Import torchaudio #1755 e11d27c

Summary: Import torchaudio #1755 e11d27c

Reviewed By: carolineechen

Differential Revision: D30844075

fbshipit-source-id: 1295cc142dda23cb4b029b332b4ef78bb0a67432

Author: hwangjeff

.circleci/config.yml

@@ -60,10 +60,6 @@ commands:
           name: Install CUDA
           command: |
               packaging/windows/internal/cuda_install.bat
-              # cudatoolkit >= 11 isn't available for windows in the nvidia channel
-              if [[ "${CU_VERSION}" =~ cu11.* ]]; then
-                export CONDA_CHANNEL_FLAGS="-c conda-forge"
-              fi
 
 binary_common: &binary_common
   parameters:
@@ -262,11 +258,16 @@ jobs:
       - windows_install_cuda
       - run:
           name: Build conda packages
+          no_output_timeout: 20m
           command: |
             set -ex
             eval "$('/C/tools/miniconda3/Scripts/conda.exe' 'shell.bash' 'hook')"
             conda activate base
             conda install -yq conda-build "conda-package-handling!=1.5.0"
+            # cudatoolkit >= 11 isn't available for windows in the nvidia channel
+            if [[ "${CU_VERSION}" =~ cu11.* ]]; then
+              export CONDA_CHANNEL_FLAGS="-c conda-forge"
+            fi
             bash packaging/build_conda.sh
       - store_artifacts:
           path: C:/tools/miniconda3/conda-bld/win-64
@@ -292,6 +293,10 @@ jobs:
 
   # Requires org-member context
   binary_wheel_upload:
+    parameters:
+      subfolder:
+        description: "What whl subfolder to upload to, e.g., blank or cu100/ (trailing slash is important)"
+        type: string
     docker:
       - image: circleci/python:3.8
     steps:
@@ -309,7 +314,7 @@ jobs:
             export AWS_SECRET_ACCESS_KEY="${PYTORCH_BINARY_AWS_SECRET_ACCESS_KEY}"
             set -x
             for pkg in ~/workspace/*.whl; do
-              aws s3 cp "$pkg" "s3://pytorch/whl/${UPLOAD_CHANNEL}/" --acl public-read
+              aws s3 cp "$pkg" "s3://pytorch/whl/${UPLOAD_CHANNEL}/<< parameters.subfolder >>" --acl public-read
             done
 
   smoke_test_linux_conda:

.circleci/config.yml.in

@@ -21,7 +21,7 @@
 
 
 PYTHON_VERSIONS = ["3.6", "3.7", "3.8", "3.9"]
-CU_VERSIONS_DICT = {"linux": ["cpu", "cu102", "cu111","rocm4.1"],
+CU_VERSIONS_DICT = {"linux": ["cpu", "cu102", "cu111","cu113", "rocm4.1"],
                     "windows": ["cpu", "cu102", "cu111"],
                     "macos": ["cpu"]}
 
@@ -75,7 +75,7 @@ def build_workflow_pair(btype, os_type, python_version, cu_version, filter_branc
 
     if upload:
 
-        w.append(generate_upload_workflow(base_workflow_name, filter_branch, btype))
+        w.append(generate_upload_workflow(base_workflow_name, filter_branch, os_type, btype, cu_version))
 
         if filter_branch == 'nightly' and os_type != 'macos':
             pydistro = 'pip' if btype == 'wheel' else 'conda'
@@ -157,13 +157,17 @@ def gen_filter_branch_tree(*branches):
     }
 
 
-def generate_upload_workflow(base_workflow_name, filter_branch, btype):
+def generate_upload_workflow(base_workflow_name, filter_branch, os_type, btype, cu_version):
     d = {
         "name": "{base_workflow_name}_upload".format(base_workflow_name=base_workflow_name),
         "context": "org-member",
         "requires": [base_workflow_name],
     }
 
+    if btype == 'wheel':
+        d["subfolder"] = "" if os_type == 'macos' else cu_version + "/"
+
+
     if filter_branch:
         d["filters"] = gen_filter_branch_tree(filter_branch)
 

.circleci/regenerate.py

@@ -68,5 +68,5 @@ fi
 # Install fairseq
 git clone https://github.com/pytorch/fairseq
 cd fairseq
-git checkout e6eddd80
+git checkout e47a4c8
 pip install .

.circleci/unittest/linux/scripts/install.sh

@@ -49,5 +49,5 @@ fi
 # Install fairseq
 git clone https://github.com/pytorch/fairseq
 cd fairseq
-git checkout e6eddd80
+git checkout e47a4c8
 pip install .

.circleci/unittest/windows/scripts/install.sh

@@ -267,6 +267,14 @@ rnnt_loss
 
 .. autofunction:: rnnt_loss
 
+:hidden:`Metric`
+~~~~~~~~~~~~~~~~
+
+edit_distance
+-------------
+
+.. autofunction:: edit_distance
+
 References
 ~~~~~~~~~~
 

docs/source/functional.rst

@@ -1,5 +1,7 @@
 This is an example pipeline for text-to-speech using Tacotron2.
 
+Here is a [colab example](https://colab.research.google.com/drive/1MPcn1_G5lKozxZ7v8b9yucOD5X5cLK4j?usp=sharing)
+that shows how the text-to-speech pipeline is used during inference with the built-in pretrained models.
 
 ## Install required packages
 

examples/pipeline_tacotron2/README.md

@@ -131,7 +131,7 @@ def text_mel_collate_fn(batch: Tuple[Tensor, Tensor],
     Args:
         batch (tuple of two tensors): the first tensor is the mel spectrogram with shape
             (n_batch, n_mels, n_frames), the second tensor is the text with shape (n_batch, ).
-        n_frames_per_step (int): The number of frames to advance every step.
+        n_frames_per_step (int, optional): The number of frames to advance every step.
 
     Returns:
         text_padded (Tensor): The input text to Tacotron2 with shape (n_batch, max of ``text_lengths``).

examples/pipeline_tacotron2/datasets.py

@@ -123,12 +123,12 @@ def text_to_sequence(sent: str,
         symbol_list (str or List of string, optional): When the input is a string, available options include
             "english_characters" and "english_phonemes". When the input is a list of string, ``symbol_list`` will
             directly be used as the symbol to encode. (Default: "english_characters")
-        phonemizer (str, optional): The phonemizer to use. Only used when ``symbol_list`` is "english_phonemes".
+        phonemizer (str or None, optional): The phonemizer to use. Only used when ``symbol_list`` is "english_phonemes".
             Available options include "DeepPhonemizer". (Default: "DeepPhonemizer")
-        checkpoint (str, optional): The path to the checkpoint of the phonemizer. Only used when ``symbol_list`` is
-            "english_phonemes". (Default: "./en_us_cmudict_forward.pt")
-        cmudict_root (str, optional): The path to the directory where the CMUDict dataset is found or downloaded.
-            Only used when ``symbol_list`` is "english_phonemes". (Default: "./")
+        checkpoint (str or None, optional): The path to the checkpoint of the phonemizer. Only used when
+            ``symbol_list`` is "english_phonemes". (Default: "./en_us_cmudict_forward.pt")
+        cmudict_root (str or None, optional): The path to the directory where the CMUDict dataset is found or
+            downloaded. Only used when ``symbol_list`` is "english_phonemes". (Default: "./")
 
     Returns:
         List of integers corresponding to the symbols in the sentence.

examples/pipeline_tacotron2/text/text_preprocessing.py

@@ -272,6 +272,7 @@ def main(args):
             n_mels=args.n_freq,
             f_min=args.f_min,
             mel_scale='slaney',
+            norm='slaney',
             **melkwargs,
         ),
         NormalizeDB(min_level_db=args.min_level_db, normalization=args.normalization),