Debug forced align on cpu

Author: pearu

.github/scripts/unittest-windows/run_test.sh

@@ -13,8 +13,8 @@ env | grep TORCHAUDIO || true
 
 cd test
 if [ -z "${CUDA_VERSION:-}" ] ; then
-    pytest --continue-on-collection-errors --cov=torchaudio --junitxml=${RUNNER_TEST_RESULTS_DIR}/junit.xml -v --durations 20 torchaudio_unittest -k "not fairseq and not demucs and not librosa"
+    pytest -sv --continue-on-collection-errors --cov=torchaudio --junitxml=${RUNNER_TEST_RESULTS_DIR}/junit.xml -v --durations 20 torchaudio_unittest -k "not fairseq and not demucs and not librosa"
 else
-    pytest --continue-on-collection-errors --cov=torchaudio --junitxml=${RUNNER_TEST_RESULTS_DIR}/junit.xml -v --durations 20 torchaudio_unittest -k "not cpu and (cuda or gpu) and not fairseq and not demucs and not librosa"
+    pytest -sv --continue-on-collection-errors --cov=torchaudio --junitxml=${RUNNER_TEST_RESULTS_DIR}/junit.xml -v --durations 20 torchaudio_unittest -k "not cpu and (cuda or gpu) and not fairseq and not demucs and not librosa"
 fi
 coverage html

.github/workflows/unittest-linux-cpu.yml

@@ -35,6 +35,7 @@ jobs:
         export FFMPEG_VERSION="${{ matrix.ffmpeg-version }}"
         export PIP_PROGRESS_BAR=off
         export CONDA_QUIET=1
+        export NO_COLOR=1
 
         # Set UPLOAD_CHANNEL
         if [[(${GITHUB_EVENT_NAME} = 'pull_request' && (${GITHUB_BASE_REF} = 'release'*)) || (${GITHUB_REF} = 'refs/heads/release'*) ]]; then

.github/workflows/unittest-linux-gpu.yml

@@ -37,6 +37,7 @@ jobs:
         export CU_VERSION_WITHOUT_PERIODS="$(echo "${{ matrix.cuda-version }}" | sed 's/\.//g')"
         export PIP_PROGRESS_BAR=off
         export CONDA_QUIET=1
+        export NO_COLOR=1
         export TORCHAUDIO_TEST_ALLOW_SKIP_IF_CUDA_SMALL_MEMORY=true
         export TORCHAUDIO_TEST_ALLOW_SKIP_IF_ON_PYTHON_310=true
         export TORCHAUDIO_TEST_ALLOW_SKIP_IF_TEMPORARY_DISABLED=true

.github/workflows/unittest-macos-cpu.yml

@@ -36,6 +36,7 @@ jobs:
         export CONDA_QUIET=1
         export CU_VERSION=""
         export USE_OPENMP="0"
+        export NO_COLOR=1
 
         # Set CHANNEL
         if [[(${GITHUB_EVENT_NAME} = 'pull_request' && (${GITHUB_BASE_REF} = 'release'*)) || (${GITHUB_REF} = 'refs/heads/release'*) ]]; then

.github/workflows/unittest-windows-cpu.yml

@@ -25,6 +25,7 @@ jobs:
         export PIP_PROGRESS_BAR=off
         export CONDA_QUIET=1
         unset CUDA_VERSION
+        export NO_COLOR=1
 
         # Set CHANNEL
         if [[(${GITHUB_EVENT_NAME} = 'pull_request' && (${GITHUB_BASE_REF} = 'release'*)) || (${GITHUB_REF} = 'refs/heads/release'*) ]]; then

.github/workflows/unittest-windows-gpu.yml

@@ -27,6 +27,7 @@ jobs:
         export CONDA_QUIET=1
         export USE_CUDA=1
         export CUBLAS_WORKSPACE_CONFIG=:4096:8
+        export NO_COLOR=1
 
         #  Set CHANNEL
         if [[(${GITHUB_EVENT_NAME} = 'pull_request' && (${GITHUB_BASE_REF} = 'release'*)) || (${GITHUB_REF} = 'refs/heads/release'*) ]]; then

src/libtorchaudio/forced_align/cpu/compute.cpp

@@ -5,7 +5,7 @@
 #include <torch/csrc/stable/tensor.h>
 #include <torch/headeronly/core/Dispatch_v2.h>
 #include <torch/headeronly/core/ScalarType.h>
-
+#include <iostream>
 namespace torchaudio {
 namespace alignment {
 namespace cpu {
@@ -21,6 +21,8 @@ void forced_align_impl(
     const Tensor& targets,
     const int64_t blank,
     Tensor& paths) {
+  std::cout << "ENTERING forced_align_impl: " << target_scalar_type
+            << std::endl;
   const scalar_t kNegInfinity = -std::numeric_limits<scalar_t>::infinity();
   using target_t = typename std::
       conditional<target_scalar_type == ScalarType::Int, int, int64_t>::type;
@@ -153,22 +155,26 @@ std::tuple<Tensor, Tensor> compute(
     Tensor inputLengths,
     Tensor targetLengths,
     const int64_t blank) {
+  std::cout << "compute:1" << std::endl;
   STD_TORCH_CHECK(logProbs.is_cpu(), "log_probs must be a CPU tensor");
   STD_TORCH_CHECK(targets.is_cpu(), "targets must be a CPU tensor");
   STD_TORCH_CHECK(inputLengths.is_cpu(), "input_lengths must be a CPU tensor");
   STD_TORCH_CHECK(
       targetLengths.is_cpu(), "target_lengths must be a CPU tensor");
+  std::cout << "compute:2" << std::endl;
   STD_TORCH_CHECK(
       logProbs.scalar_type() == ScalarType::Double ||
           logProbs.scalar_type() == ScalarType::Float ||
           logProbs.scalar_type() == ScalarType::Half,
       "log_probs must be float64, float32 or float16 (half) type");
+  std::cout << "compute:3" << std::endl;
   STD_TORCH_CHECK(
       targets.scalar_type() == ScalarType::Int ||
           targets.scalar_type() == ScalarType::Long,
       "targets must be int32 or int64 type");
   STD_TORCH_CHECK(logProbs.is_contiguous(), "log_probs must be contiguous");
   STD_TORCH_CHECK(targets.is_contiguous(), "targets must be contiguous");
+  std::cout << "compute:4" << std::endl;
   STD_TORCH_CHECK(
       logProbs.dim() == 3,
       "log_probs must be 3-D (batch_size, input length, num classes)");
@@ -178,37 +184,52 @@ std::tuple<Tensor, Tensor> compute(
       inputLengths.dim() == 1, "input_lengths must be 1-D (batch_size,)");
   STD_TORCH_CHECK(
       targetLengths.dim() == 1, "target_lengths must be 1-D (batch_size,)");
+  std::cout << "compute:5" << std::endl;
   STD_TORCH_CHECK(
       logProbs.size(0) == 1,
-      "The batch dimension for log_probs must be 1 at the current version.")
+      "The batch dimension for log_probs must be 1 at the current version.");
   STD_TORCH_CHECK(
       targets.size(0) == 1,
-      "The batch dimension for targets must be 1 at the current version.")
+      "The batch dimension for targets must be 1 at the current version.");
+  std::cout << "compute:6" << std::endl;
   STD_TORCH_CHECK(
       blank >= 0 && blank < logProbs.size(-1),
       "blank must be within [0, num classes)");
+  std::cout << "compute:7" << std::endl;
   STD_TORCH_CHECK(
       logProbs.size(1) == torchaudio::util::max<int64_t>(inputLengths),
       "input length mismatch");
+  std::cout << "compute:8" << std::endl;
   STD_TORCH_CHECK(
       targets.size(1) == torchaudio::util::max<int64_t>(targetLengths),
       "target length mismatch");
+  std::cout << "compute:9" << std::endl;
   const auto B = logProbs.size(0);
   const auto T = logProbs.size(1);
+  std::cout << "compute:10" << std::endl;
   Tensor paths = torch::stable::empty({B, T}, targets.scalar_type());
+  std::cout << "compute:11" << std::endl;
   torch::stable::zero_(paths);
+  std::cout << "compute:12" << std::endl;
   THO_DISPATCH_V2(
       logProbs.scalar_type(),
       "forced_align_impl",
       AT_WRAP([&] {
+        std::cout << "compute:13" << std::endl;
         if (targets.scalar_type() == ScalarType::Long) {
+          std::cout << "compute:14" << std::endl;
           forced_align_long_impl<scalar_t>(logProbs, targets, blank, paths);
-        } else {
+        } else if (targets.scalar_type() == ScalarType::Int) {
+          std::cout << "compute:15" << std::endl;
           forced_align_int_impl<scalar_t>(logProbs, targets, blank, paths);
+        } else {
+          STD_TORCH_CHECK(
+              false, "unexpected targets dtype:", targets.scalar_type());
         }
       }),
       AT_EXPAND(AT_FLOATING_TYPES),
       ScalarType::Half);
+  std::cout << "compute:16" << std::endl;
   return std::make_tuple(paths, logProbs);
 }
 

src/torchaudio/functional/_alignment.py

@@ -69,7 +69,17 @@ def forced_align(
     assert input_lengths is not None
     assert target_lengths is not None
 
+    def _show(t):
+        return f"{t.shape=} {t.dtype=} {t.device=}"
+
+    print(f"{_show(log_probs)=}")
+    print(f"{_show(targets)=}")
+    print(f"{_show(input_lengths)=}")
+    print(f"{_show(target_lengths)=}")
+    print(f"{blank=}", flush=True)
     paths, scores = torch.ops.torchaudio.forced_align(log_probs, targets, input_lengths, target_lengths, blank)
+    print(f"{_show(paths)=}")
+    print(f"{_show(scores)=}")
     return paths, scores[:, torch.arange(scores.shape[1]), paths[0]]
 
 

tools/setup_helpers/extension.py

@@ -50,45 +50,6 @@ def get_ext_fullpath(self, ext_name: str) -> str:
         print(f"XXXXXXXXXXXX get_ext_fullpath({ext_name}) -> {result}")
         return result
 
-    def _get_ext_filename(self, fullname: str) -> str:
-        import inspect
-
-        import setuptools.command.build_ext as m
-
-        orig = inspect.getsource(super().get_ext_filename)
-        print("AAAAAAAAAAAAAAAAAAAAA")
-        print(f"{orig}")
-        print("AAAAAAAAAAAAAAAAAAAAA")
-        print(f"OOOOOOOOOOOOOOOOOO {super().get_ext_filename(fullname)=}")
-
-        _build_ext = m._build_ext
-        so_ext = os.getenv("SETUPTOOLS_EXT_SUFFIX")
-        if so_ext:
-            filename = os.path.join(*fullname.split(".")) + so_ext
-        else:
-            filename = _build_ext.get_ext_filename(self, fullname)
-            ext_suffix = m.get_config_var("EXT_SUFFIX")
-            if not isinstance(ext_suffix, str):
-                raise OSError(
-                    "Configuration variable EXT_SUFFIX not found for this platform "
-                    "and environment variable SETUPTOOLS_EXT_SUFFIX is missing"
-                )
-            so_ext = ext_suffix
-
-        if fullname in self.ext_map:
-            ext = self.ext_map[fullname]
-            abi3_suffix = m.get_abi3_suffix()
-            if ext.py_limited_api and abi3_suffix:  # Use abi3
-                filename = filename[: -len(so_ext)] + abi3_suffix
-            if isinstance(ext, m.Library):
-                fn, ext = os.path.splitext(filename)
-                filename = self.shlib_compiler.library_filename(fn, m.libtype)
-            elif m.use_stubs and ext._links_to_dynamic:
-                d, fn = os.path.split(filename)
-                filename = os.path.join(d, "dl-" + fn)
-        print(f"YYYYYYYYYYYYYYYYYYYYYYY {filename=}")
-        return filename
-
     def get_ext_filename(self, ext_name):
         # Origin: torch/utils/cpp_extension.py
         # Fixes a bug: don't try to remove ABI part that does not