Regression example of Clay (#285)

- Add datamodule & model class for biomasters regression example
- Add notebooks to show the inference for biomasters & chesapeake bay

.pre-commit-config.yaml

@@ -15,7 +15,10 @@ repos:
   hooks:
     - id: ruff  # Run the linter
       args: [ --fix ]
-      types_or: [ python, pyi, jupyter ]
+      types_or: [ python, pyi ]
+    - id: ruff  # Run the linter for Jupyter notebooks with the PLR0913 rule ignored
+      args: [ --fix, --ignore=PLR0913 ]
+      types: [ jupyter ]
     - id: ruff-format  # Run the formatter
       types_or: [ python, pyi, jupyter ]
 

classify.py

@@ -19,7 +19,7 @@
 # %%
 def cli_main():
     """
-    Command-line inteface to run ClayMAE with ClayDataModule.
+    Command-line inteface to run Clasifier model with EuroSATDataModule.
     """
     cli = LightningCLI(EuroSATClassifier, EuroSATDataModule)
     return cli

configs/regression_biomasters.yaml

@@ -0,0 +1,63 @@
+# lightning.pytorch==2.1.2
+seed_everything: 42
+data:
+  metadata_path: configs/metadata.yaml
+  batch_size: 10
+  num_workers: 8
+  train_chip_dir: data/biomasters/train_cube
+  train_label_dir: data/biomasters/train_agbm
+  val_chip_dir: data/biomasters/test_cube
+  val_label_dir: data/biomasters/test_agbm
+model:
+  ckpt_path: checkpoints/clay-v1-base.ckpt
+  lr: 1e-3
+  wd: 0.05
+  b1: 0.9
+  b2: 0.95
+  feature_maps:
+    - 2
+    - 5
+    - 7
+    - 9
+    - 11
+trainer:
+  accelerator: auto
+  strategy: ddp
+  devices: auto
+  num_nodes: 1
+  precision: bf16-mixed
+  log_every_n_steps: 5
+  max_epochs: 100
+  default_root_dir: checkpoints/regression
+  fast_dev_run: False
+  num_sanity_val_steps: 0
+  # limit_train_batches: 0.25
+  # limit_val_batches: 0.25
+  accumulate_grad_batches: 4
+  logger:
+    - class_path: lightning.pytorch.loggers.WandbLogger
+      init_args:
+        entity: developmentseed
+        project: clay-regression
+        log_model: false
+  callbacks:
+    - class_path: lightning.pytorch.callbacks.ModelCheckpoint
+      init_args:
+        dirpath: checkpoints/regression
+        auto_insert_metric_name: False
+        filename: biomasters_epoch-{epoch:02d}_val-score-{val/score:.3f}
+        monitor: val/score
+        mode: min
+        save_last: False
+        save_top_k: 2
+        save_weights_only: True
+        verbose: True
+    - class_path: lightning.pytorch.callbacks.LearningRateMonitor
+      init_args:
+        logging_interval: step
+    - class_path: src.callbacks.LayerwiseFinetuning
+      init_args:
+        phase: 10
+        train_bn: True
+  plugins:
+    - class_path: lightning.pytorch.plugins.io.AsyncCheckpointIO

finetune/regression/README.md

@@ -0,0 +1,177 @@
+## Download data
+The data comes as multifile zip, it can be downloaded from the
+[BioMassters](https://huggingface.co/datasets/nascetti-a/BioMassters/)
+huggingface repository. Grab a coffee, this is about 250GB in size.
+
+The next step is to unzip training data. The data comes in a multi-file
+zip archive. So it needs to be unzipped using a library that can handle
+the format. 7z works quite well in this case. Grabb another coffee, this
+will take a while.
+
+```bash
+sudo apt install p7zip-full
+```
+
+### Extract train feature
+
+
+```bash
+7z e -o/home/tam/Desktop/biomasters/train_features/ /datadisk/biomasters/raw/train_features.zip
+```
+
+Should look something like this
+
+```
+7-Zip [64] 16.02 : Copyright (c) 1999-2016 Igor Pavlov : 2016-05-21
+p7zip Version 16.02 (locale=en_US.UTF-8,Utf16=on,HugeFiles=on,64 bits,16 CPUs Intel(R) Core(TM) i7-10875H CPU @ 2.30GHz (A0652),ASM,AES-NI)
+
+Scanning the drive for archives:
+1 file, 10247884383 bytes (9774 MiB)
+
+Extracting archive: /datadisk/biomasters/raw/train_features.zip
+--
+Path = /datadisk/biomasters/raw/train_features.zip
+Type = zip
+Physical Size = 10247884383
+Embedded Stub Size = 4
+64-bit = +
+Total Physical Size = 149834321503
+Multivolume = +
+Volume Index = 13
+Volumes = 14
+
+Everything is Ok
+
+Folders: 1
+Files: 189078
+Size:       231859243932
+Compressed: 149834321503
+```
+
+### Extract train AGBM
+
+```bash
+7z e -o/home/tam/Desktop/biomasters/train_agbm/ /datadisk/biomasters/raw/train_agbm.zip
+```
+
+Should look something like this
+
+```
+7-Zip [64] 16.02 : Copyright (c) 1999-2016 Igor Pavlov : 2016-05-21
+p7zip Version 16.02 (locale=en_US.UTF-8,Utf16=on,HugeFiles=on,64 bits,16 CPUs Intel(R) Core(TM) i7-10875H CPU @ 2.30GHz (A0652),ASM,AES-NI)
+
+Scanning the drive for archives:
+1 file, 575973495 bytes (550 MiB)
+
+Extracting archive: /datadisk/biomasters/raw/train_agbm.zip
+--
+Path = /datadisk/biomasters/raw/train_agbm.zip
+Type = zip
+Physical Size = 575973495
+
+Everything is Ok
+
+Folders: 1
+Files: 8689
+Size:       2280706098
+Compressed: 575973495
+```
+
+### Extract test features
+
+```bash
+7z e -o/home/tam/Desktop/biomasters/test_features/ /datadisk/biomasters/raw/test_features_splits.zip
+```
+
+Should look something like this
+
+```
+7-Zip [64] 16.02 : Copyright (c) 1999-2016 Igor Pavlov : 2016-05-21
+p7zip Version 16.02 (locale=en_US.UTF-8,Utf16=on,HugeFiles=on,64 bits,16 CPUs Intel(R) Core(TM) i7-10875H CPU @ 2.30GHz (A0652),ASM,AES-NI)
+
+Scanning the drive for archives:
+1 file, 6912625480 bytes (6593 MiB)
+
+Extracting archive: /datadisk/biomasters/raw/test_features_splits.zip
+--
+Path = /datadisk/biomasters/raw/test_features_splits.zip
+Type = zip
+Physical Size = 6912625480
+Embedded Stub Size = 4
+64-bit = +
+Total Physical Size = 49862298440
+Multivolume = +
+Volume Index = 4
+Volumes = 5
+
+Everything is Ok
+
+Folders: 1
+Files: 63348
+Size:       78334396224
+Compressed: 49862298440
+```
+
+### Extract test AGBM
+
+```bash
+7z e -o/home/tam/Desktop/biomasters/test_agbm/ /datadisk/biomasters/raw/test_agbm.tar
+```
+
+Should look something like this
+
+```
+7-Zip [64] 16.02 : Copyright (c) 1999-2016 Igor Pavlov : 2016-05-21
+p7zip Version 16.02 (locale=en_US.UTF-8,Utf16=on,HugeFiles=on,64 bits,16 CPUs Intel(R) Core(TM) i7-10875H CPU @ 2.30GHz (A0652),ASM,AES-NI)
+
+Scanning the drive for archives:
+1 file, 729766400 bytes (696 MiB)
+
+Extracting archive: /datadisk/biomasters/raw/test_agbm.tar
+--
+Path = /datadisk/biomasters/raw/test_agbm.tar
+Type = tar
+Physical Size = 729766400
+Headers Size = 1421312
+Code Page = UTF-8
+
+Everything is Ok
+
+Folders: 1
+Files: 2773
+Size:       727862586
+Compressed: 729766400
+```
+
+## Prepare data
+
+This will take the average of all timesteps available for each tile.
+The time steps for Sentinel-2 are not complete, not all months are
+provided for all tiles. In addtion, the Clay model does not take time
+series as input. So aggregating the time element is simplifying but
+ok for the purpose of this example.
+
+**In addition, we skip the one orbit because it nodata most of the time**
+
+
+### Prepare training features
+
+```bash
+python finetune/regression/preprocess_data.py \
+  --features=/home/tam/Desktop/biomasters/train_features/ \
+  --cubes=/home/tam/Desktop/biomasters/train_cubes/ \
+  --processes=12 \
+  --sample=1 \
+  --overwrite
+```
+
+### Prepare test features
+
+```bash
+python finetune/regression/preprocess_data.py \
+  --features=/home/tam/Desktop/biomasters/test_features/ \
+  --cubes=/home/tam/Desktop/biomasters/test_cubes/ \
+  --processes=12 \
+  --sample=1 \
+  --overwrite
+```