Skip to content

GitLab

Unverified Commit 814de8fa authored by Matt's avatar Matt Committed by GitHub
Browse files

Overhaul TF serving signatures + dummy inputs (#23234) 

* Let's try autodetecting serving sigs

* Don't clobber existing sigs

* Change shapes for multiplechoice models

* Make default dummy inputs smarter too

* Fix missing f-string

* Let's YOLO a serving output too

* Read __class__.__name__ properly

* Don't just pass naked lists in there and expect it to be okay

* Code cleanup

* Update default serving sig

* Clearer error messages

* Further updates to the default serving output

* make fixup

* Update the serving output a bit more

* Cleanups and renames, raise errors appropriately when we can't infer inputs

* More renames

* we're building in a functional context again, yolo

* import DUMMY_INPUTS from the right place

* import DUMMY_INPUTS from the right place

* Support cross-attention in the dummies

* Support cross-attention in the dummies

* Complete removal of dummy/serving overrides in BERT

* Complete removal of dummy/serving overrides in RoBERTa

* Obliterate lots and lots of serving sig and dummy overrides

* merge type hint changes

* Fix for token_type_ids with vocab_size 1

* Add missing property decorator

* Fix T5 and hopefully some models that take conv inputs

* More signature pruning

* Fix T5's signature

* Fix Wav2Vec2 signature

* Fix LongformerForMultipleChoice input signature

* Fix BLIP and LED

* Better default serving output error handling

* Fix BART dummies

* Fix dummies for cross-attention, esp encoder-decoder models

* Fix visionencoderdecoder signature

* Fix BLIP serving output

* Small tweak to BART dummies

* Cleanup the ugly parameter inspection line that I used in a few places

* committed a breakpoint again

* Move the text_dims check

* Remove blip_text serving_output

* Add decoder_input_ids to the default input sig

* Remove all the manual overrides for encoder-decoder model signatures

* Tweak longformer/led input sigs

* Tweak default serving output

* output.keys() -> output

* make fixup
parent 3d7baef1
Hide whitespace changes
Inline Side-by-side
Showing with 71 additions and 1297 deletions
src/transformers/models/electra/modeling_tf_electra.py
View file @ 814de8fa
...@@ -20,7 +20,7 @@ from __future__ import annotations ...@@ -20,7 +20,7 @@ from __future__ import annotations
import math import math
import warnings import warnings
from dataclasses import dataclass from dataclasses import dataclass
from typing import Dict, Optional, Tuple, Union from typing import Optional, Tuple, Union
import numpy as np import numpy as np
import tensorflow as tf import tensorflow as tf
...@@ -49,8 +49,6 @@ from ...modeling_tf_utils import ( ...@@ -49,8 +49,6 @@ from ...modeling_tf_utils import (
) )
from ...tf_utils import check_embeddings_within_bounds, shape_list, stable_softmax from ...tf_utils import check_embeddings_within_bounds, shape_list, stable_softmax
from ...utils import ( from ...utils import (
DUMMY_INPUTS,
MULTIPLE_CHOICE_DUMMY_INPUTS,
ModelOutput, ModelOutput,
add_code_sample_docstrings, add_code_sample_docstrings,
add_start_docstrings, add_start_docstrings,
...@@ -596,25 +594,6 @@ class TFElectraPreTrainedModel(TFPreTrainedModel): ...@@ -596,25 +594,6 @@ class TFElectraPreTrainedModel(TFPreTrainedModel):
_keys_to_ignore_on_load_unexpected = [r"generator_lm_head.weight"] _keys_to_ignore_on_load_unexpected = [r"generator_lm_head.weight"]
_keys_to_ignore_on_load_missing = [r"dropout"] _keys_to_ignore_on_load_missing = [r"dropout"]
@property
# Copied from transformers.models.bert.modeling_tf_bert.TFBertPreTrainedModel.dummy_inputs
def dummy_inputs(self):
"""
Dummy inputs to build the network.
Returns:
`Dict[str, tf.Tensor]`: The dummy inputs.
"""
dummy = {"input_ids": tf.constant(DUMMY_INPUTS, dtype=tf.int32)}
# Add `encoder_hidden_states` to make the cross-attention layers' weights initialized
if self.config.add_cross_attention:
batch_size, seq_len = tf.constant(DUMMY_INPUTS).shape
shape = (batch_size, seq_len) + (self.config.hidden_size,)
h = tf.random.uniform(shape=shape)
dummy["encoder_hidden_states"] = h
return dummy
@keras_serializable @keras_serializable
class TFElectraMainLayer(tf.keras.layers.Layer): class TFElectraMainLayer(tf.keras.layers.Layer):
...@@ -998,23 +977,6 @@ class TFElectraModel(TFElectraPreTrainedModel): ...@@ -998,23 +977,6 @@ class TFElectraModel(TFElectraPreTrainedModel):
return outputs return outputs
def serving_output(self, output):
output_cache = self.config.use_cache and self.config.is_decoder
pkv = tf.convert_to_tensor(output.past_key_values) if output_cache else None
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
cross_attns = tf.convert_to_tensor(output.cross_attentions) if output.cross_attentions is not None else None
if not (self.config.output_attentions and self.config.add_cross_attention):
cross_attns = None
return TFBaseModelOutputWithPastAndCrossAttentions(
last_hidden_state=output.last_hidden_state,
past_key_values=pkv,
hidden_states=hs,
attentions=attns,
cross_attentions=cross_attns,
)
@add_start_docstrings( @add_start_docstrings(
""" """
...@@ -1087,12 +1049,6 @@ class TFElectraForPreTraining(TFElectraPreTrainedModel): ...@@ -1087,12 +1049,6 @@ class TFElectraForPreTraining(TFElectraPreTrainedModel):
attentions=discriminator_hidden_states.attentions, attentions=discriminator_hidden_states.attentions,
) )
def serving_output(self, output):
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFElectraForPreTrainingOutput(logits=output.logits, hidden_states=hs, attentions=attns)
class TFElectraMaskedLMHead(tf.keras.layers.Layer): class TFElectraMaskedLMHead(tf.keras.layers.Layer):
def __init__(self, config, input_embeddings, **kwargs): def __init__(self, config, input_embeddings, **kwargs):
...@@ -1221,13 +1177,6 @@ class TFElectraForMaskedLM(TFElectraPreTrainedModel, TFMaskedLanguageModelingLos ...@@ -1221,13 +1177,6 @@ class TFElectraForMaskedLM(TFElectraPreTrainedModel, TFMaskedLanguageModelingLos
attentions=generator_hidden_states.attentions, attentions=generator_hidden_states.attentions,
) )
# Copied from transformers.models.bert.modeling_tf_bert.TFBertForMaskedLM.serving_output
def serving_output(self, output: TFMaskedLMOutput) -> TFMaskedLMOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFMaskedLMOutput(logits=output.logits, hidden_states=hs, attentions=attns)
class TFElectraClassificationHead(tf.keras.layers.Layer): class TFElectraClassificationHead(tf.keras.layers.Layer):
"""Head for sentence-level classification tasks.""" """Head for sentence-level classification tasks."""
...@@ -1329,13 +1278,6 @@ class TFElectraForSequenceClassification(TFElectraPreTrainedModel, TFSequenceCla ...@@ -1329,13 +1278,6 @@ class TFElectraForSequenceClassification(TFElectraPreTrainedModel, TFSequenceCla
attentions=outputs.attentions, attentions=outputs.attentions,
) )
# Copied from transformers.models.bert.modeling_tf_bert.TFBertForSequenceClassification.serving_output
def serving_output(self, output: TFSequenceClassifierOutput) -> TFSequenceClassifierOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFSequenceClassifierOutput(logits=output.logits, hidden_states=hs, attentions=attns)
@add_start_docstrings( @add_start_docstrings(
""" """
...@@ -1356,16 +1298,6 @@ class TFElectraForMultipleChoice(TFElectraPreTrainedModel, TFMultipleChoiceLoss) ...@@ -1356,16 +1298,6 @@ class TFElectraForMultipleChoice(TFElectraPreTrainedModel, TFMultipleChoiceLoss)
1, kernel_initializer=get_initializer(config.initializer_range), name="classifier" 1, kernel_initializer=get_initializer(config.initializer_range), name="classifier"
) )
@property
def dummy_inputs(self):
"""
Dummy inputs to build the network.
Returns:
tf.Tensor with dummy inputs
"""
return {"input_ids": tf.constant(MULTIPLE_CHOICE_DUMMY_INPUTS, dtype=tf.int32)}
@unpack_inputs @unpack_inputs
@add_start_docstrings_to_model_forward(ELECTRA_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length")) @add_start_docstrings_to_model_forward(ELECTRA_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length"))
@add_code_sample_docstrings( @add_code_sample_docstrings(
...@@ -1438,28 +1370,6 @@ class TFElectraForMultipleChoice(TFElectraPreTrainedModel, TFMultipleChoiceLoss) ...@@ -1438,28 +1370,6 @@ class TFElectraForMultipleChoice(TFElectraPreTrainedModel, TFMultipleChoiceLoss)
attentions=outputs.attentions, attentions=outputs.attentions,
) )
@tf.function(
input_signature=[
{
"input_ids": tf.TensorSpec((None, None, None), tf.int32, name="input_ids"),
"attention_mask": tf.TensorSpec((None, None, None), tf.int32, name="attention_mask"),
"token_type_ids": tf.TensorSpec((None, None, None), tf.int32, name="token_type_ids"),
}
]
)
# Copied from transformers.models.bert.modeling_tf_bert.TFBertForMultipleChoice.serving
def serving(self, inputs: Dict[str, tf.Tensor]):
output = self.call(input_ids=inputs)
return self.serving_output(output)
# Copied from transformers.models.bert.modeling_tf_bert.TFBertForMultipleChoice.serving_output
def serving_output(self, output: TFMultipleChoiceModelOutput) -> TFMultipleChoiceModelOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFMultipleChoiceModelOutput(logits=output.logits, hidden_states=hs, attentions=attns)
@add_start_docstrings( @add_start_docstrings(
""" """
...@@ -1538,13 +1448,6 @@ class TFElectraForTokenClassification(TFElectraPreTrainedModel, TFTokenClassific ...@@ -1538,13 +1448,6 @@ class TFElectraForTokenClassification(TFElectraPreTrainedModel, TFTokenClassific
attentions=discriminator_hidden_states.attentions, attentions=discriminator_hidden_states.attentions,
) )
# Copied from transformers.models.bert.modeling_tf_bert.TFBertForTokenClassification.serving_output
def serving_output(self, output: TFTokenClassifierOutput) -> TFTokenClassifierOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFTokenClassifierOutput(logits=output.logits, hidden_states=hs, attentions=attns)
@add_start_docstrings( @add_start_docstrings(
""" """
...@@ -1638,12 +1541,3 @@ class TFElectraForQuestionAnswering(TFElectraPreTrainedModel, TFQuestionAnswerin ...@@ -1638,12 +1541,3 @@ class TFElectraForQuestionAnswering(TFElectraPreTrainedModel, TFQuestionAnswerin
hidden_states=discriminator_hidden_states.hidden_states, hidden_states=discriminator_hidden_states.hidden_states,
attentions=discriminator_hidden_states.attentions, attentions=discriminator_hidden_states.attentions,
) )
# Copied from transformers.models.bert.modeling_tf_bert.TFBertForQuestionAnswering.serving_output
def serving_output(self, output: TFQuestionAnsweringModelOutput) -> TFQuestionAnsweringModelOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFQuestionAnsweringModelOutput(
start_logits=output.start_logits, end_logits=output.end_logits, hidden_states=hs, attentions=attns
)
src/transformers/models/encoder_decoder/modeling_tf_encoder_decoder.py
View file @ 814de8fa
...@@ -36,7 +36,6 @@ from ...modeling_tf_utils import ( ...@@ -36,7 +36,6 @@ from ...modeling_tf_utils import (
) )
from ...tf_utils import shape_list from ...tf_utils import shape_list
from ...utils import ( from ...utils import (
DUMMY_INPUTS,
ModelOutput, ModelOutput,
add_start_docstrings, add_start_docstrings,
add_start_docstrings_to_model_forward, add_start_docstrings_to_model_forward,
...@@ -276,19 +275,6 @@ class TFEncoderDecoderModel(TFPreTrainedModel, TFCausalLanguageModelingLoss): ...@@ -276,19 +275,6 @@ class TFEncoderDecoderModel(TFPreTrainedModel, TFCausalLanguageModelingLoss):
"following discussion on GitHub: https://github.com/huggingface/transformers/issues/23350" "following discussion on GitHub: https://github.com/huggingface/transformers/issues/23350"
) )
@property
def dummy_inputs(self):
"""
Dummy inputs to build the network.
Returns:
`Dict[str, tf.Tensor]`: The dummy inputs.
"""
# Add `decoder_input_ids` because `self.decoder` requires it.
input_ids = tf.constant(DUMMY_INPUTS, dtype=tf.int32)
dummy = {"input_ids": input_ids, "decoder_input_ids": input_ids}
return dummy
def get_encoder(self): def get_encoder(self):
return self.encoder return self.encoder
...@@ -642,33 +628,6 @@ class TFEncoderDecoderModel(TFPreTrainedModel, TFCausalLanguageModelingLoss): ...@@ -642,33 +628,6 @@ class TFEncoderDecoderModel(TFPreTrainedModel, TFCausalLanguageModelingLoss):
encoder_attentions=encoder_outputs.attentions, encoder_attentions=encoder_outputs.attentions,
) )
def serving_output(self, output):
pkv = tf.tuple(output.past_key_values)[1] if self.config.decoder.use_cache else None
dec_hs = (
tf.convert_to_tensor(output.decoder_hidden_states) if self.config.decoder.output_hidden_states else None
)
dec_attns = tf.convert_to_tensor(output.decoder_attentions) if self.config.decoder.output_attentions else None
enc_hs = (
tf.convert_to_tensor(output.encoder_hidden_states) if self.config.encoder.output_hidden_states else None
)
enc_attns = tf.convert_to_tensor(output.encoder_attentions) if self.config.encoder.output_attentions else None
cross_attns = (
tf.convert_to_tensor(output.cross_attentions)
if self.config.decoder.output_attentions and output.cross_attentions is not None
else None
)
return TFSeq2SeqLMOutput(
logits=output.logits,
past_key_values=pkv,
decoder_hidden_states=dec_hs,
decoder_attentions=dec_attns,
encoder_last_hidden_state=output.encoder_last_hidden_state,
encoder_hidden_states=enc_hs,
encoder_attentions=enc_attns,
cross_attentions=cross_attns,
)
def prepare_inputs_for_generation( def prepare_inputs_for_generation(
self, input_ids, past_key_values=None, attention_mask=None, use_cache=None, encoder_outputs=None, **kwargs self, input_ids, past_key_values=None, attention_mask=None, use_cache=None, encoder_outputs=None, **kwargs
): ):
......
src/transformers/models/esm/modeling_tf_esm.py
View file @ 814de8fa
...@@ -1038,39 +1038,6 @@ class TFEsmModel(TFEsmPreTrainedModel): ...@@ -1038,39 +1038,6 @@ class TFEsmModel(TFEsmPreTrainedModel):
) )
return outputs return outputs
@tf.function(
input_signature=[
{
"input_ids": tf.TensorSpec((None, None), tf.int32, name="input_ids"),
"attention_mask": tf.TensorSpec((None, None), tf.int32, name="attention_mask"),
}
]
)
def serving(self, inputs):
output = self.call(inputs)
return self.serving_output(output)
def serving_output(
self, output: TFBaseModelOutputWithPoolingAndCrossAttentions
) -> TFBaseModelOutputWithPoolingAndCrossAttentions:
output_cache = self.config.use_cache and self.config.is_decoder
pkv = tf.convert_to_tensor(output.past_key_values) if output_cache else None
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
cross_attns = tf.convert_to_tensor(output.cross_attentions) if output.cross_attentions is not None else None
if not (self.config.output_attentions and self.config.add_cross_attention):
cross_attns = None
return TFBaseModelOutputWithPoolingAndCrossAttentions(
last_hidden_state=output.last_hidden_state,
pooler_output=output.pooler_output,
past_key_values=pkv,
hidden_states=hs,
attentions=attns,
cross_attentions=cross_attns,
)
def predict_contacts(self, tokens, attention_mask): def predict_contacts(self, tokens, attention_mask):
return self.esm.predict_contacts(tokens, attention_mask) return self.esm.predict_contacts(tokens, attention_mask)
...@@ -1170,26 +1137,6 @@ class TFEsmForMaskedLM(TFEsmPreTrainedModel, TFMaskedLanguageModelingLoss): ...@@ -1170,26 +1137,6 @@ class TFEsmForMaskedLM(TFEsmPreTrainedModel, TFMaskedLanguageModelingLoss):
attentions=outputs.attentions, attentions=outputs.attentions,
) )
# Copied from transformers.models.bert.modeling_tf_bert.TFBertForMaskedLM.serving_output
def serving_output(self, output: TFMaskedLMOutput) -> TFMaskedLMOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFMaskedLMOutput(logits=output.logits, hidden_states=hs, attentions=attns)
@tf.function(
input_signature=[
{
"input_ids": tf.TensorSpec((None, None), tf.int32, name="input_ids"),
"attention_mask": tf.TensorSpec((None, None), tf.int32, name="attention_mask"),
}
]
)
def serving(self, inputs):
output = self.call(inputs)
return self.serving_output(output)
def predict_contacts(self, tokens, attention_mask): def predict_contacts(self, tokens, attention_mask):
return self.esm.predict_contacts(tokens, attention_mask) return self.esm.predict_contacts(tokens, attention_mask)
...@@ -1310,26 +1257,6 @@ class TFEsmForSequenceClassification(TFEsmPreTrainedModel, TFSequenceClassificat ...@@ -1310,26 +1257,6 @@ class TFEsmForSequenceClassification(TFEsmPreTrainedModel, TFSequenceClassificat
attentions=outputs.attentions, attentions=outputs.attentions,
) )
# Copied from transformers.models.bert.modeling_tf_bert.TFBertForSequenceClassification.serving_output
def serving_output(self, output: TFSequenceClassifierOutput) -> TFSequenceClassifierOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFSequenceClassifierOutput(logits=output.logits, hidden_states=hs, attentions=attns)
@tf.function(
input_signature=[
{
"input_ids": tf.TensorSpec((None, None, None), tf.int32, name="input_ids"),
"attention_mask": tf.TensorSpec((None, None, None), tf.int32, name="attention_mask"),
}
]
)
def serving(self, inputs):
output = self.call(inputs)
return self.serving_output(output)
@add_start_docstrings( @add_start_docstrings(
""" """
...@@ -1406,26 +1333,6 @@ class TFEsmForTokenClassification(TFEsmPreTrainedModel, TFTokenClassificationLos ...@@ -1406,26 +1333,6 @@ class TFEsmForTokenClassification(TFEsmPreTrainedModel, TFTokenClassificationLos
attentions=outputs.attentions, attentions=outputs.attentions,
) )
# Copied from transformers.models.bert.modeling_tf_bert.TFBertForTokenClassification.serving_output
def serving_output(self, output: TFTokenClassifierOutput) -> TFTokenClassifierOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFTokenClassifierOutput(logits=output.logits, hidden_states=hs, attentions=attns)
@tf.function(
input_signature=[
{
"input_ids": tf.TensorSpec((None, None), tf.int32, name="input_ids"),
"attention_mask": tf.TensorSpec((None, None), tf.int32, name="attention_mask"),
}
]
)
def serving(self, inputs):
output = self.call(inputs)
return self.serving_output(output)
class TFEsmClassificationHead(Layer): class TFEsmClassificationHead(Layer):
"""Head for sentence-level classification tasks.""" """Head for sentence-level classification tasks."""
......
src/transformers/models/flaubert/modeling_tf_flaubert.py
View file @ 814de8fa
...@@ -290,13 +290,6 @@ class TFFlaubertModel(TFFlaubertPreTrainedModel): ...@@ -290,13 +290,6 @@ class TFFlaubertModel(TFFlaubertPreTrainedModel):
return outputs return outputs
# Copied from transformers.models.distilbert.modeling_tf_distilbert.TFDistilBertModel.serving_output
def serving_output(self, output):
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFBaseModelOutput(last_hidden_state=output.last_hidden_state, hidden_states=hs, attentions=attns)
# Copied from transformers.models.xlm.modeling_tf_xlm.TFXLMMultiHeadAttention with XLM->Flaubert # Copied from transformers.models.xlm.modeling_tf_xlm.TFXLMMultiHeadAttention with XLM->Flaubert
class TFFlaubertMultiHeadAttention(tf.keras.layers.Layer): class TFFlaubertMultiHeadAttention(tf.keras.layers.Layer):
...@@ -845,12 +838,6 @@ class TFFlaubertWithLMHeadModel(TFFlaubertPreTrainedModel): ...@@ -845,12 +838,6 @@ class TFFlaubertWithLMHeadModel(TFFlaubertPreTrainedModel):
logits=outputs, hidden_states=transformer_outputs.hidden_states, attentions=transformer_outputs.attentions logits=outputs, hidden_states=transformer_outputs.hidden_states, attentions=transformer_outputs.attentions
) )
def serving_output(self, output):
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFFlaubertWithLMHeadModelOutput(logits=output.logits, hidden_states=hs, attentions=attns)
@add_start_docstrings( @add_start_docstrings(
""" """
...@@ -930,13 +917,6 @@ class TFFlaubertForSequenceClassification(TFFlaubertPreTrainedModel, TFSequenceC ...@@ -930,13 +917,6 @@ class TFFlaubertForSequenceClassification(TFFlaubertPreTrainedModel, TFSequenceC
attentions=transformer_outputs.attentions, attentions=transformer_outputs.attentions,
) )
# Copied from transformers.models.bert.modeling_tf_bert.TFBertForSequenceClassification.serving_output
def serving_output(self, output: TFSequenceClassifierOutput) -> TFSequenceClassifierOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFSequenceClassifierOutput(logits=output.logits, hidden_states=hs, attentions=attns)
@add_start_docstrings( @add_start_docstrings(
""" """
...@@ -1029,15 +1009,6 @@ class TFFlaubertForQuestionAnsweringSimple(TFFlaubertPreTrainedModel, TFQuestion ...@@ -1029,15 +1009,6 @@ class TFFlaubertForQuestionAnsweringSimple(TFFlaubertPreTrainedModel, TFQuestion
attentions=transformer_outputs.attentions, attentions=transformer_outputs.attentions,
) )
# Copied from transformers.models.bert.modeling_tf_bert.TFBertForQuestionAnswering.serving_output
def serving_output(self, output: TFQuestionAnsweringModelOutput) -> TFQuestionAnsweringModelOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFQuestionAnsweringModelOutput(
start_logits=output.start_logits, end_logits=output.end_logits, hidden_states=hs, attentions=attns
)
@add_start_docstrings( @add_start_docstrings(
""" """
...@@ -1119,13 +1090,6 @@ class TFFlaubertForTokenClassification(TFFlaubertPreTrainedModel, TFTokenClassif ...@@ -1119,13 +1090,6 @@ class TFFlaubertForTokenClassification(TFFlaubertPreTrainedModel, TFTokenClassif
attentions=transformer_outputs.attentions, attentions=transformer_outputs.attentions,
) )
# Copied from transformers.models.bert.modeling_tf_bert.TFBertForTokenClassification.serving_output
def serving_output(self, output: TFTokenClassifierOutput) -> TFTokenClassifierOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFTokenClassifierOutput(logits=output.logits, hidden_states=hs, attentions=attns)
@add_start_docstrings( @add_start_docstrings(
""" """
...@@ -1247,25 +1211,3 @@ class TFFlaubertForMultipleChoice(TFFlaubertPreTrainedModel, TFMultipleChoiceLos ...@@ -1247,25 +1211,3 @@ class TFFlaubertForMultipleChoice(TFFlaubertPreTrainedModel, TFMultipleChoiceLos
hidden_states=transformer_outputs.hidden_states, hidden_states=transformer_outputs.hidden_states,
attentions=transformer_outputs.attentions, attentions=transformer_outputs.attentions,
) )
@tf.function(
input_signature=[
{
"input_ids": tf.TensorSpec((None, None, None), tf.int32, name="input_ids"),
"attention_mask": tf.TensorSpec((None, None, None), tf.int32, name="attention_mask"),
"token_type_ids": tf.TensorSpec((None, None, None), tf.int32, name="token_type_ids"),
}
]
)
# Copied from transformers.models.bert.modeling_tf_bert.TFBertForMultipleChoice.serving
def serving(self, inputs: Dict[str, tf.Tensor]):
output = self.call(input_ids=inputs)
return self.serving_output(output)
# Copied from transformers.models.bert.modeling_tf_bert.TFBertForMultipleChoice.serving_output
def serving_output(self, output: TFMultipleChoiceModelOutput) -> TFMultipleChoiceModelOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFMultipleChoiceModelOutput(logits=output.logits, hidden_states=hs, attentions=attns)
src/transformers/models/funnel/modeling_tf_funnel.py
View file @ 814de8fa
...@@ -19,7 +19,7 @@ from __future__ import annotations ...@@ -19,7 +19,7 @@ from __future__ import annotations
import warnings import warnings
from dataclasses import dataclass from dataclasses import dataclass
from typing import Dict, Optional, Tuple, Union from typing import Optional, Tuple, Union
import numpy as np import numpy as np
import tensorflow as tf import tensorflow as tf
...@@ -47,7 +47,6 @@ from ...modeling_tf_utils import ( ...@@ -47,7 +47,6 @@ from ...modeling_tf_utils import (
) )
from ...tf_utils import check_embeddings_within_bounds, shape_list, stable_softmax from ...tf_utils import check_embeddings_within_bounds, shape_list, stable_softmax
from ...utils import ( from ...utils import (
MULTIPLE_CHOICE_DUMMY_INPUTS,
ModelOutput, ModelOutput,
add_code_sample_docstrings, add_code_sample_docstrings,
add_start_docstrings, add_start_docstrings,
...@@ -1425,16 +1424,6 @@ class TFFunnelForMultipleChoice(TFFunnelPreTrainedModel, TFMultipleChoiceLoss): ...@@ -1425,16 +1424,6 @@ class TFFunnelForMultipleChoice(TFFunnelPreTrainedModel, TFMultipleChoiceLoss):
self.funnel = TFFunnelBaseLayer(config, name="funnel") self.funnel = TFFunnelBaseLayer(config, name="funnel")
self.classifier = TFFunnelClassificationHead(config, 1, name="classifier") self.classifier = TFFunnelClassificationHead(config, 1, name="classifier")
@property
def dummy_inputs(self):
"""
Dummy inputs to build the network.
Returns:
tf.Tensor with dummy inputs
"""
return {"input_ids": tf.constant(MULTIPLE_CHOICE_DUMMY_INPUTS, dtype=tf.int32)}
@unpack_inputs @unpack_inputs
@add_start_docstrings_to_model_forward(FUNNEL_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length")) @add_start_docstrings_to_model_forward(FUNNEL_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length"))
@add_code_sample_docstrings( @add_code_sample_docstrings(
...@@ -1504,20 +1493,6 @@ class TFFunnelForMultipleChoice(TFFunnelPreTrainedModel, TFMultipleChoiceLoss): ...@@ -1504,20 +1493,6 @@ class TFFunnelForMultipleChoice(TFFunnelPreTrainedModel, TFMultipleChoiceLoss):
attentions=outputs.attentions, attentions=outputs.attentions,
) )
@tf.function(
input_signature=[
{
"input_ids": tf.TensorSpec((None, None), tf.int32, name="input_ids"),
"attention_mask": tf.TensorSpec((None, None), tf.float32, name="attention_mask"),
"token_type_ids": tf.TensorSpec((None, None), tf.int32, name="token_type_ids"),
}
]
)
def serving(self, inputs: Dict[str, tf.Tensor]) -> TFMultipleChoiceModelOutput:
output = self.call(input_ids=inputs)
return self.serving_output(output=output)
def serving_output(self, output: TFMultipleChoiceModelOutput) -> TFMultipleChoiceModelOutput: def serving_output(self, output: TFMultipleChoiceModelOutput) -> TFMultipleChoiceModelOutput:
# hidden_states and attentions not converted to Tensor with tf.convert_to_tensor as they are all of # hidden_states and attentions not converted to Tensor with tf.convert_to_tensor as they are all of
# different dimensions # different dimensions
......
src/transformers/models/gpt2/modeling_tf_gpt2.py
View file @ 814de8fa
...@@ -42,7 +42,6 @@ from ...modeling_tf_utils import ( ...@@ -42,7 +42,6 @@ from ...modeling_tf_utils import (
) )
from ...tf_utils import check_embeddings_within_bounds, shape_list, stable_softmax from ...tf_utils import check_embeddings_within_bounds, shape_list, stable_softmax
from ...utils import ( from ...utils import (
DUMMY_INPUTS,
ModelOutput, ModelOutput,
add_code_sample_docstrings, add_code_sample_docstrings,
add_start_docstrings, add_start_docstrings,
...@@ -522,37 +521,6 @@ class TFGPT2PreTrainedModel(TFPreTrainedModel): ...@@ -522,37 +521,6 @@ class TFGPT2PreTrainedModel(TFPreTrainedModel):
# names with a '.' represents the authorized unexpected/missing layers when a TF model is loaded from a PT model # names with a '.' represents the authorized unexpected/missing layers when a TF model is loaded from a PT model
_keys_to_ignore_on_load_unexpected = [r"h.\d+.attn.bias", r"h.\d+.crossattention.bias"] _keys_to_ignore_on_load_unexpected = [r"h.\d+.attn.bias", r"h.\d+.crossattention.bias"]
@property
def dummy_inputs(self):
"""
Dummy inputs to build the network.
Returns:
`Dict[str, tf.Tensor]`: The dummy inputs.
"""
dummy = {"input_ids": tf.constant(DUMMY_INPUTS, dtype=tf.int32)}
# Add `encoder_hidden_states` to make the cross-attention layers' weights initialized
if self.config.add_cross_attention:
batch_size, seq_len = tf.constant(DUMMY_INPUTS).shape
shape = (batch_size, seq_len) + (self.config.hidden_size,)
h = tf.random.uniform(shape=shape)
dummy["encoder_hidden_states"] = h
return dummy
@tf.function(
input_signature=[
{
"input_ids": tf.TensorSpec((None, None), tf.int32, name="input_ids"),
"attention_mask": tf.TensorSpec((None, None), tf.int32, name="attention_mask"),
}
]
)
def serving(self, inputs):
output = self.call(inputs)
return self.serving_output(output)
@dataclass @dataclass
class TFGPT2DoubleHeadsModelOutput(ModelOutput): class TFGPT2DoubleHeadsModelOutput(ModelOutput):
...@@ -773,26 +741,6 @@ class TFGPT2Model(TFGPT2PreTrainedModel): ...@@ -773,26 +741,6 @@ class TFGPT2Model(TFGPT2PreTrainedModel):
return outputs return outputs
def serving_output(self, output):
pkv = tf.convert_to_tensor(output.past_key_values) if self.config.use_cache else None
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
cross_attns = (
tf.convert_to_tensor(output.cross_attentions)
if self.config.output_attentions
and self.config.add_cross_attention
and output.cross_attentions is not None
else None
)
return TFBaseModelOutputWithPastAndCrossAttentions(
last_hidden_state=output.last_hidden_state,
past_key_values=pkv,
hidden_states=hs,
attentions=attns,
cross_attentions=cross_attns,
)
@add_start_docstrings( @add_start_docstrings(
""" """
...@@ -925,22 +873,6 @@ class TFGPT2LMHeadModel(TFGPT2PreTrainedModel, TFCausalLanguageModelingLoss): ...@@ -925,22 +873,6 @@ class TFGPT2LMHeadModel(TFGPT2PreTrainedModel, TFCausalLanguageModelingLoss):
cross_attentions=transformer_outputs.cross_attentions, cross_attentions=transformer_outputs.cross_attentions,
) )
def serving_output(self, output):
pkv = tf.convert_to_tensor(output.past_key_values) if self.config.use_cache else None
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
cross_attns = (
tf.convert_to_tensor(output.cross_attentions)
if self.config.output_attentions
and self.config.add_cross_attention
and output.cross_attentions is not None
else None
)
return TFCausalLMOutputWithCrossAttentions(
logits=output.logits, past_key_values=pkv, hidden_states=hs, attentions=attns, cross_attentions=cross_attns
)
@add_start_docstrings( @add_start_docstrings(
""" """
...@@ -1062,32 +994,13 @@ class TFGPT2DoubleHeadsModel(TFGPT2PreTrainedModel): ...@@ -1062,32 +994,13 @@ class TFGPT2DoubleHeadsModel(TFGPT2PreTrainedModel):
attentions=transformer_outputs.attentions, attentions=transformer_outputs.attentions,
) )
@tf.function( @property
input_signature=[ def input_signature(self):
{ return {
"input_ids": tf.TensorSpec((None, None, None), tf.int32, name="input_ids"), "input_ids": tf.TensorSpec((None, None, None), tf.int32, name="input_ids"),
"attention_mask": tf.TensorSpec((None, None, None), tf.int32, name="attention_mask"), "attention_mask": tf.TensorSpec((None, None, None), tf.int32, name="attention_mask"),
"mc_token_ids": tf.TensorSpec((None, None), tf.int32, name="mc_token_ids"), "mc_token_ids": tf.TensorSpec((None, None), tf.int32, name="mc_token_ids"),
} }
]
)
def serving(self, inputs):
output = self.call(inputs)
return self.serving_output(output)
def serving_output(self, output):
pkv = tf.convert_to_tensor(output.past_key_values) if self.config.use_cache else None
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFGPT2DoubleHeadsModelOutput(
logits=output.logits,
mc_logits=output.mc_logits,
past_key_values=pkv,
hidden_states=hs,
attentions=attns,
)
@add_start_docstrings( @add_start_docstrings(
...@@ -1210,12 +1123,3 @@ class TFGPT2ForSequenceClassification(TFGPT2PreTrainedModel, TFSequenceClassific ...@@ -1210,12 +1123,3 @@ class TFGPT2ForSequenceClassification(TFGPT2PreTrainedModel, TFSequenceClassific
hidden_states=transformer_outputs.hidden_states, hidden_states=transformer_outputs.hidden_states,
attentions=transformer_outputs.attentions, attentions=transformer_outputs.attentions,
) )
def serving_output(self, output):
pkv = tf.convert_to_tensor(output.past_key_values) if self.config.use_cache else None
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFSequenceClassifierOutputWithPast(
logits=output.logits, past_key_values=pkv, hidden_states=hs, attentions=attns
)
src/transformers/models/gptj/modeling_tf_gptj.py
View file @ 814de8fa
...@@ -23,7 +23,6 @@ import tensorflow as tf ...@@ -23,7 +23,6 @@ import tensorflow as tf
from ...activations_tf import get_tf_activation from ...activations_tf import get_tf_activation
from ...file_utils import ( from ...file_utils import (
DUMMY_INPUTS,
add_code_sample_docstrings, add_code_sample_docstrings,
add_start_docstrings, add_start_docstrings,
add_start_docstrings_to_model_forward, add_start_docstrings_to_model_forward,
...@@ -513,30 +512,6 @@ class TFGPTJPreTrainedModel(TFPreTrainedModel): ...@@ -513,30 +512,6 @@ class TFGPTJPreTrainedModel(TFPreTrainedModel):
# names with a '.' represents the authorized unexpected/missing layers when a TF model is loaded from a PT model # names with a '.' represents the authorized unexpected/missing layers when a TF model is loaded from a PT model
_keys_to_ignore_on_load_unexpected = [r"h.\d+.attn.bias"] _keys_to_ignore_on_load_unexpected = [r"h.\d+.attn.bias"]
@property
def dummy_inputs(self):
"""
Dummy inputs to build the network.
Returns:
`Dict[str, tf.Tensor]`: The dummy inputs.
"""
dummy = {"input_ids": tf.constant(DUMMY_INPUTS, dtype=tf.int32)}
return dummy
@tf.function(
input_signature=[
{
"input_ids": tf.TensorSpec((None, None), tf.int32, name="input_ids"),
"attention_mask": tf.TensorSpec((None, None), tf.int32, name="attention_mask"),
}
]
)
def serving(self, inputs):
output = self.call(inputs)
return self.serving_output(output)
GPTJ_START_DOCSTRING = r""" GPTJ_START_DOCSTRING = r"""
...@@ -697,18 +672,6 @@ class TFGPTJModel(TFGPTJPreTrainedModel): ...@@ -697,18 +672,6 @@ class TFGPTJModel(TFGPTJPreTrainedModel):
return outputs return outputs
def serving_output(self, output):
pkv = tf.convert_to_tensor(output.past_key_values) if self.config.use_cache else None
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFBaseModelOutputWithPast(
last_hidden_state=output.last_hidden_state,
past_key_values=pkv,
hidden_states=hs,
attentions=attns,
)
@add_start_docstrings( @add_start_docstrings(
""" """
...@@ -821,13 +784,6 @@ class TFGPTJForCausalLM(TFGPTJPreTrainedModel, TFCausalLanguageModelingLoss): ...@@ -821,13 +784,6 @@ class TFGPTJForCausalLM(TFGPTJPreTrainedModel, TFCausalLanguageModelingLoss):
attentions=transformer_outputs.attentions, attentions=transformer_outputs.attentions,
) )
def serving_output(self, output):
pkv = tf.convert_to_tensor(output.past_key_values) if self.config.use_cache else None
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFCausalLMOutputWithPast(logits=output.logits, past_key_values=pkv, hidden_states=hs, attentions=attns)
@add_start_docstrings( @add_start_docstrings(
""" """
...@@ -952,15 +908,6 @@ class TFGPTJForSequenceClassification(TFGPTJPreTrainedModel, TFSequenceClassific ...@@ -952,15 +908,6 @@ class TFGPTJForSequenceClassification(TFGPTJPreTrainedModel, TFSequenceClassific
attentions=transformer_outputs.attentions, attentions=transformer_outputs.attentions,
) )
def serving_output(self, output):
pkv = tf.convert_to_tensor(output.past_key_values) if self.config.use_cache else None
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFSequenceClassifierOutputWithPast(
logits=output.logits, past_key_values=pkv, hidden_states=hs, attentions=attns
)
@add_start_docstrings( @add_start_docstrings(
""" """
...@@ -1051,11 +998,3 @@ class TFGPTJForQuestionAnswering(TFGPTJPreTrainedModel, TFQuestionAnsweringLoss) ...@@ -1051,11 +998,3 @@ class TFGPTJForQuestionAnswering(TFGPTJPreTrainedModel, TFQuestionAnsweringLoss)
hidden_states=transformer_outputs.hidden_states, hidden_states=transformer_outputs.hidden_states,
attentions=transformer_outputs.attentions, attentions=transformer_outputs.attentions,
) )
def serving_output(self, output: TFQuestionAnsweringModelOutput) -> TFQuestionAnsweringModelOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFQuestionAnsweringModelOutput(
start_logits=output.start_logits, end_logits=output.end_logits, hidden_states=hs, attentions=attns
)
src/transformers/models/groupvit/modeling_tf_groupvit.py
View file @ 814de8fa
...@@ -20,7 +20,7 @@ from __future__ import annotations ...@@ -20,7 +20,7 @@ from __future__ import annotations
import collections.abc import collections.abc
import math import math
from dataclasses import dataclass from dataclasses import dataclass
from typing import Any, Dict, Optional, Tuple, Union from typing import Any, Optional, Tuple, Union
import numpy as np import numpy as np
import tensorflow as tf import tensorflow as tf
...@@ -28,7 +28,6 @@ import tensorflow as tf ...@@ -28,7 +28,6 @@ import tensorflow as tf
from ...activations_tf import get_tf_activation from ...activations_tf import get_tf_activation
from ...modeling_tf_outputs import TFBaseModelOutput, TFBaseModelOutputWithPooling from ...modeling_tf_outputs import TFBaseModelOutput, TFBaseModelOutputWithPooling
from ...modeling_tf_utils import ( from ...modeling_tf_utils import (
DUMMY_INPUTS,
TFModelInputType, TFModelInputType,
TFPreTrainedModel, TFPreTrainedModel,
get_initializer, get_initializer,
...@@ -1608,30 +1607,6 @@ class TFGroupViTTextModel(TFGroupViTPreTrainedModel): ...@@ -1608,30 +1607,6 @@ class TFGroupViTTextModel(TFGroupViTPreTrainedModel):
self.groupvit = TFGroupViTTextMainLayer(config, name="groupvit") self.groupvit = TFGroupViTTextMainLayer(config, name="groupvit")
@property
def dummy_inputs(self) -> Dict[str, tf.Tensor]:
"""
Dummy inputs to build the network.
Returns:
`Dict[str, tf.Tensor]`: The dummy inputs.
"""
return {
"input_ids": tf.constant(DUMMY_INPUTS, dtype=tf.int32),
}
@tf.function(
input_signature=[
{
"input_ids": tf.TensorSpec((None, None), tf.int32, name="input_ids"),
"attention_mask": tf.TensorSpec((None, None), tf.int32, name="attention_mask"),
}
]
)
def serving(self, inputs: Dict[str, tf.Tensor]) -> TFBaseModelOutputWithPooling:
output = self.call(inputs)
return self.serving_output(output)
@unpack_inputs @unpack_inputs
@add_start_docstrings_to_model_forward(GROUPVIT_TEXT_INPUTS_DOCSTRING.format("batch_size, sequence_length")) @add_start_docstrings_to_model_forward(GROUPVIT_TEXT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
@replace_return_docstrings(output_type=TFBaseModelOutputWithPooling, config_class=GroupViTTextConfig) @replace_return_docstrings(output_type=TFBaseModelOutputWithPooling, config_class=GroupViTTextConfig)
...@@ -1675,17 +1650,6 @@ class TFGroupViTTextModel(TFGroupViTPreTrainedModel): ...@@ -1675,17 +1650,6 @@ class TFGroupViTTextModel(TFGroupViTPreTrainedModel):
return outputs return outputs
def serving_output(self, output: TFBaseModelOutputWithPooling) -> TFBaseModelOutputWithPooling:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFBaseModelOutputWithPooling(
last_hidden_state=output.last_hidden_state,
pooler_output=output.pooler_output,
hidden_states=hs,
attentions=attns,
)
class TFGroupViTVisionModel(TFGroupViTPreTrainedModel): class TFGroupViTVisionModel(TFGroupViTPreTrainedModel):
config_class = GroupViTVisionConfig config_class = GroupViTVisionConfig
...@@ -1696,38 +1660,6 @@ class TFGroupViTVisionModel(TFGroupViTPreTrainedModel): ...@@ -1696,38 +1660,6 @@ class TFGroupViTVisionModel(TFGroupViTPreTrainedModel):
self.groupvit = TFGroupViTVisionMainLayer(config, name="groupvit") self.groupvit = TFGroupViTVisionMainLayer(config, name="groupvit")
@property
def dummy_inputs(self) -> Dict[str, tf.Tensor]:
"""
Dummy inputs to build the network.
Returns:
`Dict[str, tf.Tensor]`: The dummy inputs.
"""
VISION_DUMMY_INPUTS = tf.random.uniform(
shape=(len(DUMMY_INPUTS), 3, self.config.image_size, self.config.image_size), dtype=tf.float32
)
return {"pixel_values": VISION_DUMMY_INPUTS}
@tf.function(
input_signature=[
{
"pixel_values": tf.TensorSpec((None, None, None, None), tf.float32, name="pixel_values"),
}
]
)
def serving(self, inputs: Dict[str, tf.Tensor]) -> TFBaseModelOutputWithPooling:
"""
Method used for serving the model.
Args:
inputs (`Dict[str, tf.Tensor]`):
The input of the saved model as a dictionary of tensors.
"""
output = self.call(inputs)
return self.serving_output(output)
@unpack_inputs @unpack_inputs
@add_start_docstrings_to_model_forward(GROUPVIT_VISION_INPUTS_DOCSTRING) @add_start_docstrings_to_model_forward(GROUPVIT_VISION_INPUTS_DOCSTRING)
@replace_return_docstrings(output_type=TFBaseModelOutputWithPooling, config_class=GroupViTVisionConfig) @replace_return_docstrings(output_type=TFBaseModelOutputWithPooling, config_class=GroupViTVisionConfig)
...@@ -1772,15 +1704,6 @@ class TFGroupViTVisionModel(TFGroupViTPreTrainedModel): ...@@ -1772,15 +1704,6 @@ class TFGroupViTVisionModel(TFGroupViTPreTrainedModel):
return outputs return outputs
def serving_output(self, output: TFBaseModelOutputWithPooling) -> TFBaseModelOutputWithPooling:
# hidden_states and attentions not converted to Tensor with tf.convert_to_tensor as they are all of different dimensions
return TFBaseModelOutputWithPooling(
last_hidden_state=output.last_hidden_state,
pooler_output=output.pooler_output,
hidden_states=output.hidden_states,
attentions=output.attentions,
)
@add_start_docstrings(GROUPVIT_START_DOCSTRING) @add_start_docstrings(GROUPVIT_START_DOCSTRING)
class TFGroupViTModel(TFGroupViTPreTrainedModel): class TFGroupViTModel(TFGroupViTPreTrainedModel):
...@@ -1791,44 +1714,6 @@ class TFGroupViTModel(TFGroupViTPreTrainedModel): ...@@ -1791,44 +1714,6 @@ class TFGroupViTModel(TFGroupViTPreTrainedModel):
self.groupvit = TFGroupViTMainLayer(config, name="groupvit") self.groupvit = TFGroupViTMainLayer(config, name="groupvit")
@property
def dummy_inputs(self) -> Dict[str, tf.Tensor]:
"""
Dummy inputs to build the network.
Returns:
`Dict[str, tf.Tensor]`: The dummy inputs.
"""
VISION_DUMMY_INPUTS = tf.random.uniform(
shape=(len(DUMMY_INPUTS), 3, self.config.vision_config.image_size, self.config.vision_config.image_size),
dtype=tf.float32,
)
return {
"input_ids": tf.constant(DUMMY_INPUTS, dtype=tf.int32),
"pixel_values": VISION_DUMMY_INPUTS,
}
@tf.function(
input_signature=[
{
"input_ids": tf.TensorSpec((None, None), tf.int32, name="input_ids"),
"pixel_values": tf.TensorSpec((None, None, None, None), tf.float64, name="pixel_values"),
"attention_mask": tf.TensorSpec((None, None), tf.int32, name="attention_mask"),
}
]
)
def serving(self, inputs: Dict[str, tf.Tensor]) -> TFGroupViTModelOutput:
"""
Method used for serving the model.
Args:
inputs (`Dict[str, tf.Tensor]`):
The input of the saved model as a dictionary of tensors.
"""
output = self.call(inputs)
return self.serving_output(output)
@unpack_inputs @unpack_inputs
@add_start_docstrings_to_model_forward(GROUPVIT_TEXT_INPUTS_DOCSTRING.format("batch_size, sequence_length")) @add_start_docstrings_to_model_forward(GROUPVIT_TEXT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
def get_text_features( def get_text_features(
......
src/transformers/models/hubert/modeling_tf_hubert.py
View file @ 814de8fa
...@@ -17,7 +17,7 @@ ...@@ -17,7 +17,7 @@
from __future__ import annotations from __future__ import annotations
import warnings import warnings
from typing import Any, Dict, Optional, Tuple, Union from typing import Any, Optional, Tuple, Union
import numpy as np import numpy as np
import tensorflow as tf import tensorflow as tf
...@@ -1157,14 +1157,12 @@ class TFHubertPreTrainedModel(TFPreTrainedModel): ...@@ -1157,14 +1157,12 @@ class TFHubertPreTrainedModel(TFPreTrainedModel):
main_input_name = "input_values" main_input_name = "input_values"
@property @property
def dummy_inputs(self) -> Dict[str, tf.Tensor]: def input_signature(self):
pad_token = 0.0 return {
input_values = tf.convert_to_tensor(np.random.rand(1, 16000), tf.float32) "input_values": tf.TensorSpec((None, 16000), tf.float32, name="input_values"),
dummy_inputs = { "attention_mask": tf.TensorSpec((None, None), tf.int32, name="attention_mask"),
"input_values": input_values, "token_type_ids": tf.TensorSpec((None, None), tf.int32, name="token_type_ids"),
"attention_mask": tf.cast(tf.not_equal(input_values, pad_token), tf.float32),
} }
return dummy_inputs
def __init__(self, config, *inputs, **kwargs): def __init__(self, config, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs) super().__init__(config, *inputs, **kwargs)
...@@ -1173,20 +1171,6 @@ class TFHubertPreTrainedModel(TFPreTrainedModel): ...@@ -1173,20 +1171,6 @@ class TFHubertPreTrainedModel(TFPreTrainedModel):
"to train/fine-tine this model, you need a GPU or a TPU" "to train/fine-tine this model, you need a GPU or a TPU"
) )
@tf.function(
input_signature=[
{
"input_values": tf.TensorSpec((None, None), tf.float32, name="input_values"),
"attention_mask": tf.TensorSpec((None, None), tf.int32, name="attention_mask"),
"token_type_ids": tf.TensorSpec((None, None), tf.int32, name="token_type_ids"),
}
]
)
def serving(self, inputs):
output = self.call(input_values=inputs, training=False)
return self.serving_output(output)
HUBERT_START_DOCSTRING = r""" HUBERT_START_DOCSTRING = r"""
...@@ -1359,13 +1343,6 @@ class TFHubertModel(TFHubertPreTrainedModel): ...@@ -1359,13 +1343,6 @@ class TFHubertModel(TFHubertPreTrainedModel):
return outputs return outputs
def serving_output(self, output):
hidden_states = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attentions = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFBaseModelOutput(
last_hidden_state=output.last_hidden_state, hidden_states=hidden_states, attentions=attentions
)
@add_start_docstrings( @add_start_docstrings(
"""TFHubert Model with a `language modeling` head on top for Connectionist Temporal Classification (CTC).""", """TFHubert Model with a `language modeling` head on top for Connectionist Temporal Classification (CTC).""",
...@@ -1518,8 +1495,3 @@ class TFHubertForCTC(TFHubertPreTrainedModel): ...@@ -1518,8 +1495,3 @@ class TFHubertForCTC(TFHubertPreTrainedModel):
hidden_states=outputs.hidden_states, hidden_states=outputs.hidden_states,
attentions=outputs.attentions, attentions=outputs.attentions,
) )
def serving_output(self, output: TFCausalLMOutput) -> TFCausalLMOutput:
hidden_states = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attentions = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFCausalLMOutput(logits=output.logits, hidden_states=hidden_states, attentions=attentions)
src/transformers/models/layoutlm/modeling_tf_layoutlm.py
View file @ 814de8fa
...@@ -986,27 +986,6 @@ class TFLayoutLMModel(TFLayoutLMPreTrainedModel): ...@@ -986,27 +986,6 @@ class TFLayoutLMModel(TFLayoutLMPreTrainedModel):
return outputs return outputs
# Copied from transformers.models.bert.modeling_tf_bert.TFBertModel.serving_output
def serving_output(
self, output: TFBaseModelOutputWithPoolingAndCrossAttentions
) -> TFBaseModelOutputWithPoolingAndCrossAttentions:
output_cache = self.config.use_cache and self.config.is_decoder
pkv = tf.convert_to_tensor(output.past_key_values) if output_cache else None
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
cross_attns = tf.convert_to_tensor(output.cross_attentions) if output.cross_attentions is not None else None
if not (self.config.output_attentions and self.config.add_cross_attention):
cross_attns = None
return TFBaseModelOutputWithPoolingAndCrossAttentions(
last_hidden_state=output.last_hidden_state,
pooler_output=output.pooler_output,
past_key_values=pkv,
hidden_states=hs,
attentions=attns,
cross_attentions=cross_attns,
)
@add_start_docstrings("""LayoutLM Model with a `language modeling` head on top.""", LAYOUTLM_START_DOCSTRING) @add_start_docstrings("""LayoutLM Model with a `language modeling` head on top.""", LAYOUTLM_START_DOCSTRING)
class TFLayoutLMForMaskedLM(TFLayoutLMPreTrainedModel, TFMaskedLanguageModelingLoss): class TFLayoutLMForMaskedLM(TFLayoutLMPreTrainedModel, TFMaskedLanguageModelingLoss):
...@@ -1128,12 +1107,6 @@ class TFLayoutLMForMaskedLM(TFLayoutLMPreTrainedModel, TFMaskedLanguageModelingL ...@@ -1128,12 +1107,6 @@ class TFLayoutLMForMaskedLM(TFLayoutLMPreTrainedModel, TFMaskedLanguageModelingL
attentions=outputs.attentions, attentions=outputs.attentions,
) )
def serving_output(self, output: TFMaskedLMOutput) -> TFMaskedLMOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFMaskedLMOutput(logits=output.logits, hidden_states=hs, attentions=attns)
@add_start_docstrings( @add_start_docstrings(
""" """
...@@ -1252,12 +1225,6 @@ class TFLayoutLMForSequenceClassification(TFLayoutLMPreTrainedModel, TFSequenceC ...@@ -1252,12 +1225,6 @@ class TFLayoutLMForSequenceClassification(TFLayoutLMPreTrainedModel, TFSequenceC
attentions=outputs.attentions, attentions=outputs.attentions,
) )
def serving_output(self, output: TFSequenceClassifierOutput) -> TFSequenceClassifierOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFSequenceClassifierOutput(logits=output.logits, hidden_states=hs, attentions=attns)
@add_start_docstrings( @add_start_docstrings(
""" """
...@@ -1380,12 +1347,6 @@ class TFLayoutLMForTokenClassification(TFLayoutLMPreTrainedModel, TFTokenClassif ...@@ -1380,12 +1347,6 @@ class TFLayoutLMForTokenClassification(TFLayoutLMPreTrainedModel, TFTokenClassif
attentions=outputs.attentions, attentions=outputs.attentions,
) )
def serving_output(self, output: TFTokenClassifierOutput) -> TFTokenClassifierOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFTokenClassifierOutput(logits=output.logits, hidden_states=hs, attentions=attns)
@add_start_docstrings( @add_start_docstrings(
""" """
...@@ -1524,11 +1485,3 @@ class TFLayoutLMForQuestionAnswering(TFLayoutLMPreTrainedModel, TFQuestionAnswer ...@@ -1524,11 +1485,3 @@ class TFLayoutLMForQuestionAnswering(TFLayoutLMPreTrainedModel, TFQuestionAnswer
hidden_states=outputs.hidden_states, hidden_states=outputs.hidden_states,
attentions=outputs.attentions, attentions=outputs.attentions,
) )
def serving_output(self, output: TFTokenClassifierOutput) -> TFTokenClassifierOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFQuestionAnsweringModelOutput(
start_logits=output.start_logits, end_logits=output.end_logits, hidden_states=hs, attentions=attns
)
src/transformers/models/layoutlmv3/modeling_tf_layoutlmv3.py
View file @ 814de8fa
...@@ -19,7 +19,7 @@ from __future__ import annotations ...@@ -19,7 +19,7 @@ from __future__ import annotations
import collections import collections
import math import math
from typing import Dict, List, Optional, Tuple, Union from typing import List, Optional, Tuple, Union
import tensorflow as tf import tensorflow as tf
...@@ -980,37 +980,10 @@ class TFLayoutLMv3PreTrainedModel(TFPreTrainedModel): ...@@ -980,37 +980,10 @@ class TFLayoutLMv3PreTrainedModel(TFPreTrainedModel):
base_model_prefix = "layoutlmv3" base_model_prefix = "layoutlmv3"
@property @property
def dummy_inputs(self) -> Dict[str, tf.Tensor]: def input_signature(self):
size = self.config.input_size sig = super().input_signature
image_shape = (2, self.config.num_channels, size, size) sig["bbox"] = tf.TensorSpec((None, None, 4), tf.int32, name="bbox")
pixel_values = tf.random.uniform(shape=image_shape, minval=-1, maxval=1) return sig
return {
"input_ids": tf.constant(_DUMMY_INPUT_IDS, dtype=tf.int32),
"bbox": tf.constant(_DUMMY_BBOX, dtype=tf.int32),
"pixel_values": pixel_values,
}
@tf.function(
input_signature=[
{
"input_ids": tf.TensorSpec((None, None), tf.int32, name="input_ids"),
"bbox": tf.TensorSpec((None, None, 4), tf.int32, name="bbox"),
"pixel_values": tf.TensorSpec((None, None, None, None), tf.float32, name="pixel_values"),
"attention_mask": tf.TensorSpec((None, None), tf.int32, name="attention_mask"),
}
]
)
def serving(self, inputs):
"""
Method used for serving the model.
Args:
inputs (`Dict[str, tf.Tensor]`):
The input of the saved model as a dictionary of tensors.
"""
output = self.call(inputs)
return self.serving_output(output)
LAYOUTLMV3_START_DOCSTRING = r""" LAYOUTLMV3_START_DOCSTRING = r"""
...@@ -1207,16 +1180,6 @@ class TFLayoutLMv3Model(TFLayoutLMv3PreTrainedModel): ...@@ -1207,16 +1180,6 @@ class TFLayoutLMv3Model(TFLayoutLMv3PreTrainedModel):
return outputs return outputs
def serving_output(self, output: TFBaseModelOutput) -> TFBaseModelOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFBaseModelOutput(
last_hidden_state=output.last_hidden_state,
hidden_states=hs,
attentions=attns,
)
class TFLayoutLMv3ClassificationHead(tf.keras.layers.Layer): class TFLayoutLMv3ClassificationHead(tf.keras.layers.Layer):
""" """
...@@ -1354,13 +1317,6 @@ class TFLayoutLMv3ForSequenceClassification(TFLayoutLMv3PreTrainedModel, TFSeque ...@@ -1354,13 +1317,6 @@ class TFLayoutLMv3ForSequenceClassification(TFLayoutLMv3PreTrainedModel, TFSeque
attentions=outputs.attentions, attentions=outputs.attentions,
) )
# Copied from transformers.models.bert.modeling_tf_bert.TFBertForSequenceClassification.serving_output
def serving_output(self, output: TFSequenceClassifierOutput) -> TFSequenceClassifierOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFSequenceClassifierOutput(logits=output.logits, hidden_states=hs, attentions=attns)
@add_start_docstrings( @add_start_docstrings(
""" """
...@@ -1484,13 +1440,6 @@ class TFLayoutLMv3ForTokenClassification(TFLayoutLMv3PreTrainedModel, TFTokenCla ...@@ -1484,13 +1440,6 @@ class TFLayoutLMv3ForTokenClassification(TFLayoutLMv3PreTrainedModel, TFTokenCla
attentions=outputs.attentions, attentions=outputs.attentions,
) )
# Copied from transformers.models.bert.modeling_tf_bert.TFBertForTokenClassification.serving_output
def serving_output(self, output: TFTokenClassifierOutput) -> TFTokenClassifierOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFTokenClassifierOutput(logits=output.logits, hidden_states=hs, attentions=attns)
@add_start_docstrings( @add_start_docstrings(
""" """
...@@ -1618,12 +1567,3 @@ class TFLayoutLMv3ForQuestionAnswering(TFLayoutLMv3PreTrainedModel, TFQuestionAn ...@@ -1618,12 +1567,3 @@ class TFLayoutLMv3ForQuestionAnswering(TFLayoutLMv3PreTrainedModel, TFQuestionAn
hidden_states=outputs.hidden_states, hidden_states=outputs.hidden_states,
attentions=outputs.attentions, attentions=outputs.attentions,
) )
# Copied from transformers.models.bert.modeling_tf_bert.TFBertForQuestionAnswering.serving_output
def serving_output(self, output: TFQuestionAnsweringModelOutput) -> TFQuestionAnsweringModelOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFQuestionAnsweringModelOutput(
start_logits=output.start_logits, end_logits=output.end_logits, hidden_states=hs, attentions=attns
)
src/transformers/models/led/modeling_tf_led.py
View file @ 814de8fa
...@@ -1323,33 +1323,10 @@ class TFLEDPreTrainedModel(TFPreTrainedModel): ...@@ -1323,33 +1323,10 @@ class TFLEDPreTrainedModel(TFPreTrainedModel):
base_model_prefix = "led" base_model_prefix = "led"
@property @property
def dummy_inputs(self): def input_signature(self):
input_ids = tf.convert_to_tensor([[7, 6, 0, 0, 1], [1, 2, 3, 0, 0]], dtype=tf.int32) sig = super().input_signature
# make sure global layers are initialized sig["global_attention_mask"] = tf.TensorSpec((None, None), tf.int32, name="global_attention_mask")
attention_mask = tf.convert_to_tensor([[1, 1, 0, 0, 1], [1, 1, 1, 0, 0]], dtype=tf.int32) return sig
global_attention_mask = tf.convert_to_tensor([[0, 0, 0, 0, 1], [0, 0, 1, 0, 0]], dtype=tf.int32)
dummy_inputs = {
"input_ids": input_ids,
"attention_mask": attention_mask,
"global_attention_mask": global_attention_mask,
"decoder_input_ids": input_ids,
}
return dummy_inputs
@tf.function(
input_signature=[
{
"input_ids": tf.TensorSpec((None, None), tf.int32, name="input_ids"),
"attention_mask": tf.TensorSpec((None, None), tf.int32, name="attention_mask"),
"decoder_input_ids": tf.TensorSpec((None, None), tf.int32, name="decoder_input_ids"),
"decoder_attention_mask": tf.TensorSpec((None, None), tf.int32, name="decoder_attention_mask"),
}
]
)
def serving(self, inputs):
output = self.call(inputs)
return self.serving_output(output)
@dataclass @dataclass
......
src/transformers/models/longformer/modeling_tf_longformer.py
View file @ 814de8fa
...@@ -39,7 +39,6 @@ from ...modeling_tf_utils import ( ...@@ -39,7 +39,6 @@ from ...modeling_tf_utils import (
) )
from ...tf_utils import check_embeddings_within_bounds, shape_list, stable_softmax from ...tf_utils import check_embeddings_within_bounds, shape_list, stable_softmax
from ...utils import ( from ...utils import (
MULTIPLE_CHOICE_DUMMY_INPUTS,
ModelOutput, ModelOutput,
add_code_sample_docstrings, add_code_sample_docstrings,
add_start_docstrings, add_start_docstrings,
...@@ -1874,31 +1873,10 @@ class TFLongformerPreTrainedModel(TFPreTrainedModel): ...@@ -1874,31 +1873,10 @@ class TFLongformerPreTrainedModel(TFPreTrainedModel):
base_model_prefix = "longformer" base_model_prefix = "longformer"
@property @property
def dummy_inputs(self): def input_signature(self):
input_ids = tf.convert_to_tensor([[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]], dtype=tf.int32) sig = super().input_signature
# make sure global layers are initialized sig["global_attention_mask"] = tf.TensorSpec((None, None), tf.int32, name="global_attention_mask")
attention_mask = tf.convert_to_tensor([[1, 1, 0, 0, 1], [1, 1, 1, 0, 0], [1, 0, 0, 1, 1]], dtype=tf.int32) return sig
global_attention_mask = tf.convert_to_tensor(
[[0, 0, 0, 0, 1], [0, 0, 1, 0, 0], [0, 0, 0, 0, 1]], dtype=tf.int32
)
return {
"input_ids": input_ids,
"attention_mask": attention_mask,
"global_attention_mask": global_attention_mask,
}
@tf.function(
input_signature=[
{
"input_ids": tf.TensorSpec((None, None), tf.int32, name="input_ids"),
"attention_mask": tf.TensorSpec((None, None), tf.int32, name="attention_mask"),
}
]
)
def serving(self, inputs):
output = self.call(inputs)
return self.serving_output(output)
LONGFORMER_START_DOCSTRING = r""" LONGFORMER_START_DOCSTRING = r"""
...@@ -2069,19 +2047,6 @@ class TFLongformerModel(TFLongformerPreTrainedModel): ...@@ -2069,19 +2047,6 @@ class TFLongformerModel(TFLongformerPreTrainedModel):
return outputs return outputs
def serving_output(self, output):
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
g_attns = tf.convert_to_tensor(output.global_attentions) if self.config.output_attentions else None
return TFLongformerBaseModelOutputWithPooling(
last_hidden_state=output.last_hidden_state,
pooler_output=output.pooler_output,
hidden_states=hs,
attentions=attns,
global_attentions=g_attns,
)
@add_start_docstrings( @add_start_docstrings(
"""Longformer Model with a `language modeling` head on top.""", """Longformer Model with a `language modeling` head on top.""",
...@@ -2166,15 +2131,6 @@ class TFLongformerForMaskedLM(TFLongformerPreTrainedModel, TFMaskedLanguageModel ...@@ -2166,15 +2131,6 @@ class TFLongformerForMaskedLM(TFLongformerPreTrainedModel, TFMaskedLanguageModel
global_attentions=outputs.global_attentions, global_attentions=outputs.global_attentions,
) )
def serving_output(self, output):
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
g_attns = tf.convert_to_tensor(output.global_attentions) if self.config.output_attentions else None
return TFLongformerMaskedLMOutput(
logits=output.logits, hidden_states=hs, attentions=attns, global_attentions=g_attns
)
@add_start_docstrings( @add_start_docstrings(
""" """
...@@ -2305,19 +2261,6 @@ class TFLongformerForQuestionAnswering(TFLongformerPreTrainedModel, TFQuestionAn ...@@ -2305,19 +2261,6 @@ class TFLongformerForQuestionAnswering(TFLongformerPreTrainedModel, TFQuestionAn
global_attentions=outputs.global_attentions, global_attentions=outputs.global_attentions,
) )
def serving_output(self, output):
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
g_attns = tf.convert_to_tensor(output.global_attentions) if self.config.output_attentions else None
return TFLongformerQuestionAnsweringModelOutput(
start_logits=output.start_logits,
end_logits=output.end_logits,
hidden_states=hs,
attentions=attns,
global_attentions=g_attns,
)
class TFLongformerClassificationHead(tf.keras.layers.Layer): class TFLongformerClassificationHead(tf.keras.layers.Layer):
"""Head for sentence-level classification tasks.""" """Head for sentence-level classification tasks."""
...@@ -2446,15 +2389,6 @@ class TFLongformerForSequenceClassification(TFLongformerPreTrainedModel, TFSeque ...@@ -2446,15 +2389,6 @@ class TFLongformerForSequenceClassification(TFLongformerPreTrainedModel, TFSeque
global_attentions=outputs.global_attentions, global_attentions=outputs.global_attentions,
) )
def serving_output(self, output):
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
g_attns = tf.convert_to_tensor(output.global_attentions) if self.config.output_attentions else None
return TFLongformerSequenceClassifierOutput(
logits=output.logits, hidden_states=hs, attentions=attns, global_attentions=g_attns
)
@add_start_docstrings( @add_start_docstrings(
""" """
...@@ -2477,11 +2411,12 @@ class TFLongformerForMultipleChoice(TFLongformerPreTrainedModel, TFMultipleChoic ...@@ -2477,11 +2411,12 @@ class TFLongformerForMultipleChoice(TFLongformerPreTrainedModel, TFMultipleChoic
) )
@property @property
def dummy_inputs(self): def input_signature(self):
input_ids = tf.convert_to_tensor(MULTIPLE_CHOICE_DUMMY_INPUTS, dtype=tf.int32) return {
# make sure global layers are initialized "input_ids": tf.TensorSpec((None, None, None), tf.int32, name="input_ids"),
global_attention_mask = tf.convert_to_tensor([[[0, 0, 0, 1], [0, 0, 0, 1]]] * 2, dtype=tf.int32) "attention_mask": tf.TensorSpec((None, None, None), tf.int32, name="attention_mask"),
return {"input_ids": input_ids, "global_attention_mask": global_attention_mask} "global_attention_mask": tf.TensorSpec((None, None, None), tf.int32, name="global_attention_mask"),
}
@unpack_inputs @unpack_inputs
@add_start_docstrings_to_model_forward( @add_start_docstrings_to_model_forward(
...@@ -2568,28 +2503,6 @@ class TFLongformerForMultipleChoice(TFLongformerPreTrainedModel, TFMultipleChoic ...@@ -2568,28 +2503,6 @@ class TFLongformerForMultipleChoice(TFLongformerPreTrainedModel, TFMultipleChoic
global_attentions=outputs.global_attentions, global_attentions=outputs.global_attentions,
) )
@tf.function(
input_signature=[
{
"input_ids": tf.TensorSpec((None, None, None), tf.int32, name="input_ids"),
"attention_mask": tf.TensorSpec((None, None, None), tf.int32, name="attention_mask"),
}
]
)
def serving(self, inputs):
output = self.call(inputs)
return self.serving_output(output)
def serving_output(self, output):
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
g_attns = tf.convert_to_tensor(output.global_attentions) if self.config.output_attentions else None
return TFLongformerMultipleChoiceModelOutput(
logits=output.logits, hidden_states=hs, attentions=attns, global_attentions=g_attns
)
@add_start_docstrings( @add_start_docstrings(
""" """
...@@ -2669,12 +2582,3 @@ class TFLongformerForTokenClassification(TFLongformerPreTrainedModel, TFTokenCla ...@@ -2669,12 +2582,3 @@ class TFLongformerForTokenClassification(TFLongformerPreTrainedModel, TFTokenCla
attentions=outputs.attentions, attentions=outputs.attentions,
global_attentions=outputs.global_attentions, global_attentions=outputs.global_attentions,
) )
def serving_output(self, output):
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
g_attns = tf.convert_to_tensor(output.global_attentions) if self.config.output_attentions else None
return TFLongformerTokenClassifierOutput(
logits=output.logits, hidden_states=hs, attentions=attns, global_attentions=g_attns
)
src/transformers/models/lxmert/modeling_tf_lxmert.py
View file @ 814de8fa
...@@ -636,26 +636,6 @@ class TFLxmertEncoder(tf.keras.layers.Layer): ...@@ -636,26 +636,6 @@ class TFLxmertEncoder(tf.keras.layers.Layer):
class TFLxmertMainLayer(tf.keras.layers.Layer): class TFLxmertMainLayer(tf.keras.layers.Layer):
config_class = LxmertConfig config_class = LxmertConfig
@property
def dummy_inputs(self):
"""
Dummy inputs to build the network.
Returns:
tf.Tensor with dummy inputs
"""
batch_size = 2
num_visual_features = 10
input_ids = tf.constant([[3, 5, 6], [2, 3, 4]], dtype=tf.int32)
visual_feats = tf.random.uniform((batch_size, num_visual_features, self.config.visual_feat_dim))
visual_pos = tf.random.uniform((batch_size, num_visual_features, 4))
return {
"input_ids": input_ids,
"visual_feats": visual_feats,
"visual_pos": visual_pos,
}
def __init__(self, config, **kwargs): def __init__(self, config, **kwargs):
super().__init__(**kwargs) super().__init__(**kwargs)
...@@ -802,25 +782,35 @@ class TFLxmertPreTrainedModel(TFPreTrainedModel): ...@@ -802,25 +782,35 @@ class TFLxmertPreTrainedModel(TFPreTrainedModel):
base_model_prefix = "lxmert" base_model_prefix = "lxmert"
@property @property
def dummy_inputs(self) -> Dict[str, tf.Tensor]: def dummy_inputs(self):
return getattr(self, self.base_model_prefix).dummy_inputs """
Dummy inputs to build the network.
@tf.function(
input_signature=[
{
"input_ids": tf.TensorSpec((None, None), tf.int32, name="input_ids"),
"attention_mask": tf.TensorSpec((None, None), tf.int32, name="attention_mask"),
"visual_feats": tf.TensorSpec((None, None, None), tf.float32, name="visual_feats"),
"visual_pos": tf.TensorSpec((None, None, None), tf.float32, name="visual_pos"),
"visual_attention_mask": tf.TensorSpec((None, None), tf.int32, name="visual_attention_mask"),
"token_type_ids": tf.TensorSpec((None, None), tf.int32, name="token_type_ids"),
}
]
)
def serving(self, inputs):
output = self.call(inputs)
return self.serving_output(output) Returns:
tf.Tensor with dummy inputs
"""
batch_size = 2
num_visual_features = 10
input_ids = tf.constant([[3, 5, 6], [2, 3, 4]], dtype=tf.int32)
visual_feats = tf.random.uniform((batch_size, num_visual_features, self.config.visual_feat_dim))
visual_pos = tf.random.uniform((batch_size, num_visual_features, 4))
return {
"input_ids": input_ids,
"visual_feats": visual_feats,
"visual_pos": visual_pos,
}
@property
def input_signature(self):
return {
"input_ids": tf.TensorSpec((None, None), tf.int32, name="input_ids"),
"attention_mask": tf.TensorSpec((None, None), tf.int32, name="attention_mask"),
"visual_feats": tf.TensorSpec((None, None, self.config.visual_feat_dim), tf.float32, name="visual_feats"),
"visual_pos": tf.TensorSpec((None, None, 4), tf.float32, name="visual_pos"),
"visual_attention_mask": tf.TensorSpec((None, None), tf.int32, name="visual_attention_mask"),
"token_type_ids": tf.TensorSpec((None, None), tf.int32, name="token_type_ids"),
}
LXMERT_START_DOCSTRING = r""" LXMERT_START_DOCSTRING = r"""
...@@ -976,24 +966,6 @@ class TFLxmertModel(TFLxmertPreTrainedModel): ...@@ -976,24 +966,6 @@ class TFLxmertModel(TFLxmertPreTrainedModel):
return outputs return outputs
def serving_output(self, output):
l_hs = tf.convert_to_tensor(output.language_hidden_states) if self.config.output_hidden_states else None
v_hs = tf.convert_to_tensor(output.vision_hidden_states) if self.config.output_hidden_states else None
l_attns = tf.convert_to_tensor(output.language_attentions) if self.config.output_attentions else None
v_attns = tf.convert_to_tensor(output.vision_attentions) if self.config.output_attentions else None
c_enc_attns = tf.convert_to_tensor(output.cross_encoder_attentions) if self.config.output_attentions else None
return TFLxmertModelOutput(
pooled_output=output.pooled_output,
language_output=output.language_output,
vision_output=output.vision_output,
language_hidden_states=l_hs,
vision_hidden_states=v_hs,
language_attentions=l_attns,
vision_attentions=v_attns,
cross_encoder_attentions=c_enc_attns,
)
class TFLxmertPooler(tf.keras.layers.Layer): class TFLxmertPooler(tf.keras.layers.Layer):
def __init__(self, config, **kwargs): def __init__(self, config, **kwargs):
...@@ -1415,21 +1387,3 @@ class TFLxmertForPreTraining(TFLxmertPreTrainedModel): ...@@ -1415,21 +1387,3 @@ class TFLxmertForPreTraining(TFLxmertPreTrainedModel):
vision_attentions=lxmert_output.vision_attentions, vision_attentions=lxmert_output.vision_attentions,
cross_encoder_attentions=lxmert_output.cross_encoder_attentions, cross_encoder_attentions=lxmert_output.cross_encoder_attentions,
) )
def serving_output(self, output):
l_hs = tf.convert_to_tensor(output.language_hidden_states) if self.config.output_hidden_states else None
v_hs = tf.convert_to_tensor(output.vision_hidden_states) if self.config.output_hidden_states else None
l_attns = tf.convert_to_tensor(output.language_attentions) if self.config.output_attentions else None
v_attns = tf.convert_to_tensor(output.vision_attentions) if self.config.output_attentions else None
c_enc_attns = tf.convert_to_tensor(output.cross_encoder_attentions) if self.config.output_attentions else None
return TFLxmertForPreTrainingOutput(
prediction_logits=output.prediction_logits,
cross_relationship_score=output.cross_relationship_score,
question_answering_score=output.question_answering_score,
language_hidden_states=l_hs,
vision_hidden_states=v_hs,
language_attentions=l_attns,
vision_attentions=v_attns,
cross_encoder_attentions=c_enc_attns,
)
src/transformers/models/marian/modeling_tf_marian.py
View file @ 814de8fa
...@@ -33,7 +33,6 @@ from ...modeling_tf_outputs import ( ...@@ -33,7 +33,6 @@ from ...modeling_tf_outputs import (
# Public API # Public API
from ...modeling_tf_utils import ( from ...modeling_tf_utils import (
DUMMY_INPUTS,
TFCausalLanguageModelingLoss, TFCausalLanguageModelingLoss,
TFPreTrainedModel, TFPreTrainedModel,
keras_serializable, keras_serializable,
...@@ -501,34 +500,6 @@ class TFMarianPreTrainedModel(TFPreTrainedModel): ...@@ -501,34 +500,6 @@ class TFMarianPreTrainedModel(TFPreTrainedModel):
config_class = MarianConfig config_class = MarianConfig
base_model_prefix = "model" base_model_prefix = "model"
@property
def dummy_inputs(self):
pad_token = 1
input_ids = tf.cast(tf.convert_to_tensor(DUMMY_INPUTS), tf.int32)
decoder_input_ids = tf.cast(tf.convert_to_tensor(DUMMY_INPUTS), tf.int32)
dummy_inputs = {
"decoder_input_ids": decoder_input_ids,
"attention_mask": tf.cast(input_ids != pad_token, tf.int32),
"input_ids": input_ids,
}
return dummy_inputs
@tf.function(
input_signature=[
{
"input_ids": tf.TensorSpec((None, None), tf.int32, name="input_ids"),
"attention_mask": tf.TensorSpec((None, None), tf.int32, name="attention_mask"),
"decoder_input_ids": tf.TensorSpec((None, None), tf.int32, name="decoder_input_ids"),
"decoder_attention_mask": tf.TensorSpec((None, None), tf.int32, name="decoder_attention_mask"),
}
]
)
# Copied from transformers.models.bart.modeling_tf_bart.TFBartPretrainedModel.serving
def serving(self, inputs):
output = self.call(inputs)
return self.serving_output(output)
MARIAN_START_DOCSTRING = r""" MARIAN_START_DOCSTRING = r"""
This model inherits from [`TFPreTrainedModel`]. Check the superclass documentation for the generic methods the This model inherits from [`TFPreTrainedModel`]. Check the superclass documentation for the generic methods the
......
src/transformers/models/mbart/modeling_tf_mbart.py
View file @ 814de8fa
...@@ -32,7 +32,6 @@ from ...modeling_tf_outputs import ( ...@@ -32,7 +32,6 @@ from ...modeling_tf_outputs import (
# Public API # Public API
from ...modeling_tf_utils import ( from ...modeling_tf_utils import (
DUMMY_INPUTS,
TFCausalLanguageModelingLoss, TFCausalLanguageModelingLoss,
TFModelInputType, TFModelInputType,
TFPreTrainedModel, TFPreTrainedModel,
...@@ -468,34 +467,6 @@ class TFMBartPreTrainedModel(TFPreTrainedModel): ...@@ -468,34 +467,6 @@ class TFMBartPreTrainedModel(TFPreTrainedModel):
config_class = MBartConfig config_class = MBartConfig
base_model_prefix = "model" base_model_prefix = "model"
@property
def dummy_inputs(self):
pad_token = 1
input_ids = tf.cast(tf.convert_to_tensor(DUMMY_INPUTS), tf.int32)
decoder_input_ids = tf.cast(tf.convert_to_tensor(DUMMY_INPUTS), tf.int32)
dummy_inputs = {
"decoder_input_ids": decoder_input_ids,
"attention_mask": tf.cast(input_ids != pad_token, tf.int32),
"input_ids": input_ids,
}
return dummy_inputs
@tf.function(
input_signature=[
{
"input_ids": tf.TensorSpec((None, None), tf.int32, name="input_ids"),
"attention_mask": tf.TensorSpec((None, None), tf.int32, name="attention_mask"),
"decoder_input_ids": tf.TensorSpec((None, None), tf.int32, name="decoder_input_ids"),
"decoder_attention_mask": tf.TensorSpec((None, None), tf.int32, name="decoder_attention_mask"),
}
]
)
# Copied from transformers.models.bart.modeling_tf_bart.TFBartPretrainedModel.serving
def serving(self, inputs):
output = self.call(inputs)
return self.serving_output(output)
MBART_START_DOCSTRING = r""" MBART_START_DOCSTRING = r"""
This model inherits from [`TFPreTrainedModel`]. Check the superclass documentation for the generic methods the This model inherits from [`TFPreTrainedModel`]. Check the superclass documentation for the generic methods the
......
src/transformers/models/mobilebert/modeling_tf_mobilebert.py
View file @ 814de8fa
...@@ -20,7 +20,7 @@ from __future__ import annotations ...@@ -20,7 +20,7 @@ from __future__ import annotations
import warnings import warnings
from dataclasses import dataclass from dataclasses import dataclass
from typing import Dict, Optional, Tuple, Union from typing import Optional, Tuple, Union
import numpy as np import numpy as np
import tensorflow as tf import tensorflow as tf
...@@ -51,7 +51,6 @@ from ...modeling_tf_utils import ( ...@@ -51,7 +51,6 @@ from ...modeling_tf_utils import (
) )
from ...tf_utils import check_embeddings_within_bounds, shape_list, stable_softmax from ...tf_utils import check_embeddings_within_bounds, shape_list, stable_softmax
from ...utils import ( from ...utils import (
MULTIPLE_CHOICE_DUMMY_INPUTS,
ModelOutput, ModelOutput,
add_code_sample_docstrings, add_code_sample_docstrings,
add_start_docstrings, add_start_docstrings,
...@@ -998,17 +997,6 @@ class TFMobileBertModel(TFMobileBertPreTrainedModel): ...@@ -998,17 +997,6 @@ class TFMobileBertModel(TFMobileBertPreTrainedModel):
return outputs return outputs
def serving_output(self, output: TFBaseModelOutputWithPooling) -> TFBaseModelOutputWithPooling:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFBaseModelOutputWithPooling(
last_hidden_state=output.last_hidden_state,
pooler_output=output.pooler_output,
hidden_states=hs,
attentions=attns,
)
@add_start_docstrings( @add_start_docstrings(
""" """
...@@ -1099,17 +1087,6 @@ class TFMobileBertForPreTraining(TFMobileBertPreTrainedModel, TFMobileBertPreTra ...@@ -1099,17 +1087,6 @@ class TFMobileBertForPreTraining(TFMobileBertPreTrainedModel, TFMobileBertPreTra
attentions=outputs.attentions, attentions=outputs.attentions,
) )
def serving_output(self, output):
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFMobileBertForPreTrainingOutput(
prediction_logits=output.prediction_logits,
seq_relationship_logits=output.seq_relationship_logits,
hidden_states=hs,
attentions=attns,
)
@add_start_docstrings("""MobileBert Model with a `language modeling` head on top.""", MOBILEBERT_START_DOCSTRING) @add_start_docstrings("""MobileBert Model with a `language modeling` head on top.""", MOBILEBERT_START_DOCSTRING)
class TFMobileBertForMaskedLM(TFMobileBertPreTrainedModel, TFMaskedLanguageModelingLoss): class TFMobileBertForMaskedLM(TFMobileBertPreTrainedModel, TFMaskedLanguageModelingLoss):
...@@ -1190,13 +1167,6 @@ class TFMobileBertForMaskedLM(TFMobileBertPreTrainedModel, TFMaskedLanguageModel ...@@ -1190,13 +1167,6 @@ class TFMobileBertForMaskedLM(TFMobileBertPreTrainedModel, TFMaskedLanguageModel
attentions=outputs.attentions, attentions=outputs.attentions,
) )
# Copied from transformers.models.bert.modeling_tf_bert.TFBertForMaskedLM.serving_output
def serving_output(self, output: TFMaskedLMOutput) -> TFMaskedLMOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFMaskedLMOutput(logits=output.logits, hidden_states=hs, attentions=attns)
class TFMobileBertOnlyNSPHead(tf.keras.layers.Layer): class TFMobileBertOnlyNSPHead(tf.keras.layers.Layer):
def __init__(self, config, **kwargs): def __init__(self, config, **kwargs):
...@@ -1289,13 +1259,6 @@ class TFMobileBertForNextSentencePrediction(TFMobileBertPreTrainedModel, TFNextS ...@@ -1289,13 +1259,6 @@ class TFMobileBertForNextSentencePrediction(TFMobileBertPreTrainedModel, TFNextS
attentions=outputs.attentions, attentions=outputs.attentions,
) )
# Copied from transformers.models.bert.modeling_tf_bert.TFBertForNextSentencePrediction.serving_output
def serving_output(self, output: TFNextSentencePredictorOutput) -> TFNextSentencePredictorOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFNextSentencePredictorOutput(logits=output.logits, hidden_states=hs, attentions=attns)
@add_start_docstrings( @add_start_docstrings(
""" """
...@@ -1386,13 +1349,6 @@ class TFMobileBertForSequenceClassification(TFMobileBertPreTrainedModel, TFSeque ...@@ -1386,13 +1349,6 @@ class TFMobileBertForSequenceClassification(TFMobileBertPreTrainedModel, TFSeque
attentions=outputs.attentions, attentions=outputs.attentions,
) )
# Copied from transformers.models.bert.modeling_tf_bert.TFBertForSequenceClassification.serving_output
def serving_output(self, output: TFSequenceClassifierOutput) -> TFSequenceClassifierOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFSequenceClassifierOutput(logits=output.logits, hidden_states=hs, attentions=attns)
@add_start_docstrings( @add_start_docstrings(
""" """
...@@ -1492,15 +1448,6 @@ class TFMobileBertForQuestionAnswering(TFMobileBertPreTrainedModel, TFQuestionAn ...@@ -1492,15 +1448,6 @@ class TFMobileBertForQuestionAnswering(TFMobileBertPreTrainedModel, TFQuestionAn
attentions=outputs.attentions, attentions=outputs.attentions,
) )
# Copied from transformers.models.bert.modeling_tf_bert.TFBertForQuestionAnswering.serving_output
def serving_output(self, output: TFQuestionAnsweringModelOutput) -> TFQuestionAnsweringModelOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFQuestionAnsweringModelOutput(
start_logits=output.start_logits, end_logits=output.end_logits, hidden_states=hs, attentions=attns
)
@add_start_docstrings( @add_start_docstrings(
""" """
...@@ -1528,16 +1475,6 @@ class TFMobileBertForMultipleChoice(TFMobileBertPreTrainedModel, TFMultipleChoic ...@@ -1528,16 +1475,6 @@ class TFMobileBertForMultipleChoice(TFMobileBertPreTrainedModel, TFMultipleChoic
1, kernel_initializer=get_initializer(config.initializer_range), name="classifier" 1, kernel_initializer=get_initializer(config.initializer_range), name="classifier"
) )
@property
def dummy_inputs(self):
"""
Dummy inputs to build the network.
Returns:
tf.Tensor with dummy inputs
"""
return {"input_ids": tf.constant(MULTIPLE_CHOICE_DUMMY_INPUTS, dtype=tf.int32)}
@unpack_inputs @unpack_inputs
@add_start_docstrings_to_model_forward( @add_start_docstrings_to_model_forward(
MOBILEBERT_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length") MOBILEBERT_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length")
...@@ -1612,28 +1549,6 @@ class TFMobileBertForMultipleChoice(TFMobileBertPreTrainedModel, TFMultipleChoic ...@@ -1612,28 +1549,6 @@ class TFMobileBertForMultipleChoice(TFMobileBertPreTrainedModel, TFMultipleChoic
attentions=outputs.attentions, attentions=outputs.attentions,
) )
@tf.function(
input_signature=[
{
"input_ids": tf.TensorSpec((None, None, None), tf.int32, name="input_ids"),
"attention_mask": tf.TensorSpec((None, None, None), tf.int32, name="attention_mask"),
"token_type_ids": tf.TensorSpec((None, None, None), tf.int32, name="token_type_ids"),
}
]
)
# Copied from transformers.models.bert.modeling_tf_bert.TFBertForMultipleChoice.serving
def serving(self, inputs: Dict[str, tf.Tensor]):
output = self.call(input_ids=inputs)
return self.serving_output(output)
# Copied from transformers.models.bert.modeling_tf_bert.TFBertForMultipleChoice.serving_output
def serving_output(self, output: TFMultipleChoiceModelOutput) -> TFMultipleChoiceModelOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFMultipleChoiceModelOutput(logits=output.logits, hidden_states=hs, attentions=attns)
@add_start_docstrings( @add_start_docstrings(
""" """
...@@ -1722,10 +1637,3 @@ class TFMobileBertForTokenClassification(TFMobileBertPreTrainedModel, TFTokenCla ...@@ -1722,10 +1637,3 @@ class TFMobileBertForTokenClassification(TFMobileBertPreTrainedModel, TFTokenCla
hidden_states=outputs.hidden_states, hidden_states=outputs.hidden_states,
attentions=outputs.attentions, attentions=outputs.attentions,
) )
# Copied from transformers.models.bert.modeling_tf_bert.TFBertForTokenClassification.serving_output
def serving_output(self, output: TFTokenClassifierOutput) -> TFTokenClassifierOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFTokenClassifierOutput(logits=output.logits, hidden_states=hs, attentions=attns)
src/transformers/models/mobilevit/modeling_tf_mobilevit.py
View file @ 814de8fa
...@@ -735,38 +735,6 @@ class TFMobileViTPreTrainedModel(TFPreTrainedModel): ...@@ -735,38 +735,6 @@ class TFMobileViTPreTrainedModel(TFPreTrainedModel):
base_model_prefix = "mobilevit" base_model_prefix = "mobilevit"
main_input_name = "pixel_values" main_input_name = "pixel_values"
@property
def dummy_inputs(self) -> Dict[str, tf.Tensor]:
"""
Dummy inputs to build the network.
Returns:
`Dict[str, tf.Tensor]`: The dummy inputs.
"""
VISION_DUMMY_INPUTS = tf.random.uniform(
shape=(3, self.config.num_channels, self.config.image_size, self.config.image_size),
dtype=tf.float32,
)
return {"pixel_values": tf.constant(VISION_DUMMY_INPUTS)}
@tf.function(
input_signature=[
{
"pixel_values": tf.TensorSpec((None, None, None, None), tf.float32, name="pixel_values"),
}
]
)
def serving(self, inputs):
"""
Method used for serving the model.
Args:
inputs (`Dict[str, tf.Tensor]`):
The input of the saved model as a dictionary of tensors.
"""
output = self.call(inputs)
return self.serving_output(output)
MOBILEVIT_START_DOCSTRING = r""" MOBILEVIT_START_DOCSTRING = r"""
This model inherits from [`TFPreTrainedModel`]. Check the superclass documentation for the generic methods the This model inherits from [`TFPreTrainedModel`]. Check the superclass documentation for the generic methods the
...@@ -856,14 +824,6 @@ class TFMobileViTModel(TFMobileViTPreTrainedModel): ...@@ -856,14 +824,6 @@ class TFMobileViTModel(TFMobileViTPreTrainedModel):
output = self.mobilevit(pixel_values, output_hidden_states, return_dict, training=training) output = self.mobilevit(pixel_values, output_hidden_states, return_dict, training=training)
return output return output
def serving_output(self, output: TFBaseModelOutputWithPooling) -> TFBaseModelOutputWithPooling:
# hidden_states not converted to Tensor with tf.convert_to_tensor as they are all of different dimensions
return TFBaseModelOutputWithPooling(
last_hidden_state=output.last_hidden_state,
pooler_output=output.pooler_output,
hidden_states=output.hidden_states,
)
@add_start_docstrings( @add_start_docstrings(
""" """
...@@ -924,10 +884,6 @@ class TFMobileViTForImageClassification(TFMobileViTPreTrainedModel, TFSequenceCl ...@@ -924,10 +884,6 @@ class TFMobileViTForImageClassification(TFMobileViTPreTrainedModel, TFSequenceCl
return TFImageClassifierOutputWithNoAttention(loss=loss, logits=logits, hidden_states=outputs.hidden_states) return TFImageClassifierOutputWithNoAttention(loss=loss, logits=logits, hidden_states=outputs.hidden_states)
def serving_output(self, output: TFImageClassifierOutputWithNoAttention) -> TFImageClassifierOutputWithNoAttention:
# hidden_states and attention not converted to Tensor with tf.convert_to_tensor as they are all of different dimensions
return TFImageClassifierOutputWithNoAttention(logits=output.logits, hidden_states=output.hidden_states)
class TFMobileViTASPPPooling(tf.keras.layers.Layer): class TFMobileViTASPPPooling(tf.keras.layers.Layer):
def __init__(self, config: MobileViTConfig, out_channels: int, **kwargs) -> None: def __init__(self, config: MobileViTConfig, out_channels: int, **kwargs) -> None:
...@@ -1157,8 +1113,3 @@ class TFMobileViTForSemanticSegmentation(TFMobileViTPreTrainedModel): ...@@ -1157,8 +1113,3 @@ class TFMobileViTForSemanticSegmentation(TFMobileViTPreTrainedModel):
logits=logits, logits=logits,
hidden_states=outputs.hidden_states if output_hidden_states else None, hidden_states=outputs.hidden_states if output_hidden_states else None,
) )
def serving_output(
self, output: TFSemanticSegmenterOutputWithNoAttention
) -> TFSemanticSegmenterOutputWithNoAttention:
return TFSemanticSegmenterOutputWithNoAttention(logits=output.logits, hidden_states=output.hidden_states)
src/transformers/models/mpnet/modeling_tf_mpnet.py
View file @ 814de8fa
...@@ -49,7 +49,6 @@ from ...modeling_tf_utils import ( ...@@ -49,7 +49,6 @@ from ...modeling_tf_utils import (
) )
from ...tf_utils import check_embeddings_within_bounds, shape_list, stable_softmax from ...tf_utils import check_embeddings_within_bounds, shape_list, stable_softmax
from ...utils import ( from ...utils import (
MULTIPLE_CHOICE_DUMMY_INPUTS,
add_code_sample_docstrings, add_code_sample_docstrings,
add_start_docstrings, add_start_docstrings,
add_start_docstrings_to_model_forward, add_start_docstrings_to_model_forward,
...@@ -77,19 +76,6 @@ class TFMPNetPreTrainedModel(TFPreTrainedModel): ...@@ -77,19 +76,6 @@ class TFMPNetPreTrainedModel(TFPreTrainedModel):
config_class = MPNetConfig config_class = MPNetConfig
base_model_prefix = "mpnet" base_model_prefix = "mpnet"
@tf.function(
input_signature=[
{
"input_ids": tf.TensorSpec((None, None), tf.int32, name="input_ids"),
"attention_mask": tf.TensorSpec((None, None), tf.int32, name="attention_mask"),
}
]
)
def serving(self, inputs):
output = self.call(inputs)
return self.serving_output(output)
class TFMPNetEmbeddings(tf.keras.layers.Layer): class TFMPNetEmbeddings(tf.keras.layers.Layer):
"""Construct the embeddings from word, position embeddings.""" """Construct the embeddings from word, position embeddings."""
...@@ -707,17 +693,6 @@ class TFMPNetModel(TFMPNetPreTrainedModel): ...@@ -707,17 +693,6 @@ class TFMPNetModel(TFMPNetPreTrainedModel):
) )
return outputs return outputs
def serving_output(self, output: TFBaseModelOutputWithPooling) -> TFBaseModelOutputWithPooling:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFBaseModelOutputWithPooling(
last_hidden_state=output.last_hidden_state,
pooler_output=output.pooler_output,
hidden_states=hs,
attentions=attns,
)
class TFMPNetLMHead(tf.keras.layers.Layer): class TFMPNetLMHead(tf.keras.layers.Layer):
"""MPNet head for masked and permuted language modeling""" """MPNet head for masked and permuted language modeling"""
...@@ -841,13 +816,6 @@ class TFMPNetForMaskedLM(TFMPNetPreTrainedModel, TFMaskedLanguageModelingLoss): ...@@ -841,13 +816,6 @@ class TFMPNetForMaskedLM(TFMPNetPreTrainedModel, TFMaskedLanguageModelingLoss):
attentions=outputs.attentions, attentions=outputs.attentions,
) )
# Copied from transformers.models.bert.modeling_tf_bert.TFBertForMaskedLM.serving_output
def serving_output(self, output: TFMaskedLMOutput) -> TFMaskedLMOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFMaskedLMOutput(logits=output.logits, hidden_states=hs, attentions=attns)
class TFMPNetClassificationHead(tf.keras.layers.Layer): class TFMPNetClassificationHead(tf.keras.layers.Layer):
"""Head for sentence-level classification tasks.""" """Head for sentence-level classification tasks."""
...@@ -945,13 +913,6 @@ class TFMPNetForSequenceClassification(TFMPNetPreTrainedModel, TFSequenceClassif ...@@ -945,13 +913,6 @@ class TFMPNetForSequenceClassification(TFMPNetPreTrainedModel, TFSequenceClassif
attentions=outputs.attentions, attentions=outputs.attentions,
) )
# Copied from transformers.models.bert.modeling_tf_bert.TFBertForSequenceClassification.serving_output
def serving_output(self, output: TFSequenceClassifierOutput) -> TFSequenceClassifierOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFSequenceClassifierOutput(logits=output.logits, hidden_states=hs, attentions=attns)
@add_start_docstrings( @add_start_docstrings(
""" """
...@@ -970,16 +931,6 @@ class TFMPNetForMultipleChoice(TFMPNetPreTrainedModel, TFMultipleChoiceLoss): ...@@ -970,16 +931,6 @@ class TFMPNetForMultipleChoice(TFMPNetPreTrainedModel, TFMultipleChoiceLoss):
1, kernel_initializer=get_initializer(config.initializer_range), name="classifier" 1, kernel_initializer=get_initializer(config.initializer_range), name="classifier"
) )
@property
def dummy_inputs(self):
"""
Dummy inputs to build the network.
Returns:
tf.Tensor with dummy inputs
"""
return {"input_ids": tf.constant(MULTIPLE_CHOICE_DUMMY_INPUTS, dtype=tf.int32)}
@unpack_inputs @unpack_inputs
@add_start_docstrings_to_model_forward(MPNET_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length")) @add_start_docstrings_to_model_forward(MPNET_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length"))
@add_code_sample_docstrings( @add_code_sample_docstrings(
...@@ -1048,26 +999,6 @@ class TFMPNetForMultipleChoice(TFMPNetPreTrainedModel, TFMultipleChoiceLoss): ...@@ -1048,26 +999,6 @@ class TFMPNetForMultipleChoice(TFMPNetPreTrainedModel, TFMultipleChoiceLoss):
attentions=outputs.attentions, attentions=outputs.attentions,
) )
@tf.function(
input_signature=[
{
"input_ids": tf.TensorSpec((None, None, None), tf.int32, name="input_ids"),
"attention_mask": tf.TensorSpec((None, None, None), tf.int32, name="attention_mask"),
}
]
)
def serving(self, inputs):
output = self.call(inputs)
return self.serving_output(output)
# Copied from transformers.models.bert.modeling_tf_bert.TFBertForMultipleChoice.serving_output
def serving_output(self, output: TFMultipleChoiceModelOutput) -> TFMultipleChoiceModelOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFMultipleChoiceModelOutput(logits=output.logits, hidden_states=hs, attentions=attns)
@add_start_docstrings( @add_start_docstrings(
""" """
...@@ -1142,13 +1073,6 @@ class TFMPNetForTokenClassification(TFMPNetPreTrainedModel, TFTokenClassificatio ...@@ -1142,13 +1073,6 @@ class TFMPNetForTokenClassification(TFMPNetPreTrainedModel, TFTokenClassificatio
attentions=outputs.attentions, attentions=outputs.attentions,
) )
# Copied from transformers.models.bert.modeling_tf_bert.TFBertForTokenClassification.serving_output
def serving_output(self, output: TFTokenClassifierOutput) -> TFTokenClassifierOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFTokenClassifierOutput(logits=output.logits, hidden_states=hs, attentions=attns)
@add_start_docstrings( @add_start_docstrings(
""" """
...@@ -1235,12 +1159,3 @@ class TFMPNetForQuestionAnswering(TFMPNetPreTrainedModel, TFQuestionAnsweringLos ...@@ -1235,12 +1159,3 @@ class TFMPNetForQuestionAnswering(TFMPNetPreTrainedModel, TFQuestionAnsweringLos
hidden_states=outputs.hidden_states, hidden_states=outputs.hidden_states,
attentions=outputs.attentions, attentions=outputs.attentions,
) )
# Copied from transformers.models.bert.modeling_tf_bert.TFBertForQuestionAnswering.serving_output
def serving_output(self, output: TFQuestionAnsweringModelOutput) -> TFQuestionAnsweringModelOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFQuestionAnsweringModelOutput(
start_logits=output.start_logits, end_logits=output.end_logits, hidden_states=hs, attentions=attns
)
src/transformers/models/openai/modeling_tf_openai.py
View file @ 814de8fa
...@@ -357,19 +357,6 @@ class TFOpenAIGPTPreTrainedModel(TFPreTrainedModel): ...@@ -357,19 +357,6 @@ class TFOpenAIGPTPreTrainedModel(TFPreTrainedModel):
config_class = OpenAIGPTConfig config_class = OpenAIGPTConfig
base_model_prefix = "transformer" base_model_prefix = "transformer"
@tf.function(
input_signature=[
{
"input_ids": tf.TensorSpec((None, None), tf.int32, name="input_ids"),
"attention_mask": tf.TensorSpec((None, None), tf.int32, name="attention_mask"),
}
]
)
def serving(self, inputs):
output = self.call(inputs)
return self.serving_output(output)
@dataclass @dataclass
class TFOpenAIGPTDoubleHeadsModelOutput(ModelOutput): class TFOpenAIGPTDoubleHeadsModelOutput(ModelOutput):
...@@ -541,13 +528,6 @@ class TFOpenAIGPTModel(TFOpenAIGPTPreTrainedModel): ...@@ -541,13 +528,6 @@ class TFOpenAIGPTModel(TFOpenAIGPTPreTrainedModel):
) )
return outputs return outputs
# Copied from transformers.models.distilbert.modeling_tf_distilbert.TFDistilBertModel.serving_output
def serving_output(self, output):
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFBaseModelOutput(last_hidden_state=output.last_hidden_state, hidden_states=hs, attentions=attns)
@add_start_docstrings( @add_start_docstrings(
""" """
...@@ -630,12 +610,6 @@ class TFOpenAIGPTLMHeadModel(TFOpenAIGPTPreTrainedModel, TFCausalLanguageModelin ...@@ -630,12 +610,6 @@ class TFOpenAIGPTLMHeadModel(TFOpenAIGPTPreTrainedModel, TFCausalLanguageModelin
attentions=transformer_outputs.attentions, attentions=transformer_outputs.attentions,
) )
def serving_output(self, output: TFCausalLMOutput) -> TFCausalLMOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFCausalLMOutput(logits=output.logits, hidden_states=hs, attentions=attns)
def prepare_inputs_for_generation(self, inputs, **kwargs): def prepare_inputs_for_generation(self, inputs, **kwargs):
return {"input_ids": inputs} return {"input_ids": inputs}
...@@ -752,27 +726,13 @@ class TFOpenAIGPTDoubleHeadsModel(TFOpenAIGPTPreTrainedModel): ...@@ -752,27 +726,13 @@ class TFOpenAIGPTDoubleHeadsModel(TFOpenAIGPTPreTrainedModel):
attentions=transformer_outputs.attentions, attentions=transformer_outputs.attentions,
) )
@tf.function( @property
input_signature=[ def input_signature(self):
{ return {
"input_ids": tf.TensorSpec((None, None, None), tf.int32, name="input_ids"), "input_ids": tf.TensorSpec((None, None, None), tf.int32, name="input_ids"),
"attention_mask": tf.TensorSpec((None, None, None), tf.int32, name="attention_mask"), "attention_mask": tf.TensorSpec((None, None, None), tf.int32, name="attention_mask"),
"mc_token_ids": tf.TensorSpec((None, None), tf.int32, name="token_type_ids"), "mc_token_ids": tf.TensorSpec((None, None), tf.int32, name="token_type_ids"),
} }
]
)
def serving(self, inputs):
output = self.call(inputs)
return self.serving_output(output)
def serving_output(self, output):
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFOpenAIGPTDoubleHeadsModelOutput(
logits=output.logits, mc_logits=output.mc_logits, hidden_states=hs, attentions=attns
)
@add_start_docstrings( @add_start_docstrings(
...@@ -894,10 +854,3 @@ class TFOpenAIGPTForSequenceClassification(TFOpenAIGPTPreTrainedModel, TFSequenc ...@@ -894,10 +854,3 @@ class TFOpenAIGPTForSequenceClassification(TFOpenAIGPTPreTrainedModel, TFSequenc
hidden_states=transformer_outputs.hidden_states, hidden_states=transformer_outputs.hidden_states,
attentions=transformer_outputs.attentions, attentions=transformer_outputs.attentions,
) )
# Copied from transformers.models.bert.modeling_tf_bert.TFBertForSequenceClassification.serving_output
def serving_output(self, output: TFSequenceClassifierOutput) -> TFSequenceClassifierOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFSequenceClassifierOutput(logits=output.logits, hidden_states=hs, attentions=attns)
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment