Merge pull request #55 from espnet/notebooks/pin-released-espnet

Install the release the class needs, not whatever master is today

Author: sw005320

.github/workflows/run_notebooks.yml

@@ -1,10 +1,12 @@
 # These notebooks rot quietly: a Colab page still opens, the cells still look
 # right, and the failure only appears when somebody runs them in class. So run
-# them here on a schedule instead, against the released espnet.
+# them here on a schedule instead, against the release the notebooks pin.
 #
-# Only the inference notebooks run: the two fine-tuning ones need a GPU to
-# finish in reasonable time, and the TTS one needs a vocoder package that does
-# not build on every runner. Add more as they prove they fit.
+# One notebook runs. speaker_verification is the one that fits a free runner:
+# no GPU, no package that has to be built, and it finishes well inside the
+# timeout. The others are each blocked on something of their own - see this
+# directory's README - and adding one means finding that out first, not
+# putting it in the matrix and hoping.
 
 name: run notebooks
 
@@ -44,9 +46,10 @@ jobs:
           set -euo pipefail
           python -m pip install --upgrade pip
           # the runner skips the notebooks' install cells, so install what
-          # they would have installed - from the same place they install it
-          # from, or a break on master would pass here and fail in Colab
-          python -m pip install "espnet[enh] @ git+https://github.com/espnet/espnet"
+          # they would have installed, pinned to the same release they pin -
+          # testing a different espnet from the one Colab gets would make a
+          # pass here mean nothing
+          python -m pip install "espnet[enh]==202610.post1"
           python -m pip install nbclient nbformat ipykernel \
             librosa scikit-learn matplotlib tqdm soundfile editdistance
 

Courses/CMUSpeechTechnology26S/README.md

@@ -28,12 +28,28 @@ the packages are already present, and replaces `wget` and `tar` with their
 Python equivalents.
 
 ```sh
-pip install espnet nbclient nbformat ipykernel librosa scikit-learn matplotlib
+pip install "espnet==202610.post1" nbclient nbformat ipykernel librosa \
+    scikit-learn matplotlib
 python ../../tools/run_notebook.py speaker_verification.ipynb
 ```
 
-`.github/workflows/run_notebooks.yml` runs the inference notebooks weekly, so
-a notebook that stops working is noticed here rather than in class.
+The notebooks pin that release rather than installing espnet from git. A course
+notebook is worth having because it does the same thing in April that it does
+today, and installing from `master` gives neither that nor a fast install: it
+builds from source, and a break on `master` the night before class is a break
+in class. Each release, the pin moves and the notebooks are run again.
+
+`.github/workflows/run_notebooks.yml` runs `speaker_verification.ipynb` every
+Sunday, so a notebook that stops working is noticed here rather than in class.
+
+**One of the five, not all five.** It is the one that fits a free runner: no
+GPU, nothing that has to be built, and it finishes inside the timeout. The
+others are each blocked on something of their own — `text_to_speech` builds a
+vocoder from source, `neural_codec` and `speech_translation` pin an old numpy
+and TensorFlow against the rest of the environment, `speech_enhancement` runs
+VERSA over several models. Until one of those is sorted out, the other four are
+checked by running them by hand, which is a worse guarantee and should be said
+plainly rather than implied.
 
 ## What was changed
 

Courses/CMUSpeechTechnology26S/neural_codec.ipynb

@@ -44,7 +44,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "%pip install -q git+https://github.com/espnet/espnet\n",
+    "%pip install -q \"espnet==202610.post1\"\n",
     "!git clone https://github.com/wavlab-speech/versa\n",
     "!cd versa && pip install .\n",
     "!pip install espnet_model_zoo\n",

Courses/CMUSpeechTechnology26S/speaker_verification.ipynb

@@ -59,7 +59,7 @@
     "import locale\n",
     "locale.getpreferredencoding = lambda *a, **kw: \"UTF-8\"\n",
     "\n",
-    "%pip install -q git+https://github.com/espnet/espnet\n",
+    "%pip install -q \"espnet==202610.post1\"\n",
     "!pip install ffmpeg-python"
    ]
   },

Courses/CMUSpeechTechnology26S/speech_enhancement.ipynb

@@ -61,7 +61,7 @@
     "\n",
     "# [enh] brings fast_bss_eval, which the enhancement losses need at load time;\n",
     "# editdistance is used by the word error rate comparison further down\n",
-    "%pip install -q \"espnet[enh] @ git+https://github.com/espnet/espnet\"\n",
+    "%pip install -q \"espnet[enh]==202610.post1\"\n",
     "%pip install -q editdistance\n"
    ],
    "execution_count": null,

Courses/CMUSpeechTechnology26S/speech_translation.ipynb

@@ -35,7 +35,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "%pip install -q git+https://github.com/espnet/espnet\n",
+    "%pip install -q \"espnet==202610.post1\"\n",
     "!pip install -q espnet_model_zoo\n",
     "!pip install tensorflow jax jaxlib ml_dtypes transformers"
    ]

Courses/CMUSpeechTechnology26S/text_to_speech.ipynb

@@ -45,7 +45,7 @@
     "# Around 5-6 minutes\n",
     "\n",
     "# NOTE: pip shows imcompatible errors due to preinstalled libraries but you do not need to care\n",
-    "%pip install -q \"espnet[tts,spk] @ git+https://github.com/espnet/espnet\"\n",
+    "%pip install -q \"espnet[tts,spk]==202610.post1\"\n",
     "\n",
     "# Vocoder\n",
     "!git clone --depth 5 https://github.com/ftshijt/ParallelWaveGAN.git\n",