Upload 24 files

.gitattributes

@@ -1,35 +1,35 @@
-*.7z filter=lfs diff=lfs merge=lfs -text
-*.arrow filter=lfs diff=lfs merge=lfs -text
-*.bin filter=lfs diff=lfs merge=lfs -text
-*.bz2 filter=lfs diff=lfs merge=lfs -text
-*.ckpt filter=lfs diff=lfs merge=lfs -text
-*.ftz filter=lfs diff=lfs merge=lfs -text
-*.gz filter=lfs diff=lfs merge=lfs -text
-*.h5 filter=lfs diff=lfs merge=lfs -text
-*.joblib filter=lfs diff=lfs merge=lfs -text
-*.lfs.* filter=lfs diff=lfs merge=lfs -text
-*.mlmodel filter=lfs diff=lfs merge=lfs -text
-*.model filter=lfs diff=lfs merge=lfs -text
-*.msgpack filter=lfs diff=lfs merge=lfs -text
-*.npy filter=lfs diff=lfs merge=lfs -text
-*.npz filter=lfs diff=lfs merge=lfs -text
-*.onnx filter=lfs diff=lfs merge=lfs -text
-*.ot filter=lfs diff=lfs merge=lfs -text
-*.parquet filter=lfs diff=lfs merge=lfs -text
-*.pb filter=lfs diff=lfs merge=lfs -text
-*.pickle filter=lfs diff=lfs merge=lfs -text
-*.pkl filter=lfs diff=lfs merge=lfs -text
-*.pt filter=lfs diff=lfs merge=lfs -text
-*.pth filter=lfs diff=lfs merge=lfs -text
-*.rar filter=lfs diff=lfs merge=lfs -text
-*.safetensors filter=lfs diff=lfs merge=lfs -text
-saved_model/**/* filter=lfs diff=lfs merge=lfs -text
-*.tar.* filter=lfs diff=lfs merge=lfs -text
-*.tar filter=lfs diff=lfs merge=lfs -text
-*.tflite filter=lfs diff=lfs merge=lfs -text
-*.tgz filter=lfs diff=lfs merge=lfs -text
-*.wasm filter=lfs diff=lfs merge=lfs -text
-*.xz filter=lfs diff=lfs merge=lfs -text
-*.zip filter=lfs diff=lfs merge=lfs -text
-*.zst filter=lfs diff=lfs merge=lfs -text
-*tfevents* filter=lfs diff=lfs merge=lfs -text
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,5 @@
+process.ipynb
+clean_speech/
+noise/
+__pycache__/
+test.py

README.md

@@ -1,12 +1,12 @@
----
-title: Speech Enhancement
-emoji: 🏆
-colorFrom: pink
-colorTo: green
-sdk: streamlit
-sdk_version: 1.36.0
-app_file: app.py
-pinned: false
----
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+---
+title: Speech Denoising
+emoji: ⚡
+colorFrom: red
+colorTo: red
+sdk: streamlit
+sdk_version: 1.36.0
+app_file: app.py
+pinned: false
+---
+
+Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

app.py

@@ -0,0 +1,181 @@
+import io
+import os
+import base64
+import librosa
+import numpy as np
+from io import BytesIO
+import streamlit as st
+from pydub import AudioSegment
+import matplotlib.pyplot as plt
+from scipy.io.wavfile import write
+from src.denoise import denoise
+from myrecorder import recorder
+
+
+SR = 16000
+CONTAINER_HEIGHT = 380
+
+
+def np_audio_to_bytesio(np_audio, np_audio_sr):
+    _bytes = bytes()
+    byte_io = io.BytesIO(_bytes)
+    write(byte_io, np_audio_sr, np_audio)
+    bytes_audio = byte_io.read()
+    return bytes_audio
+
+
+def autoplay_audio(audio: str):
+    audio_base64 = base64.b64encode(audio).decode('utf-8')
+    audio_tag = f'<audio autoplay="true" src="data:audio/wav;base64,{audio_base64}">'
+    st.markdown(audio_tag, unsafe_allow_html=True)
+
+
+def load_noisy_speech(root=os.path.join(os.getcwd(), 'noisy_speech')):
+    noisy_speech_paths = {'EN':{}, 'JA': {}}
+    noisy_speech_names = os.listdir(root)
+    for name in noisy_speech_names:
+        splt = name.split('_')
+        lang, snr = splt[0].upper(), int(splt[1][:2])
+        noisy_speech_paths[lang][snr] = os.path.join(root, name)
+        
+    en_keys = list(noisy_speech_paths['EN'].keys())
+    en_keys.sort()
+    en_keys.reverse()
+    noisy_speech_paths['EN'] = {f'{key}dB': noisy_speech_paths['EN'][key] for key in en_keys}
+    
+    ja_keys = list(noisy_speech_paths['JA'].keys())
+    ja_keys.sort()
+    ja_keys.reverse()
+    noisy_speech_paths['JA'] = {f'{key}dB': noisy_speech_paths['JA'][key] for key in ja_keys}
+    
+    return noisy_speech_paths
+
+
+def load_wav(wav_path):
+    wav_22k, sr = librosa.load(wav_path)
+    wav_16k = librosa.resample(wav_22k, orig_sr=sr, target_sr=SR)
+    return wav_22k, wav_16k
+
+
+def wav_to_spec(wav, sr):
+    if sr == 16000:
+        wav = librosa.resample(wav, orig_sr=sr, target_sr=22050)
+    spec = np.abs(librosa.stft(wav))
+    spec = librosa.amplitude_to_db(spec, ref=np.max)
+    return spec
+
+
+def export_spec_to_buffer(spec):
+    plt.rcParams['figure.figsize'] = (16, 4.5)
+    plt.rc('axes', labelsize=15)
+    plt.rc('xtick', labelsize=15)
+    plt.rc('ytick', labelsize=15)
+    librosa.display.specshow(spec, y_axis='log', x_axis='time')
+    img_buffer = BytesIO()
+    plt.savefig(img_buffer, format='JPEG', bbox_inches='tight', pad_inches=0)
+    return img_buffer
+
+
+def process_recorded_wav_bytes(wav_bytes, sr):
+    file = BytesIO(wav_bytes)
+    audio = AudioSegment.from_file(file=file, format='wav')
+    audio = audio.set_sample_width(2)
+    audio = audio.set_channels(1)
+    audio_22k = audio.set_frame_rate(sr)
+    audio_16k = audio.set_frame_rate(SR)
+    audio_22k = np.array(audio_22k.get_array_of_samples(), dtype=np.float32)
+    audio_16k = np.array(audio_16k.get_array_of_samples(), dtype=np.float32)
+    return audio_22k, audio_16k
+
+
+def main():
+    
+    st.set_page_config(
+        page_title="speech-denoising-app",
+        layout="wide"
+    )
+    
+    logo_space, title_space, _ = st.columns([1, 5, 1], gap="small")
+    
+    with logo_space:
+        st.write(
+            """
+            <div style="display: flex; justify-content: left;">
+                <b><span style="text-align: center; color: #101414; font-size: 14px">FPT Corporation</span></b>
+            </div>
+            """,
+            unsafe_allow_html=True
+        )
+        st.image('aic-logo.png')
+    
+    with title_space:
+        st.image('logo.png')
+        
+    noisy_speech_files = load_noisy_speech()
+
+    input_space, output_space = st.columns([1, 1], gap="medium")
+    _, record_space, _, compute_space= st.columns([0.7, 1, 1, 1], gap="small")
+    
+    with record_space:
+        record = recorder(
+            start_prompt="Start Recording",
+            stop_prompt="Stop Recording",
+            just_once=False,
+            use_container_width=False,
+            format="wav",
+            callback=None,
+            args=(),
+            kwargs={},
+            key=None
+        )
+        
+    with compute_space:
+        compute = st.button('Denoise')
+    
+    with input_space.container(height=CONTAINER_HEIGHT, border=True):
+        lang_select_space, snr_select_space = st.columns([1, 1], gap="small")
+        with lang_select_space:
+            language_select = st.selectbox("Language", list(noisy_speech_files.keys()))
+        with snr_select_space:
+            if language_select:
+                snr_select = st.selectbox("SNR Level", list(noisy_speech_files[language_select].keys()))
+        
+        if record:
+            wav_bytes_record = record['bytes']
+            sr = record['sample_rate']
+            noisy_wav_22k, noisy_wav = process_recorded_wav_bytes(wav_bytes_record, sr=22050)
+            noisy_spec = wav_to_spec(noisy_wav_22k, sr=22050)
+            noisy_spec_buff = export_spec_to_buffer(noisy_spec)
+            
+            st.audio(wav_bytes_record, format="wav")
+            st.image(image=noisy_spec_buff)
+
+        elif language_select and snr_select:
+            audio_path = noisy_speech_files[language_select][snr_select]
+            noisy_wav_22k, noisy_wav = load_wav(audio_path)
+            noisy_spec = wav_to_spec(noisy_wav_22k, sr=22050)
+            noisy_spec_buff = export_spec_to_buffer(noisy_spec)
+        
+            st.audio(audio_path, format="wav")
+            st.image(image=noisy_spec_buff)
+    
+    with output_space.container(height=CONTAINER_HEIGHT, border=True):
+        st.write(
+            """
+            <div style="display: flex; justify-content: center;">
+                <b><span style="text-align: center; color: #808080; font-size: 51.5px">Output</span></b>
+            </div>
+            """,
+            unsafe_allow_html=True
+        )
+        if noisy_wav.any() and compute:
+            denoised_wav = denoise(noisy_wav)
+            st.audio(denoised_wav, sample_rate=SR, format="audio/wav")
+            denoised_spec = wav_to_spec(denoised_wav, sr=SR)
+            denoised_spec_buff = export_spec_to_buffer(denoised_spec)
+            st.image(image=denoised_spec_buff)
+            record = None
+        
+
+if __name__ == '__main__':
+    main()

ckpt/full.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:145c101eb5bbfa3ba52fb2b4ec7e5b64a361c102f89291f75e1dd42601d95dc9
+size 184336765

ckpt/high.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:513d9e4f69483bf2bcc3059dd6b3644140763bf3f22df41d7ee366cc2cbd1829
+size 184336765

configs/full.json

@@ -0,0 +1,54 @@
+{   
+    "network_config": {
+        "channels_input": 1,
+        "channels_output": 1,
+        "channels_H": 64,
+        "max_H": 768,
+        "encoder_n_layers": 8,
+        "kernel_size": 4,
+        "stride": 2,
+        "tsfm_n_layers": 5, 
+        "tsfm_n_head": 8,
+        "tsfm_d_model": 512, 
+        "tsfm_d_inner": 2048
+    },
+    "train_config": {
+        "exp_path": "DNS-large-full",
+        "log":{
+            "directory": "./exp",
+            "ckpt_iter": "max",
+            "iters_per_ckpt": 10000,
+            "iters_per_valid": 500
+        },
+        "optimization":{
+            "n_iters": 250000,
+            "learning_rate": 2e-4,
+            "batch_size_per_gpu": 8
+        },
+        "loss_config":{
+            "ell_p": 1,
+            "ell_p_lambda": 1,
+            "stft_lambda": 1,
+            "stft_config":{
+                "sc_lambda": 0.5,
+                "mag_lambda": 0.5,
+                "band": "full",
+                "hop_sizes": [50, 120, 240],
+                "win_lengths": [240, 600, 1200],
+                "fft_sizes": [512, 1024, 2048]
+            }
+        }
+    },
+    "trainset_config": {
+        "root": "./dns",
+        "crop_length_sec": 10,
+        "sample_rate": 16000
+    },
+    "gen_config":{
+        "output_directory": "./exp"
+    },
+    "dist_config": {
+        "dist_backend": "nccl",
+        "dist_url": "tcp://localhost:54321"
+    }
+}

configs/high.json

@@ -0,0 +1,54 @@
+{   
+    "network_config": {
+        "channels_input": 1,
+        "channels_output": 1,
+        "channels_H": 64,
+        "max_H": 768,
+        "encoder_n_layers": 8,
+        "kernel_size": 4,
+        "stride": 2,
+        "tsfm_n_layers": 5, 
+        "tsfm_n_head": 8,
+        "tsfm_d_model": 512, 
+        "tsfm_d_inner": 2048
+    },
+    "train_config": {
+        "exp_path": "DNS-large-high",
+        "log":{
+            "directory": "./exp",
+            "ckpt_iter": "max",
+            "iters_per_ckpt": 10000,
+            "iters_per_valid": 500
+        },
+        "optimization":{
+            "n_iters": 250000,
+            "learning_rate": 2e-4,
+            "batch_size_per_gpu": 8
+        },
+        "loss_config":{
+            "ell_p": 1,
+            "ell_p_lambda": 1,
+            "stft_lambda": 1,
+            "stft_config":{
+                "sc_lambda": 0.5,
+                "mag_lambda": 0.5,
+                "band": "high",
+                "hop_sizes": [50, 120, 240],
+                "win_lengths": [240, 600, 1200],
+                "fft_sizes": [512, 1024, 2048]
+            }
+        }
+    },
+    "trainset_config": {
+        "root": "./dns",
+        "crop_length_sec": 10,
+        "sample_rate": 16000
+    },
+    "gen_config":{
+        "output_directory": "./exp"
+    },
+    "dist_config": {
+        "dist_backend": "nccl",
+        "dist_url": "tcp://localhost:54321"
+    }
+}

requirements.txt

@@ -0,0 +1,6 @@
+numpy <= 1.25
+streamlit
+scipy
+myrecorder
+librosa
+torch

src/denoise.py

@@ -0,0 +1,33 @@
+import os
+import json
+import torch
+import random
+random.seed(0)
+torch.manual_seed(0)
+import numpy as np
+np.random.seed(0)
+from src.model import DenoisingModel
+
+
+def denoise(
+    wav: np.ndarray, 
+    ckpt_path: str = os.path.join(os.getcwd(), 'ckpt', 'full.pkl'),
+    cfg_path: str = os.path.join(os.getcwd(), 'configs', 'full.json'),
+):
+
+    with open(cfg_path) as f:
+        data = f.read()
+    config = json.loads(data)
+
+    net = DenoisingModel(**config['network_config']).to('cpu')
+
+    # load checkpoint
+    checkpoint = torch.load(ckpt_path, map_location='cpu')
+    net.load_state_dict(checkpoint['model_state_dict'])
+    net.eval()
+
+    # inference
+    wav = torch.from_numpy(wav).unsqueeze(0)
+    wav_denoised = net(wav).squeeze(0).detach().numpy().reshape(-1)
+
+    return wav_denoised

src/model.py

@@ -0,0 +1,385 @@
+# Copyright (c) 2022 NVIDIA CORPORATION. 
+#   Licensed under the MIT license.
+
+import numpy as np
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from src.utils import weight_scaling_init
+
+
+# Transformer (encoder) https://github.com/jadore801120/attention-is-all-you-need-pytorch
+# Original Copyright 2017 Victor Huang
+#  MIT License (https://opensource.org/licenses/MIT)
+
+class ScaledDotProductAttention(nn.Module):
+    ''' Scaled Dot-Product Attention '''
+
+    def __init__(self, temperature, attn_dropout=0.1):
+        super().__init__()
+        self.temperature = temperature
+        self.dropout = nn.Dropout(attn_dropout)
+
+    def forward(self, q, k, v, mask=None):
+
+        attn = torch.matmul(q / self.temperature, k.transpose(2, 3))
+
+        if mask is not None:
+            attn = attn.masked_fill(mask == 0, -1e9)
+
+        attn = self.dropout(F.softmax(attn, dim=-1))
+        output = torch.matmul(attn, v)
+
+        return output, attn
+
+
+class MultiHeadAttention(nn.Module):
+    ''' Multi-Head Attention module '''
+
+    def __init__(self, n_head, d_model, d_k, d_v, dropout=0.1):
+        super().__init__()
+
+        self.n_head = n_head
+        self.d_k = d_k
+        self.d_v = d_v
+
+        self.w_qs = nn.Linear(d_model, n_head * d_k, bias=False)
+        self.w_ks = nn.Linear(d_model, n_head * d_k, bias=False)
+        self.w_vs = nn.Linear(d_model, n_head * d_v, bias=False)
+        self.fc = nn.Linear(n_head * d_v, d_model, bias=False)
+
+        self.attention = ScaledDotProductAttention(temperature=d_k ** 0.5)
+
+        self.dropout = nn.Dropout(dropout)
+        self.layer_norm = nn.LayerNorm(d_model, eps=1e-6)
+
+
+    def forward(self, q, k, v, mask=None):
+
+        d_k, d_v, n_head = self.d_k, self.d_v, self.n_head
+        sz_b, len_q, len_k, len_v = q.size(0), q.size(1), k.size(1), v.size(1)
+
+        residual = q
+
+        # Pass through the pre-attention projection: b x lq x (n*dv)
+        # Separate different heads: b x lq x n x dv
+        q = self.w_qs(q).view(sz_b, len_q, n_head, d_k)
+        k = self.w_ks(k).view(sz_b, len_k, n_head, d_k)
+        v = self.w_vs(v).view(sz_b, len_v, n_head, d_v)
+
+        # Transpose for attention dot product: b x n x lq x dv
+        q, k, v = q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2)
+
+        if mask is not None:
+            mask = mask.unsqueeze(1)   # For head axis broadcasting.
+
+        q, attn = self.attention(q, k, v, mask=mask)
+
+        # Transpose to move the head dimension back: b x lq x n x dv
+        # Combine the last two dimensions to concatenate all the heads together: b x lq x (n*dv)
+        q = q.transpose(1, 2).contiguous().view(sz_b, len_q, -1)
+        q = self.dropout(self.fc(q))
+        q += residual
+
+        q = self.layer_norm(q)
+
+        return q, attn
+
+
+class PositionwiseFeedForward(nn.Module):
+    ''' A two-feed-forward-layer module '''
+
+    def __init__(self, d_in, d_hid, dropout=0.1):
+        super().__init__()
+        self.w_1 = nn.Linear(d_in, d_hid) # position-wise
+        self.w_2 = nn.Linear(d_hid, d_in) # position-wise
+        self.layer_norm = nn.LayerNorm(d_in, eps=1e-6)
+        self.dropout = nn.Dropout(dropout)
+
+    def forward(self, x):
+
+        residual = x
+
+        x = self.w_2(F.relu(self.w_1(x)))
+        x = self.dropout(x)
+        x += residual
+
+        x = self.layer_norm(x)
+
+        return x
+
+
+def get_subsequent_mask(seq):
+    ''' For masking out the subsequent info. '''
+    sz_b, len_s = seq.size()
+    subsequent_mask = (1 - torch.triu(
+        torch.ones((1, len_s, len_s), device=seq.device), diagonal=1)).bool()
+    return subsequent_mask
+
+
+class PositionalEncoding(nn.Module):
+
+    def __init__(self, d_hid, n_position=200):
+        super(PositionalEncoding, self).__init__()
+
+        # Not a parameter
+        self.register_buffer('pos_table', self._get_sinusoid_encoding_table(n_position, d_hid))
+
+    def _get_sinusoid_encoding_table(self, n_position, d_hid):
+        ''' Sinusoid position encoding table '''
+        # TODO: make it with torch instead of numpy
+
+        def get_position_angle_vec(position):
+            return [position / np.power(10000, 2 * (hid_j // 2) / d_hid) for hid_j in range(d_hid)]
+
+        sinusoid_table = np.array([get_position_angle_vec(pos_i) for pos_i in range(n_position)])
+        sinusoid_table[:, 0::2] = np.sin(sinusoid_table[:, 0::2])  # dim 2i
+        sinusoid_table[:, 1::2] = np.cos(sinusoid_table[:, 1::2])  # dim 2i+1
+
+        return torch.FloatTensor(sinusoid_table).unsqueeze(0)
+
+    def forward(self, x):
+        return x + self.pos_table[:, :x.size(1)].clone().detach()
+
+
+class EncoderLayer(nn.Module):
+    ''' Compose with two layers '''
+
+    def __init__(self, d_model, d_inner, n_head, d_k, d_v, dropout=0.0):
+        super(EncoderLayer, self).__init__()
+        self.slf_attn = MultiHeadAttention(n_head, d_model, d_k, d_v, dropout=dropout)
+        self.pos_ffn = PositionwiseFeedForward(d_model, d_inner, dropout=dropout)
+
+    def forward(self, enc_input, slf_attn_mask=None):
+        enc_output, enc_slf_attn = self.slf_attn(
+            enc_input, enc_input, enc_input, mask=slf_attn_mask)
+        enc_output = self.pos_ffn(enc_output)
+        return enc_output, enc_slf_attn
+
+
+class TransformerEncoder(nn.Module):
+    ''' A encoder model with self attention mechanism. '''
+
+    def __init__(
+            self, d_word_vec=512, n_layers=2, n_head=8, d_k=64, d_v=64,
+            d_model=512, d_inner=2048, dropout=0.1, n_position=624, scale_emb=False):
+
+        super().__init__()
+
+        # self.src_word_emb = nn.Embedding(n_src_vocab, d_word_vec, padding_idx=pad_idx)
+        if n_position > 0:
+            self.position_enc = PositionalEncoding(d_word_vec, n_position=n_position)
+        else:
+            self.position_enc = lambda x: x
+        self.dropout = nn.Dropout(p=dropout)
+        self.layer_stack = nn.ModuleList([
+            EncoderLayer(d_model, d_inner, n_head, d_k, d_v, dropout=dropout)
+            for _ in range(n_layers)])
+        self.layer_norm = nn.LayerNorm(d_model, eps=1e-6)
+        self.scale_emb = scale_emb
+        self.d_model = d_model
+
+    def forward(self, src_seq, src_mask, return_attns=False):
+
+        enc_slf_attn_list = []
+
+        # -- Forward
+        # enc_output = self.src_word_emb(src_seq)
+        enc_output = src_seq
+        if self.scale_emb:
+            enc_output *= self.d_model ** 0.5
+        enc_output = self.dropout(self.position_enc(enc_output))
+        enc_output = self.layer_norm(enc_output)
+
+        for enc_layer in self.layer_stack:
+            enc_output, enc_slf_attn = enc_layer(enc_output, slf_attn_mask=src_mask)
+            enc_slf_attn_list += [enc_slf_attn] if return_attns else []
+
+        if return_attns:
+            return enc_output, enc_slf_attn_list
+        return enc_output
+
+
+# CleanUNet architecture
+
+
+def padding(x, D, K, S):
+    """padding zeroes to x so that denoised audio has the same length"""
+
+    L = x.shape[-1]
+    for _ in range(D):
+        if L < K:
+            L = 1
+        else:
+            L = 1 + np.ceil((L - K) / S)
+
+    for _ in range(D):
+        L = (L - 1) * S + K
+    
+    L = int(L)
+    x = F.pad(x, (0, L - x.shape[-1]))
+    return x
+
+
+class DenoisingModel(nn.Module):
+    """ CleanUNet architecture. """
+
+    def __init__(self, channels_input=1, channels_output=1,
+                 channels_H=64, max_H=768,
+                 encoder_n_layers=8, kernel_size=4, stride=2,
+                 tsfm_n_layers=3, 
+                 tsfm_n_head=8,
+                 tsfm_d_model=512, 
+                 tsfm_d_inner=2048):
+        
+        """
+        Parameters:
+        channels_input (int):   input channels
+        channels_output (int):  output channels
+        channels_H (int):       middle channels H that controls capacity
+        max_H (int):            maximum H
+        encoder_n_layers (int): number of encoder/decoder layers D
+        kernel_size (int):      kernel size K
+        stride (int):           stride S
+        tsfm_n_layers (int):    number of self attention blocks N
+        tsfm_n_head (int):      number of heads in each self attention block
+        tsfm_d_model (int):     d_model of self attention
+        tsfm_d_inner (int):     d_inner of self attention
+        """
+
+        super(DenoisingModel, self).__init__()
+
+        self.channels_input = channels_input
+        self.channels_output = channels_output
+        self.channels_H = channels_H
+        self.max_H = max_H
+        self.encoder_n_layers = encoder_n_layers
+        self.kernel_size = kernel_size
+        self.stride = stride
+
+        self.tsfm_n_layers = tsfm_n_layers
+        self.tsfm_n_head = tsfm_n_head
+        self.tsfm_d_model = tsfm_d_model
+        self.tsfm_d_inner = tsfm_d_inner
+
+        # encoder and decoder
+        self.encoder = nn.ModuleList()
+        self.decoder = nn.ModuleList()
+
+        for i in range(encoder_n_layers):
+            self.encoder.append(nn.Sequential(
+                nn.Conv1d(channels_input, channels_H, kernel_size, stride),
+                nn.ReLU(),
+                nn.Conv1d(channels_H, channels_H * 2, 1), 
+                nn.GLU(dim=1)
+            ))
+            channels_input = channels_H
+
+            if i == 0:
+                # no relu at end
+                self.decoder.append(nn.Sequential(
+                    nn.Conv1d(channels_H, channels_H * 2, 1), 
+                    nn.GLU(dim=1),
+                    nn.ConvTranspose1d(channels_H, channels_output, kernel_size, stride)
+                ))
+            else:
+                self.decoder.insert(0, nn.Sequential(
+                    nn.Conv1d(channels_H, channels_H * 2, 1), 
+                    nn.GLU(dim=1),
+                    nn.ConvTranspose1d(channels_H, channels_output, kernel_size, stride),
+                    nn.ReLU()
+                ))
+            channels_output = channels_H
+            
+            # double H but keep below max_H
+            channels_H *= 2
+            channels_H = min(channels_H, max_H)
+        
+        # self attention block
+        self.tsfm_conv1 = nn.Conv1d(channels_output, tsfm_d_model, kernel_size=1)
+        self.tsfm_encoder = TransformerEncoder(d_word_vec=tsfm_d_model, 
+                                               n_layers=tsfm_n_layers, 
+                                               n_head=tsfm_n_head, 
+                                               d_k=tsfm_d_model // tsfm_n_head, 
+                                               d_v=tsfm_d_model // tsfm_n_head, 
+                                               d_model=tsfm_d_model, 
+                                               d_inner=tsfm_d_inner, 
+                                               dropout=0.0, 
+                                               n_position=0, 
+                                               scale_emb=False)
+        self.tsfm_conv2 = nn.Conv1d(tsfm_d_model, channels_output, kernel_size=1)
+
+        # weight scaling initialization
+        for layer in self.modules():
+            if isinstance(layer, (nn.Conv1d, nn.ConvTranspose1d)):
+                weight_scaling_init(layer)
+
+    def forward(self, noisy_audio):
+        # (B, L) -> (B, C, L)
+        if len(noisy_audio.shape) == 2:
+            noisy_audio = noisy_audio.unsqueeze(1)
+        B, C, L = noisy_audio.shape
+        assert C == 1
+        
+        # normalization and padding
+        std = noisy_audio.std(dim=2, keepdim=True) + 1e-3
+        noisy_audio /= std
+        x = padding(noisy_audio, self.encoder_n_layers, self.kernel_size, self.stride)
+        
+        # encoder
+        skip_connections = []
+        for downsampling_block in self.encoder:
+            x = downsampling_block(x)
+            skip_connections.append(x)
+        skip_connections = skip_connections[::-1]
+
+        # attention mask for causal inference; for non-causal, set attn_mask to None
+        len_s = x.shape[-1]  # length at bottleneck
+        attn_mask = (1 - torch.triu(torch.ones((1, len_s, len_s), device=x.device), diagonal=1)).bool()
+
+        x = self.tsfm_conv1(x)  # C 1024 -> 512
+        x = x.permute(0, 2, 1)
+        x = self.tsfm_encoder(x, src_mask=attn_mask)
+        x = x.permute(0, 2, 1)
+        x = self.tsfm_conv2(x)  # C 512 -> 1024
+
+        # decoder
+        for i, upsampling_block in enumerate(self.decoder):
+            skip_i = skip_connections[i]
+            x += skip_i[:, :, :x.shape[-1]]
+            x = upsampling_block(x)
+
+        x = x[:, :, :L] * std
+        return x
+
+
+if __name__ == '__main__':
+    import json
+    import argparse 
+    import os
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument('-c', '--config', type=str, default='configs/DNS-large-full.json', 
+                        help='JSON file for configuration')
+    args = parser.parse_args()
+
+    with open(args.config) as f:
+        data = f.read()
+    config = json.loads(data)
+    network_config = config["network_config"]
+
+    model = CleanUNet(**network_config).cuda()
+    from util import print_size
+    print_size(model, keyword="tsfm")
+    
+    input_data = torch.ones([4,1,int(4.5*16000)]).cuda()
+    output = model(input_data)
+    print(output.shape)
+
+    y = torch.rand([4,1,int(4.5*16000)]).cuda()
+    loss = torch.nn.MSELoss()(y, output)
+    loss.backward()
+    print(loss.item())
+    

src/utils.py

@@ -0,0 +1,223 @@
+import os
+import time 
+import functools
+import numpy as np
+from math import cos, pi, floor, sin
+from tqdm import tqdm
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+# from stft_loss import MultiResolutionSTFTLoss
+
+
+def flatten(v):
+    return [x for y in v for x in y]
+
+
+def rescale(x):
+    return (x - x.min()) / (x.max() - x.min())
+
+
+def find_max_epoch(path):
+    """
+    Find latest checkpoint
+    
+    Returns:
+    maximum iteration, -1 if there is no (valid) checkpoint
+    """
+
+    files = os.listdir(path)
+    epoch = -1
+    for f in files:
+        if len(f) <= 4:
+            continue
+        if f[-4:]  == '.pkl':
+            number = f[:-4]
+            try:
+                epoch = max(epoch, int(number))
+            except:
+                continue
+    return epoch
+
+
+def print_size(net, keyword=None):
+    """
+    Print the number of parameters of a network
+    """
+
+    if net is not None and isinstance(net, torch.nn.Module):
+        module_parameters = filter(lambda p: p.requires_grad, net.parameters())
+        params = sum([np.prod(p.size()) for p in module_parameters])
+        
+        print("{} Parameters: {:.6f}M".format(
+            net.__class__.__name__, params / 1e6), flush=True, end="; ")
+        
+        if keyword is not None:
+            keyword_parameters = [p for name, p in net.named_parameters() if p.requires_grad and keyword in name]
+            params = sum([np.prod(p.size()) for p in keyword_parameters])
+            print("{} Parameters: {:.6f}M".format(
+                keyword, params / 1e6), flush=True, end="; ")
+        
+        print(" ")
+
+
+####################### lr scheduler: Linear Warmup then Cosine Decay #############################
+
+# Adapted from https://github.com/rosinality/vq-vae-2-pytorch
+
+# Original Copyright 2019 Kim Seonghyeon
+#  MIT License (https://opensource.org/licenses/MIT)
+
+
+def anneal_linear(start, end, proportion):
+    return start + proportion * (end - start)
+
+
+def anneal_cosine(start, end, proportion):
+    cos_val = cos(pi * proportion) + 1
+    return end + (start - end) / 2 * cos_val
+
+
+class Phase:
+    def __init__(self, start, end, n_iter, cur_iter, anneal_fn):
+        self.start, self.end = start, end
+        self.n_iter = n_iter
+        self.anneal_fn = anneal_fn
+        self.n = cur_iter
+
+    def step(self):
+        self.n += 1
+
+        return self.anneal_fn(self.start, self.end, self.n / self.n_iter)
+
+    def reset(self):
+        self.n = 0
+
+    @property
+    def is_done(self):
+        return self.n >= self.n_iter
+
+
+class LinearWarmupCosineDecay:
+    def __init__(
+        self,
+        optimizer,
+        lr_max,
+        n_iter,
+        iteration=0,
+        divider=25,
+        warmup_proportion=0.3,
+        phase=('linear', 'cosine'),
+    ):
+        self.optimizer = optimizer
+
+        phase1 = int(n_iter * warmup_proportion)
+        phase2 = n_iter - phase1
+        lr_min = lr_max / divider
+
+        phase_map = {'linear': anneal_linear, 'cosine': anneal_cosine}
+
+        cur_iter_phase1 = iteration
+        cur_iter_phase2 = max(0, iteration - phase1)
+        self.lr_phase = [
+            Phase(lr_min, lr_max, phase1, cur_iter_phase1, phase_map[phase[0]]),
+            Phase(lr_max, lr_min / 1e4, phase2, cur_iter_phase2, phase_map[phase[1]]),
+        ]
+
+        if iteration < phase1:
+            self.phase = 0
+        else:
+            self.phase = 1
+
+    def step(self):
+        lr = self.lr_phase[self.phase].step()
+
+        for group in self.optimizer.param_groups:
+            group['lr'] = lr
+
+        if self.lr_phase[self.phase].is_done:
+            self.phase += 1
+
+        if self.phase >= len(self.lr_phase):
+            for phase in self.lr_phase:
+                phase.reset()
+
+            self.phase = 0
+
+        return lr
+
+
+####################### model util #############################
+
+def std_normal(size):
+    """
+    Generate the standard Gaussian variable of a certain size
+    """
+
+    return torch.normal(0, 1, size=size).cuda()
+
+
+def weight_scaling_init(layer):
+    """
+    weight rescaling initialization from https://arxiv.org/abs/1911.13254
+    """
+    w = layer.weight.detach()
+    alpha = 10.0 * w.std()
+    layer.weight.data /= torch.sqrt(alpha)
+    layer.bias.data /= torch.sqrt(alpha)
+
+
+@torch.no_grad()
+def sampling(net, noisy_audio):
+    """
+    Perform denoising (forward) step
+    """
+
+    return net(noisy_audio)
+
+
+def loss_fn(net, X, ell_p, ell_p_lambda, stft_lambda, mrstftloss, **kwargs):
+    """
+    Loss function in CleanUNet
+
+    Parameters:
+    net: network
+    X: training data pair (clean audio, noisy_audio)
+    ell_p: \ell_p norm (1 or 2) of the AE loss
+    ell_p_lambda: factor of the AE loss
+    stft_lambda: factor of the STFT loss
+    mrstftloss: multi-resolution STFT loss function
+
+    Returns:
+    loss: value of objective function
+    output_dic: values of each component of loss
+    """
+
+    assert type(X) == tuple and len(X) == 2
+    
+    clean_audio, noisy_audio = X
+    B, C, L = clean_audio.shape
+    output_dic = {}
+    loss = 0.0
+    
+    # AE loss
+    denoised_audio = net(noisy_audio)  
+
+    if ell_p == 2:
+        ae_loss = nn.MSELoss()(denoised_audio, clean_audio)
+    elif ell_p == 1:
+        ae_loss = F.l1_loss(denoised_audio, clean_audio)
+    else:
+        raise NotImplementedError
+    loss += ae_loss * ell_p_lambda
+    output_dic["reconstruct"] = ae_loss.data * ell_p_lambda
+
+    if stft_lambda > 0:
+        sc_loss, mag_loss = mrstftloss(denoised_audio.squeeze(1), clean_audio.squeeze(1))
+        loss += (sc_loss + mag_loss) * stft_lambda
+        output_dic["stft_sc"] = sc_loss.data * stft_lambda
+        output_dic["stft_mag"] = mag_loss.data * stft_lambda
+
+    return loss, output_dic