You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

245 lines
12 KiB

# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
# This file was automatically generated from src/transformers/models/lasr/modular_lasr.py.
# Do NOT edit this file manually as any edits will be overwritten by the generation of
# the file from the modular. If any change should be done, please apply the change to the
# modular_lasr.py file directly. One of our CI enforces this.
# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
# Copyright 2025 The HuggingFace Inc. team and Google LLC. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from ...configuration_utils import PreTrainedConfig
class LasrEncoderConfig(PreTrainedConfig):
r"""
This is the configuration class to store the configuration of a [`LasrEncoder`]. It is used to instantiate a
`LasrEncoder` model according to the specified arguments, defining the model architecture.
Configuration objects inherit from [`PreTrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PreTrainedConfig`] for more information.
Args:
hidden_size (`int`, *optional*, defaults to 512):
Dimension of the layers and the hidden states.
num_hidden_layers (`int`, *optional*, defaults to 17):
Number of hidden layers in the Transformer encoder.
num_attention_heads (`int`, *optional*, defaults to 8):
Number of attention heads for each attention layer in the Transformer encoder.
intermediate_size (`int`, *optional*, defaults to 2048):
Dimension of the "intermediate" (often named feed-forward) layer in the Transformer encoder.
hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
The non-linear activation function (function or string) in the encoder and pooler.
attention_bias (`bool`, *optional*, defaults to `False`):
Whether to use bias in the attention layers.
convolution_bias (`bool`, *optional*, defaults to `False`):
Whether to use bias in convolutions of the conformer's convolution module.
conv_kernel_size (`int`, *optional*, defaults to 32):
The kernel size of the convolution layers in the Conformer block.
subsampling_conv_channels (`int`, *optional*, defaults to 256):
The number of channels in the subsampling convolution layers.
subsampling_conv_kernel_size (`int`, *optional*, defaults to 5):
The kernel size of the subsampling convolution layers.
subsampling_conv_stride (`int`, *optional*, defaults to 2):
The stride of the subsampling convolution layers.
num_mel_bins (`int`, *optional*, defaults to 128):
Number of mel features.
dropout (`float`, *optional*, defaults to 0.1):
The dropout ratio for all fully connected layers in the embeddings, encoder, and pooler.
dropout_positions (`float`, *optional*, defaults to 0.0):
The dropout ratio for the positions in the input sequence.
layerdrop (`float`, *optional*, defaults to 0.1):
The dropout ratio for the layers in the encoder.
activation_dropout (`float`, *optional*, defaults to 0.1):
The dropout ratio for activations inside the fully connected layer.
attention_dropout (`float`, *optional*, defaults to 0.1):
The dropout ratio for the attention layers.
max_position_embeddings (`int`, *optional*, defaults to 10000):
The maximum sequence length that this model might ever be used with.
initializer_range (`float`, *optional*, defaults to 0.02):
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
layer_norm_eps (`float`, *optional*, defaults to 1e-06):
The epsilon used by the layer normalization layers.
feed_forward_residual_weights (`tuple[float, float]`, *optional*, defaults to `[1.5, 0.5]`):
The residual weights for the feed forward layers.
conv_residual_weights (`tuple[float, float]`, *optional*, defaults to `[2.0, 1.0]`):
The residual weights for the convolution layers.
batch_norm_momentum (`float`, *optional*, defaults to 0.01):
The momentum for the batch normalization layers.
rope_parameters (`RopeParameters`, *optional*):
Dictionary containing the configuration parameters for the RoPE embeddings. The dictionary should contain
a value for `rope_theta` and optionally parameters used for scaling in case you want to use RoPE
with longer `max_position_embeddings`.
Example:
```python
>>> from transformers import LasrEncoderModel, LasrEncoderConfig
>>> # Initializing a `LasrEncoder` configuration
>>> configuration = LasrEncoderConfig()
>>> # Initializing a model from the configuration
>>> model = LasrEncoderModel(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
```
This configuration class is based on the LasrEncoder architecture from Google Health AI. You can find more details
and pre-trained models at [TODO/TODO](https://huggingface.co/TODO/TODO).
"""
model_type = "lasr_encoder"
keys_to_ignore_at_inference = ["past_key_values"]
def __init__(
self,
hidden_size=512,
num_hidden_layers=17,
num_attention_heads=8,
intermediate_size=2048,
hidden_act="silu",
attention_bias=False,
convolution_bias=False,
conv_kernel_size=32,
subsampling_conv_channels=256,
subsampling_conv_kernel_size=5,
subsampling_conv_stride=2,
num_mel_bins=128,
dropout=0.1,
dropout_positions=0.0,
layerdrop=0.1,
activation_dropout=0.1,
attention_dropout=0.1,
max_position_embeddings=10000,
initializer_range=0.02,
layer_norm_eps=1e-6,
feed_forward_residual_weights=[1.5, 0.5],
conv_residual_weights=[2.0, 1.0],
batch_norm_momentum=0.01,
rope_parameters=None,
**kwargs,
):
self.rope_parameters = rope_parameters
self.layer_norm_eps = layer_norm_eps
self.feed_forward_residual_weights = feed_forward_residual_weights
self.conv_residual_weights = conv_residual_weights
self.batch_norm_momentum = batch_norm_momentum
self.hidden_size = hidden_size
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.num_key_value_heads = num_attention_heads # LlamaAttention compatibility
self.intermediate_size = intermediate_size
self.hidden_act = hidden_act
self.attention_bias = attention_bias
self.convolution_bias = convolution_bias
self.conv_kernel_size = conv_kernel_size
self.subsampling_conv_kernel_size = subsampling_conv_kernel_size
self.subsampling_conv_stride = subsampling_conv_stride
self.subsampling_conv_channels = subsampling_conv_channels
self.num_mel_bins = num_mel_bins
self.dropout = dropout
self.dropout_positions = dropout_positions
self.layerdrop = layerdrop
self.activation_dropout = activation_dropout
self.attention_dropout = attention_dropout
self.max_position_embeddings = max_position_embeddings
self.initializer_range = initializer_range
super().__init__(
**kwargs,
)
class LasrCTCConfig(PreTrainedConfig):
r"""
This is the configuration class to store the configuration of a [`LasrForCTC`]. It is used to instantiate a
Lasr CTC model according to the specified arguments, defining the model architecture.
Configuration objects inherit from [`PreTrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PreTrainedConfig`] for more information.
Args:
vocab_size (`int`, *optional*, defaults to 512):
Vocabulary size of the model.
ctc_loss_reduction (`str`, *optional*, defaults to `"mean"`):
Specifies the reduction to apply to the output of `torch.nn.CTCLoss`. Only relevant when training an
instance of [`LasrForCTC`].
ctc_zero_infinity (`bool`, *optional*, defaults to `True`):
Whether to zero infinite losses and the associated gradients of `torch.nn.CTCLoss`. Infinite losses mainly
occur when the inputs are too short to be aligned to the targets. Only relevant when training an instance
of [`LasrForCTC`].
encoder_config (`Union[dict, LasrEncoderConfig]`, *optional*):
The config object or dictionary of the encoder.
pad_token_id (`int`, *optional*, defaults to 0):
Padding token id. Also used as blank token id.
Example:
```python
>>> from transformers import LasrForCTC, LasrCTCConfig
>>> # Initializing a Lasr configuration
>>> configuration = LasrCTCConfig()
>>> # Initializing a model from the configuration
>>> model = LasrForCTC(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
```
This configuration class is based on the Lasr CTC architecture from Google Health AI. You can find more details
and pre-trained models at [TODO/TODO](https://huggingface.co/TODO/TODO).
"""
model_type = "lasr_ctc"
sub_configs = {"encoder_config": LasrEncoderConfig}
def __init__(
self,
vocab_size=512,
ctc_loss_reduction="mean",
ctc_zero_infinity=True,
encoder_config: dict | LasrEncoderConfig = None,
pad_token_id=0,
**kwargs,
):
self.vocab_size = vocab_size
self.ctc_loss_reduction = ctc_loss_reduction
self.ctc_zero_infinity = ctc_zero_infinity
if isinstance(encoder_config, dict):
self.encoder_config = LasrEncoderConfig(**encoder_config)
elif encoder_config is None:
self.encoder_config = LasrEncoderConfig()
self.encoder_config = self.encoder_config
self.initializer_range = self.encoder_config.initializer_range
self.pad_token_id = pad_token_id
super().__init__(**kwargs)
@classmethod
def from_encoder_config(cls, encoder_config: LasrEncoderConfig, **kwargs):
r"""
Instantiate a [`LasrCTCConfig`] (or a derived class) from lasr encoder model configuration.
Returns:
[`LasrCTCConfig`]: An instance of a configuration object
"""
return cls(encoder_config=encoder_config.to_dict(), **kwargs)
@property
def inputs_to_logits_ratio(self):
return self.encoder_config.subsampling_conv_stride**2
__all__ = ["LasrEncoderConfig", "LasrCTCConfig"]