Merge branch 'main' of github.com:pytorch/torchcodec into encoding_num_channels

Author: NicolasHug

src/torchcodec/_core/Encoder.cpp

@@ -126,14 +126,18 @@ AudioEncoder::AudioEncoder(
   TORCH_CHECK(
       avFormatContext != nullptr,
       "Couldn't allocate AVFormatContext. ",
-      "Check the desired extension? ",
+      "The destination file is ",
+      fileName,
+      ", check the desired extension? ",
       getFFMPEGErrorStringFromErrorCode(status));
   avFormatContext_.reset(avFormatContext);
 
   status = avio_open(&avFormatContext_->pb, fileName.data(), AVIO_FLAG_WRITE);
   TORCH_CHECK(
       status >= 0,
-      "avio_open failed: ",
+      "avio_open failed. The destination file is ",
+      fileName,
+      ", make sure it's a valid path? ",
       getFFMPEGErrorStringFromErrorCode(status));
 
   initializeEncoder(sampleRate, bitRate, numChannels);
@@ -155,7 +159,9 @@ AudioEncoder::AudioEncoder(
   TORCH_CHECK(
       avFormatContext != nullptr,
       "Couldn't allocate AVFormatContext. ",
-      "Check the desired extension? ",
+      "Check the desired format? Got format=",
+      formatName,
+      ". ",
       getFFMPEGErrorStringFromErrorCode(status));
   avFormatContext_.reset(avFormatContext);
 

src/torchcodec/_core/ops.py

@@ -161,6 +161,7 @@ def create_from_file_abstract(filename: str, seek_mode: Optional[str]) -> torch.
     return torch.empty([], dtype=torch.long)
 
 
+# TODO-ENCODING: rename wf to samples
 @register_fake("torchcodec_ns::encode_audio_to_file")
 def encode_audio_to_file_abstract(
     wf: torch.Tensor,

src/torchcodec/encoders/__init__.py

@@ -0,0 +1 @@
+from ._audio_encoder import AudioEncoder  # noqa

src/torchcodec/encoders/_audio_encoder.py

@@ -0,0 +1,53 @@
+from pathlib import Path
+from typing import Optional, Union
+
+import torch
+from torch import Tensor
+
+from torchcodec import _core
+
+
+class AudioEncoder:
+    def __init__(self, samples: Tensor, *, sample_rate: int):
+        # Some of these checks are also done in C++: it's OK, they're cheap, and
+        # doing them here allows to surface them when the AudioEncoder is
+        # instantiated, rather than later when the encoding methods are called.
+        if not isinstance(samples, Tensor):
+            raise ValueError(
+                f"Expected samples to be a Tensor, got {type(samples) = }."
+            )
+        if samples.ndim != 2:
+            raise ValueError(f"Expected 2D samples, got {samples.shape = }.")
+        if samples.dtype != torch.float32:
+            raise ValueError(f"Expected float32 samples, got {samples.dtype = }.")
+        if sample_rate <= 0:
+            raise ValueError(f"{sample_rate = } must be > 0.")
+
+        self._samples = samples
+        self._sample_rate = sample_rate
+
+    def to_file(
+        self,
+        dest: Union[str, Path],
+        *,
+        bit_rate: Optional[int] = None,
+    ) -> None:
+        _core.encode_audio_to_file(
+            wf=self._samples,
+            sample_rate=self._sample_rate,
+            filename=dest,
+            bit_rate=bit_rate,
+        )
+
+    def to_tensor(
+        self,
+        format: str,
+        *,
+        bit_rate: Optional[int] = None,
+    ) -> Tensor:
+        return _core.encode_audio_to_tensor(
+            wf=self._samples,
+            sample_rate=self._sample_rate,
+            format=format,
+            bit_rate=bit_rate,
+        )

test/test_encoders.py

@@ -0,0 +1,41 @@
+import re
+
+import pytest
+import torch
+
+from torchcodec.encoders import AudioEncoder
+
+
+class TestAudioEncoder:
+
+    def test_bad_input(self):
+        with pytest.raises(ValueError, match="Expected samples to be a Tensor"):
+            AudioEncoder(samples=123, sample_rate=32_000)
+        with pytest.raises(ValueError, match="Expected 2D samples"):
+            AudioEncoder(samples=torch.rand(10), sample_rate=32_000)
+        with pytest.raises(ValueError, match="Expected float32 samples"):
+            AudioEncoder(
+                samples=torch.rand(10, 10, dtype=torch.float64), sample_rate=32_000
+            )
+        with pytest.raises(ValueError, match="sample_rate = 0 must be > 0"):
+            AudioEncoder(samples=torch.rand(10, 10), sample_rate=0)
+
+        encoder = AudioEncoder(samples=torch.rand(2, 100), sample_rate=32_000)
+
+        bad_path = "/bad/path.mp3"
+        with pytest.raises(
+            RuntimeError,
+            match=f"avio_open failed. The destination file is {bad_path}, make sure it's a valid path",
+        ):
+            encoder.to_file(dest=bad_path)
+
+        bad_extension = "output.bad_extension"
+        with pytest.raises(RuntimeError, match="check the desired extension"):
+            encoder.to_file(dest=bad_extension)
+
+        bad_format = "bad_format"
+        with pytest.raises(
+            RuntimeError,
+            match=re.escape(f"Check the desired format? Got format={bad_format}"),
+        ):
+            encoder.to_tensor(format=bad_format)

test/test_ops.py

@@ -1134,7 +1134,7 @@ def test_bad_input(self, tmp_path):
             encode_audio_to_file(
                 wf=torch.rand(2, 10), sample_rate=10, filename="./bad/path.mp3"
             )
-        with pytest.raises(RuntimeError, match="Check the desired extension"):
+        with pytest.raises(RuntimeError, match="check the desired extension"):
             encode_audio_to_file(
                 wf=torch.rand(2, 10), sample_rate=10, filename="./file.bad_extension"
             )