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 131 additions and 1313 deletions
src/transformers/modeling_tf_utils.py
View file @ 814de8fa
...@@ -42,7 +42,6 @@ from .dynamic_module_utils import custom_object_save ...@@ -42,7 +42,6 @@ from .dynamic_module_utils import custom_object_save
from .generation import GenerationConfig, TFGenerationMixin from .generation import GenerationConfig, TFGenerationMixin
from .tf_utils import expand_1d, load_attributes_from_hdf5_group, save_attributes_to_hdf5_group, shape_list from .tf_utils import expand_1d, load_attributes_from_hdf5_group, save_attributes_to_hdf5_group, shape_list
from .utils import ( from .utils import (
DUMMY_INPUTS,
SAFE_WEIGHTS_INDEX_NAME, SAFE_WEIGHTS_INDEX_NAME,
SAFE_WEIGHTS_NAME, SAFE_WEIGHTS_NAME,
TF2_WEIGHTS_INDEX_NAME, TF2_WEIGHTS_INDEX_NAME,
...@@ -1114,9 +1113,25 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin, TFGenerationMixin, Pu ...@@ -1114,9 +1113,25 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin, TFGenerationMixin, Pu
Returns: Returns:
`Dict[str, tf.Tensor]`: The dummy inputs. `Dict[str, tf.Tensor]`: The dummy inputs.
""" """
return { dummies = {}
"input_ids": tf.constant(DUMMY_INPUTS, dtype=tf.int32), sig = self._prune_signature(self.input_signature)
} for key, spec in sig.items():
# 3 is the most correct arbitrary size. I will not be taking questions
dummies[key] = tf.ones(shape=[dim if dim is not None else 3 for dim in spec.shape], dtype=spec.dtype)
if key == "token_type_ids":
# Some models have token_type_ids but with a vocab_size of 1
dummies[key] = tf.zeros_like(dummies[key])
if self.config.add_cross_attention and "encoder_hidden_states" in inspect.signature(self.call).parameters:
if "encoder_hidden_states" not in dummies:
if self.main_input_name == "input_ids":
dummies["encoder_hidden_states"] = tf.ones(
shape=(3, 3, self.config.hidden_size), dtype=tf.float32, name="encoder_hidden_states"
)
else:
raise NotImplementedError(
"Model has cross-attention but we couldn't infer the shape for the encoder hidden states. Please manually override dummy_inputs!"
)
return dummies
@property @property
def framework(self) -> str: def framework(self) -> str:
...@@ -1137,6 +1152,10 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin, TFGenerationMixin, Pu ...@@ -1137,6 +1152,10 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin, TFGenerationMixin, Pu
self.config = config self.config = config
self.name_or_path = config.name_or_path self.name_or_path = config.name_or_path
self.generation_config = GenerationConfig.from_model_config(config) if self.can_generate() else None self.generation_config = GenerationConfig.from_model_config(config) if self.can_generate() else None
if not hasattr(self, "serving"): # Don't overwrite existing serving signatures
self.serving = tf.function(
self.eager_serving, input_signature=[self._prune_signature(self.input_signature)]
)
# Set the serving spec quickly to ensure that Keras doesn't use the specific dummy input shapes as the spec # Set the serving spec quickly to ensure that Keras doesn't use the specific dummy input shapes as the spec
self._set_save_spec(self.serving.input_signature[0]) self._set_save_spec(self.serving.input_signature[0])
...@@ -1201,36 +1220,82 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin, TFGenerationMixin, Pu ...@@ -1201,36 +1220,82 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin, TFGenerationMixin, Pu
return self.serving_output(output) return self.serving_output(output)
@tf.function( @property
input_signature=[ def input_signature(self) -> Dict[str, tf.TensorSpec]:
{
"input_ids": tf.TensorSpec((None, None), tf.int32, name="input_ids"),
"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):
""" """
Method used for serving the model. This property should return a dict mapping input names to tf.TensorSpec objects, representing the expected
shape and dtype for model inputs. It is used for both serving and for generating the dummy inputs used to build
Args: the model.
inputs (`Dict[str, tf.Tensor]`):
The input of the saved model as a dictionary of tensors.
""" """
output = self.call(inputs) model_inputs = list(inspect.signature(self.call).parameters)
sig = {}
if "input_ids" in model_inputs:
if self.__class__.__name__.endswith("ForMultipleChoice"):
text_dims = 3
else:
text_dims = 2
for input_name in (
"input_ids",
"attention_mask",
"token_type_ids",
"decoder_input_ids",
"decoder_attention_mask",
):
if input_name in model_inputs:
sig[input_name] = tf.TensorSpec([None] * text_dims, tf.int32, name=input_name)
if "pixel_values" in model_inputs:
pixel_values_shape = [None, None, None, None]
if hasattr(self.config, "vision_config"):
vision_config = self.config.vision_config
else:
vision_config = self.config
if hasattr(vision_config, "num_channels"):
pixel_values_shape[1] = vision_config.num_channels
else:
raise NotImplementedError(
"Could not infer number of channels from config, please override input_signature to specify input shapes."
)
if hasattr(vision_config, "image_size"):
pixel_values_shape[2] = pixel_values_shape[3] = vision_config.image_size
elif hasattr(vision_config, "input_size"):
pixel_values_shape[2] = pixel_values_shape[3] = vision_config.input_size
else:
raise NotImplementedError(
"Could not infer input image shape from config, please override input_signature to specify input shapes."
)
sig["pixel_values"] = tf.TensorSpec(pixel_values_shape, tf.float32, name="pixel_values")
if "input_features" in model_inputs:
raise NotImplementedError("Audio models need a manually defined input_signature")
return sig
return self.serving_output(output) def _prune_signature(self, signature):
"""Keeps only the keys of a given input signature that are valid for this model."""
model_inputs = list(inspect.signature(self.call).parameters)
return {key: val for key, val in signature.items() if key in model_inputs}
def serving_output(self, output): def serving_output(self, output):
""" """
Prepare the output of the saved model. Each model must implement this function. Prepare the output of the saved model. Can be overridden if specific serving modifications are required.
"""
Args: if not isinstance(output, ModelOutput):
output ([`TFBaseModelOutput`]): return output
The output returned by the model. for key in output:
""" if key.endswith("hidden_states") and not getattr(self.config, "output_hidden_states", False):
raise NotImplementedError output[key] = None
elif key.endswith("attentions") and not getattr(self.config, "output_attentions", False):
output[key] = None
elif key == "past_key_values" and not getattr(self.config, "use_cache", False):
output[key] = None
elif key == "cross_attentions" and not (
getattr(self.config, "output_attentions", False) and getattr(self.config, "add_cross_attention", False)
):
output[key] = None
if isinstance(output[key], (tuple, list)):
try:
output[key] = tf.convert_to_tensor(output[key])
except (ValueError, tf.errors.InvalidArgumentError):
pass # Layers may not have the same dimensions
return output
def can_generate(self) -> bool: def can_generate(self) -> bool:
""" """
...@@ -1384,7 +1449,7 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin, TFGenerationMixin, Pu ...@@ -1384,7 +1449,7 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin, TFGenerationMixin, Pu
if not isinstance(dataset, datasets.Dataset): if not isinstance(dataset, datasets.Dataset):
raise TypeError("Dataset argument should be a datasets.Dataset!") raise TypeError("Dataset argument should be a datasets.Dataset!")
model_inputs = list(dict(inspect.signature(self.call).parameters).keys()) model_inputs = list(inspect.signature(self.call).parameters)
model_labels = find_labels(self.__class__) model_labels = find_labels(self.__class__)
if "cols_to_retain" in list(inspect.signature(dataset._get_output_signature).parameters.keys()): if "cols_to_retain" in list(inspect.signature(dataset._get_output_signature).parameters.keys()):
output_signature, _ = dataset._get_output_signature( output_signature, _ = dataset._get_output_signature(
...@@ -1496,7 +1561,7 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin, TFGenerationMixin, Pu ...@@ -1496,7 +1561,7 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin, TFGenerationMixin, Pu
return self.hf_compute_loss(*args, **kwargs) return self.hf_compute_loss(*args, **kwargs)
def get_label_to_output_name_mapping(self): def get_label_to_output_name_mapping(self):
arg_names = list(dict(inspect.signature(self.call).parameters).keys()) arg_names = list(inspect.signature(self.call).parameters)
if self._label_to_output_map is not None: if self._label_to_output_map is not None:
return self._label_to_output_map return self._label_to_output_map
elif "start_positions" in arg_names: elif "start_positions" in arg_names:
...@@ -1519,7 +1584,7 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin, TFGenerationMixin, Pu ...@@ -1519,7 +1584,7 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin, TFGenerationMixin, Pu
""" """
# We hardcode the most common renamings; models with weirder names can set `self._label_to_output_map` # We hardcode the most common renamings; models with weirder names can set `self._label_to_output_map`
arg_names = list(dict(inspect.signature(self.call).parameters).keys()) arg_names = list(inspect.signature(self.call).parameters)
label_kwargs = find_labels(self.__class__) label_kwargs = find_labels(self.__class__)
label_to_output = self.get_label_to_output_name_mapping() label_to_output = self.get_label_to_output_name_mapping()
output_to_label = {val: key for key, val in label_to_output.items()} output_to_label = {val: key for key, val in label_to_output.items()}
...@@ -1626,7 +1691,7 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin, TFGenerationMixin, Pu ...@@ -1626,7 +1691,7 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin, TFGenerationMixin, Pu
that they are available to the model during the forward pass. that they are available to the model during the forward pass.
""" """
# We hardcode the most common renamings; models with weirder names can set `self._label_to_output_map` # We hardcode the most common renamings; models with weirder names can set `self._label_to_output_map`
arg_names = list(dict(inspect.signature(self.call).parameters).keys()) arg_names = list(inspect.signature(self.call).parameters)
label_kwargs = find_labels(self.__class__) label_kwargs = find_labels(self.__class__)
label_to_output = self.get_label_to_output_name_mapping() label_to_output = self.get_label_to_output_name_mapping()
output_to_label = {val: key for key, val in label_to_output.items()} output_to_label = {val: key for key, val in label_to_output.items()}
...@@ -1645,7 +1710,7 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin, TFGenerationMixin, Pu ...@@ -1645,7 +1710,7 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin, TFGenerationMixin, Pu
# When using a dummy loss, we ensure that separate labels are copied to the correct model arguments, # When using a dummy loss, we ensure that separate labels are copied to the correct model arguments,
# if those keys are not already present in the input dict # if those keys are not already present in the input dict
if self._using_dummy_loss and y is not None: if self._using_dummy_loss and y is not None:
arg_names = list(dict(inspect.signature(self.call).parameters).keys()) arg_names = list(inspect.signature(self.call).parameters)
# If y is a tensor and the model only has one label-like input, map y to that input # If y is a tensor and the model only has one label-like input, map y to that input
if len(label_kwargs) == 1 and isinstance(y, tf.Tensor): if len(label_kwargs) == 1 and isinstance(y, tf.Tensor):
if isinstance(x, tf.Tensor): if isinstance(x, tf.Tensor):
......
src/transformers/models/albert/modeling_tf_albert.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,
ModelOutput, ModelOutput,
add_code_sample_docstrings, add_code_sample_docstrings,
add_start_docstrings, add_start_docstrings,
...@@ -826,17 +825,6 @@ class TFAlbertModel(TFAlbertPreTrainedModel): ...@@ -826,17 +825,6 @@ class TFAlbertModel(TFAlbertPreTrainedModel):
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(
""" """
...@@ -933,17 +921,6 @@ class TFAlbertForPreTraining(TFAlbertPreTrainedModel, TFAlbertPreTrainingLoss): ...@@ -933,17 +921,6 @@ class TFAlbertForPreTraining(TFAlbertPreTrainedModel, TFAlbertPreTrainingLoss):
attentions=outputs.attentions, attentions=outputs.attentions,
) )
def serving_output(self, output: TFAlbertForPreTrainingOutput) -> TFAlbertForPreTrainingOutput:
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 TFAlbertForPreTrainingOutput(
prediction_logits=output.prediction_logits,
sop_logits=output.sop_logits,
hidden_states=hs,
attentions=attns,
)
class TFAlbertSOPHead(tf.keras.layers.Layer): class TFAlbertSOPHead(tf.keras.layers.Layer):
def __init__(self, config: AlbertConfig, **kwargs): def __init__(self, config: AlbertConfig, **kwargs):
...@@ -1058,13 +1035,6 @@ class TFAlbertForMaskedLM(TFAlbertPreTrainedModel, TFMaskedLanguageModelingLoss) ...@@ -1058,13 +1035,6 @@ class TFAlbertForMaskedLM(TFAlbertPreTrainedModel, 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)
@add_start_docstrings( @add_start_docstrings(
""" """
...@@ -1147,13 +1117,6 @@ class TFAlbertForSequenceClassification(TFAlbertPreTrainedModel, TFSequenceClass ...@@ -1147,13 +1117,6 @@ class TFAlbertForSequenceClassification(TFAlbertPreTrainedModel, TFSequenceClass
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(
""" """
...@@ -1237,13 +1200,6 @@ class TFAlbertForTokenClassification(TFAlbertPreTrainedModel, TFTokenClassificat ...@@ -1237,13 +1200,6 @@ class TFAlbertForTokenClassification(TFAlbertPreTrainedModel, TFTokenClassificat
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(
""" """
...@@ -1339,15 +1295,6 @@ class TFAlbertForQuestionAnswering(TFAlbertPreTrainedModel, TFQuestionAnsweringL ...@@ -1339,15 +1295,6 @@ class TFAlbertForQuestionAnswering(TFAlbertPreTrainedModel, TFQuestionAnsweringL
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(
""" """
...@@ -1370,16 +1317,6 @@ class TFAlbertForMultipleChoice(TFAlbertPreTrainedModel, TFMultipleChoiceLoss): ...@@ -1370,16 +1317,6 @@ class TFAlbertForMultipleChoice(TFAlbertPreTrainedModel, TFMultipleChoiceLoss):
units=1, kernel_initializer=get_initializer(config.initializer_range), name="classifier" units=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(ALBERT_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length")) @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length"))
@add_code_sample_docstrings( @add_code_sample_docstrings(
...@@ -1457,25 +1394,3 @@ class TFAlbertForMultipleChoice(TFAlbertPreTrainedModel, TFMultipleChoiceLoss): ...@@ -1457,25 +1394,3 @@ class TFAlbertForMultipleChoice(TFAlbertPreTrainedModel, TFMultipleChoiceLoss):
hidden_states=outputs.hidden_states, hidden_states=outputs.hidden_states,
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]) -> TFMultipleChoiceModelOutput:
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/bart/modeling_tf_bart.py
View file @ 814de8fa
...@@ -34,7 +34,6 @@ from ...modeling_tf_outputs import ( ...@@ -34,7 +34,6 @@ 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,
...@@ -487,31 +486,14 @@ class TFBartPretrainedModel(TFPreTrainedModel): ...@@ -487,31 +486,14 @@ class TFBartPretrainedModel(TFPreTrainedModel):
@property @property
def dummy_inputs(self): def dummy_inputs(self):
pad_token = 1 dummy_inputs = super().dummy_inputs
input_ids = tf.convert_to_tensor(DUMMY_INPUTS, dtype=tf.int32) # Dummy inputs should not contain the default val of 1
decoder_input_ids = tf.convert_to_tensor(DUMMY_INPUTS, dtype=tf.int32) # as this is the padding token and some assertions check it
dummy_inputs = { dummy_inputs["input_ids"] = dummy_inputs["input_ids"] * 2
"decoder_input_ids": decoder_input_ids, if "decoder_input_ids" in dummy_inputs:
"attention_mask": tf.cast(input_ids != pad_token, tf.int32), dummy_inputs["decoder_input_ids"] = dummy_inputs["decoder_input_ids"] * 2
"input_ids": input_ids,
}
return dummy_inputs 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)
BART_START_DOCSTRING = r""" BART_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
...@@ -1461,16 +1443,6 @@ class TFBartForConditionalGeneration(TFBartPretrainedModel, TFCausalLanguageMode ...@@ -1461,16 +1443,6 @@ class TFBartForConditionalGeneration(TFBartPretrainedModel, TFCausalLanguageMode
BART_START_DOCSTRING, BART_START_DOCSTRING,
) )
class TFBartForSequenceClassification(TFBartPretrainedModel, TFSequenceClassificationLoss): class TFBartForSequenceClassification(TFBartPretrainedModel, TFSequenceClassificationLoss):
@property
def dummy_inputs(self):
pad_token = self.config.pad_token_id
input_ids = tf.constant([[0, 6, 10, 4, 2], [0, 8, 12, 2, pad_token]])
dummy_inputs = {
"attention_mask": tf.cast(tf.math.not_equal(input_ids, (pad_token)), dtype=tf.int32),
"input_ids": input_ids,
}
return dummy_inputs
def __init__(self, config: BartConfig, load_weight_prefix=None, *inputs, **kwargs): def __init__(self, config: BartConfig, load_weight_prefix=None, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs) super().__init__(config, *inputs, **kwargs)
self.model = TFBartMainLayer(config, load_weight_prefix=load_weight_prefix, name="model") self.model = TFBartMainLayer(config, load_weight_prefix=load_weight_prefix, name="model")
......
src/transformers/models/bert/modeling_tf_bert.py
View file @ 814de8fa
...@@ -54,8 +54,6 @@ from ...modeling_tf_utils import ( ...@@ -54,8 +54,6 @@ 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,
...@@ -903,24 +901,6 @@ class TFBertPreTrainedModel(TFPreTrainedModel): ...@@ -903,24 +901,6 @@ class TFBertPreTrainedModel(TFPreTrainedModel):
config_class = BertConfig config_class = BertConfig
base_model_prefix = "bert" base_model_prefix = "bert"
@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
@dataclass @dataclass
class TFBertForPreTrainingOutput(ModelOutput): class TFBertForPreTrainingOutput(ModelOutput):
...@@ -1123,26 +1103,6 @@ class TFBertModel(TFBertPreTrainedModel): ...@@ -1123,26 +1103,6 @@ class TFBertModel(TFBertPreTrainedModel):
) )
return outputs return outputs
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( @add_start_docstrings(
""" """
...@@ -1255,17 +1215,6 @@ class TFBertForPreTraining(TFBertPreTrainedModel, TFBertPreTrainingLoss): ...@@ -1255,17 +1215,6 @@ class TFBertForPreTraining(TFBertPreTrainedModel, TFBertPreTrainingLoss):
attentions=outputs.attentions, attentions=outputs.attentions,
) )
def serving_output(self, output: TFBertForPreTrainingOutput) -> TFBertForPreTrainingOutput:
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 TFBertForPreTrainingOutput(
prediction_logits=output.prediction_logits,
seq_relationship_logits=output.seq_relationship_logits,
hidden_states=hs,
attentions=attns,
)
@add_start_docstrings("""Bert Model with a `language modeling` head on top.""", BERT_START_DOCSTRING) @add_start_docstrings("""Bert Model with a `language modeling` head on top.""", BERT_START_DOCSTRING)
class TFBertForMaskedLM(TFBertPreTrainedModel, TFMaskedLanguageModelingLoss): class TFBertForMaskedLM(TFBertPreTrainedModel, TFMaskedLanguageModelingLoss):
...@@ -1352,12 +1301,6 @@ class TFBertForMaskedLM(TFBertPreTrainedModel, TFMaskedLanguageModelingLoss): ...@@ -1352,12 +1301,6 @@ class TFBertForMaskedLM(TFBertPreTrainedModel, TFMaskedLanguageModelingLoss):
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)
class TFBertLMHeadModel(TFBertPreTrainedModel, TFCausalLanguageModelingLoss): class TFBertLMHeadModel(TFBertPreTrainedModel, TFCausalLanguageModelingLoss):
# 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
...@@ -1483,19 +1426,6 @@ class TFBertLMHeadModel(TFBertPreTrainedModel, TFCausalLanguageModelingLoss): ...@@ -1483,19 +1426,6 @@ class TFBertLMHeadModel(TFBertPreTrainedModel, TFCausalLanguageModelingLoss):
cross_attentions=outputs.cross_attentions, cross_attentions=outputs.cross_attentions,
) )
def serving_output(self, output: TFCausalLMOutputWithCrossAttentions) -> TFCausalLMOutputWithCrossAttentions:
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 TFCausalLMOutputWithCrossAttentions(
logits=output.logits, past_key_values=pkv, hidden_states=hs, attentions=attns, cross_attentions=cross_attns
)
@add_start_docstrings( @add_start_docstrings(
"""Bert Model with a `next sentence prediction (classification)` head on top.""", """Bert Model with a `next sentence prediction (classification)` head on top.""",
...@@ -1578,12 +1508,6 @@ class TFBertForNextSentencePrediction(TFBertPreTrainedModel, TFNextSentencePredi ...@@ -1578,12 +1508,6 @@ class TFBertForNextSentencePrediction(TFBertPreTrainedModel, TFNextSentencePredi
attentions=outputs.attentions, attentions=outputs.attentions,
) )
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(
""" """
...@@ -1670,12 +1594,6 @@ class TFBertForSequenceClassification(TFBertPreTrainedModel, TFSequenceClassific ...@@ -1670,12 +1594,6 @@ class TFBertForSequenceClassification(TFBertPreTrainedModel, TFSequenceClassific
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(
""" """
...@@ -1698,16 +1616,6 @@ class TFBertForMultipleChoice(TFBertPreTrainedModel, TFMultipleChoiceLoss): ...@@ -1698,16 +1616,6 @@ class TFBertForMultipleChoice(TFBertPreTrainedModel, TFMultipleChoiceLoss):
units=1, kernel_initializer=get_initializer(config.initializer_range), name="classifier" units=1, kernel_initializer=get_initializer(config.initializer_range), name="classifier"
) )
@property
def dummy_inputs(self) -> Dict[str, tf.Tensor]:
"""
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(BERT_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length")) @add_start_docstrings_to_model_forward(BERT_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length"))
@add_code_sample_docstrings( @add_code_sample_docstrings(
...@@ -1785,26 +1693,6 @@ class TFBertForMultipleChoice(TFBertPreTrainedModel, TFMultipleChoiceLoss): ...@@ -1785,26 +1693,6 @@ class TFBertForMultipleChoice(TFBertPreTrainedModel, 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"),
}
]
)
def serving(self, inputs: Dict[str, tf.Tensor]) -> TFMultipleChoiceModelOutput:
output = self.call(input_ids=inputs)
return self.serving_output(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(
""" """
...@@ -1895,12 +1783,6 @@ class TFBertForTokenClassification(TFBertPreTrainedModel, TFTokenClassificationL ...@@ -1895,12 +1783,6 @@ class TFBertForTokenClassification(TFBertPreTrainedModel, TFTokenClassificationL
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(
""" """
...@@ -2002,11 +1884,3 @@ class TFBertForQuestionAnswering(TFBertPreTrainedModel, TFQuestionAnsweringLoss) ...@@ -2002,11 +1884,3 @@ class TFBertForQuestionAnswering(TFBertPreTrainedModel, TFQuestionAnsweringLoss)
hidden_states=outputs.hidden_states, hidden_states=outputs.hidden_states,
attentions=outputs.attentions, attentions=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/blenderbot/modeling_tf_blenderbot.py
View file @ 814de8fa
...@@ -34,7 +34,6 @@ from ...modeling_tf_outputs import ( ...@@ -34,7 +34,6 @@ 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,
...@@ -464,34 +463,6 @@ class TFBlenderbotPreTrainedModel(TFPreTrainedModel): ...@@ -464,34 +463,6 @@ class TFBlenderbotPreTrainedModel(TFPreTrainedModel):
config_class = BlenderbotConfig config_class = BlenderbotConfig
base_model_prefix = "model" base_model_prefix = "model"
@property
def dummy_inputs(self):
pad_token = 1
input_ids = tf.convert_to_tensor(DUMMY_INPUTS, dtype=tf.int32)
decoder_input_ids = tf.convert_to_tensor(DUMMY_INPUTS, dtype=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)
BLENDERBOT_START_DOCSTRING = r""" BLENDERBOT_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/blenderbot_small/modeling_tf_blenderbot_small.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,
...@@ -464,34 +463,6 @@ class TFBlenderbotSmallPreTrainedModel(TFPreTrainedModel): ...@@ -464,34 +463,6 @@ class TFBlenderbotSmallPreTrainedModel(TFPreTrainedModel):
config_class = BlenderbotSmallConfig config_class = BlenderbotSmallConfig
base_model_prefix = "model" base_model_prefix = "model"
@property
def dummy_inputs(self):
pad_token = 1
input_ids = tf.convert_to_tensor(DUMMY_INPUTS, dtype=tf.int32)
decoder_input_ids = tf.convert_to_tensor(DUMMY_INPUTS, dtype=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)
BLENDERBOT_SMALL_START_DOCSTRING = r""" BLENDERBOT_SMALL_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/blip/modeling_tf_blip.py
View file @ 814de8fa
...@@ -17,13 +17,12 @@ ...@@ -17,13 +17,12 @@
from __future__ import annotations from __future__ import annotations
from dataclasses import dataclass from dataclasses import dataclass
from typing import Any, Dict, Optional, Tuple, Union from typing import Any, Optional, Tuple, Union
import tensorflow as tf import tensorflow as tf
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,
TFPreTrainedModel, TFPreTrainedModel,
get_initializer, get_initializer,
get_tf_activation, get_tf_activation,
...@@ -648,38 +647,6 @@ class TFBlipVisionModel(TFBlipPreTrainedModel): ...@@ -648,38 +647,6 @@ class TFBlipVisionModel(TFBlipPreTrainedModel):
self.encoder = TFBlipEncoder(config, name="encoder") self.encoder = TFBlipEncoder(config, name="encoder")
self.post_layernorm = tf.keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="post_layernorm") self.post_layernorm = tf.keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="post_layernorm")
@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)
def serving_output(self, output: TFBaseModelOutputWithPooling) -> TFBaseModelOutputWithPooling: def serving_output(self, output: TFBaseModelOutputWithPooling) -> TFBaseModelOutputWithPooling:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states 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 attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
...@@ -881,44 +848,6 @@ class TFBlipModel(TFBlipPreTrainedModel): ...@@ -881,44 +848,6 @@ class TFBlipModel(TFBlipPreTrainedModel):
self.blip = TFBlipMainLayer(config, name="blip") self.blip = TFBlipMainLayer(config, name="blip")
@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.float32, name="pixel_values"),
"attention_mask": tf.TensorSpec((None, None), tf.int32, name="attention_mask"),
}
]
)
def serving(self, inputs: Dict[str, tf.Tensor]) -> TFBlipOutput:
"""
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)
def serving_output(self, output: TFBlipOutput) -> TFBlipOutput: def serving_output(self, output: TFBlipOutput) -> TFBlipOutput:
return TFBlipOutput( return TFBlipOutput(
logits_per_image=output.logits_per_image, logits_per_image=output.logits_per_image,
...@@ -1082,48 +1011,6 @@ class TFBlipForConditionalGeneration(TFBlipPreTrainedModel): ...@@ -1082,48 +1011,6 @@ class TFBlipForConditionalGeneration(TFBlipPreTrainedModel):
def get_input_embeddings(self) -> tf.keras.layers.Layer: def get_input_embeddings(self) -> tf.keras.layers.Layer:
return self.vision_model.embeddings.patch_embedding return self.vision_model.embeddings.patch_embedding
@property
def dummy_inputs(self):
input_ids = tf.constant(DUMMY_INPUTS, dtype=tf.int32)
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": input_ids, "pixel_values": VISION_DUMMY_INPUTS}
@tf.function(
input_signature=[
{
"pixel_values": tf.TensorSpec((None, None, None, None), tf.float32, name="pixel_values"),
"input_ids": tf.TensorSpec((None, None), tf.int32, name="input_ids"),
}
]
)
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)
def serving_output(
self, output: TFBlipForConditionalGenerationModelOutput
) -> TFBlipForConditionalGenerationModelOutput:
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 TFBlipForConditionalGenerationModelOutput(
last_hidden_state=output.last_hidden_state,
image_embeds=output.image_embeds,
hidden_states=hs,
attentions=attns,
)
@unpack_inputs @unpack_inputs
@add_start_docstrings_to_model_forward(BLIP_VISION_INPUTS_DOCSTRING) @add_start_docstrings_to_model_forward(BLIP_VISION_INPUTS_DOCSTRING)
@replace_return_docstrings(output_type=TFBlipForConditionalGenerationModelOutput, config_class=BlipConfig) @replace_return_docstrings(output_type=TFBlipForConditionalGenerationModelOutput, config_class=BlipConfig)
...@@ -1297,46 +1184,30 @@ class TFBlipForQuestionAnswering(TFBlipPreTrainedModel): ...@@ -1297,46 +1184,30 @@ class TFBlipForQuestionAnswering(TFBlipPreTrainedModel):
def get_input_embeddings(self) -> tf.keras.layers.Layer: def get_input_embeddings(self) -> tf.keras.layers.Layer:
return self.vision_model.embeddings.patch_embedding return self.vision_model.embeddings.patch_embedding
@property # Adapted from transformers.models.t5.modeling_tf_t5.TFT5PreTrainedModel._shift_right
def dummy_inputs(self): def _shift_right(self, input_ids):
input_ids = tf.constant(DUMMY_INPUTS, dtype=tf.int32) decoder_start_token_id = self.decoder_start_token_id
VISION_DUMMY_INPUTS = tf.random.uniform( pad_token_id = self.decoder_pad_token_id
shape=(len(DUMMY_INPUTS), 3, self.config.vision_config.image_size, self.config.vision_config.image_size),
dtype=tf.float32,
)
return {"input_ids": input_ids, "pixel_values": VISION_DUMMY_INPUTS, "decoder_input_ids": input_ids}
@tf.function(
input_signature=[
{
"pixel_values": tf.TensorSpec((None, None, None, None), tf.float32, name="pixel_values"),
"input_ids": tf.TensorSpec((None, None), tf.int32, name="input_ids"),
}
]
)
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) if decoder_start_token_id is None or pad_token_id is None:
raise ValueError("decoder_start_token_id and pad_token_id must be defined!")
def serving_output(self, output: TFBlipTextVisionModelOutput) -> TFBlipTextVisionModelOutput: start_tokens = tf.fill((shape_list(input_ids)[0], 1), decoder_start_token_id)
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None start_tokens = tf.cast(start_tokens, input_ids.dtype) # Ensure compatible dtypes for concatenation
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None shifted_input_ids = tf.concat([start_tokens, input_ids[:, :-1]], -1)
return TFBlipTextVisionModelOutput( # replace possible -100 values in labels by `pad_token_id`
image_embeds=output.image_embeds, shifted_input_ids = tf.where(
last_hidden_state=output.last_hidden_state, shifted_input_ids == -100,
hidden_states=hs, tf.cast(tf.fill(shape_list(shifted_input_ids), pad_token_id), shifted_input_ids.dtype),
attentions=attns, shifted_input_ids,
) )
# "Verify that `labels` has only positive values and -100"
tf.debugging.assert_greater_equal(shifted_input_ids, tf.constant(0, dtype=shifted_input_ids.dtype))
return shifted_input_ids
@unpack_inputs @unpack_inputs
@add_start_docstrings_to_model_forward(BLIP_VISION_INPUTS_DOCSTRING) @add_start_docstrings_to_model_forward(BLIP_VISION_INPUTS_DOCSTRING)
@replace_return_docstrings(output_type=TFBlipTextVisionModelOutput, config_class=BlipVisionConfig) @replace_return_docstrings(output_type=TFBlipTextVisionModelOutput, config_class=BlipVisionConfig)
...@@ -1389,7 +1260,7 @@ class TFBlipForQuestionAnswering(TFBlipPreTrainedModel): ...@@ -1389,7 +1260,7 @@ class TFBlipForQuestionAnswering(TFBlipPreTrainedModel):
```""" ```"""
if labels is None and decoder_input_ids is None: if labels is None and decoder_input_ids is None:
raise ValueError( raise ValueError(
"Either `decoder_input_ids` or `labels` should be passed when calling `forward` with" "Either `decoder_input_ids` or `labels` should be passed when calling"
" `TFBlipForQuestionAnswering`. if you are training the model make sure that `labels` is passed, if you" " `TFBlipForQuestionAnswering`. if you are training the model make sure that `labels` is passed, if you"
" are using the model for inference make sure that `decoder_input_ids` is passed or call `generate`" " are using the model for inference make sure that `decoder_input_ids` is passed or call `generate`"
) )
...@@ -1579,47 +1450,6 @@ class TFBlipForImageTextRetrieval(TFBlipPreTrainedModel): ...@@ -1579,47 +1450,6 @@ class TFBlipForImageTextRetrieval(TFBlipPreTrainedModel):
def get_input_embeddings(self) -> tf.keras.layers.Layer: def get_input_embeddings(self) -> tf.keras.layers.Layer:
return self.vision_model.embeddings.patch_embedding return self.vision_model.embeddings.patch_embedding
@property
def dummy_inputs(self):
input_ids = tf.constant(DUMMY_INPUTS, dtype=tf.int32)
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": input_ids, "pixel_values": VISION_DUMMY_INPUTS}
@tf.function(
input_signature=[
{
"pixel_values": tf.TensorSpec((None, None, None, None), tf.float32, name="pixel_values"),
"input_ids": tf.TensorSpec((None, None), tf.int32, name="input_ids"),
}
]
)
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)
def serving_output(self, output: TFBlipImageTextMatchingModelOutput) -> TFBlipImageTextMatchingModelOutput:
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 TFBlipImageTextMatchingModelOutput(
itm_score=output.itm_score,
last_hidden_state=hs,
hidden_states=output.hidden_states,
attentions=attns,
question_embeds=output.question_embeds,
)
@unpack_inputs @unpack_inputs
@add_start_docstrings_to_model_forward(BLIP_VISION_INPUTS_DOCSTRING) @add_start_docstrings_to_model_forward(BLIP_VISION_INPUTS_DOCSTRING)
@replace_return_docstrings(output_type=TFBlipImageTextMatchingModelOutput, config_class=BlipVisionConfig) @replace_return_docstrings(output_type=TFBlipImageTextMatchingModelOutput, config_class=BlipVisionConfig)
......
src/transformers/models/blip/modeling_tf_blip_text.py
View file @ 814de8fa
...@@ -17,7 +17,7 @@ ...@@ -17,7 +17,7 @@
from __future__ import annotations from __future__ import annotations
import math import math
from typing import Dict, Optional, Tuple from typing import Optional, Tuple
import tensorflow as tf import tensorflow as tf
...@@ -27,7 +27,6 @@ from ...modeling_tf_outputs import ( ...@@ -27,7 +27,6 @@ from ...modeling_tf_outputs import (
TFCausalLMOutputWithCrossAttentions, TFCausalLMOutputWithCrossAttentions,
) )
from ...modeling_tf_utils import ( from ...modeling_tf_utils import (
DUMMY_INPUTS,
TFPreTrainedModel, TFPreTrainedModel,
get_initializer, get_initializer,
get_tf_activation, get_tf_activation,
...@@ -593,31 +592,6 @@ class TFBlipTextModel(TFBlipTextPreTrainedModel): ...@@ -593,31 +592,6 @@ class TFBlipTextModel(TFBlipTextPreTrainedModel):
self.encoder = TFBlipTextEncoder(config, name="encoder") self.encoder = TFBlipTextEncoder(config, name="encoder")
self.pooler = TFBlipTextPooler(config, name="pooler") if add_pooling_layer else None self.pooler = TFBlipTextPooler(config, name="pooler") if add_pooling_layer else None
@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]) -> TFBaseModelOutputWithPoolingAndCrossAttentions:
output = self.call(inputs)
return self.serving_output(output)
def serving_output(
self, output: TFBaseModelOutputWithPoolingAndCrossAttentions
) -> TFBaseModelOutputWithPoolingAndCrossAttentions:
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 TFBaseModelOutputWithPoolingAndCrossAttentions(
last_hidden_state=output.last_hidden_state,
pooler_output=output.pooler_output,
hidden_states=hs,
attentions=attns,
)
def get_input_embeddings(self): def get_input_embeddings(self):
return self.embeddings.word_embeddings return self.embeddings.word_embeddings
...@@ -844,46 +818,6 @@ class TFBlipTextLMHeadModel(TFBlipTextPreTrainedModel): ...@@ -844,46 +818,6 @@ class TFBlipTextLMHeadModel(TFBlipTextPreTrainedModel):
def set_output_embeddings(self, new_embeddings): def set_output_embeddings(self, new_embeddings):
self.cls.predictions.decoder = new_embeddings self.cls.predictions.decoder = new_embeddings
@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"),
}
]
)
def serving(self, inputs: Dict[str, tf.Tensor]) -> TFCausalLMOutputWithCrossAttentions:
"""
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)
def serving_output(self, output: TFCausalLMOutputWithCrossAttentions) -> TFCausalLMOutputWithCrossAttentions:
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 TFCausalLMOutputWithCrossAttentions(
logits=output.logits,
cross_attentions=output.cross_attentions,
hidden_states=hs,
attentions=attns,
)
@add_start_docstrings_to_model_forward(BLIP_TEXT_INPUTS_DOCSTRING) @add_start_docstrings_to_model_forward(BLIP_TEXT_INPUTS_DOCSTRING)
@unpack_inputs @unpack_inputs
def call( def call(
......
src/transformers/models/camembert/modeling_tf_camembert.py
View file @ 814de8fa
...@@ -51,8 +51,6 @@ from ...modeling_tf_utils import ( ...@@ -51,8 +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 (
DUMMY_INPUTS,
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,
...@@ -873,38 +871,6 @@ class TFCamembertPreTrainedModel(TFPreTrainedModel): ...@@ -873,38 +871,6 @@ class TFCamembertPreTrainedModel(TFPreTrainedModel):
config_class = CamembertConfig config_class = CamembertConfig
base_model_prefix = "roberta" base_model_prefix = "roberta"
@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
@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)
@add_start_docstrings( @add_start_docstrings(
"The bare CamemBERT Model transformer outputting raw hidden-states without any specific head on top.", "The bare CamemBERT Model transformer outputting raw hidden-states without any specific head on top.",
...@@ -979,27 +945,6 @@ class TFCamembertModel(TFCamembertPreTrainedModel): ...@@ -979,27 +945,6 @@ class TFCamembertModel(TFCamembertPreTrainedModel):
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,
)
# Copied from transformers.models.roberta.modeling_tf_roberta.TFRobertaLMHead with Roberta->Camembert # Copied from transformers.models.roberta.modeling_tf_roberta.TFRobertaLMHead with Roberta->Camembert
class TFCamembertLMHead(tf.keras.layers.Layer): class TFCamembertLMHead(tf.keras.layers.Layer):
...@@ -1135,13 +1080,6 @@ class TFCamembertForMaskedLM(TFCamembertPreTrainedModel, TFMaskedLanguageModelin ...@@ -1135,13 +1080,6 @@ class TFCamembertForMaskedLM(TFCamembertPreTrainedModel, TFMaskedLanguageModelin
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)
# Copied from transformers.models.roberta.modeling_tf_roberta.TFRobertaClassificationHead # Copied from transformers.models.roberta.modeling_tf_roberta.TFRobertaClassificationHead
class TFCamembertClassificationHead(tf.keras.layers.Layer): class TFCamembertClassificationHead(tf.keras.layers.Layer):
...@@ -1248,13 +1186,6 @@ class TFCamembertForSequenceClassification(TFCamembertPreTrainedModel, TFSequenc ...@@ -1248,13 +1186,6 @@ class TFCamembertForSequenceClassification(TFCamembertPreTrainedModel, TFSequenc
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(
""" """
...@@ -1339,13 +1270,6 @@ class TFCamembertForTokenClassification(TFCamembertPreTrainedModel, TFTokenClass ...@@ -1339,13 +1270,6 @@ class TFCamembertForTokenClassification(TFCamembertPreTrainedModel, TFTokenClass
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(
""" """
...@@ -1369,16 +1293,6 @@ class TFCamembertForMultipleChoice(TFCamembertPreTrainedModel, TFMultipleChoiceL ...@@ -1369,16 +1293,6 @@ class TFCamembertForMultipleChoice(TFCamembertPreTrainedModel, TFMultipleChoiceL
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(
CAMEMBERT_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length") CAMEMBERT_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length")
...@@ -1449,26 +1363,6 @@ class TFCamembertForMultipleChoice(TFCamembertPreTrainedModel, TFMultipleChoiceL ...@@ -1449,26 +1363,6 @@ class TFCamembertForMultipleChoice(TFCamembertPreTrainedModel, TFMultipleChoiceL
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(
""" """
...@@ -1562,15 +1456,6 @@ class TFCamembertForQuestionAnswering(TFCamembertPreTrainedModel, TFQuestionAnsw ...@@ -1562,15 +1456,6 @@ class TFCamembertForQuestionAnswering(TFCamembertPreTrainedModel, TFQuestionAnsw
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(
"""CamemBERT Model with a `language modeling` head on top for CLM fine-tuning.""", CAMEMBERT_START_DOCSTRING """CamemBERT Model with a `language modeling` head on top for CLM fine-tuning.""", CAMEMBERT_START_DOCSTRING
...@@ -1696,17 +1581,3 @@ class TFCamembertForCausalLM(TFCamembertPreTrainedModel, TFCausalLanguageModelin ...@@ -1696,17 +1581,3 @@ class TFCamembertForCausalLM(TFCamembertPreTrainedModel, TFCausalLanguageModelin
attentions=outputs.attentions, attentions=outputs.attentions,
cross_attentions=outputs.cross_attentions, cross_attentions=outputs.cross_attentions,
) )
# Copied from transformers.models.bert.modeling_tf_bert.TFBertLMHeadModel.serving_output
def serving_output(self, output: TFCausalLMOutputWithCrossAttentions) -> TFCausalLMOutputWithCrossAttentions:
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 TFCausalLMOutputWithCrossAttentions(
logits=output.logits, past_key_values=pkv, hidden_states=hs, attentions=attns, cross_attentions=cross_attns
)
src/transformers/models/clip/modeling_tf_clip.py
View file @ 814de8fa
...@@ -19,7 +19,7 @@ from __future__ import annotations ...@@ -19,7 +19,7 @@ from __future__ import annotations
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
...@@ -29,7 +29,6 @@ from ...modeling_tf_outputs import TFBaseModelOutput, TFBaseModelOutputWithPooli ...@@ -29,7 +29,6 @@ from ...modeling_tf_outputs import TFBaseModelOutput, TFBaseModelOutputWithPooli
# Public API # Public API
from ...modeling_tf_utils import ( from ...modeling_tf_utils import (
DUMMY_INPUTS,
TFModelInputType, TFModelInputType,
TFPreTrainedModel, TFPreTrainedModel,
get_initializer, get_initializer,
...@@ -1090,29 +1089,6 @@ class TFCLIPTextModel(TFCLIPPreTrainedModel): ...@@ -1090,29 +1089,6 @@ class TFCLIPTextModel(TFCLIPPreTrainedModel):
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: Dict[str, tf.Tensor]) -> TFBaseModelOutputWithPooling:
output = self.call(inputs)
return self.serving_output(output)
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 TFCLIPVisionModel(TFCLIPPreTrainedModel): class TFCLIPVisionModel(TFCLIPPreTrainedModel):
config_class = CLIPVisionConfig config_class = CLIPVisionConfig
...@@ -1123,38 +1099,6 @@ class TFCLIPVisionModel(TFCLIPPreTrainedModel): ...@@ -1123,38 +1099,6 @@ class TFCLIPVisionModel(TFCLIPPreTrainedModel):
self.clip = TFCLIPVisionMainLayer(config, name="clip") self.clip = TFCLIPVisionMainLayer(config, name="clip")
@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(CLIP_VISION_INPUTS_DOCSTRING) @add_start_docstrings_to_model_forward(CLIP_VISION_INPUTS_DOCSTRING)
@replace_return_docstrings(output_type=TFBaseModelOutputWithPooling, config_class=CLIPVisionConfig) @replace_return_docstrings(output_type=TFBaseModelOutputWithPooling, config_class=CLIPVisionConfig)
...@@ -1199,17 +1143,6 @@ class TFCLIPVisionModel(TFCLIPPreTrainedModel): ...@@ -1199,17 +1143,6 @@ class TFCLIPVisionModel(TFCLIPPreTrainedModel):
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(CLIP_START_DOCSTRING) @add_start_docstrings(CLIP_START_DOCSTRING)
class TFCLIPModel(TFCLIPPreTrainedModel): class TFCLIPModel(TFCLIPPreTrainedModel):
...@@ -1220,44 +1153,6 @@ class TFCLIPModel(TFCLIPPreTrainedModel): ...@@ -1220,44 +1153,6 @@ class TFCLIPModel(TFCLIPPreTrainedModel):
self.clip = TFCLIPMainLayer(config, name="clip") self.clip = TFCLIPMainLayer(config, name="clip")
@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.float32, name="pixel_values"),
"attention_mask": tf.TensorSpec((None, None), tf.int32, name="attention_mask"),
}
]
)
def serving(self, inputs: Dict[str, tf.Tensor]) -> TFCLIPOutput:
"""
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(CLIP_TEXT_INPUTS_DOCSTRING.format("batch_size, sequence_length")) @add_start_docstrings_to_model_forward(CLIP_TEXT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
def get_text_features( def get_text_features(
......
src/transformers/models/convbert/modeling_tf_convbert.py
View file @ 814de8fa
...@@ -46,7 +46,6 @@ from ...modeling_tf_utils import ( ...@@ -46,7 +46,6 @@ 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,
...@@ -770,12 +769,6 @@ class TFConvBertModel(TFConvBertPreTrainedModel): ...@@ -770,12 +769,6 @@ class TFConvBertModel(TFConvBertPreTrainedModel):
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
return TFBaseModelOutput(last_hidden_state=output.last_hidden_state, hidden_states=hs, attentions=attns)
class TFConvBertMaskedLMHead(tf.keras.layers.Layer): class TFConvBertMaskedLMHead(tf.keras.layers.Layer):
def __init__(self, config, input_embeddings, **kwargs): def __init__(self, config, input_embeddings, **kwargs):
...@@ -907,13 +900,6 @@ class TFConvBertForMaskedLM(TFConvBertPreTrainedModel, TFMaskedLanguageModelingL ...@@ -907,13 +900,6 @@ class TFConvBertForMaskedLM(TFConvBertPreTrainedModel, TFMaskedLanguageModelingL
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):
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 TFConvBertClassificationHead(tf.keras.layers.Layer): class TFConvBertClassificationHead(tf.keras.layers.Layer):
"""Head for sentence-level classification tasks.""" """Head for sentence-level classification tasks."""
...@@ -1012,12 +998,6 @@ class TFConvBertForSequenceClassification(TFConvBertPreTrainedModel, TFSequenceC ...@@ -1012,12 +998,6 @@ class TFConvBertForSequenceClassification(TFConvBertPreTrainedModel, TFSequenceC
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 TFSequenceClassifierOutput(logits=output.logits, hidden_states=hs, attentions=attns)
@add_start_docstrings( @add_start_docstrings(
""" """
...@@ -1038,16 +1018,6 @@ class TFConvBertForMultipleChoice(TFConvBertPreTrainedModel, TFMultipleChoiceLos ...@@ -1038,16 +1018,6 @@ class TFConvBertForMultipleChoice(TFConvBertPreTrainedModel, TFMultipleChoiceLos
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.convert_to_tensor(MULTIPLE_CHOICE_DUMMY_INPUTS, dtype=tf.int32)}
@unpack_inputs @unpack_inputs
@add_start_docstrings_to_model_forward( @add_start_docstrings_to_model_forward(
CONVBERT_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length") CONVBERT_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length")
...@@ -1121,26 +1091,6 @@ class TFConvBertForMultipleChoice(TFConvBertPreTrainedModel, TFMultipleChoiceLos ...@@ -1121,26 +1091,6 @@ class TFConvBertForMultipleChoice(TFConvBertPreTrainedModel, TFMultipleChoiceLos
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"),
}
]
)
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 TFMultipleChoiceModelOutput(logits=output.logits, hidden_states=hs, attentions=attns)
@add_start_docstrings( @add_start_docstrings(
""" """
...@@ -1216,12 +1166,6 @@ class TFConvBertForTokenClassification(TFConvBertPreTrainedModel, TFTokenClassif ...@@ -1216,12 +1166,6 @@ class TFConvBertForTokenClassification(TFConvBertPreTrainedModel, TFTokenClassif
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 TFTokenClassifierOutput(logits=output.logits, hidden_states=hs, attentions=attns)
@add_start_docstrings( @add_start_docstrings(
""" """
...@@ -1307,11 +1251,3 @@ class TFConvBertForQuestionAnswering(TFConvBertPreTrainedModel, TFQuestionAnswer ...@@ -1307,11 +1251,3 @@ class TFConvBertForQuestionAnswering(TFConvBertPreTrainedModel, TFQuestionAnswer
hidden_states=outputs.hidden_states, hidden_states=outputs.hidden_states,
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 TFQuestionAnsweringModelOutput(
start_logits=output.start_logits, end_logits=output.end_logits, hidden_states=hs, attentions=attns
)
src/transformers/models/convnext/modeling_tf_convnext.py
View file @ 814de8fa
...@@ -17,7 +17,7 @@ ...@@ -17,7 +17,7 @@
from __future__ import annotations from __future__ import annotations
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
...@@ -351,43 +351,6 @@ class TFConvNextPreTrainedModel(TFPreTrainedModel): ...@@ -351,43 +351,6 @@ class TFConvNextPreTrainedModel(TFPreTrainedModel):
base_model_prefix = "convnext" base_model_prefix = "convnext"
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)
CONVNEXT_START_DOCSTRING = r""" CONVNEXT_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
...@@ -509,14 +472,6 @@ class TFConvNextModel(TFConvNextPreTrainedModel): ...@@ -509,14 +472,6 @@ class TFConvNextModel(TFConvNextPreTrainedModel):
hidden_states=outputs.hidden_states, hidden_states=outputs.hidden_states,
) )
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(
""" """
...@@ -609,7 +564,3 @@ class TFConvNextForImageClassification(TFConvNextPreTrainedModel, TFSequenceClas ...@@ -609,7 +564,3 @@ class TFConvNextForImageClassification(TFConvNextPreTrainedModel, TFSequenceClas
logits=logits, logits=logits,
hidden_states=outputs.hidden_states, hidden_states=outputs.hidden_states,
) )
def serving_output(self, output: TFSequenceClassifierOutput) -> TFSequenceClassifierOutput:
# hidden_states not converted to Tensor with tf.convert_to_tensor as they are all of different dimensions
return TFSequenceClassifierOutput(logits=output.logits, hidden_states=output.hidden_states)
src/transformers/models/ctrl/modeling_tf_ctrl.py
View file @ 814de8fa
...@@ -564,15 +564,6 @@ class TFCTRLModel(TFCTRLPreTrainedModel): ...@@ -564,15 +564,6 @@ class TFCTRLModel(TFCTRLPreTrainedModel):
) )
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
)
class TFCTRLLMHead(tf.keras.layers.Layer): class TFCTRLLMHead(tf.keras.layers.Layer):
def __init__(self, config, input_embeddings, **kwargs): def __init__(self, config, input_embeddings, **kwargs):
...@@ -705,13 +696,6 @@ class TFCTRLLMHeadModel(TFCTRLPreTrainedModel, TFCausalLanguageModelingLoss): ...@@ -705,13 +696,6 @@ class TFCTRLLMHeadModel(TFCTRLPreTrainedModel, 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(
""" """
...@@ -839,10 +823,3 @@ class TFCTRLForSequenceClassification(TFCTRLPreTrainedModel, TFSequenceClassific ...@@ -839,10 +823,3 @@ class TFCTRLForSequenceClassification(TFCTRLPreTrainedModel, TFSequenceClassific
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)
src/transformers/models/cvt/modeling_tf_cvt.py
View file @ 814de8fa
...@@ -19,7 +19,7 @@ from __future__ import annotations ...@@ -19,7 +19,7 @@ from __future__ import annotations
import collections.abc import collections.abc
from dataclasses import dataclass from dataclasses import dataclass
from typing import Dict, Optional, Tuple, Union from typing import Optional, Tuple, Union
import tensorflow as tf import tensorflow as tf
...@@ -707,35 +707,6 @@ class TFCvtPreTrainedModel(TFPreTrainedModel): ...@@ -707,35 +707,6 @@ class TFCvtPreTrainedModel(TFPreTrainedModel):
base_model_prefix = "cvt" base_model_prefix = "cvt"
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, 224, 224), 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)
TFCVT_START_DOCSTRING = r""" TFCVT_START_DOCSTRING = r"""
...@@ -844,13 +815,6 @@ class TFCvtModel(TFCvtPreTrainedModel): ...@@ -844,13 +815,6 @@ class TFCvtModel(TFCvtPreTrainedModel):
hidden_states=outputs.hidden_states, hidden_states=outputs.hidden_states,
) )
def serving_output(self, output: TFBaseModelOutputWithCLSToken) -> TFBaseModelOutputWithCLSToken:
return TFBaseModelOutputWithCLSToken(
last_hidden_state=output.last_hidden_state,
cls_token_value=output.cls_token_value,
hidden_states=output.hidden_states,
)
@add_start_docstrings( @add_start_docstrings(
""" """
...@@ -945,6 +909,3 @@ class TFCvtForImageClassification(TFCvtPreTrainedModel, TFSequenceClassification ...@@ -945,6 +909,3 @@ class TFCvtForImageClassification(TFCvtPreTrainedModel, TFSequenceClassification
return ((loss,) + output) if loss is not None else output return ((loss,) + output) if loss is not None else output
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:
return TFImageClassifierOutputWithNoAttention(logits=output.logits, hidden_states=output.hidden_states)
src/transformers/models/data2vec/modeling_tf_data2vec_vision.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 Dict, List, Optional, Tuple, Union from typing import List, Optional, Tuple, Union
import numpy as np import numpy as np
import tensorflow as tf import tensorflow as tf
...@@ -774,36 +774,6 @@ class TFData2VecVisionPreTrainedModel(TFPreTrainedModel): ...@@ -774,36 +774,6 @@ class TFData2VecVisionPreTrainedModel(TFPreTrainedModel):
main_input_name = "pixel_values" main_input_name = "pixel_values"
_keys_to_ignore_on_load_unexpected = [r"relative_position_index"] _keys_to_ignore_on_load_unexpected = [r"relative_position_index"]
@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)
DATA2VEC_VISION_START_DOCSTRING = r""" DATA2VEC_VISION_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
...@@ -926,17 +896,6 @@ class TFData2VecVisionModel(TFData2VecVisionPreTrainedModel): ...@@ -926,17 +896,6 @@ class TFData2VecVisionModel(TFData2VecVisionPreTrainedModel):
return outputs return outputs
def serving_output(self, output: TFData2VecVisionModelOutputWithPooling) -> TFData2VecVisionModelOutputWithPooling:
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 TFData2VecVisionModelOutputWithPooling(
last_hidden_state=output.last_hidden_state,
pooler_output=output.pooler_output,
hidden_states=hidden_states,
attentions=attentions,
)
@add_start_docstrings( @add_start_docstrings(
""" """
...@@ -1009,12 +968,6 @@ class TFData2VecVisionForImageClassification(TFData2VecVisionPreTrainedModel, TF ...@@ -1009,12 +968,6 @@ class TFData2VecVisionForImageClassification(TFData2VecVisionPreTrainedModel, TF
attentions=outputs.attentions, attentions=outputs.attentions,
) )
def serving_output(self, output: TFSequenceClassifierOutput) -> TFSequenceClassifierOutput:
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 TFSequenceClassifierOutput(logits=output.logits, hidden_states=hidden_states, attentions=attentions)
class TFData2VecVisionConvModule(tf.keras.layers.Layer): class TFData2VecVisionConvModule(tf.keras.layers.Layer):
""" """
...@@ -1475,9 +1428,3 @@ class TFData2VecVisionForSemanticSegmentation(TFData2VecVisionPreTrainedModel): ...@@ -1475,9 +1428,3 @@ class TFData2VecVisionForSemanticSegmentation(TFData2VecVisionPreTrainedModel):
hidden_states=outputs.hidden_states if output_hidden_states else None, hidden_states=outputs.hidden_states if output_hidden_states else None,
attentions=outputs.attentions, attentions=outputs.attentions,
) )
def serving_output(self, output: TFSemanticSegmenterOutput) -> TFSemanticSegmenterOutput:
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 TFSemanticSegmenterOutput(logits=output.logits, hidden_states=hidden_states, attentions=attentions)
src/transformers/models/deberta/modeling_tf_deberta.py
View file @ 814de8fa
...@@ -1118,12 +1118,6 @@ class TFDebertaModel(TFDebertaPreTrainedModel): ...@@ -1118,12 +1118,6 @@ class TFDebertaModel(TFDebertaPreTrainedModel):
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)
@add_start_docstrings("""DeBERTa Model with a `language modeling` head on top.""", DEBERTA_START_DOCSTRING) @add_start_docstrings("""DeBERTa Model with a `language modeling` head on top.""", DEBERTA_START_DOCSTRING)
class TFDebertaForMaskedLM(TFDebertaPreTrainedModel, TFMaskedLanguageModelingLoss): class TFDebertaForMaskedLM(TFDebertaPreTrainedModel, TFMaskedLanguageModelingLoss):
...@@ -1194,12 +1188,6 @@ class TFDebertaForMaskedLM(TFDebertaPreTrainedModel, TFMaskedLanguageModelingLos ...@@ -1194,12 +1188,6 @@ class TFDebertaForMaskedLM(TFDebertaPreTrainedModel, TFMaskedLanguageModelingLos
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(
""" """
...@@ -1281,12 +1269,6 @@ class TFDebertaForSequenceClassification(TFDebertaPreTrainedModel, TFSequenceCla ...@@ -1281,12 +1269,6 @@ class TFDebertaForSequenceClassification(TFDebertaPreTrainedModel, TFSequenceCla
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(
""" """
...@@ -1358,12 +1340,6 @@ class TFDebertaForTokenClassification(TFDebertaPreTrainedModel, TFTokenClassific ...@@ -1358,12 +1340,6 @@ class TFDebertaForTokenClassification(TFDebertaPreTrainedModel, TFTokenClassific
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(
""" """
...@@ -1448,11 +1424,3 @@ class TFDebertaForQuestionAnswering(TFDebertaPreTrainedModel, TFQuestionAnswerin ...@@ -1448,11 +1424,3 @@ class TFDebertaForQuestionAnswering(TFDebertaPreTrainedModel, TFQuestionAnswerin
hidden_states=outputs.hidden_states, hidden_states=outputs.hidden_states,
attentions=outputs.attentions, attentions=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/deberta_v2/modeling_tf_deberta_v2.py
View file @ 814de8fa
...@@ -1212,12 +1212,6 @@ class TFDebertaV2Model(TFDebertaV2PreTrainedModel): ...@@ -1212,12 +1212,6 @@ class TFDebertaV2Model(TFDebertaV2PreTrainedModel):
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)
@add_start_docstrings("""DeBERTa Model with a `language modeling` head on top.""", DEBERTA_START_DOCSTRING) @add_start_docstrings("""DeBERTa Model with a `language modeling` head on top.""", DEBERTA_START_DOCSTRING)
# Copied from transformers.models.deberta.modeling_tf_deberta.TFDebertaForMaskedLM with Deberta->DebertaV2 # Copied from transformers.models.deberta.modeling_tf_deberta.TFDebertaForMaskedLM with Deberta->DebertaV2
...@@ -1289,12 +1283,6 @@ class TFDebertaV2ForMaskedLM(TFDebertaV2PreTrainedModel, TFMaskedLanguageModelin ...@@ -1289,12 +1283,6 @@ class TFDebertaV2ForMaskedLM(TFDebertaV2PreTrainedModel, TFMaskedLanguageModelin
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(
""" """
...@@ -1377,12 +1365,6 @@ class TFDebertaV2ForSequenceClassification(TFDebertaV2PreTrainedModel, TFSequenc ...@@ -1377,12 +1365,6 @@ class TFDebertaV2ForSequenceClassification(TFDebertaV2PreTrainedModel, TFSequenc
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(
""" """
...@@ -1455,12 +1437,6 @@ class TFDebertaV2ForTokenClassification(TFDebertaV2PreTrainedModel, TFTokenClass ...@@ -1455,12 +1437,6 @@ class TFDebertaV2ForTokenClassification(TFDebertaV2PreTrainedModel, TFTokenClass
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(
""" """
...@@ -1546,11 +1522,3 @@ class TFDebertaV2ForQuestionAnswering(TFDebertaV2PreTrainedModel, TFQuestionAnsw ...@@ -1546,11 +1522,3 @@ class TFDebertaV2ForQuestionAnswering(TFDebertaV2PreTrainedModel, TFQuestionAnsw
hidden_states=outputs.hidden_states, hidden_states=outputs.hidden_states,
attentions=outputs.attentions, attentions=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/deit/modeling_tf_deit.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 Dict, Optional, Tuple, Union from typing import Optional, Tuple, Union
import tensorflow as tf import tensorflow as tf
...@@ -568,38 +568,6 @@ class TFDeiTPreTrainedModel(TFPreTrainedModel): ...@@ -568,38 +568,6 @@ class TFDeiTPreTrainedModel(TFPreTrainedModel):
base_model_prefix = "deit" base_model_prefix = "deit"
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)
DEIT_START_DOCSTRING = r""" DEIT_START_DOCSTRING = r"""
This model is a TensorFlow This model is a TensorFlow
...@@ -679,17 +647,6 @@ class TFDeiTModel(TFDeiTPreTrainedModel): ...@@ -679,17 +647,6 @@ class TFDeiTModel(TFDeiTPreTrainedModel):
) )
return outputs return outputs
def serving_output(self, output: TFBaseModelOutputWithPooling) -> TFBaseModelOutputWithPooling:
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 TFBaseModelOutputWithPooling(
last_hidden_state=output.last_hidden_state,
pooler_output=output.pooler_output,
hidden_states=hidden_states,
attentions=attentions,
)
# Copied from transformers.models.vit.modeling_tf_vit.TFViTPooler with ViT->DeiT # Copied from transformers.models.vit.modeling_tf_vit.TFViTPooler with ViT->DeiT
class TFDeiTPooler(tf.keras.layers.Layer): class TFDeiTPooler(tf.keras.layers.Layer):
...@@ -865,14 +822,6 @@ class TFDeiTForMaskedImageModeling(TFDeiTPreTrainedModel): ...@@ -865,14 +822,6 @@ class TFDeiTForMaskedImageModeling(TFDeiTPreTrainedModel):
attentions=outputs.attentions, attentions=outputs.attentions,
) )
def serving_output(self, output: TFMaskedImageModelingOutput) -> TFMaskedImageModelingOutput:
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 TFMaskedImageModelingOutput(
reconstruction=output.reconstruction, hidden_states=hidden_states, attentions=attentions
)
@add_start_docstrings( @add_start_docstrings(
""" """
...@@ -970,12 +919,6 @@ class TFDeiTForImageClassification(TFDeiTPreTrainedModel, TFSequenceClassificati ...@@ -970,12 +919,6 @@ class TFDeiTForImageClassification(TFDeiTPreTrainedModel, TFSequenceClassificati
attentions=outputs.attentions, attentions=outputs.attentions,
) )
def serving_output(self, output: TFImageClassifierOutput) -> TFImageClassifierOutput:
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 TFImageClassifierOutput(logits=output.logits, hidden_states=hidden_states, attentions=attentions)
@add_start_docstrings( @add_start_docstrings(
""" """
...@@ -1055,17 +998,3 @@ class TFDeiTForImageClassificationWithTeacher(TFDeiTPreTrainedModel): ...@@ -1055,17 +998,3 @@ class TFDeiTForImageClassificationWithTeacher(TFDeiTPreTrainedModel):
hidden_states=outputs.hidden_states, hidden_states=outputs.hidden_states,
attentions=outputs.attentions, attentions=outputs.attentions,
) )
def serving_output(
self, output: TFDeiTForImageClassificationWithTeacherOutput
) -> TFDeiTForImageClassificationWithTeacherOutput:
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 TFDeiTForImageClassificationWithTeacherOutput(
logits=output.logits,
cls_logits=output.cls_logits,
distillation_logits=output.distillation_logits,
hidden_states=hidden_states,
attentions=attentions,
)
src/transformers/models/distilbert/modeling_tf_distilbert.py
View file @ 814de8fa
...@@ -48,7 +48,6 @@ from ...modeling_tf_utils import ( ...@@ -48,7 +48,6 @@ 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,
...@@ -424,19 +423,6 @@ class TFDistilBertPreTrainedModel(TFPreTrainedModel): ...@@ -424,19 +423,6 @@ class TFDistilBertPreTrainedModel(TFPreTrainedModel):
config_class = DistilBertConfig config_class = DistilBertConfig
base_model_prefix = "distilbert" base_model_prefix = "distilbert"
@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)
DISTILBERT_START_DOCSTRING = r""" DISTILBERT_START_DOCSTRING = r"""
...@@ -562,12 +548,6 @@ class TFDistilBertModel(TFDistilBertPreTrainedModel): ...@@ -562,12 +548,6 @@ class TFDistilBertModel(TFDistilBertPreTrainedModel):
) )
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
return TFBaseModelOutput(last_hidden_state=output.last_hidden_state, hidden_states=hs, attentions=attns)
class TFDistilBertLMHead(tf.keras.layers.Layer): class TFDistilBertLMHead(tf.keras.layers.Layer):
def __init__(self, config, input_embeddings, **kwargs): def __init__(self, config, input_embeddings, **kwargs):
...@@ -687,13 +667,6 @@ class TFDistilBertForMaskedLM(TFDistilBertPreTrainedModel, TFMaskedLanguageModel ...@@ -687,13 +667,6 @@ class TFDistilBertForMaskedLM(TFDistilBertPreTrainedModel, TFMaskedLanguageModel
attentions=distilbert_output.attentions, attentions=distilbert_output.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)
@add_start_docstrings( @add_start_docstrings(
""" """
...@@ -773,13 +746,6 @@ class TFDistilBertForSequenceClassification(TFDistilBertPreTrainedModel, TFSeque ...@@ -773,13 +746,6 @@ class TFDistilBertForSequenceClassification(TFDistilBertPreTrainedModel, TFSeque
attentions=distilbert_output.attentions, attentions=distilbert_output.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(
""" """
...@@ -848,13 +814,6 @@ class TFDistilBertForTokenClassification(TFDistilBertPreTrainedModel, TFTokenCla ...@@ -848,13 +814,6 @@ class TFDistilBertForTokenClassification(TFDistilBertPreTrainedModel, 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(
""" """
...@@ -879,16 +838,6 @@ class TFDistilBertForMultipleChoice(TFDistilBertPreTrainedModel, TFMultipleChoic ...@@ -879,16 +838,6 @@ class TFDistilBertForMultipleChoice(TFDistilBertPreTrainedModel, 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(
DISTILBERT_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length") DISTILBERT_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length")
...@@ -959,26 +908,6 @@ class TFDistilBertForMultipleChoice(TFDistilBertPreTrainedModel, TFMultipleChoic ...@@ -959,26 +908,6 @@ class TFDistilBertForMultipleChoice(TFDistilBertPreTrainedModel, TFMultipleChoic
attentions=distilbert_output.attentions, attentions=distilbert_output.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(
""" """
...@@ -1062,12 +991,3 @@ class TFDistilBertForQuestionAnswering(TFDistilBertPreTrainedModel, TFQuestionAn ...@@ -1062,12 +991,3 @@ class TFDistilBertForQuestionAnswering(TFDistilBertPreTrainedModel, TFQuestionAn
hidden_states=distilbert_output.hidden_states, hidden_states=distilbert_output.hidden_states,
attentions=distilbert_output.attentions, attentions=distilbert_output.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/dpr/modeling_tf_dpr.py
View file @ 814de8fa
...@@ -372,19 +372,6 @@ class TFDPRPretrainedReader(TFPreTrainedModel): ...@@ -372,19 +372,6 @@ class TFDPRPretrainedReader(TFPreTrainedModel):
config_class = DPRConfig config_class = DPRConfig
base_model_prefix = "reader" base_model_prefix = "reader"
@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)
############### ###############
# Actual Models # Actual Models
...@@ -612,12 +599,6 @@ class TFDPRContextEncoder(TFDPRPretrainedContextEncoder): ...@@ -612,12 +599,6 @@ class TFDPRContextEncoder(TFDPRPretrainedContextEncoder):
pooler_output=outputs.pooler_output, hidden_states=outputs.hidden_states, attentions=outputs.attentions pooler_output=outputs.pooler_output, hidden_states=outputs.hidden_states, 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 TFDPRContextEncoderOutput(pooler_output=output.pooler_output, hidden_states=hs, attentions=attns)
@add_start_docstrings( @add_start_docstrings(
"The bare DPRQuestionEncoder transformer outputting pooler outputs as question representations.", "The bare DPRQuestionEncoder transformer outputting pooler outputs as question representations.",
...@@ -698,12 +679,6 @@ class TFDPRQuestionEncoder(TFDPRPretrainedQuestionEncoder): ...@@ -698,12 +679,6 @@ class TFDPRQuestionEncoder(TFDPRPretrainedQuestionEncoder):
pooler_output=outputs.pooler_output, hidden_states=outputs.hidden_states, attentions=outputs.attentions pooler_output=outputs.pooler_output, hidden_states=outputs.hidden_states, 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 TFDPRQuestionEncoderOutput(pooler_output=output.pooler_output, hidden_states=hs, attentions=attns)
@add_start_docstrings( @add_start_docstrings(
"The bare DPRReader transformer outputting span predictions.", "The bare DPRReader transformer outputting span predictions.",
...@@ -777,15 +752,3 @@ class TFDPRReader(TFDPRPretrainedReader): ...@@ -777,15 +752,3 @@ class TFDPRReader(TFDPRPretrainedReader):
return_dict=return_dict, return_dict=return_dict,
training=training, training=training,
) )
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 TFDPRReaderOutput(
start_logits=output.start_logits,
end_logits=output.end_logits,
relevance_logits=output.relevance_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