[Jasper/PyT, QuartzNet/PyT] Fix Ada L40 on 23.06 base container

Author: alancucki

PyTorch/SpeechRecognition/Jasper/Dockerfile

@@ -24,7 +24,7 @@ COPY requirements.txt .
 RUN if [[ ! -z "$(command -v conda)" ]]; then conda install -y pyyaml==5.4.1; fi
 RUN pip install --disable-pip-version-check -U -r requirements.txt
 
-RUN pip install --force-reinstall --extra-index-url https://developer.download.nvidia.com/compute/redist nvidia-dali-cuda110==1.9.0
+RUN pip install --force-reinstall --extra-index-url https://developer.download.nvidia.com/compute/redist nvidia-dali-cuda110==1.27.0
 
 # Copy rest of files
 COPY . .

PyTorch/SpeechRecognition/Jasper/common/audio.py

@@ -45,7 +45,7 @@ def __init__(self, filename, target_sr=None, int_values=False, offset=0,
                  duration=0, trim=False, trim_db=60):
         """Create audio segment from samples.
 
-        Samples are convert float32 internally, with int scaled to [-1, 1].
+        Samples are converted to float32 internally, with int scaled to [-1, 1].
         Load a file supported by librosa and return as an AudioSegment.
         :param filename: path of file to load
         :param target_sr: the desired sample rate
@@ -67,10 +67,11 @@ def __init__(self, filename, target_sr=None, int_values=False, offset=0,
 
         samples = self._convert_samples_to_float32(samples)
         if target_sr is not None and target_sr != sample_rate:
-            samples = librosa.core.resample(samples, sample_rate, target_sr)
+            samples = librosa.resample(samples, orig_sr=sample_rate,
+                                       target_sr=target_sr)
             sample_rate = target_sr
         if trim:
-            samples, _ = librosa.effects.trim(samples, trim_db)
+            samples, _ = librosa.effects.trim(samples, top_db=trim_db)
         self._samples = samples
         self._sample_rate = sample_rate
         if self._samples.ndim >= 2:

PyTorch/SpeechRecognition/Jasper/common/features.py

@@ -233,7 +233,7 @@ def __init__(self, spec_augment=None, cutout_augment=None,
         window_tensor = window_fn(self.win_length,
                                   periodic=False) if window_fn else None
         filterbanks = torch.tensor(
-            librosa.filters.mel(sample_rate, self.n_fft, n_mels=n_filt,
+            librosa.filters.mel(sr=sample_rate, n_fft=self.n_fft, n_mels=n_filt,
                                 fmin=lowfreq, fmax=highfreq),
             dtype=torch.float).unsqueeze(0)
         # torchscript

PyTorch/SpeechRecognition/Jasper/requirements.txt

@@ -1,6 +1,6 @@
 inflect==5.3.0
 ipdb
-librosa==0.8.0
+librosa==0.9.0
 pandas==1.5.2
 pyyaml>=5.4
 soundfile

PyTorch/SpeechRecognition/Jasper/train.py

@@ -54,7 +54,7 @@ def parse_args():
     training.add_argument('--amp', '--fp16', action='store_true', default=False,
                           help='Use pytorch native mixed precision training')
     training.add_argument('--seed', default=42, type=int, help='Random seed')
-    training.add_argument('--local_rank', default=os.getenv('LOCAL_RANK', 0),
+    training.add_argument('--local_rank', '--local-rank', default=os.getenv('LOCAL_RANK', 0),
                           type=int, help='GPU id used for distributed training')
     training.add_argument('--pre_allocate_range', default=None, type=int, nargs=2,
                           help='Warmup with batches of length [min, max] before training')

PyTorch/SpeechRecognition/Jasper/utils/preprocessing_utils.py

@@ -15,7 +15,6 @@
 #!/usr/bin/env python
 import os
 import multiprocessing
-import librosa
 import functools
 
 import sox

PyTorch/SpeechRecognition/QuartzNet/Dockerfile

@@ -24,7 +24,7 @@ COPY requirements.txt .
 RUN if [[ ! -z "$(command -v conda)" ]]; then conda install -y pyyaml==5.4.1; fi
 RUN pip install --disable-pip-version-check -U -r requirements.txt
 
-RUN pip install --extra-index-url https://developer.download.nvidia.com/compute/redist nvidia-dali-cuda110==1.9.0
+RUN pip install --extra-index-url https://developer.download.nvidia.com/compute/redist nvidia-dali-cuda110==1.27.0
 
 # Copy rest of files
 COPY . .

PyTorch/SpeechRecognition/QuartzNet/train.py

@@ -56,7 +56,7 @@ def parse_args():
     training.add_argument('--amp', '--fp16', action='store_true', default=False,
                           help='Use pytorch native mixed precision training')
     training.add_argument('--seed', default=None, type=int, help='Random seed')
-    training.add_argument('--local_rank', default=os.getenv('LOCAL_RANK', 0), type=int,
+    training.add_argument('--local_rank', '--local-rank', default=os.getenv('LOCAL_RANK', 0), type=int,
                           help='GPU id used for distributed training')
     training.add_argument('--pre_allocate_range', default=None, type=int, nargs=2,
                           help='Warmup with batches of length [min, max] before training')