### Set up Virtual Environment and Install Dependencies

Source: https://github.com/terrastackai/terratorch/blob/main/examples/custom_modules/README.md

These commands set up a Python virtual environment and install the necessary packages for the project. Activate the environment before running the example.

```bash
python -m venv .venv
source .venv/bin/activate
pip install -e .
```

```bash
source .venv/bin/activate
```

--------------------------------

### Install Dependencies

Source: https://github.com/terrastackai/terratorch/blob/main/examples/utils/prithvi_v2_eo_300_tl_unet_burnscars_tensorrt_metrics.ipynb

Install the required packages for the deprecated example.

```bash
!pip install terratorch==1.0.1 gdown tensorrt onnx onnxruntime polygraphy numpy pycuda numba
```

--------------------------------

### Install Necessary Packages

Source: https://github.com/terrastackai/terratorch/blob/main/docs/tutorials/using_datamodule_multitemporalclassificationModule.md

Installs terratorch, gdown, and tensorboard. Output is redirected to install.log.

```python
!pip install terratorch gdown tensorboard >& install.log
```

--------------------------------

### Setup and Inspect DataModule

Source: https://github.com/terrastackai/terratorch/blob/main/examples/object_detection/ft_aed_elephants_terramind_tiny.ipynb

Prepares the DataModule and inspects the training dataloader.

```python
dm.setup()
```

```python
tdl = dm.train_dataloader()
```

```python
len(tdl.dataset)
```

```python
tdl.dataset[0].keys()
```

--------------------------------

### Setup Pipeline Components

Source: https://github.com/terrastackai/terratorch/blob/main/examples/models/models-abstraction.ipynb

Initializes the necessary components for a full training pipeline.

```python
# Level 4: Working example with fake data
from torch.utils.data import Dataset, DataLoader
from lightning import Trainer
import torch
```

--------------------------------

### Setup Training Data and Visualize Samples

Source: https://github.com/terrastackai/terratorch/blob/main/examples/weather_climate/WxCTutorialDownscalingECCC.ipynb

Prepares the datamodule for the fitting stage and visualizes a sample batch.

```python
datamodule.setup(stage='fit')
train_dl = datamodule.train_dataloader()
plot_sample(next(iter(train_dl)))
```

--------------------------------

### Install Dependencies

Source: https://github.com/terrastackai/terratorch/blob/main/examples/weather_climate/WxCTutorialDownscalingECCC.ipynb

Installs necessary libraries for TerraTorch, Prithvi-WxC, and granite-wxc.

```python
!pip install -q git+https://github.com/IBM/terratorch.git
!pip install -q git+https://github.com/NASA-IMPACT/Prithvi-WxC.git
!pip install -q git+https://github.com/IBM/granite-wxc.git
```

--------------------------------

### Install Dependencies

Source: https://github.com/terrastackai/terratorch/blob/main/examples/weather_climate/WxCTutorialGravityWave.ipynb

Install the required TerraTorch, Hugging Face, and Prithvi WxC packages.

```bash
!pip install terratorch==0.99.9 huggingface_hub PrithviWxC
```

```bash
!pip install  git+https://github.com/romeokienzler/gravity-wave-finetuning.git
```

--------------------------------

### Setup Test Dataset

Source: https://github.com/terrastackai/terratorch/blob/main/examples/utils/prithvi_v2_eo_300_tl_unet_burnscars_tensorrt_metrics.ipynb

Prepare the test dataset and check its length.

```python
datamodule.setup("test")
test_dataset = datamodule.test_dataset
len(test_dataset)
```

--------------------------------

### Install Dependencies

Source: https://github.com/terrastackai/terratorch/blob/main/examples/utils/visualize_model.ipynb

Install the necessary libraries, torchview and graphviz, using pip.

```python
!pip install torchview
!pip install graphviz
```

--------------------------------

### Setup and Load Batch from DataModule

Source: https://github.com/terrastackai/terratorch/blob/main/examples/object_detection/od_data_augmentation.ipynb

Sets up the data module for training and then retrieves the first batch of data from the training dataloader.

```python
dm.setup("fit")

train_loader = dm.train_dataloader()
batch = next(iter(train_loader))
```

--------------------------------

### Launch TensorBoard for Training Analysis

Source: https://github.com/terrastackai/terratorch/blob/main/examples/weather_climate/WxCTutorialDownscalingECCC.ipynb

Command to start the TensorBoard server for monitoring training logs.

```python
# !tensorboard --logdir logs/ --port 9010  # add --host $(hostname -f) for clusters
```

--------------------------------

### Run the TerraTorch Example

Source: https://github.com/terrastackai/terratorch/blob/main/examples/custom_modules/README.md

Execute the TerraTorch training command from the `examples/custom_modules` directory using the specified configuration file.

```bash
terratorch fit -c ./config.yaml
```

--------------------------------

### Initialization and Camera Setup

Source: https://github.com/terrastackai/terratorch/blob/main/examples/webapps/prithvi.tiny.onnx.demo/model_tm_tiny_xview.html

Initializes the web worker, retrieves available camera devices, and logs the initialization status.

```javascript
log("Initializing..."); initWorker(); getCameras();
```

--------------------------------

### Setup DataModule

Source: https://github.com/terrastackai/terratorch/blob/main/examples/weather_climate/InspectMerra2Data.ipynb

Configure and setup the Merra2DownscaleNonGeoDataModule for prediction.

```python
from terratorch.datamodules.merra2_downscale import Merra2DownscaleNonGeoDataModule
dm = Merra2DownscaleNonGeoDataModule(
    time_range=('2001-01-01', '2001-01-02'),
    data_path_surface = config.data.data_path_surface,
    data_path_vertical = config.data.data_path_vertical,
    climatology_path_surface = config.data.climatology_path_surface,
    climatology_path_vertical = config.data.climatology_path_vertical,
    input_surface_vars = config.data.input_surface_vars,
    input_static_surface_vars = config.data.input_static_surface_vars,
    input_vertical_vars = config.data.input_vertical_vars,
    input_levels = config.data.input_levels,
    n_input_timestamps = config.data.n_input_timestamps,
    output_vars=config.data.output_vars,
    transforms=_get_transforms(config),
)
dm.setup('predict')
```

--------------------------------

### Download sample images via wget

Source: https://github.com/terrastackai/terratorch/blob/main/examples/webapps/terramind.tiny.onnx.demo/README.md

Use these commands to fetch the required example tiles for the TerraMind.tiny model demo.

```bash
wget https://terramind-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/tile_0_0_0.png

wget https://terramind-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/tile_0_1152_3072.png

wget https://terramind-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/tile_0_768_3072.png

wget https://terramind-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/tile_1260_0_2688.png

wget https://terramind-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/tile_100_3840_1536.png

wget https://terramind-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/tile_100_4224_1152.png

wget https://terramind-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/tile_100_4224_1536.png

wget https://terramind-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/tile_100_4224_768.png

wget https://terramind-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/tile_100_4608_0.png

wget https://terramind-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/tile_1261_4224_768.png

wget https://terramind-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/tile_1013_3072_3072.png

wget https://terramind-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/tile_101_3072_384.png
```

--------------------------------

### Commented Out Execution Examples

Source: https://github.com/terrastackai/terratorch/blob/main/examples/object_detection/ft_aed_elephants_terramind_tiny.ipynb

Examples of how to call visualization or iteration logic, currently commented out.

```python
#m.setup("fit")
#show_from_datamodule(dm, n=30, split="train")
```

```python
""" train_loader = dm.train_dataloader()

# Iterate and count
total_elements = 0
for batch in train_loader:
    # 'batch' is a tensor or list of tensors depending on your dataset
    batch_size = len(batch)
    print(f"Processing a batch of size: {batch_size}")
    total_elements += batch_size

print(f"All elements read. Total count: {total_elements}") """
```

--------------------------------

### Setup Data Module and Predict

Source: https://github.com/terrastackai/terratorch/blob/main/examples/weather_climate/WxCTutorialDownscaling.ipynb

Initializes the data module for prediction and runs the trainer to generate results.

```python
dm = Merra2DownscaleNonGeoDataModule(
    time_range=('2020-01-01T00:00:00', '2020-01-01T23:59:59'),
    data_path_surface = config.data.data_path_surface,
    data_path_vertical = config.data.data_path_vertical,
    climatology_path_surface = config.data.climatology_path_surface,
    climatology_path_vertical = config.data.climatology_path_vertical,
    input_surface_vars = config.data.input_surface_vars,
    input_static_surface_vars = config.data.input_static_surface_vars,
    input_vertical_vars = config.data.input_vertical_vars,
    input_levels = config.data.input_levels,
    n_input_timestamps = config.data.n_input_timestamps,
    output_vars=config.data.output_vars,
    transforms=_get_transforms(config),
)
dm.setup('predict')
first_element = dm.predict_dataloader().dataset[0]
```

```python
dm = Merra2DownscaleNonGeoDataModule(
    time_range=('2020-01-01T00:00:00', '2020-01-01T23:59:59'),
    data_path_surface = config.data.data_path_surface,
    data_path_vertical = config.data.data_path_vertical,
    climatology_path_surface = config.data.climatology_path_surface,
    climatology_path_vertical = config.data.climatology_path_vertical,
    input_surface_vars = config.data.input_surface_vars,
    input_static_surface_vars = config.data.input_static_surface_vars,
    input_vertical_vars = config.data.input_vertical_vars,
    input_levels = config.data.input_levels,
    n_input_timestamps = config.data.n_input_timestamps,
    output_vars=config.data.output_vars,
    transforms=_get_transforms(config),
)
dm.setup('predict')

trainer = Trainer(
    max_epochs=1,
    limit_predict_batches=1,
)
results = trainer.predict(model=task, datamodule=dm)
```

--------------------------------

### Start vLLM Serving Instance

Source: https://github.com/terrastackai/terratorch/blob/main/docs/guide/vllm/serving_a_model_tensor.md

Use this command to start a vLLM serving instance with a specified HuggingFace model. Ensure the model is prepared for vLLM serving. Options like `--trust-remote-code`, `--dtype float16`, `--skip-tokenizer-init`, `--enforce-eager`, and `--enable-mm-embeds` are used for model loading and configuration.

```bash
vllm serve \
--model='ibm-nasa-geospatial/Prithvi-EO-2.0-300M-TL-Sen1Floods11' \
--trust-remote-code \
--dtype float16 \
--skip-tokenizer-init \
--enforce-eager \
--enable-mm-embeds
```

--------------------------------

### Install PrithviWxC and Dependencies

Source: https://github.com/terrastackai/terratorch/blob/main/examples/weather_climate/WxCTutorialDownscaling.ipynb

Install the PrithviWxC library and its dependencies, including granitewxc, huggingface_hub, and terratorch, using pip.

```python
!pip install PrithviWxC granitewxc==0.1.1rc1 huggingface_hub terratorch==0.99.9
```

--------------------------------

### Local development and build commands

Source: https://github.com/terrastackai/terratorch/blob/main/examples/webapps/terramind.tiny.onnx.demo/README.md

Commands for installing dependencies, running the development server, and building the project.

```bash
pnpm i
npm install vite-plugin-pwa --save-dev
pnpm dev
pnpm build
python3 -m http.server
```

--------------------------------

### Execute Training

Source: https://github.com/terrastackai/terratorch/blob/main/examples/classification/classification_eurosat.ipynb

Start the training process using the configured trainer and datamodule.

```python
# Start training
trainer.fit(model, datamodule=datamodule)
```

--------------------------------

### Multimodal Data Organization Examples

Source: https://github.com/terrastackai/terratorch/blob/main/docs/guide/vllm/plugins/tm_segmentation_io_plugin.md

Examples showing how to structure the data field for URL and path-based inputs, including the required directory structure.

```json
{
  "data_format": "url",
  "data": {
    "DEM": "https://example.com/path/to/dem_file",
    "S1RTC": "https://example.com/path/to/S1RTC_file",
    "S2L2A": "https://example.com/path/to/S2L2A_file"
  }
}
```

```json
{
  "data_format": "path",
  "data": "/path/to/input/directory/"
}
```

```text
/path/to/input/directory/
├── DEM/
│   └── FILE_NAME_DEM.tiff
├── S1RTC/
│   └── FILE_NAME_S1RTC.zarr.zip
└── S2L2A/
    └── FILE_NAME_S2L2A.zarr.zip
```

--------------------------------

### Install TerraTorch from GitHub

Source: https://github.com/terrastackai/terratorch/blob/main/docs/guide/quick_start.md

Install the most recent version from the main branch of the GitHub repository.

```shell
pip install git+https://github.com/IBM/terratorch.git
```

--------------------------------

### Start vLLM Serving Instance

Source: https://github.com/terrastackai/terratorch/blob/main/docs/guide/vllm/serving_a_model_image.md

Use this command to start a vLLM serving instance for a TerraTorch model. Ensure you have prepared your model for vLLM and identified the correct IOProcessor plugin.

```bash
vllm serve ibm-nasa-geospatial/Prithvi-EO-2.0-300M-TL-Sen1Floods11 \
--io-processor-plugin terratorch_segmentation \
--model-impl terratorch \
--skip-tokenizer-init \
--enforce-eager \
--enable-mm-embeds
```

--------------------------------

### Setup Test Dataset

Source: https://github.com/terrastackai/terratorch/blob/main/examples/pixelwise_regression/pixelwise_regression_agb.ipynb

Configure the datamodule for testing and check the test dataset size.

```python
# checking datasets testing split size
datamodule.setup("test")
test_dataset = datamodule.test_dataset

print(f"Available samples in the test dataset: {len(test_dataset)}")
```

--------------------------------

### Download Example Input Image

Source: https://github.com/terrastackai/terratorch/blob/main/examples/embeddings/embedding_generation_manual_backbone_registry.ipynb

Downloads a sample Sentinel-2 TIFF file from the Hugging Face Hub if it does not exist locally.

```python
# Download example input image
image_path = Path('examples/S2L2A/Santiago.tif')

if not image_path.exists():
    hf_hub_download(repo_id='ibm-esa-geospatial/Examples', filename='S2L2A/Santiago.tif', repo_type='dataset', local_dir='examples/')
```

--------------------------------

### Install TerraTorch for Development

Source: https://github.com/terrastackai/terratorch/blob/main/README.md

Commands to clone the repository and install it in editable mode with optional dependencies.

```bash
git clone https://github.com/terrastackai/terratorch.git
cd terratorch
pip install -e .[test]
```

```bash
pip install -e .[wxc]
```

--------------------------------

### Setup Lightning Trainer

Source: https://github.com/terrastackai/terratorch/blob/main/examples/segmentation/segmentation_sen1floods11.ipynb

Configures the PyTorch Lightning Trainer with logging, checkpointing, and progress tracking callbacks.

```python
import os
from lightning.pytorch import Trainer
from lightning.pytorch.callbacks import LearningRateMonitor, ModelCheckpoint, RichProgressBar
from lightning.pytorch.loggers import TensorBoardLogger

experiment = "tutorial"

os.makedirs("tutorial_experiments", exist_ok=True)
default_root_dir = os.path.join("tutorial_experiments", experiment)

logger = TensorBoardLogger(save_dir=default_root_dir, name=experiment)

checkpoint_callback = ModelCheckpoint(
    save_top_k=1,
    save_last=True,
)

trainer = Trainer(
    accelerator="auto",  # or specify cpu or gpu
    max_epochs=1,  # for demo purposes
    default_root_dir=default_root_dir,
    logger=logger,
    callbacks=[
        RichProgressBar(),
        checkpoint_callback,
        LearningRateMonitor(logging_interval="epoch"),
    ],
    log_every_n_steps=1,
)
```

--------------------------------

### Initialize Trainer and Start Training

Source: https://github.com/terrastackai/terratorch/blob/main/examples/datasets_and_benchmarks/forestnet_dataset_timm_model.ipynb

Configure the PyTorch Lightning Trainer with callbacks and loggers, then execute the training loop.

```python
from lightning.pytorch import Trainer
from lightning.pytorch.callbacks import ModelCheckpoint, RichProgressBar
from lightning.pytorch.loggers import TensorBoardLogger

checkpoint = ModelCheckpoint(monitor=task.monitor, save_top_k=1, save_last=True)
logger = TensorBoardLogger(save_dir="output", name="tutorial")

trainer = Trainer(
    devices=1,
    precision="16-mixed",
    max_epochs=1,  # demo only
    callbacks=[RichProgressBar(), checkpoint],
    logger=logger,
    default_root_dir="output/tutorial",
    log_every_n_steps=1,
)

_ = trainer.fit(task, datamodule)
```

--------------------------------

### Specify example input image file

Source: https://github.com/terrastackai/terratorch/blob/main/docs/tutorials/burn_scars_inference_simplified.md

Sets the path to a single image file for testing direct inference.

```python
example_file = "data/examples/subsetted_512x512_HLS.S30.T10SEH.2018190.v1.4_merged.tif"
```

--------------------------------

### Fine-tuning via CLI

Source: https://github.com/terrastackai/terratorch/blob/main/examples/multitemporal_data/temporalwrapper_segementation_crop.ipynb

Execute fine-tuning using a configuration file for a no-code setup. This command requires a YAML configuration file specifying the training components.

```bash
!terratorch fit -c temporalwrapper_segementation_crop.yaml
```

--------------------------------

### Start vLLM server with TerraTorch plugin

Source: https://github.com/terrastackai/terratorch/blob/main/docs/guide/vllm/vllm_io_plugins.md

Use the --io-processor-plugin flag to enable the terratorch_segmentation plugin during vLLM server startup.

```bash
vllm serve \
    --model=ibm-nasa-geospatial/Prithvi-EO-2.0-300M-TL-Sen1Floods11 \
    --model-impl terratorch \
    --task embed --trust-remote-code \
    --skip-tokenizer-init --enforce-eager \
    --io-processor-plugin terratorch_segmentation
```

--------------------------------

### Debug and Fix Broken Model URLs

Source: https://github.com/terrastackai/terratorch/blob/main/examples/models/models-abstraction.ipynb

This example guides you on how to find and fix broken model URLs by locating the model's configuration file and checking for `hf_hub_id` or `url` settings. It uses Python's `inspect` module to find the file location.

```python
# How to find and fix broken model URLs
# 
# Step 1: Find where the model is defined
#   - Search the codebase for the model name
#   - For Clay: terratorch/models/backbones/clay_v1/embedder.py
#
# Step 2: Look for `default_cfgs` or `hf_hub_id`
#   - This is where HuggingFace URLs are configured

import inspect
from terratorch.models.backbones.clay_v1 import embedder

# Find the file location
print("📁 FILE TO EDIT:")
print(f"   {inspect.getfile(embedder)}")
print()

# Show the current HuggingFace configuration
print("🔗 CURRENT HUGGINGFACE CONFIG:")
if hasattr(embedder, 'default_cfgs'):
    for model_name, cfg in embedder.default_cfgs.items():
        hf_url = getattr(cfg.default, 'hf_hub_id', 'Not set')
        print(f"   {model_name}: {hf_url}")
else:
    print("   (Check the file manually for hf_hub_id or url settings)")
```

--------------------------------

### Install TerraTorch as a developer

Source: https://github.com/terrastackai/terratorch/blob/main/docs/guide/quick_start.md

Clone the repository and install in editable mode for development.

```shell
git clone https://github.com/<your_username>/terratorch.git
pip install -e .
```

--------------------------------

### Install TerraTorch via Conda

Source: https://github.com/terrastackai/terratorch/blob/main/README.md

Installation command for users of the conda-forge channel.

```bash
conda install -c conda-forge terratorch
```

--------------------------------

### Initialize Environment and Device

Source: https://github.com/terrastackai/terratorch/blob/main/examples/embeddings/embedding_generation_manual_backbone_registry.ipynb

Sets up the necessary imports and detects the available hardware device (CUDA, MPS, or CPU).

```python
from pathlib import Path

import torch
import rioxarray as rxr
from huggingface_hub import hf_hub_download

import numpy as np

# Select device
if torch.cuda.is_available():
    device = 'cuda'
elif torch.backends.mps.is_available():
    device = 'mps'
else:
    device = 'cpu'
```

--------------------------------

### Install GDAL via Conda

Source: https://github.com/terrastackai/terratorch/blob/main/README.md

Recommended method for installing the required GDAL dependency.

```bash
conda install -c conda-forge gdal
```

--------------------------------

### Install TerraTorch via Pipx

Source: https://github.com/terrastackai/terratorch/blob/main/README.md

Installs the library in an isolated environment for CLI usage.

```bash
pipx install terratorch
```

--------------------------------

### Load and Inspect Backbones

Source: https://github.com/terrastackai/terratorch/blob/main/examples/models/models-abstraction.ipynb

Demonstrates how to instantiate a backbone from the registry and process an image to extract features.

```python
# Level 1: Load a backbone
from terratorch.registry import BACKBONE_REGISTRY
import torch

# Build a backbone
backbone = BACKBONE_REGISTRY.build("prithvi_eo_v2_100_tl", pretrained=True)
print(f"Type: {type(backbone).__name__}")
```

```python
# See how the backbone processes an image
fake_image = torch.randn(1, 6, 224, 224)  # [batch, channels, height, width]

backbone.eval()
with torch.no_grad():
    features = backbone(fake_image)

# Print output shapes
print("Features shape:")
for i, f in enumerate(features):
    print(f"  tensor {i}: {f.shape}")
```

--------------------------------

### Download Sample Data Assets

Source: https://github.com/terrastackai/terratorch/blob/main/examples/webapps/prithvi.tiny.onnx.demo/README.md

Use wget to retrieve individual sample images and binary files from the cloud storage bucket.

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_51_y.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_51_pred_mask.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_80_rgb.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_80_x.bin
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_80_y.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_80_pred_mask.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_108_rgb.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_108_x.bin
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_108_y.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_108_pred_mask.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_111_rgb.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_111_x.bin
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_111_y.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_111_pred_mask.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_112_rgb.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_112_x.bin
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_112_y.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_112_pred_mask.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_113_rgb.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_113_x.bin
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_113_y.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_113_pred_mask.png
```

--------------------------------

### Install TerraTorch Dependencies

Source: https://github.com/terrastackai/terratorch/blob/main/examples/datasets_and_benchmarks/forestnet_dataset_timm_model.ipynb

Install the required packages for running TerraTorch in a Google Colab environment.

```python
!pip install terratorch==1.2.1
!pip install gdown tensorboard
!pip install -U jupyter ipywidgets
```

--------------------------------

### Install TerraTorch and Dependencies

Source: https://github.com/terrastackai/terratorch/blob/main/examples/classification/classification_eurosat.ipynb

Install the required packages for running TerraTorch in a Google Colab environment.

```bash
!pip install terratorch==1.2.1
!pip install tensorboard
!pip install -U jupyter ipywidgets
```

--------------------------------

### Install TerraTorch via Pip

Source: https://github.com/terrastackai/terratorch/blob/main/README.md

Commands for installing stable releases or the latest development branch.

```bash
pip install terratorch==<version>
```

```bash
pip install git+https://github.com/terrastackai/terratorch.git
```

--------------------------------

### Local Development and Deployment Commands

Source: https://github.com/terrastackai/terratorch/blob/main/examples/webapps/prithvi.tiny.onnx.demo/README.md

Commands for setting up the development environment, building the project, and serving the application locally.

```bash
python3 -m http.server
```

```bash
pnpm i
```

```bash
npm install vite-plugin-pwa --save-dev
```

```bash
pnpm dev
```

```bash
pnpm build
```

--------------------------------

### Install TerraTorch using pip

Source: https://github.com/terrastackai/terratorch/blob/main/docs/guide/quick_start.md

Use this command to install the latest stable release of TerraTorch.

```shell
pip install terratorch
```

--------------------------------

### Request Payload Examples

Source: https://github.com/terrastackai/terratorch/blob/main/docs/guide/vllm/plugins/tm_segmentation_io_plugin.md

Examples of request payloads demonstrating different data formats and output configurations.

```json
{
  "data_format": "url",
  "out_data_format": "b64_json",
  "data": {
    "DEM": "https://example.com/path/to/dem_file",
    "S1RTC": "https://example.com/path/to/S1RTC_file",
    "S2L2A": "https://example.com/path/to/S2L2A_file"
  }
}
```

```json
{
  "data_format": "path",
  "out_data_format": "path",
  "data": "/path/to/input/directory"
}
```

```json
{
  "data_format": "url",
  "out_data_format": "path",
  "out_path": "/custom/output/directory",
  "data": {
    "DEM": "https://example.com/path/to/dem_file",
    "S1RTC": "https://example.com/path/to/S1RTC_file",
    "S2L2A": "https://example.com/path/to/S2L2A_file"
  }
}
```

--------------------------------

### Download Sample Data for Prithvi.tiny Demo

Source: https://github.com/terrastackai/terratorch/blob/main/examples/webapps/prithvi.tiny.onnx.demo/README.md

Use these wget commands to download the necessary sample images and data files for the Prithvi.tiny web demo. Ensure you copy them to the current directory and the 'dist' directory as needed.

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_3_rgb.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_3_x.bin
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_3_y.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_3_pred_mask.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_16_rgb.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_16_x.bin
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_16_y.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_16_pred_mask.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_18_rgb.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_18_x.bin
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_18_y.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_18_pred_mask.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_21_rgb.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_21_x.bin
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_21_y.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_21_pred_mask.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_23_rgb.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_23_x.bin
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_23_y.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_23_pred_mask.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_37_rgb.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_37_x.bin
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_37_y.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_37_pred_mask.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_39_rgb.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_39_x.bin
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_39_y.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_39_pred_mask.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_40_rgb.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_40_x.bin
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_40_y.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_40_pred_mask.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_42_rgb.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_42_x.bin
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_42_y.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_42_pred_mask.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_49_rgb.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_49_x.bin
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_49_y.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_49_pred_mask.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_51_rgb.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_51_x.bin
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_51_y.png
```

```bash
wget https://prithvi-demo-onnx-mobile.s3-web.eu-de.cloud-object-storage.appdomain.cloud/sample_51_pred_mask.png
```

--------------------------------

### Generate vLLM Config with Input Spec as File

Source: https://github.com/terrastackai/terratorch/blob/main/docs/guide/vllm/prepare_your_model.md

This example demonstrates generating a vLLM configuration by providing the input specification via a JSON file. First, create the input JSON file, then pass its path to the script.

```bash
cat << EOF > ./input_sample.json
{
  "target": "pixel_values",
  "data": {
    "pixel_values": {
      "type": "torch.Tensor",
      "shape": [6,512,512]
    },
    "location_coords": {
      "type": "torch.Tensor",
      "shape": [1,2]
    }
  }
}
EOF

python vllm_config_generator.py \
   --ttconfig config.yaml \
   -i ./input_sample.json
```

--------------------------------

### Using MultiSourceRegistries for Decoders

Source: https://github.com/terrastackai/terratorch/blob/main/docs/guide/registry.md

Illustrates how to use `MultiSourceRegistries` like `DECODER_REGISTRY` to instantiate decoders. It shows how to specify the registry prefix for clarity or rely on the default search order. The `keys()` method can be used to inspect available prefixes.

```python
from terratorch import DECODER_REGISTRY

# decoder registries always take at least one extra argument, the channel list with the channel dimension of each embedding passed to it
DECODER_REGISTRY.build("FCNDecoder", [32, 64, 128])

DECODER_REGISTRY.build("terratorch_FCNDecoder", [32, 64, 128])

# Find all prefixes
DECODER_REGISTRY.keys()
>>> odict_keys(['terratorch', 'smp', 'mmseg'])
```

--------------------------------

### Install MMSegmentation dependencies

Source: https://github.com/terrastackai/terratorch/blob/main/docs/guide/encoder_decoder_factory.md

Commands to install the required MMSegmentation environment, noting the specific version constraints for mmcv and torch.

```sh
pip install -U openmim
mim install mmengine
mim install mmcv==2.1.0
pip install regex ftfy mmsegmentation
```

--------------------------------

### Install TerraTorch and Dependencies

Source: https://github.com/terrastackai/terratorch/blob/main/examples/embeddings/embedding_generation_burnscars.ipynb

Installs the TerraTorch library and other necessary packages like gdown and tensorboard. Recommended for Google Colab environments.

```python
!pip install terratorch==1.2.4
!pip install gdown tensorboard
!pip install -U jupyter ipywidgets
```

--------------------------------

### Configure Training Pipeline

Source: https://github.com/terrastackai/terratorch/blob/main/examples/pixelwise_regression/pixelwise_regression_agb.ipynb

Set up the Lightning Trainer and ModelCheckpoint for fine-tuning the Prithvi model.

```python
from terratorch.tasks import PixelwiseRegressionTask
import lightning.pytorch as pl

pl.seed_everything(0)

checkpoint_callback = pl.callbacks.ModelCheckpoint(
    dirpath="output/agb/checkpoints/",
    monitor="val/RMSE",          # metric to monitor
    mode="min",
    filename="best-{epoch:02d}",
)

# Lightning Trainer
trainer = pl.Trainer(
    accelerator="auto",
    strategy="auto",
    devices=1,                  # single device for notebook stability
    precision="16-mixed",        # speed up training
    num_nodes=1,
    logger=True,                # uses TensorBoard by default
    max_epochs=1,               # demo only
    log_every_n_steps=1,
    enable_checkpointing=True,
    callbacks=[checkpoint_callback, pl.callbacks.RichProgressBar()],
    default_root_dir="output/agb",
    detect_anomaly=True,
)
```

--------------------------------

### Perform Inference and Get Predictions

Source: https://github.com/terrastackai/terratorch/blob/main/examples/pixelwise_regression/pixelwise_regression_agb.ipynb

Loads data, applies augmentation, moves images to the model's device, and performs inference to get model predictions. Ensure the augmentation matches training.

```python
import torch 

test_loader = datamodule.test_dataloader()

with torch.no_grad():
    batch = next(iter(test_loader))
    # Apply the same normalization used during training
    batch = datamodule.aug(batch)

    images = batch["image"].to(model.device)

    outputs = model(images)
    preds = outputs.output

```

--------------------------------

### Initialize Environment and Imports

Source: https://github.com/terrastackai/terratorch/blob/main/examples/weather_climate/WxCTutorialGravityWave.ipynb

Import necessary libraries and initialize the distributed process group for training.

```python
import terratorch # this import is needed to initialize TT's factories
from lightning.pytorch import Trainer
import os
import torch
from huggingface_hub import hf_hub_download, snapshot_download
from terratorch.models.wxc_model_factory import WxCModelFactory
from terratorch.tasks.wxc_task import WxCTask
import torch.distributed as dist
```

```python
os.environ['MASTER_ADDR'] = 'localhost'
os.environ['MASTER_PORT'] = '12355' 

if dist.is_initialized():
    dist.destroy_process_group()

dist.init_process_group(
    backend='gloo',
    init_method='env://', 
    rank=0,
    world_size=1
)
```

--------------------------------

### Model Pipeline Output

Source: https://github.com/terrastackai/terratorch/blob/main/examples/multitemporal_data/temporalwrapper_segementation_crop.ipynb

Example output from the pipeline inspection script.

```text
Output:
Model Pipeline — Diff Subsets Pooling
--------------------------------------
Input shape:          (1, 6, 3, 224, 224)  [B, C, T, H, W] → Batch size 1, 6 channels, 3 timesteps
Backbone:             terramind_v1_small
Temporal pooling:     diff
Encoder feature maps: 12
Example feature map:  (1, 196, 384)

The 12 intermediate layers of the TerraMind ViT-Small backbone each produce a feature map of 196 tokens with 384 dimensions.
Diff subsets pooling computes the difference between the mean of two temporal subsets and returns a single feature representation with the same shape as a single timestep.This keeps the decoder size and parameter count identical to mean pooling.

Model Parameters
----------------
Total parameters:     31.32M
Trainable (decoder):  9.44M
```

--------------------------------

### Fine-tune via CLI

Source: https://github.com/terrastackai/terratorch/blob/main/examples/datasets_and_benchmarks/burnscars_dataset_prithvi.ipynb

Executes training using configuration files via the command line interface.

```bash
!terratorch fit -c burnscars_dataset_prithvi.yaml
```

```bash
!terratorch fit -c burnscars_dataset_smp_unet.yaml
```

--------------------------------

### Using Registries to Find and Instantiate Models

Source: https://github.com/terrastackai/terratorch/blob/main/docs/guide/registry.md

Demonstrates how to search for available models within a registry, check for model existence, and instantiate models with optional arguments like pretrained weights or custom checkpoint paths. The registry prefix is optional during instantiation.

```python
from terratorch import BACKBONE_REGISTRY

# find available prithvi models
print([model_name for model_name in BACKBONE_REGISTRY if "terratorch_prithvi" in model_name])
>>> ['terratorch_prithvi_eo_tiny', 'terratorch_prithvi_eo_v1_100', 'terratorch_prithvi_eo_v2_300', 'terratorch_prithvi_eo_v2_600', 'terratorch_prithvi_eo_v2_300_tl', 'terratorch_prithvi_eo_v2_600_tl']

# show all models with list(BACKBONE_REGISTRY)

# check a model is in the registry
"terratorch_prithvi_eo_v2_300" in BACKBONE_REGISTRY
>>> True

# without the prefix, all internal registries will be searched until the first match is found
"prithvi_eo_v1_100" in BACKBONE_REGISTRY
>>> True

# instantiate your desired model
# the backbone registry prefix (e.g. `terratorch` or `timm`) is optional
# in this case, the underlying registry is terratorch.
model = BACKBONE_REGISTRY.build("prithvi_eo_v1_100", pretrained=True)

# instantiate your model with more options, for instance, passing weights from your own file
model = BACKBONE_REGISTRY.build(
    "prithvi_eo_v2_300", num_frames=1, ckpt_path='path/to/model.pt'
)
# Rest of your PyTorch / PyTorchLightning code
```

--------------------------------

### Verify Dataset Size

Source: https://github.com/terrastackai/terratorch/blob/main/examples/segmentation/segmentation_sen1floods11.ipynb

Checks the number of samples available in the test dataset after setup.

```python
# Check the size of the test dataset
datamodule.setup("test")
test_dataset = datamodule.test_dataset

print(f"Available samples in the test dataset: {len(test_dataset)}")
```

--------------------------------

### Initialize Configuration

Source: https://github.com/terrastackai/terratorch/blob/main/examples/weather_climate/InspectMerra2Data.ipynb

Load the configuration file required for the model factory.

```python
import xarray as xr
from granitewxc.utils.config import get_config
from granitewxc.utils.data import _get_transforms

config = get_config("../../integrationtests/test_prithvi_wxc_model_factory_config.yaml")
```

--------------------------------

### Get Training Dataset Size

Source: https://github.com/terrastackai/terratorch/blob/main/examples/multitemporal_data/temporalwrapper_segementation_crop.ipynb

Retrieve the training dataset object from the datamodule and display its size.

```python
train_dataset = datamodule.train_dataset
len(train_dataset)
```

--------------------------------

### Configure Model and Task

Source: https://github.com/terrastackai/terratorch/blob/main/examples/weather_climate/WxCTutorialGravityWave.ipynb

Define model arguments and initialize the WxCTask for evaluation or training.

```python
from prithviwxc.gravitywave.datamodule import ERA5DataModule

model_args = {
    "in_channels": 1280,
    "input_size_time": 1,
    "n_lats_px": 64,
    "n_lons_px": 128,
    "patch_size_px": [2, 2],
    "mask_unit_size_px": [8, 16],
    "mask_ratio_inputs": 0.5,
    "embed_dim": 2560,
    "n_blocks_encoder": 12,
    "n_blocks_decoder": 2,
    "mlp_multiplier": 4,
    "n_heads": 16,
    "dropout": 0.0,
    "drop_path": 0.05,
    "parameter_dropout": 0.0,
    "residual": "none",
    "masking_mode": "both",
    "decoder_shifting": False,
    "positional_encoding": "absolute",
    "checkpoint_encoder": [3, 6, 9, 12, 15, 18, 21, 24],
    "checkpoint_decoder": [1, 3],
    "in_channels_static": 3,
    "input_scalers_mu": torch.tensor([0] * 1280),
    "input_scalers_sigma": torch.tensor([1] * 1280),
    "input_scalers_epsilon": 0,
    "static_input_scalers_mu": torch.tensor([0] * 3),
    "static_input_scalers_sigma": torch.tensor([1] * 3),
    "static_input_scalers_epsilon": 0,
    "output_scalers": torch.tensor([0] * 1280),
    #"encoder_hidden_channels_multiplier" : [1, 2, 4, 8],
    #"encoder_num_encoder_blocks" : 4,
    #"decoder_hidden_channels_multiplier" : [(16, 8), (12, 4), (6, 2), (3, 1)],
    #"decoder_num_decoder_blocks" : 4,
    "aux_decoders": "unetpincer",
    "backbone": "prithviwxc",
    "skip_connection": "True"
}
task = WxCTask('WxCModelFactory', model_args=model_args, mode='eval')
```

```python
task = WxCTask('WxCModelFactory', model_args=model_args, mode='train')
```

--------------------------------

### Specify configuration file path

Source: https://github.com/terrastackai/terratorch/blob/main/docs/tutorials/burn_scars_inference_simplified.md

Sets the path to the YAML configuration file required for model setup.

```python
config_file = "burn_scars_config.yaml"
```

--------------------------------

### Get Validation Dataset Size

Source: https://github.com/terrastackai/terratorch/blob/main/examples/multitemporal_data/multitemporal_model_segementation_crop.ipynb

Calculates and returns the total number of samples in the validation split of the dataset.

```python
# checking datasets validation split size
val_dataset = datamodule.val_dataset
len(val_dataset)
```

--------------------------------

### Initialise a Prithvi model

Source: https://github.com/terrastackai/terratorch/blob/main/examples/segmentation/segmentation_sen1floods11.ipynb

Builds a Prithvi EO v2 model using the registry.

```python
model = BACKBONE_REGISTRY.build("prithvi_eo_v2_100_tl")
print(model)
```

--------------------------------

### Get Training Dataset Size

Source: https://github.com/terrastackai/terratorch/blob/main/examples/multitemporal_data/multitemporal_model_segementation_crop.ipynb

Calculates and returns the total number of samples in the training split of the dataset.

```python
# checking datasets train split size
train_dataset = datamodule.train_dataset
len(train_dataset)
```

--------------------------------

### vLLM Instance Ready Logs

Source: https://github.com/terrastackai/terratorch/blob/main/docs/guide/vllm/serving_a_model_image.md

These logs indicate a successfully initialized vLLM serving instance, ready to accept requests.

```log
(APIServer pid=532339) INFO 10-01 09:01:06 [launcher.py:44] Route: /scale_elastic_ep, Methods: POST
(APIServer pid=532339) INFO 10-01 09:01:06 [launcher.py:44] Route: /is_scaling_elastic_ep, Methods: POST
(APIServer pid=532339) INFO 10-01 09:01:06 [launcher.py:44] Route: /invocations, Methods: POST
(APIServer pid=532339) INFO 10-01 09:01:06 [launcher.py:44] Route: /metrics, Methods: GET
(APIServer pid=532339) INFO:     Started server process [532339]
(APIServer pid=532339) INFO:     Waiting for application startup.
(APIServer pid=532339) INFO:     Application startup complete.
```

--------------------------------

### Configure and Run PyTorch Lightning Trainer

Source: https://github.com/terrastackai/terratorch/blob/main/examples/weather_climate/WxCTutorialDownscalingECCC.ipynb

Initializes a trainer with logging and checkpointing, then executes the training process.

```python
logger = TensorBoardLogger("logs", name="eccc", log_graph=False)
checkpoint_callback = CheckpointCallback(config, save_every_n_epochs=2)

trainer = Trainer(
    max_epochs=3,
    check_val_every_n_epoch=1,
    log_every_n_steps=config.data.n_random_windows,
    logger=logger,
    callbacks=[checkpoint_callback],
    devices=1,
    precision="16-mixed",
)
```

```python
datamodule.setup(stage='val')
trainer.fit(model=task, datamodule=datamodule)
```

--------------------------------

### Example Output for Concat Pooling Inspection

Source: https://github.com/terrastackai/terratorch/blob/main/examples/multitemporal_data/temporalwrapper_segementation_crop.ipynb

Expected console output when running the inspection script for concat pooling.

```text
Model Pipeline — Concat Pooling
-------------------------------
Input shape:          (1, 6, 3, 224, 224)  [B, C, T, H, W] → Batch size 1, 6 channels, 3 timesteps
Backbone:             terramind_v1_small
Temporal pooling:     concat
Encoder feature maps: 12
Example feature map:  (1, 196, 1152)

The TerraMind ViT-Small backbone returns a feature map per transformer block (12 total), each with 196 spatial tokens and 384-dimensional embeddings.
Concat pooling stacks the features from all 3 timesteps along the channel dimension (now 1152-dim) before passing them to the U-Net decoder, increasing the decoder’s input size and number of parameters.

Model Parameters
----------------
Total parameters:     44.81M
Trainable (decoder):  22.94M
```