update ExtractTTSSpecs notebook

notebooks/ExtractTTSpectrogram.ipynb

@@ -7,15 +7,6 @@
     "This is a notebook to generate mel-spectrograms from a TTS model to be used for WaveRNN training."
    ]
   },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "TTS_PATH = \"/home/erogol/projects/\""
-   ]
-  },
   {
    "cell_type": "code",
    "execution_count": null,
@@ -26,7 +17,6 @@
     "%autoreload 2\n",
     "import os\n",
     "import sys\n",
-    "sys.path.append(TTS_PATH)\n",
     "import torch\n",
     "import importlib\n",
     "import numpy as np\n",
@@ -42,7 +32,7 @@
     "%matplotlib inline\n",
     "\n",
     "import os\n",
-    "os.environ['CUDA_VISIBLE_DEVICES']='2'"
+    "os.environ['CUDA_VISIBLE_DEVICES']='0'"
    ]
   },
   {
@@ -69,12 +59,12 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "OUT_PATH = \"/data/rw/pit/data/turkish-vocoder/\"\n",
-    "DATA_PATH = \"/data/rw/home/Turkish\"\n",
+    "OUT_PATH = \"/home/erogol/Data/LJSpeech-1.1/ljspeech-March-17-2020_01+16AM-871588c/\"\n",
+    "DATA_PATH = \"/home/erogol/Data/LJSpeech-1.1/\"\n",
     "DATASET = \"ljspeech\"\n",
-    "METADATA_FILE = \"metadata.txt\"\n",
-    "CONFIG_PATH = \"/data/rw/pit/keep/turkish-January-08-2020_01+56AM-ca5e133/config.json\"\n",
-    "MODEL_FILE = \"/data/rw/pit/keep/turkish-January-08-2020_01+56AM-ca5e133/checkpoint_255000.pth.tar\"\n",
+    "METADATA_FILE = \"metadata.csv\"\n",
+    "CONFIG_PATH = \"/home/erogol/Models/LJSpeech/ljspeech-March-17-2020_01+16AM-871588c/config.json\"\n",
+    "MODEL_FILE = \"/home/erogol/Models/LJSpeech/ljspeech-March-17-2020_01+16AM-871588c/checkpoint_420000.pth.tar\"\n",
     "BATCH_SIZE = 32\n",
     "\n",
     "QUANTIZED_WAV = False\n",
@@ -85,6 +75,7 @@
     "print(\" > CUDA enabled: \", use_cuda)\n",
     "\n",
     "C = load_config(CONFIG_PATH)\n",
+    "C.audio['do_trim_silence'] = False  # IMPORTANT!!!!!!!!!!!!!!! disable to align mel specs with the wav files\n",
     "ap = AudioProcessor(bits=QUANTIZE_BIT, **C.audio)"
    ]
   },
@@ -94,7 +85,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-   "# if the vocabulary was passed, replace the default\n",
+    "# if the vocabulary was passed, replace the default\n",
     "if 'characters' in C.keys():\n",
     "    symbols, phonemes = make_symbols(**C.characters)\n",
     "\n",
@@ -120,7 +111,7 @@
     "preprocessor = importlib.import_module('TTS.datasets.preprocess')\n",
     "preprocessor = getattr(preprocessor, DATASET.lower())\n",
     "meta_data = preprocessor(DATA_PATH,METADATA_FILE)\n",
-    "dataset = MyDataset(checkpoint['r'], C.text_cleaner, ap, meta_data,tp=C.characters if 'characters' in C.keys() else None, use_phonemes=C.use_phonemes,  phoneme_cache_path=C.phoneme_cache_path, enable_eos_bos=C.enable_eos_bos_chars)\n",
+    "dataset = MyDataset(checkpoint['r'], C.text_cleaner, False, ap, meta_data,tp=C.characters if 'characters' in C.keys() else None, use_phonemes=C.use_phonemes,  phoneme_cache_path=C.phoneme_cache_path, enable_eos_bos=C.enable_eos_bos_chars)\n",
     "loader = DataLoader(dataset, batch_size=BATCH_SIZE, num_workers=4, collate_fn=dataset.collate_fn, shuffle=False, drop_last=False)"
    ]
   },
@@ -143,7 +134,7 @@
     "metadata = []\n",
     "losses = []\n",
     "postnet_losses = []\n",
-    "criterion = L1LossMasked()\n",
+    "criterion = L1LossMasked(seq_len_norm=C.seq_len_norm)\n",
     "with torch.no_grad():\n",
     "    for data in tqdm(loader):\n",
     "        # setup input data\n",
@@ -232,7 +223,31 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "### Check model performance"
+    "### Sanity Check"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "idx = 1\n",
+    "ap.melspectrogram(ap.load_wav(item_idx[idx])).shape"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import soundfile as sf\n",
+    "wav, sr = sf.read(item_idx[idx])\n",
+    "mel_postnet = postnet_outputs[idx][:mel_lengths[idx], :]\n",
+    "mel_decoder = mel_outputs[idx][:mel_lengths[idx], :].detach().cpu().numpy()\n",
+    "mel_truth = ap.melspectrogram(wav)\n",
+    "print(mel_truth.shape)"
    ]
   },
   {
@@ -242,10 +257,8 @@
    "outputs": [],
    "source": [
     "# plot posnet output\n",
-    "idx = 1\n",
-    "mel_example = postnet_outputs[idx]\n",
-    "plot_spectrogram(mel_example[:mel_lengths[idx], :], ap);\n",
-    "print(mel_example[:mel_lengths[1], :].shape)"
+    "plot_spectrogram(mel_postnet, ap);\n",
+    "print(mel_postnet[:mel_lengths[idx], :].shape)"
    ]
   },
   {
@@ -255,9 +268,8 @@
    "outputs": [],
    "source": [
     "# plot decoder output\n",
-    "mel_example = mel_outputs[idx].data.cpu().numpy()\n",
-    "plot_spectrogram(mel_example[:mel_lengths[idx], :], ap);\n",
-    "print(mel_example[:mel_lengths[1], :].shape)"
+    "plot_spectrogram(mel_decoder, ap);\n",
+    "print(mel_decoder.shape)"
    ]
   },
   {
@@ -267,10 +279,8 @@
    "outputs": [],
    "source": [
     "# plot GT specgrogram\n",
-    "wav = ap.load_wav(item_idx[idx])\n",
-    "melt = ap.melspectrogram(wav)\n",
-    "print(melt.shape)\n",
-    "plot_spectrogram(melt.T, ap);"
+    "print(mel_truth.shape)\n",
+    "plot_spectrogram(mel_truth.T, ap);"
    ]
   },
   {
@@ -281,9 +291,9 @@
    "source": [
     "# postnet, decoder diff\n",
     "from matplotlib import pylab as plt\n",
-    "mel_diff = mel_outputs[idx] - postnet_outputs[idx]\n",
+    "mel_diff = mel_decoder - mel_postnet\n",
     "plt.figure(figsize=(16, 10))\n",
-    "plt.imshow(abs(mel_diff.detach().cpu().numpy()[:mel_lengths[idx],:]).T,aspect=\"auto\", origin=\"lower\");\n",
+    "plt.imshow(abs(mel_diff[:mel_lengths[idx],:]).T,aspect=\"auto\", origin=\"lower\");\n",
     "plt.colorbar()\n",
     "plt.tight_layout()"
    ]
@@ -294,10 +304,25 @@
    "metadata": {},
    "outputs": [],
    "source": [
+    "# PLOT GT SPECTROGRAM diff\n",
     "from matplotlib import pylab as plt\n",
-    "# mel = mel_poutputs[idx].detach().cpu().numpy()\n",
-    "mel = postnet_outputs[idx].detach().cpu().numpy()\n",
-    "mel_diff2 = melt.T - mel[:melt.shape[1]]\n",
+    "mel_diff2 = mel_truth.T - mel_decoder\n",
+    "plt.figure(figsize=(16, 10))\n",
+    "plt.imshow(abs(mel_diff2).T,aspect=\"auto\", origin=\"lower\");\n",
+    "plt.colorbar()\n",
+    "plt.tight_layout()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# PLOT GT SPECTROGRAM diff\n",
+    "from matplotlib import pylab as plt\n",
+    "mel = postnet_outputs[idx]\n",
+    "mel_diff2 = mel_truth.T - mel[:mel_truth.shape[1]]\n",
     "plt.figure(figsize=(16, 10))\n",
     "plt.imshow(abs(mel_diff2).T,aspect=\"auto\", origin=\"lower\");\n",
     "plt.colorbar()\n",

utils/audio.py

@@ -269,12 +269,11 @@ class AudioProcessor(object):
             y = self._istft(S_complex * angles)
         return y
 
-    def compute_stft_paddings(x, fsize, fshift, pad_sides=1):
+    def compute_stft_paddings(x, pad_sides=1):
         '''compute right padding (final frame) or both sides padding (first and final frames)
         '''
         assert pad_sides in (1, 2)
-        # return int(fsize // 2)
-        pad = (x.shape[0] // fshift + 1) * fshift - x.shape[0]
+        pad = (x.shape[0] // self.hop_length + 1) * self.hop_length - x.shape[0]
         if pad_sides == 1:
             return 0, pad
         else: