_model.config.eos_token_id = None
if hasattr(enc_dec_model.config, "decoder") and hasattr(enc_dec_model.config.decoder, "eos_token_id"):
enc_dec_model.config.decoder.eos_token_id = None
# Bert does not have a bos token id, so use pad_token_id instead # Bert does not have a bos token id, so use pad_token_id instead
generated_output = enc_dec_model.generate( generated_output = enc_dec_model.generate(
pixel_values, decoder_start_token_id=enc_dec_model.config.decoder.pad_token_id pixel_values, decoder_start_token_id=enc_dec_model.config.decoder.pad_token_id

6
tests/vision_encoder_decoder/test_modeling_vision_encoder_decoder.py
View File

@@ -269,6 +269,12 @@ class EncoderDecoderMixin:
def check_encoder_decoder_model_generate(self, config, decoder_config, pixel_values=None, **kwargs): def check_encoder_decoder_model_generate(self, config, decoder_config, pixel_values=None, **kwargs):
encoder_model, decoder_model = self.get_encoder_decoder_model(config, decoder_config) encoder_model, decoder_model = self.get_encoder_decoder_model(config, decoder_config)
enc_dec_model = VisionEncoderDecoderModel(encoder=encoder_model, decoder=decoder_model) enc_dec_model = VisionEncoderDecoderModel(encoder=encoder_model, decoder=decoder_model)
# Generate until max length
if hasattr(enc_dec_model.config, "eos_token_id"):
enc_dec_model.config.eos_token_id = None
if hasattr(enc_dec_model.config, "decoder") and hasattr(enc_dec_model.config.decoder, "eos_token_id"):
enc_dec_model.config.decoder.eos_token_id = None
enc_dec_model.to(torch_device) enc_dec_model.to(torch_device)
inputs = pixel_values inputs = pixel_values