Merge remote-tracking branch 'origin/main' into publication

Author: Azure Pipelines

.actions/assistant.py

@@ -8,7 +8,7 @@
 from datetime import datetime
 from shutil import copyfile
 from textwrap import wrap
-from typing import Any, Dict, List, Optional, Sequence, Tuple
+from typing import Any, Dict, List, Optional, Sequence, Tuple, Union
 from warnings import warn
 
 import fire
@@ -283,11 +283,12 @@ def _valid_accelerator(folder: str) -> bool:
         return any(ac in meta_accels for ac in device_accels)
 
     @staticmethod
-    def _parse_requirements(folder: str) -> Tuple[str, str]:
+    def _parse_requirements(folder: str, formatted: bool = True) -> Union[Tuple[str, str], Tuple[list, list]]:
         """Parse standard requirements from meta file.
 
         Args:
             folder: path to the folder with python script, meta and artefacts
+            formatted: format it into two strings
 
         """
         meta = AssistantCLI._load_meta(folder)
@@ -298,15 +299,27 @@ def _parse_requirements(folder: str) -> Tuple[str, str]:
             for k, v in meta.items()
             if k.startswith(AssistantCLI._META_PIP_KEY)
         }
-        pip_args = ['--extra-index-url="https://download.pytorch.org/whl/"' + _RUNTIME_VERSIONS.get("DEVICE")]
+        pip_args = [f'--extra-index-url="https://download.pytorch.org/whl/{_RUNTIME_VERSIONS.get("DEVICE")}"']
         for pip_key in meta_pip_args:
             if not isinstance(meta_pip_args[pip_key], (list, tuple, set)):
                 meta_pip_args[pip_key] = [meta_pip_args[pip_key]]
             for arg in meta_pip_args[pip_key]:
                 arg = arg % _RUNTIME_VERSIONS
                 pip_args.append(f"--{pip_key} {arg}")
+        if formatted:
+            return " ".join([f'"{req}"' for req in requires]), " ".join(pip_args)
+        return list(requires), pip_args
 
-        return " ".join([f'"{req}"' for req in requires]), " ".join(pip_args)
+    @staticmethod
+    def pip_install(folder: str) -> str:
+        """Print all notebook requirements to be pre-installed in format of requirements file.
+
+        Args:
+            folder: path to the folder with python script, meta and artefacts
+
+        """
+        req, args = AssistantCLI._parse_requirements(folder, formatted=False)
+        return os.linesep.join(req) + os.linesep + os.linesep.join(args)
 
     @staticmethod
     def _bash_download_data(folder: str) -> List[str]:

.azure/ipynb-publish.yml

@@ -143,9 +143,11 @@ jobs:
       - bash: |
           set -e
           pip --version
-          # todo: export requirements for notebooks to file and execute
           # todo: adjust torch ecosystem versions
           pip install -r requirements.txt
+          # export requirements for notebooks to file and execute
+          python .actions/assistant.py pip-install --folder=$(notebook) > notebook.txt
+          pip install -r notebook.txt
         displayName: "Install dependencies"
         timeoutInMinutes: "15"
 
@@ -156,14 +158,14 @@ jobs:
           python -m papermill --version
         displayName: "Sanity check"
 
-      - bash: python .actions/assistant.py convert-ipynb $(notebook)
+      - bash: python .actions/assistant.py convert-ipynb --folder=$(notebook)
         displayName: "Generate notebook"
         timeoutInMinutes: "5"
 
       - bash: |
           set -e
           mkdir $(PATH_DATASETS)
-          python .actions/assistant.py bash-render $(notebook)
+          python .actions/assistant.py bash-render --folder=$(notebook)
           cat .actions/_ipynb-render.sh
           bash .actions/_ipynb-render.sh
           git status

.azure/ipynb-validate.yml

@@ -85,7 +85,8 @@ jobs:
           set -e
           pip --version
           pip install -r requirements.txt
-          pip list
+          python .actions/assistant.py pip-install --folder=$(notebook) > notebook.txt
+          pip install -r notebook.txt
         displayName: "Install dependencies"
 
       - bash: |
@@ -94,13 +95,13 @@ jobs:
           python -c "import torch ; mgpu = torch.cuda.device_count() ; assert mgpu > 0, f'GPU: {mgpu}'"
         displayName: "Sanity check"
 
-      - bash: python .actions/assistant.py convert-ipynb $(notebook)
+      - bash: python .actions/assistant.py convert-ipynb --folder=$(notebook)
         displayName: "Generate notebook"
 
       - bash: |
           set -e
           mkdir $(PATH_DATASETS)
-          python .actions/assistant.py bash-validate $(notebook)
+          python .actions/assistant.py bash-validate --folder=$(notebook)
           cat .actions/_ipynb-validate.sh
           bash .actions/_ipynb-validate.sh
         env:

_requirements/devel.txt

@@ -6,7 +6,7 @@ virtualenv>=20.10
 ipython[notebook]>=8.0.0, <8.24.0
 urllib3  # for ipython
 jupytext>=1.10, <1.15  # converting
-papermill>=2.3.4, <2.5.0  # rendering
+papermill ==2.6.*  # rendering
 
 matplotlib
 

_requirements/docs.txt

@@ -1,7 +1,7 @@
 sphinx >5.0, <7.0
-myst-parser >=0.18.1, <3.0.0
+myst-parser >=1.0.0, <4.0.0
 nbsphinx >=0.8.5, <0.10
-pandoc >=1.0, <=2.3
+pandoc >=1.0, <2.4
 #docutils >=0.16
 sphinx-paramlinks >=0.5.1, <=0.6.0
 ipython[notebook] >=8.0.0, <8.2.0

course_UvA-DL/requirements.txt

@@ -1,7 +1,7 @@
 numpy <2.0 # needed for older Torch
-torch >=1.8.1,<2.2
+torch >=1.8.1,<2.5
 pytorch-lightning >=2.0,<2.1
-torchmetrics >=1.0,<1.3
+torchmetrics >=1.0,<1.5
 torchvision
 matplotlib
 seaborn

lightning_examples/finetuning-scheduler/finetuning-scheduler.py

@@ -554,9 +554,7 @@ def train() -> None:
 # the implicit schedule will limit fine-tuning to just the last 4 parameters of the model, which is only a small fraction
 # of the parameters you'd want to tune for maximum performance. Since the implicit schedule is quite computationally
 # intensive and most useful for exploring model behavior, leaving [max_depth](https://finetuning-scheduler.readthedocs.io/en/stable/api/finetuning_scheduler.fts.html?highlight=max_depth#finetuning_scheduler.fts.FinetuningScheduler.params.max_depth) 1 allows us to demo implicit mode
-# behavior while keeping the computational cost and runtime of this notebook reasonable. To review how a full implicit
-# mode run compares to the ``nofts_baseline`` and ``fts_explicit`` scenarios, please see the the following
-# [tensorboard experiment summary](https://tensorboard.dev/experiment/n7U8XhrzRbmvVzC4SQSpWw/).
+# behavior while keeping the computational cost and runtime of this notebook reasonable.
 
 
 # %%
@@ -579,16 +577,15 @@ def train() -> None:
 # %% [markdown]
 # ### Reviewing the Training Results
 #
-# See the [tensorboard experiment summaries](https://tensorboard.dev/experiment/n7U8XhrzRbmvVzC4SQSpWw/) to get a sense
-# of the relative computational and performance tradeoffs associated with these [FinetuningScheduler](https://finetuning-scheduler.readthedocs.io/en/stable/api/finetuning_scheduler.fts.html#finetuning_scheduler.fts.FinetuningScheduler) configurations.
-# The summary compares a full ``fts_implicit`` execution to ``fts_explicit`` and ``nofts_baseline`` scenarios using DDP
+# It's worth considering the relative computational and performance tradeoffs associated with different [FinetuningScheduler](https://finetuning-scheduler.readthedocs.io/en/stable/api/finetuning_scheduler.fts.html#finetuning_scheduler.fts.FinetuningScheduler) configurations.
+# The example below compares ``fts_implicit`` execution to ``fts_explicit`` and ``nofts_baseline`` scenarios using DDP
 # training with 2 GPUs. The full logs/schedules for all three scenarios are available
 # [here](https://drive.google.com/file/d/1LrUcisRLHeJgh_BDOOD_GUBPp5iHAkoR/view?usp=sharing) and the checkpoints
 # produced in the scenarios [here](https://drive.google.com/file/d/1t7myBgcqcZ9ax_IT9QVk-vFH_l_o5UXB/view?usp=sharing)
 # (caution, ~3.5GB).
 #
-# [![fts_explicit_accuracy](fts_explicit_accuracy.png){height="315px" width="492px"}](https://tensorboard.dev/experiment/n7U8XhrzRbmvVzC4SQSpWw/#scalars&_smoothingWeight=0&runSelectionState=eyJmdHNfZXhwbGljaXQiOnRydWUsIm5vZnRzX2Jhc2VsaW5lIjpmYWxzZSwiZnRzX2ltcGxpY2l0IjpmYWxzZX0%3D)
-# [![nofts_baseline](nofts_baseline_accuracy.png){height="316px" width="505px"}](https://tensorboard.dev/experiment/n7U8XhrzRbmvVzC4SQSpWw/#scalars&_smoothingWeight=0&runSelectionState=eyJmdHNfZXhwbGljaXQiOmZhbHNlLCJub2Z0c19iYXNlbGluZSI6dHJ1ZSwiZnRzX2ltcGxpY2l0IjpmYWxzZX0%3D)
+# ![fts_explicit_accuracy](fts_explicit_accuracy.png){height="315px" width="492px"}
+# ![nofts_baseline](nofts_baseline_accuracy.png){height="316px" width="505px"}
 #
 # Note that given execution context differences, there could be a modest variation in performance from the tensorboard summaries generated by this notebook.
 #
@@ -597,7 +594,7 @@ def train() -> None:
 # greater fine-tuning flexibility for model exploration in research. For example, glancing at DeBERTa-v3's implicit training
 # run, a critical tuning transition point is immediately apparent:
 #
-# [![implicit_training_transition](implicit_training_transition.png){height="272px" width="494px"}](https://tensorboard.dev/experiment/n7U8XhrzRbmvVzC4SQSpWw/#scalars&_smoothingWeight=0&runSelectionState=eyJmdHNfZXhwbGljaXQiOmZhbHNlLCJub2Z0c19iYXNlbGluZSI6ZmFsc2UsImZ0c19pbXBsaWNpdCI6dHJ1ZX0%3D)
+# ![implicit_training_transition](implicit_training_transition.png){height="272px" width="494px"}
 #
 # Our `val_loss` begins a precipitous decline at step 3119 which corresponds to phase 17 in the schedule. Referring to our
 # schedule, in phase 17 we're beginning tuning the attention parameters of our 10th encoder layer (of 11). Interesting!

lightning_examples/requirements.txt

@@ -1,5 +1,5 @@
 numpy <2.0 # needed for older Torch
-torch>=1.8.1, <2.2
+torch>=1.8.1, <2.5
 pytorch-lightning>=2.0, <2.2
-torchmetrics>=1.0, <1.3
+torchmetrics>=1.0, <1.5
 matplotlib