initial commit

.gitattributes

@@ -25,3 +25,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zstandard filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+epoch=3-step=7459.ckpt filter=lfs diff=lfs merge=lfs -text
+wav2vec_small.pt filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,3 @@
+demo/
+__pycache__/
+flagged/

LICENSE

@@ -0,0 +1,7 @@
+Copyright 2022 sarulab-speech
+
+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.

demo.py

@@ -0,0 +1,56 @@
+
+from random import sample
+import gradio as gr
+import torchaudio
+import torch
+import torch.nn as nn
+import lightning_module
+
+class ChangeSampleRate(nn.Module):
+    def __init__(self, input_rate: int, output_rate: int):
+        super().__init__()
+        self.output_rate = output_rate
+        self.input_rate = input_rate
+
+    def forward(self, wav: torch.tensor) -> torch.tensor:
+        # Only accepts 1-channel waveform input
+        wav = wav.view(wav.size(0), -1)
+        new_length = wav.size(-1) * self.output_rate // self.input_rate
+        indices = (torch.arange(new_length) * (self.input_rate / self.output_rate))
+        round_down = wav[:, indices.long()]
+        round_up = wav[:, (indices.long() + 1).clamp(max=wav.size(-1) - 1)]
+        output = round_down * (1. - indices.fmod(1.)).unsqueeze(0) + round_up * indices.fmod(1.).unsqueeze(0)
+        return output
+
+model = lightning_module.BaselineLightningModule.load_from_checkpoint("epoch=3-step=7459.ckpt")
+def calc_mos(audio_path):
+    wav, sr = torchaudio.load(audio_path)
+    osr = 16_000
+    batch = wav.unsqueeze(0).repeat(10, 1, 1)
+    csr = ChangeSampleRate(sr, osr)
+    out_wavs = csr(wav)
+    batch = {
+        'wav': out_wavs,
+        'domains': torch.tensor([0]),
+        'judge_id': torch.tensor([288])
+    }
+    output = model(batch)
+    return output.mean(dim=1).squeeze().detach().numpy()*2 + 3
+
+
+description ="""
+This model is trained on BVCC dataset.
+""" 
+
+iface = gr.Interface(
+  fn=calc_mos,
+  inputs=gr.inputs.Audio(type='filepath'),
+  outputs="text",
+  title="UTMOS demo page",
+  description=description,
+  allow_flagging=False,
+  
+).launch(
+    auth=("test", "sarulab-test"),
+    share=True
+)

epoch=3-step=7459.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:44c57e3e4135a243b43d2c82b6a693fcd56f15f9ad0e1eb2a8b31fdecd3a49b8
+size 1238128841

lightning_module.py

@@ -0,0 +1,41 @@
+import pytorch_lightning as pl
+import torch
+import torch.nn as nn
+import os
+import numpy as np
+import hydra
+from model import load_ssl_model, PhonemeEncoder, DomainEmbedding, LDConditioner, Projection
+
+
+class BaselineLightningModule(pl.LightningModule):
+    def __init__(self, cfg):
+        super().__init__()
+        self.cfg = cfg
+        self.construct_model()
+        self.save_hyperparameters()
+    
+    def construct_model(self):
+        self.feature_extractors = nn.ModuleList([
+            load_ssl_model(cp_path='wav2vec_small.pt'),
+            DomainEmbedding(3,128),
+        ])
+        output_dim = sum([ feature_extractor.get_output_dim() for feature_extractor in self.feature_extractors])
+        output_layers = [
+            LDConditioner(judge_dim=128,num_judges=3000,input_dim=output_dim)
+        ]
+        output_dim = output_layers[-1].get_output_dim()
+        output_layers.append(
+            Projection(hidden_dim=2048,activation=torch.nn.ReLU(),range_clipping=False,input_dim=output_dim)
+
+        )
+
+        self.output_layers = nn.ModuleList(output_layers)
+
+    def forward(self, inputs):
+        outputs = {}
+        for feature_extractor in self.feature_extractors:
+            outputs.update(feature_extractor(inputs))
+        x = outputs
+        for output_layer in self.output_layers:
+            x = output_layer(x,inputs)
+        return x

model.py

@@ -0,0 +1,191 @@
+import torch
+import torch.nn as nn
+import fairseq
+import os
+import hydra
+
+def load_ssl_model(cp_path):
+    ssl_model_type = cp_path.split("/")[-1]
+    wavlm =  "WavLM" in ssl_model_type
+    if wavlm:
+        checkpoint = torch.load(cp_path)
+        cfg = WavLMConfig(checkpoint['cfg'])
+        ssl_model = WavLM(cfg)
+        ssl_model.load_state_dict(checkpoint['model'])
+        if 'Large' in ssl_model_type:
+            SSL_OUT_DIM = 1024
+        else:
+            SSL_OUT_DIM = 768
+    else:
+        if ssl_model_type == "wav2vec_small.pt":
+            SSL_OUT_DIM = 768
+        elif ssl_model_type in ["w2v_large_lv_fsh_swbd_cv.pt", "xlsr_53_56k.pt"]:
+            SSL_OUT_DIM = 1024
+        else:
+            print("*** ERROR *** SSL model type " + ssl_model_type + " not supported.")
+            exit()
+        model, cfg, task = fairseq.checkpoint_utils.load_model_ensemble_and_task(
+            [cp_path]
+        )
+        ssl_model = model[0]
+        ssl_model.remove_pretraining_modules()
+    return SSL_model(ssl_model, SSL_OUT_DIM, wavlm)
+
+class SSL_model(nn.Module):
+    def __init__(self,ssl_model,ssl_out_dim,wavlm) -> None:
+        super(SSL_model,self).__init__()
+        self.ssl_model, self.ssl_out_dim = ssl_model, ssl_out_dim
+        self.WavLM = wavlm
+
+    def forward(self,batch):
+        wav = batch['wav'] 
+        wav = wav.squeeze(1) # [batches, audio_len]
+        if self.WavLM:
+            x = self.ssl_model.extract_features(wav)[0]
+        else:
+            res = self.ssl_model(wav, mask=False, features_only=True)
+            x = res["x"]
+        return {"ssl-feature":x}
+    def get_output_dim(self):
+        return self.ssl_out_dim
+
+
+class PhonemeEncoder(nn.Module):
+    '''
+    PhonemeEncoder consists of an embedding layer, an LSTM layer, and a linear layer.
+    Args:
+        vocab_size: the size of the vocabulary
+        hidden_dim: the size of the hidden state of the LSTM
+        emb_dim: the size of the embedding layer
+        out_dim: the size of the output of the linear layer
+        n_lstm_layers: the number of LSTM layers
+    '''
+    def __init__(self, vocab_size, hidden_dim, emb_dim, out_dim,n_lstm_layers,with_reference=True) -> None:
+        super().__init__()
+        self.with_reference = with_reference
+        self.embedding = nn.Embedding(vocab_size, emb_dim)
+        self.encoder = nn.LSTM(emb_dim, hidden_dim,
+                               num_layers=n_lstm_layers, dropout=0.1, bidirectional=True)
+        self.linear = nn.Sequential(
+                nn.Linear(hidden_dim + hidden_dim*self.with_reference, out_dim),
+                nn.ReLU()
+                )
+        self.out_dim = out_dim
+
+    def forward(self,batch):
+        seq = batch['phonemes']
+        lens = batch['phoneme_lens']
+        reference_seq = batch['reference']
+        reference_lens = batch['reference_lens']
+        emb = self.embedding(seq)
+        emb = torch.nn.utils.rnn.pack_padded_sequence(
+            emb, lens, batch_first=True, enforce_sorted=False)
+        _, (ht, _) = self.encoder(emb)
+        feature = ht[-1] + ht[0]
+        if self.with_reference:
+            if reference_seq==None or reference_lens ==None:
+                raise ValueError("reference_batch and reference_lens should not be None when with_reference is True")
+            reference_emb = self.embedding(reference_seq)
+            reference_emb = torch.nn.utils.rnn.pack_padded_sequence(
+                reference_emb, reference_lens, batch_first=True, enforce_sorted=False)
+            _, (ht_ref, _) = self.encoder(emb)
+            reference_feature = ht_ref[-1] + ht_ref[0]
+            feature = self.linear(torch.cat([feature,reference_feature],1))
+        else:
+            feature = self.linear(feature)
+        return {"phoneme-feature": feature}
+    def get_output_dim(self):
+        return self.out_dim
+
+class DomainEmbedding(nn.Module):
+    def __init__(self,n_domains,domain_dim) -> None:
+        super().__init__()
+        self.embedding = nn.Embedding(n_domains,domain_dim)
+        self.output_dim = domain_dim
+    def forward(self, batch):
+        return {"domain-feature": self.embedding(batch['domains'])}
+    def get_output_dim(self):
+        return self.output_dim
+
+
+class LDConditioner(nn.Module):
+    '''
+    Conditions ssl output by listener embedding
+    '''
+    def __init__(self,input_dim, judge_dim, num_judges=None):
+        super().__init__()
+        self.input_dim = input_dim
+        self.judge_dim = judge_dim
+        self.num_judges = num_judges
+        assert num_judges !=None
+        self.judge_embedding = nn.Embedding(num_judges, self.judge_dim)
+        # concat [self.output_layer, phoneme features]
+        
+        self.decoder_rnn = nn.LSTM(
+            input_size = self.input_dim + self.judge_dim,
+            hidden_size = 512,
+            num_layers = 1,
+            batch_first = True,
+            bidirectional = True
+        ) # linear?
+        self.out_dim = self.decoder_rnn.hidden_size*2
+
+    def get_output_dim(self):
+        return self.out_dim
+
+
+    def forward(self, x, batch):
+        judge_ids = batch['judge_id']
+        if 'phoneme-feature' in x.keys():
+            concatenated_feature = torch.cat((x['ssl-feature'], x['phoneme-feature'].unsqueeze(1).expand(-1,x['ssl-feature'].size(1) ,-1)),dim=2)
+        else:
+            concatenated_feature = x['ssl-feature']
+        if 'domain-feature' in x.keys():
+            concatenated_feature = torch.cat(
+                (
+                    concatenated_feature,
+                    x['domain-feature']
+                    .unsqueeze(1)
+                    .expand(-1, concatenated_feature.size(1), -1),
+                ),
+                dim=2,
+            )
+        if judge_ids != None:
+            concatenated_feature = torch.cat(
+                (
+                    concatenated_feature,
+                    self.judge_embedding(judge_ids)
+                    .unsqueeze(1)
+                    .expand(-1, concatenated_feature.size(1), -1),
+                ),
+                dim=2,
+            )
+            decoder_output, (h, c) = self.decoder_rnn(concatenated_feature)
+        return decoder_output
+
+class Projection(nn.Module):
+    def __init__(self, input_dim, hidden_dim, activation, range_clipping=False):
+        super(Projection, self).__init__()
+        self.range_clipping = range_clipping
+        output_dim = 1
+        if range_clipping:
+            self.proj = nn.Tanh()
+        
+        self.net = nn.Sequential(
+            nn.Linear(input_dim, hidden_dim),
+            activation,
+            nn.Dropout(0.3),
+            nn.Linear(hidden_dim, output_dim),
+        )
+        self.output_dim = output_dim
+    
+    def forward(self, x, batch):
+        output = self.net(x)
+
+        # range clipping
+        if self.range_clipping:
+            return self.proj(output) * 2.0 + 3
+        else:
+            return output
+    def get_output_dim(self):
+        return self.output_dim

requirements.txt

@@ -0,0 +1,94 @@
+absl-py==1.0.0
+aiohttp==3.8.1
+aiosignal==1.2.0
+analytics-python==1.4.0
+antlr4-python3-runtime==4.8
+anyio==3.5.0
+asgiref==3.5.0
+async-timeout==4.0.2
+attrs==21.4.0
+backoff==1.10.0
+bcrypt==3.2.0
+bitarray==2.4.0
+cachetools==5.0.0
+certifi==2021.10.8
+cffi==1.15.0
+charset-normalizer==2.0.12
+click==8.0.4
+colorama==0.4.4
+cryptography==36.0.1
+cycler==0.11.0
+Cython==0.29.28
+fairseq @ git+https://github.com/pytorch/fairseq.git@d03f4e771484a433f025f47744017c2eb6e9c6bc
+fastapi==0.75.0
+ffmpy==0.3.0
+fonttools==4.30.0
+frozenlist==1.3.0
+fsspec==2022.2.0
+future==0.18.2
+google-auth==2.6.0
+google-auth-oauthlib==0.4.6
+gradio==2.8.10
+grpcio==1.44.0
+h11==0.13.0
+hydra-core==1.0.7
+idna==3.3
+importlib-metadata==4.11.3
+Jinja2==3.0.3
+kiwisolver==1.3.2
+linkify-it-py==1.0.3
+Markdown==3.3.6
+markdown-it-py==2.0.1
+MarkupSafe==2.1.0
+matplotlib==3.5.1
+mdit-py-plugins==0.3.0
+mdurl==0.1.0
+monotonic==1.6
+multidict==6.0.2
+numpy==1.22.3
+oauthlib==3.2.0
+omegaconf==2.0.6
+orjson==3.6.7
+packaging==21.3
+pandas==1.4.1
+paramiko==2.10.1
+Pillow==9.0.1
+portalocker==2.4.0
+protobuf==3.19.4
+pyasn1==0.4.8
+pyasn1-modules==0.2.8
+pycparser==2.21
+pycryptodome==3.14.1
+pydantic==1.9.0
+pyDeprecate==0.3.1
+pydub==0.25.1
+PyNaCl==1.5.0
+pyparsing==3.0.7
+python-dateutil==2.8.2
+python-multipart==0.0.5
+pytorch-lightning==1.5.10
+pytz==2021.3
+PyYAML==6.0
+regex==2022.3.2
+requests==2.27.1
+requests-oauthlib==1.3.1
+rsa==4.8
+sacrebleu==2.0.0
+six==1.16.0
+sniffio==1.2.0
+starlette==0.17.1
+tabulate==0.8.9
+tensorboard==2.8.0
+tensorboard-data-server==0.6.1
+tensorboard-plugin-wit==1.8.1
+torch==1.11.0
+torchaudio==0.11.0
+torchmetrics==0.7.2
+tqdm==4.63.0
+typing-extensions==4.1.1
+uc-micro-py==1.0.1
+urllib3==1.26.8
+uvicorn==0.17.6
+Werkzeug==2.0.3
+yarl==1.7.2
+zipp==3.7.0

wav2vec_small.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c66c39eaed1b79a61ea8573f71e08f6641ff156b6a8f458cfaab53877dfa4a26
+size 950500491