Add support for Whisper

docs/classes/models/whisper.rst

@@ -0,0 +1,25 @@
+Whisper
+-----------------------------------------------------------------------------------------------------------------------
+
+The Whisper model was presented in `Robust Speech Recognition via Large-Scale Weak Supervision
+<https://arxiv.org/abs/2212.04356>`_ by Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine
+McLeavey, Ilya Sutskever. 
+
+Whisper is a state-of-the-art speech recognition model trained on 680,000 hours of multilingual and multitask data, presented by OpenAI.
+
+The abstract from the paper is the following:
+
+*We study the capabilities of speech processing systems trained simply to predict large amounts of
+transcripts of audio on the internet. When scaled to 680,000 hours of multilingual and multitask
+supervision, the resulting models generalize well to standard benchmarks and are often competitive
+with prior fully supervised results but in a zeroshot transfer setting without the need for any finetuning. When compared to humans, the models
+approach their accuracy and robustness. We are releasing models and inference code to serve as
+a foundation for further work on robust speech processing.*
+
+
+WhisperAdapterModel
+~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: adapters.WhisperAdapterModel
+    :members:
+    :inherited-members: WhisperPreTrainedModel

docs/index.rst

@@ -87,6 +87,7 @@ Currently, we support the PyTorch versions of all models as listed on the `Model
    classes/models/roberta
    classes/models/t5
    classes/models/vit
+   classes/models/whisper
    classes/models/xlmroberta
    classes/models/xmod
 

docs/model_overview.md

@@ -10,6 +10,7 @@ The table below further shows which model architectures support which adaptation
     E.g., for BERT, this means adapters provides a ``BertAdapterModel`` class, but you can also use ``BertModel``, ``BertForSequenceClassification`` etc. together with adapters.
 ```
 
+
 | Model                                   | (Bottleneck)<br> Adapters | Prefix<br> Tuning | LoRA | Compacter | Adapter<br> Fusion | Invertible<br> Adapters | Parallel<br> block | Prompt<br> Tuning | ReFT |
 | --------------------------------------- | -| - | - | - | - | - | - |- | - |
 | [ALBERT](classes/models/albert.html)    | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
@@ -33,6 +34,7 @@ The table below further shows which model architectures support which adaptation
 | [RoBERTa](classes/models/roberta.html)  | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
 | [T5](classes/models/t5.html)            | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | | ✅ |
 | [ViT](classes/models/vit.html)            | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
+| [Whisper](classes/models/whisper.html)  | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | | ✅ |
 | [XLM-RoBERTa](classes/models/xlmroberta.html) | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
 | [X-MOD](classes/models/xmod.html) | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
 

src/adapters/__init__.py

@@ -115,6 +115,7 @@
     "models.roberta": ["RobertaAdapterModel"],
     "models.t5": ["T5AdapterModel"],
     "models.vit": ["ViTAdapterModel"],
+    "models.whisper": ["WhisperAdapterModel"],
     "models.xlm_roberta": ["XLMRobertaAdapterModel"],
     "models.xmod": ["XmodAdapterModel"],
     "trainer": ["AdapterTrainer", "Seq2SeqAdapterTrainer"],
@@ -224,6 +225,7 @@
     from .models.roberta import RobertaAdapterModel
     from .models.t5 import T5AdapterModel
     from .models.vit import ViTAdapterModel
+    from .models.whisper import WhisperAdapterModel
     from .models.xlm_roberta import XLMRobertaAdapterModel
     from .models.xmod import XmodAdapterModel
     from .trainer import AdapterTrainer, Seq2SeqAdapterTrainer

src/adapters/composition.py

@@ -139,6 +139,7 @@ def __init__(
         "llama",
         "mistral",
         "electra",
+        "whisper",
         "xmod",
     ],
 }

src/adapters/head_utils.py

@@ -5,7 +5,6 @@
 
 logger = logging.getLogger(__name__)
 
-
 # The "layers" attributes in the configs below map from static head module names to flex head module names.
 # In this context, "None" refers to a flex-head layer without weights (e.g. dropout, acts).
 STATIC_TO_FLEX_HEAD_MAP = {
@@ -771,6 +770,16 @@
             "generator_lm_head",
         ],
     },
+    "WhisperForConditionalGeneration": {
+        "config": {
+            "head_type": "seq2seq_lm",
+            "layers": 1,
+            "activation_function": None,
+            "layer_norm": False,
+            "bias": False,
+        },
+        "layers": ["proj_out"],
+    },
 }
 
 

src/adapters/heads/language_modeling.py

@@ -131,7 +131,7 @@ def forward(self, outputs, cls_output=None, attention_mask=None, return_dict=Fal
                 )
                 labels = torch.cat((prompt_labels, labels), dim=-1)
 
-            loss = loss_fct(logits_for_loss.view(-1, self.config["vocab_size"]), labels.view(-1))
+            loss = loss_fct(logits_for_loss.reshape(-1, self.config["vocab_size"]), labels.reshape(-1))
 
         if return_dict:
             return self._create_model_output(loss, lm_logits, outputs)

src/adapters/heads/model_mixin.py

@@ -27,7 +27,6 @@
 
 logger = logging.getLogger(__name__)
 
-
 MODEL_HEAD_MAP = {
     "classification": ClassificationHead,
     "multilabel_classification": MultiLabelClassificationHead,
@@ -438,47 +437,51 @@ def add_dependency_parsing_head(self, head_name, num_labels=2, overwrite_ok=Fals
         self.add_prediction_head(head, overwrite_ok)
 
     @head_type("masked_lm")
-    def add_masked_lm_head(self, head_name, activation_function="gelu", overwrite_ok=False):
+    def add_masked_lm_head(self, head_name, activation_function="gelu", layers=2, overwrite_ok=False):
         """
         Adds a masked language modeling head on top of the model.
 
         Args:
             head_name (str): The name of the head.
             activation_function (str, optional): Activation function. Defaults to 'gelu'.
+            layers (int, optional): Number of layers. Defaults to 2.
             overwrite_ok (bool, optional): Force overwrite if a head with the same name exists. Defaults to False.
         """
-        head = BertStyleMaskedLMHead(self, head_name, activation_function=activation_function)
+        head = BertStyleMaskedLMHead(self, head_name, layers=layers, activation_function=activation_function)
         self.add_prediction_head(head, overwrite_ok=overwrite_ok)
 
     @head_type("causal_lm")
-    def add_causal_lm_head(self, head_name, activation_function="gelu", overwrite_ok=False):
+    def add_causal_lm_head(self, head_name, activation_function="gelu", layers=2, overwrite_ok=False):
         """
         Adds a causal language modeling head on top of the model.
 
         Args:
             head_name (str): The name of the head.
             activation_function (str, optional): Activation function. Defaults to 'gelu'.
+            layers (int, optional): Number of layers. Defaults to 2.
             overwrite_ok (bool, optional): Force overwrite if a head with the same name exists. Defaults to False.
         """
         head = CausalLMHead(
-            self, head_name, layers=2, activation_function=activation_function, layer_norm=True, bias=True
+            self, head_name, layers=layers, activation_function=activation_function, layer_norm=True, bias=True
         )
         self.add_prediction_head(head, overwrite_ok=overwrite_ok)
 
     @head_type("seq2seq_lm")
     def add_seq2seq_lm_head(
         self,
         head_name,
+        layers=1,
         overwrite_ok=False,
     ):
         """
         Adds a sequence-to-sequence language modeling head on top of the model.
 
         Args:
             head_name (str): The name of the head.
+            layers (int, optional): Number of layers. Defaults to 1.
             overwrite_ok (bool, optional): Force overwrite if a head with the same name exists. Defaults to False.
         """
-        head = Seq2SeqLMHead(self, head_name)
+        head = Seq2SeqLMHead(self, head_name, layers=layers)
         self.add_prediction_head(head, overwrite_ok=overwrite_ok)
 
     def delete_head(self, head_name: str):

src/adapters/methods/prefix_tuning.py

@@ -229,7 +229,14 @@ def forward(self, *args, **kwargs):
         prefix_states = {}
         if adapter_setup is not None:
             # Infer batch size
-            input_tensor_names = ["input_ids", "decoder_input_ids", "attention_mask", "inputs_embeds", "pixel_values"]
+            input_tensor_names = [
+                "input_ids",
+                "decoder_input_ids",
+                "attention_mask",
+                "inputs_embeds",
+                "pixel_values",
+                "input_features",
+            ]
             batch_size = None
             for name in input_tensor_names:
                 if kwargs.get(name, None) is not None:

src/adapters/model_mixin.py

@@ -1311,7 +1311,18 @@ def _prepare_model_inputs(self, *args, **kwargs):
             and self.adapters_config.active_setup
             and self.adapters_config.active_setup.parallel_channels > 1
         ):
-            input_ids = input_ids.repeat(self.adapters_config.active_setup.parallel_channels, 1)
+            # Extract original shape
+            input_shape = input_ids.shape
+            # Replicate input_ids to match the number of parallel channels
+            # Also works for inputs with more than 2 dimensions
+            repeat_shape = [
+                self.adapters_config.active_setup.parallel_channels
+            ] + [  # first dimension is parallel channels
+                1
+            ] * (
+                len(input_shape) - 1
+            )  # residual dims should be replicated parallel_channels times
+            input_ids = input_ids.repeat(repeat_shape)
             model_kwargs["adapter_input_parallelized"] = True
 
         return input_ids, input_name, model_kwargs

src/adapters/models/__init__.py

@@ -32,6 +32,13 @@
     T5ModelAdaptersMixin,
 )
 from .vit.mixin_vit import ViTIntermediateAdaptersMixin, ViTModelAdaptersMixin
+from .whisper.mixin_whisper import (
+    WhisperDecoderAdaptersMixin,
+    WhisperDecoderWrapperAdaptersMixin,
+    WhisperEncoderAdaptersMixin,
+    WhisperForAudioClassificationWithHeadsMixin,
+    WhisperModelAdaptersMixin,
+)
 from .xmod.mixin_xmod import XmodModelAdaptersMixin
 
 
@@ -94,6 +101,11 @@
     "BertGenerationEncoder": BertModelAdaptersMixin,
     "BertGenerationLayer": BertLayerAdaptersMixin,
     "LlamaModel": LlamaModelAdapterMixin,
+    "WhisperEncoder": WhisperEncoderAdaptersMixin,
+    "WhisperDecoder": WhisperDecoderAdaptersMixin,
+    "WhisperModel": WhisperModelAdaptersMixin,
+    "WhisperDecoderWrapper": WhisperDecoderWrapperAdaptersMixin,
+    "WhisperForAudioClassification": WhisperForAudioClassificationWithHeadsMixin,
     "LlamaForQuestionAnswering": LlamaForQuestionAnsweringAdapterMixin,
     "MistralModel": MistralModelAdapterMixin,
 }

src/adapters/models/auto/adapter_model.py

@@ -29,6 +29,7 @@
         ("roberta", "RobertaAdapterModel"),
         ("t5", "T5AdapterModel"),
         ("vit", "ViTAdapterModel"),
+        ("whisper", "WhisperAdapterModel"),
         ("xlm-roberta", "XLMRobertaAdapterModel"),
         ("xmod", "XmodAdapterModel"),
     ]

src/adapters/models/mt5/modeling_mt5.py

@@ -77,7 +77,7 @@ def forward(
         if past_key_value is not None:
             assert (
                 len(past_key_value) == 2
-            ), f"past_key_value should have 2 past states: keys and values. Got { len(past_key_value)} past states"
+            ), f"past_key_value should have 2 past states: keys and values. Got {len(past_key_value)} past states"
             real_seq_length += past_key_value[0].shape[2] if query_length is None else query_length
 
         key_length = real_seq_length if key_value_states is None else key_value_states.shape[1]

src/adapters/models/whisper/__init__.py

@@ -0,0 +1,39 @@
+# flake8: noqa
+# There's no way to ignore "F401 '...' imported but unused" warnings in this
+# module, but to preserve other warnings. So, don't check this module at all.
+
+# Copyright 2020 The Adapter-Hub Team. 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 typing import TYPE_CHECKING
+
+from transformers.utils import _LazyModule
+
+
+_import_structure = {
+    "adapter_model": ["WhisperAdapterModel"],
+}
+
+
+if TYPE_CHECKING:
+    from .adapter_model import WhisperAdapterModel
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(
+        __name__,
+        globals()["__file__"],
+        _import_structure,
+    )