feat(model) : add segmentation model based on self-supervised representation

pyannote/audio/models/blocks/selfsup.py

@@ -0,0 +1,181 @@
+# MIT License
+#
+# Copyright (c) 2020 CNRS
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+
+import sys
+import re
+from typing import Optional
+import torch
+import torchaudio
+import torch.nn as nn
+from torch.nn.functional import normalize
+import torch.nn.functional as F
+from collections import OrderedDict
+from torchaudio.models.wav2vec2 import wav2vec2_model, wavlm_model
+from torchaudio.pipelines import Wav2Vec2Bundle
+
+class SelfSupModel(nn.Module):
+
+    def __init__(self,checkpoint=None,torchaudio_ssl=None,torchaudio_cache=None,finetune=None,average_layers=None,average_all=None,name=None,layer=None,cfg=None):
+        super().__init__()
+        if torchaudio_ssl:
+            if checkpoint:
+                raise ValueError("Error : Cannot specify both a checkpoint and a torchaudio model.")
+
+            print("\nThe Self-Supervised Model "+str(torchaudio_ssl)+" is loaded from torchaudio.\n")
+            name,config,ordered_dict = self.dict_torchaudio(torchaudio_ssl,torchaudio_cache)
+        else:
+            print("A checkpoint from a Self-Supervised Model is used for training.")
+            if torch.cuda.is_available():  
+                ckpt = torch.load(checkpoint)
+            else:
+                ckpt = torch.load(checkpoint,map_location=torch.device('cpu'))
+                #Check if the checkpoint is from an already finetuned Diarization model (containing SSL), or from a SSL pretrained model only
+            if 'pyannote.audio' in ckpt: #1: Check if there is a Segmentation model attached onto or not
+                print("The checkpoint is used for finetuning. \nThe attached SSL model will be used for feature extraction.")
+                name,config,ordered_dict = self.dict_finetune(ckpt)
+
+            else: #Otherwise, load the dictionary of the SSL checkpoint
+                print("The checkpoint is a pretrained SSL model to use for Segmentation.\nBuilding the SSL model.")
+                name,config,ordered_dict = self.dict_pretrained(ckpt)
+
+        # Layer-wise pooling (same way as SUPERB)
+        if not average_all:      
+            if not average_layers :
+                if layer is None :
+                    print("\nLayer number not specified. Default to layer 1.\n")
+
+                    self.layer = 1
+                else :        
+
+                    self.layer = layer
+                    print("\nSelected layer is "+ str(layer) +". \n")
+            else:
+                print("Layers "+str(average_layers)+" selected for layer-wise pooling.")
+
+                self.W = nn.Parameter(torch.randn(len(average_layers))) #Set specific number of learnable weights
+
+                self.average_layers = average_layers
+
+                self.layer = max(average_layers)
+        else:
+            print("All layers are selected for layer-wise pooling.")
+
+            self.W = nn.Parameter(torch.randn(config['encoder_num_layers'])) #Set max number of learnable weights
+
+            self.average_layers = list(range(config['encoder_num_layers']))
+
+            self.layer = config['encoder_num_layers']
+
+        if finetune: #Finetuning not working
+            print("Self-supervised model is unfrozen.")      
+            #config['encoder_ff_interm_dropout'] = 0.3
+            config['encoder_layer_norm_first'] = True
+        else :
+            print("Self-supervised model is frozen.")
+
+        config['encoder_num_layers'] = self.layer    
+        ordered_dict = self.remove_layers_dict(ordered_dict,self.layer) #Remove weights from unused transformer encoders
+        self.model_name = name
+        self.finetune = finetune #Assign mode
+        self.average_layers = average_layers
+        self.feat_size = config['encoder_embed_dim'] #Get feature output dimension
+        self.config = config #Assign the configuration
+        SelfSupModel.__name__ = self.model_name #Assign name of the class
+
+        if name is "WAVLM_BASE" or name is "WAVLM_LARGE": #Only wavlm_model has two additional arguments
+            model = wavlm_model(**config)
+        else:
+            model = wav2vec2_model(**config)
+        model.load_state_dict(ordered_dict) #Assign state dict to the model
+
+        if finetune:
+            self.ssl_model = model.train()
+        else:
+            self.ssl_model = model.eval()
+
+
+    def dict_finetune(self, ckpt):
+        #Need to reconstruct the dictionary
+        #Get dict
+        dict_modules = list(ckpt['state_dict'].keys()) #Get the list of ssl modules
+        ssl_modules = [key for key in dict_modules if 'selfsupervised' in key] #Extract only the SSL parts
+        weights = [ckpt['state_dict'][key] for key in ssl_modules] #Get the weights corresponding to the modules
+        modules_torchaudio = ['.'.join(key.split('.')[2:]) for key in ssl_modules] #Get a new list which contains only torchaudio keywords
+        ordered_dict = OrderedDict((key,weight) for key,weight in zip(modules_torchaudio,weights)) #Recreate the state_dict
+        config = ckpt['hyper_parameters']['selfsupervised']['cfg'] #Get config
+        name = ckpt['hyper_parameters']['selfsupervised']['name'] #Get model name
+
+        return(name,config,ordered_dict)
+
+    def dict_pretrained(self, ckpt):
+        ordered_dict = ckpt['state_dict'] #Get dict
+        config = ckpt['config'] #Get config
+        name = ckpt['model_name'] #Get model name
+
+        return(ckpt['model_name'],ckpt['config'],ckpt['state_dict'])
+
+    def dict_torchaudio(self,torchaudio_ssl,torchaudio_cache):
+        bundle = getattr(torchaudio.pipelines, torchaudio_ssl)
+        #Name is torchaudio_ssl
+        name = torchaudio_ssl #Get name
+        config = bundle._params #Get config
+        if torchaudio_cache:
+            torch.hub.set_dir(torchaudio_cache) #Set cache
+        ordered_dict = bundle.get_model().state_dict() #Get the dict
+
+        return(name,config,ordered_dict)
+    def remove_layers_dict(self,state_dict,layer):
+        keys_to_delete = []
+        for key in state_dict.keys():
+            if "transformer.layers" in key:
+                nb = int(re.findall(r'\d+',key)[0])
+                if nb>(layer-1):
+                    keys_to_delete.append(key)
+        for key in keys_to_delete:
+            del state_dict[key]
+
+        return(state_dict)
+
+    def avg_pool(self,scalars,feat_list):
+        sum = 0
+        for i in range(0,len(feat_list)):
+            sum = sum + scalars[i]*feat_list[i]
+        return(sum)
+
+    def forward(self, waveforms: torch.Tensor) -> torch.Tensor:
+        waveforms = torch.squeeze(waveforms,1) #waveforms : (batch, channel, sample) -> (batch,sample)
+        if self.finetune:
+            feat,_ = self.ssl_model.extract_features(waveforms,None,self.layer)
+        else:
+            with torch.no_grad():
+                feat,_ = self.ssl_model.extract_features(waveforms,None,self.layer)
+        if self.average_layers:
+            feat_learn_list = []
+            for index in self.average_layers:
+                feat_learn_list.append(feat[index-1])
+            w = self.W.softmax(-1)
+            outputs = self.avg_pool(w,feat_learn_list)
+            #print(w)
+            #print(outputs.size())
+        else:
+            outputs = feat[self.layer-1]
+        return (outputs)

pyannote/audio/models/segmentation/PyanSup.py

@@ -0,0 +1,183 @@
+from typing import Optional
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+from pyannote.core.utils.generators import pairwise
+
+from pyannote.audio.core.model import Model
+from pyannote.audio.core.task import Task
+from pyannote.audio.models.blocks.selfsup import SelfSupModel
+from pyannote.audio.utils.params import merge_dict
+
+
+class PyanSup(Model):
+    """PyanHugg segmentation model
+
+    Self-Supervised Model > LSTM > Feed forward > Classifier
+
+    Parameters
+    ----------
+    sample_rate : int, optional
+        Audio sample rate. Defaults to 16kHz (16000).
+    num_channels : int, optional
+        Number of channels. Defaults to mono (1).
+
+    selfsupervised : dict, optional
+        Keyword arugments passed to the selfsupervised block. name and cfg are used to reconstruct the feature extractor from the dictionary of the checkpoint. Layer corresponds to the layer that serves for the feature extraction.
+        Defaults to {
+        "name": None,
+        "layer": None,
+        "cfg": None,
+    }
+    lstm : dict, optional
+        Keyword arguments passed to the LSTM layer.
+        Defaults to {"hidden_size": 128, "num_layers": 2, "bidirectional": True},
+        i.e. two bidirectional layers with 128 units each.
+        Set "monolithic" to False to split monolithic multi-layer LSTM into multiple mono-layer LSTMs.
+        This may proove useful for probing LSTM internals.
+    linear : dict, optional
+        Keyword arugments used to initialize linear layers
+        Defaults to {"hidden_size": 128, "num_layers": 2},
+        i.e. two linear layers with 128 units each.
+    """
+
+
+
+    SSL_DEFAULTS = {
+        "name": None,
+        "layer": None,
+        "average_layers": None,
+        "average_all": False,
+        "cfg": None,
+    }
+    LSTM_DEFAULTS = {
+        "hidden_size": 128,
+        "num_layers": 2,
+        "bidirectional": True,
+        "monolithic": True,
+        "dropout": 0.0,
+    }
+    LINEAR_DEFAULTS = {"hidden_size": 128, "num_layers": 2}
+    def __init__(
+        self,
+        ckpt: str = None,
+        torchaudio_ssl: str = None,
+        torchaudio_cache: str = None,
+        finetune: bool = False,
+        selfsupervised: dict = None,
+        lstm: dict = None,
+        linear: dict = None,
+        sample_rate: int = 16000,
+        num_channels: int = 1,
+        task: Optional[Task] = None,
+    ):
+
+        super().__init__(sample_rate=sample_rate, num_channels=num_channels, task=task)
+
+        selfsupervised = merge_dict(self.SSL_DEFAULTS, selfsupervised)
+        lstm = merge_dict(self.LSTM_DEFAULTS, lstm)
+        lstm["batch_first"] = True
+        linear = merge_dict(self.LINEAR_DEFAULTS, linear)
+        self.save_hyperparameters("lstm", "linear") #A first merge is done using the default parameters specified
+
+        print("\n##################################################################")
+        print("### A self-supervised model is used for the feature extraction ###")
+        print("##################################################################")
+        self.selfsupervised = SelfSupModel(ckpt,torchaudio_ssl,torchaudio_cache,finetune,**selfsupervised)
+        selfsupervised['name'] = self.selfsupervised.model_name
+        selfsupervised['cfg'] = self.selfsupervised.config
+        self.save_hyperparameters("selfsupervised")
+        feat_size = self.selfsupervised.feat_size
+
+        print("##################################################################\n")
+
+        monolithic = lstm["monolithic"]
+        if monolithic:
+            multi_layer_lstm = dict(lstm)
+            del multi_layer_lstm["monolithic"]
+            self.lstm = nn.LSTM(feat_size, **multi_layer_lstm)
+
+        else:
+            num_layers = lstm["num_layers"]
+            if num_layers > 1:
+                self.dropout = nn.Dropout(p=lstm["dropout"])
+
+            one_layer_lstm = dict(lstm)
+            one_layer_lstm["num_layers"] = 1
+            one_layer_lstm["dropout"] = 0.0
+            del one_layer_lstm["monolithic"]
+
+            self.lstm = nn.ModuleList(
+                [
+                    nn.LSTM(
+                        feat_size
+                        if i == 0
+                        else lstm["hidden_size"] * (2 if lstm["bidirectional"] else 1),
+                        **one_layer_lstm
+                    )
+                    for i in range(num_layers)
+                ]
+            )
+
+        if linear["num_layers"] < 1:
+            return
+
+        lstm_out_features: int = self.hparams.lstm["hidden_size"] * (
+            2 if self.hparams.lstm["bidirectional"] else 1
+        )
+        self.linear = nn.ModuleList(
+            [
+                nn.Linear(in_features, out_features)
+                for in_features, out_features in pairwise(
+                    [
+                        lstm_out_features,
+                    ]
+                    + [self.hparams.linear["hidden_size"]]
+                    * self.hparams.linear["num_layers"]
+                )
+            ]
+        )  
+    def build(self):
+
+        if self.hparams.linear["num_layers"] > 0:
+            in_features = self.hparams.linear["hidden_size"]
+        else:
+            in_features = self.hparams.lstm["hidden_size"] * (
+                2 if self.hparams.lstm["bidirectional"] else 1
+            )
+
+        if self.specifications.powerset:
+            out_features = self.specifications.num_powerset_classes
+        else:
+            out_features = len(self.specifications.classes)
+
+        self.classifier = nn.Linear(in_features, out_features)
+        self.activation = self.default_activation()
+
+    def forward(self, waveforms: torch.Tensor) -> torch.Tensor:
+        """Pass forward
+
+        Parameters
+        ----------
+        waveforms : (batch, channel, sample)
+
+        Returns
+        -------
+        scores : (batch, frame, classes)
+        """
+        outputs = self.selfsupervised(waveforms)
+        if self.hparams.lstm["monolithic"]:
+            outputs, _ = self.lstm(outputs)
+        else:
+            for i, lstm in enumerate(self.lstm):
+                outputs, _ = lstm(outputs)
+                if i + 1 < self.hparams.lstm["num_layers"]:
+                    outputs = self.dropout(outputs)
+
+        if self.hparams.linear["num_layers"] > 0:
+            for linear in self.linear:
+                outputs = F.leaky_relu(linear(outputs))
+
+        return self.activation(self.classifier(outputs))

pyannote/audio/models/segmentation/__init__.py

@@ -21,5 +21,6 @@
 # SOFTWARE.
 
 from .PyanNet import PyanNet
+from .PyanSup import PyanSup
 
-__all__ = ["PyanNet"]
+__all__ = ["PyanNet","PyanSup"]