Posts in category: PyTorch

At the end of March, the PyTorch Korea User Group hosted a special meetup that brought together prominent speakers for deep discussions on the PyTorch core and its broader ecosystem. With the event more than doubling in size compared to past gatherings, we were able to connect with even more developers and share insights. Huge thanks to goorm for sponsoring the fantastic venue!

This recap is for those who couldn’t attend in person, as well as for participants who want to revisit the energy and insights of the day. The event featured experts in core PyTorch, AI accelerators, inference optimization, and large language model development. Below is a quick overview of the key sessions that anchored the conference.

1⃣ Jerry Lee | PyTorch Foundation

Representing the PyTorch Foundation, part of the Linux Foundation, Jaeung provided an overview of how PyTorch is driving core open source technologies forward. He shared PyTorch’s growth story, the many global projects currently in motion, and the ecosystem’s impressive 20%+ annual growth. The session also covered how the foundation operates, how member organizations are involved, and upcoming plans that are particularly useful for practitioners.

2⃣ Alban Desmaison | PyTorch Roadmap

Alban shared the design philosophy behind PyTorch and Meta’s official contribution roadmap (link). He provided a deep technical dive into the differences between Eager and Compiled modes, especially breaking down the backend architecture of device Eager execution. Practical tools and improvements were also introduced—such as memory profilers, enhanced custom operator support, and pinned memory optimizations.

3⃣ Hongseok Kim | PyTorch on Rebellions AI Accelerators: Status

Rebellions is building runtime integration for their proprietary NPU architecture, fully aligned with the structural changes in PyTorch 2.0. This talk introduced the performance and scalability of their upcoming chip, their integration strategy with the PyTorch runtime, and challenges in supporting Eager Mode. Hongseok also previewed their roadmap toward releasing these features within the year.

4⃣ Kyujin Cho | Backend.AI: A Unified Platform for All AI Accelerators

Backend.AI abstracts and integrates various AI accelerators into a unified workflow. As the diversity of accelerator architectures grows, the need for portability and infrastructure unification becomes even more important. This session showcased features across development and operations—from NPU scheduling and resource allocation to monitoring. Backend.AI currently supports accelerators from NVIDIA, Intel, Tenstorrent, Rebellions, and more.

5⃣ Taeho Kim | Optimizing & Deploying Models Across Multiple Chipsets Using NetsPresso

This talk focused on the challenges of inference in real-world industrial applications of AI models. As new state-of-the-art models emerge rapidly, there’s a growing need for environments that can quickly validate device compatibility—ideally with one-click ease. NetsPresso is actively working on a static graph representation compatible with PyTorch, offering efficient support for model development, optimization, and testing.

6⃣ Jungyeop Lee | The Journey to Reproduce Deepseek-R1

Jungyeop took us through his journey of reproducing Deepseek, a large language model—an effort that involved 201 experiments. He shared real-world lessons from training with Korean data, tokenizer modifications, and fine-tuning strategies. His practical insights and next steps were especially valuable for those building or re-implementing large models from scratch.

7⃣ Sol Kim | A journey from TCP architecture to production-level LLMs

Sol presented an integrated optimization approach to deploying large models using the TCP(Tensor Contraction Processor) architecture, which supports tensor contraction at the hardware level. The talk highlighted optimization techniques built on hardware abstraction layers (HALs) and bottom-up integration strategies with PyTorch—offering a hybrid hardware-software perspective.

Panel Talk & Q&A

The event wrapped up with an engaging panel discussion. Attendees asked sharp questions, and the speakers offered insightful answers. It was a powerful moment that captured the community’s enthusiasm for PyTorch and their hunger for deeper technical understanding.

Final Thoughts

Since our first offline meetup in October 2022, the PyTorch Korea User Group has held five major technical conferences. Each event deepens our appreciation for the scale and depth of the PyTorch ecosystem. With perspectives from users, contributors, and ecosystem builders, the stories we share are only growing—and we’re committed to continuing this journey together.

See you at the next conference—with even more exciting talks to come!

Read More

PyTorch Day France offers a front-row seat to the future of open source AI. Taking place 7 May at Station F in Paris and co-located with GOSIM AI Paris, this one-day event will bring together developers, researchers, and industry leaders for a day of technical sessions, real-world insights, and community exchange.

A Major Milestone for the PyTorch Foundation

This event marks the very first PyTorch Day, launching a new international series hosted annually in different regions to convene AI researchers, developers, engineers, and enthusiasts. PyTorch Days are designed to spotlight open source AI advancements, foster community collaboration, and provide a forum to learn about active, high-impact AI projects built using PyTorch.

PyTorch Day France also represents a pivotal moment in the PyTorch Foundation’s journey. With its recent expansion into an umbrella foundation, PyTorch is now positioned to support a broader ecosystem of trusted, community-driven AI projects across the full AI lifecycle.

At PyTorch Day France, you’ll hear directly from PyTorch Foundation Executive Director, Matt White, about this transition—and get a first look at some exciting announcements.

Registration Details

Register now with code  PYTORCH for free access to the full day of PyTorch Day France sessions, plus GOSIM AI Paris.

Two events, one registration—double the sessions, double the innovation.
Register here

Featured Sessions

The day’s agenda includes deep technical dives and applied AI use cases from across the community, including the following talks:

• Luca Antiga (Lightning AI)
  Lightning Thunder: Supercharged PyTorch for Modern Hardware
• Erwan Gallen & Eldar Kurtic (Red Hat)
  Scaling LLM Inference with vLLM: Multi‑Accelerator Serving and Quantized LLMs
• Pierre Rouanet (Pollen Robotics)
  Real-World Robotics as the Next Frontier for AI?
• Pablo Montalvo (Hugging Face)
  PyTorch x Transformers: Pythonicity, Autodiff, and Modularity Defining Modern AI
• Pedro Ortis (Common Crawl)
  Harnessing Common Crawl for AI and ML Applications
• Meriem Bendris (NVIDIA)
  Teaching Mistral to Reason: Post-Training with PyTorch and NVIDIA
• Olatunji Ruwase (Snowflake)
  DeepSpeed – Efficient Training Scalability for Deep Learning Models

View the full schedule.

Whether you’re a contributor, practitioner, or simply curious about what’s ahead, PyTorch Day France is an opportunity to connect with the community and shape what’s next for our ecosystem.

Read More

PyTorch Day France offers a front-row seat to the future of open source AI. Taking place 7 May at Station F in Paris and co-located with GOSIM AI Paris, this one-day event will bring together developers, researchers, and industry leaders for a day of technical sessions, real-world insights, and community exchange.

A Major Milestone for the PyTorch Foundation

This event marks the very first PyTorch Day, launching a new international series hosted annually in different regions to convene AI researchers, developers, engineers, and enthusiasts. PyTorch Days are designed to spotlight open source AI advancements, foster community collaboration, and provide a forum to learn about active, high-impact AI projects built using PyTorch.

PyTorch Day France also represents a pivotal moment in the PyTorch Foundation’s journey. With its recent expansion into an umbrella foundation, PyTorch is now positioned to support a broader ecosystem of trusted, community-driven AI projects across the full AI lifecycle.

At PyTorch Day France, you’ll hear directly from PyTorch Foundation Executive Director, Matt White, about this transition—and get a first look at some exciting announcements.

Registration Details

Register now with code  PYTORCH for free access to the full day of PyTorch Day France sessions, plus GOSIM AI Paris.

Two events, one registration—double the sessions, double the innovation.
Register here

Featured Sessions

The day’s agenda includes deep technical dives and applied AI use cases from across the community, including the following talks:

• Luca Antiga (Lightning AI)
  Lightning Thunder: Supercharged PyTorch for Modern Hardware
• Erwan Gallen & Eldar Kurtic (Red Hat)
  Scaling LLM Inference with vLLM: Multi‑Accelerator Serving and Quantized LLMs
• Pierre Rouanet (Pollen Robotics)
  Real-World Robotics as the Next Frontier for AI?
• Pablo Montalvo (Hugging Face)
  PyTorch x Transformers: Pythonicity, Autodiff, and Modularity Defining Modern AI
• Pedro Ortis (Common Crawl)
  Harnessing Common Crawl for AI and ML Applications
• Meriem Bendris (NVIDIA)
  Teaching Mistral to Reason: Post-Training with PyTorch and NVIDIA
• Olatunji Ruwase (Snowflake)
  DeepSpeed – Efficient Training Scalability for Deep Learning Models

View the full schedule.

Whether you’re a contributor, practitioner, or simply curious about what’s ahead, PyTorch Day France is an opportunity to connect with the community and shape what’s next for our ecosystem.

Read More

At the end of March, the PyTorch Korea User Group hosted a special meetup that brought together prominent speakers for deep discussions on the PyTorch core and its broader ecosystem. With the event more than doubling in size compared to past gatherings, we were able to connect with even more developers and share insights. Huge thanks to goorm for sponsoring the fantastic venue! 😄

This recap is for those who couldn’t attend in person, as well as for participants who want to revisit the energy and insights of the day. The event featured experts in core PyTorch, AI accelerators, inference optimization, and large language model development. Below is a quick overview of the key sessions that anchored the conference.

1️⃣ Jerry Lee | PyTorch Foundation

Representing the PyTorch Foundation, part of the Linux Foundation, Jaeung provided an overview of how PyTorch is driving core open source technologies forward. He shared PyTorch’s growth story, the many global projects currently in motion, and the ecosystem’s impressive 20%+ annual growth. The session also covered how the foundation operates, how member organizations are involved, and upcoming plans that are particularly useful for practitioners.

2️⃣ Alban Desmaison | PyTorch Roadmap

Alban shared the design philosophy behind PyTorch and Meta’s official contribution roadmap (link). He provided a deep technical dive into the differences between Eager and Compiled modes, especially breaking down the backend architecture of device Eager execution. Practical tools and improvements were also introduced—such as memory profilers, enhanced custom operator support, and pinned memory optimizations.

3️⃣ Hongseok Kim | PyTorch on Rebellions AI Accelerators: Status

Rebellions is building runtime integration for their proprietary NPU architecture, fully aligned with the structural changes in PyTorch 2.0. This talk introduced the performance and scalability of their upcoming chip, their integration strategy with the PyTorch runtime, and challenges in supporting Eager Mode. Hongseok also previewed their roadmap toward releasing these features within the year.

4️⃣ Kyujin Cho | Backend.AI: A Unified Platform for All AI Accelerators

Backend.AI abstracts and integrates various AI accelerators into a unified workflow. As the diversity of accelerator architectures grows, the need for portability and infrastructure unification becomes even more important. This session showcased features across development and operations—from NPU scheduling and resource allocation to monitoring. Backend.AI currently supports accelerators from NVIDIA, Intel, Tenstorrent, Rebellions, and more.

5️⃣ Taeho Kim | Optimizing & Deploying Models Across Multiple Chipsets Using NetsPresso

This talk focused on the challenges of inference in real-world industrial applications of AI models. As new state-of-the-art models emerge rapidly, there’s a growing need for environments that can quickly validate device compatibility—ideally with one-click ease. NetsPresso is actively working on a static graph representation compatible with PyTorch, offering efficient support for model development, optimization, and testing.

6️⃣ Jungyeop Lee | The Journey to Reproduce Deepseek-R1

Jungyeop took us through his journey of reproducing Deepseek, a large language model—an effort that involved 201 experiments. He shared real-world lessons from training with Korean data, tokenizer modifications, and fine-tuning strategies. His practical insights and next steps were especially valuable for those building or re-implementing large models from scratch.

7️⃣ Sol Kim | A journey from TCP architecture to production-level LLMs

Sol presented an integrated optimization approach to deploying large models using the TCP(Tensor Contraction Processor) architecture, which supports tensor contraction at the hardware level. The talk highlighted optimization techniques built on hardware abstraction layers (HALs) and bottom-up integration strategies with PyTorch—offering a hybrid hardware-software perspective.

💡 Panel Talk & Q&A 💡

The event wrapped up with an engaging panel discussion. Attendees asked sharp questions, and the speakers offered insightful answers. It was a powerful moment that captured the community’s enthusiasm for PyTorch and their hunger for deeper technical understanding.

Final Thoughts

Since our first offline meetup in October 2022, the PyTorch Korea User Group has held five major technical conferences. Each event deepens our appreciation for the scale and depth of the PyTorch ecosystem. With perspectives from users, contributors, and ecosystem builders, the stories we share are only growing—and we’re committed to continuing this journey together.

See you at the next conference—with even more exciting talks to come! 🙌

Read More

Earth Observation-based analytics are becoming essential for understanding our planet — from monitoring deforestation to tracking urban development and analyzing the impacts of climate change. However, the coding and deep learning skills for applying AI models to satellite imagery and earth observation data has traditionally been a major barrier for many practitioners.

By IBM Research’s launch of TerraTorch 1.0, a PyTorch domain library for fine-tuning of Geospatial Computer Vision Foundation Models, we make geospatial AI not only more accessible but also more practical for the wider PyTorch community. Our goal: simplify the process so that any data scientist, researcher, or enthusiast can build powerful geospatial models with ease and low GPU and data processing requirements.

The power of foundation models, even with 75-95% of the input data removed, the models do a fantastic job in reconstruction of the input data – therefore learning the underlying physics of our planet in a deep, latent space

The Business Challenge

Our goal was to remove the technical barriers that prevent people from working with satellite imagery, weather and climate data at scale. Together with NASA, we’ve developed the Prithvi family of foundation models. Integrating the latest innovations of AI research using the clean API PyTorch provides has facilitated the job.

We wanted to create a framework that anyone can use to go from raw data to inference ready models in just a few steps.

How a weather and climate foundation model created and fine-tuned on PyTorch is used for weather forecasts

How IBM Research Used PyTorch

We’ve built TerraTorch on top of PyTorch, leveraging its dynamic ecosystem to integrate:

• PyTorch Lightning for clean, scalable training loops
• TorchGeo for geospatial data handling and transformations (PyTorch transforms)
• For foundation models like the leading generative multimodal foundation model ‘Terramind’, co-developed by IBM and ESA, and the ‘Prithvi’ family, co-developed by IBM and NASA, TerraTorch has been used to fine-tune all of the downstream geospatial models for satellite imagery, weather and climate data. It includes the family of fine-tuned models that IBM has released as part of Granite. In addition, other interesting foundation models and ecosystem components like Clay, SatMAE, Satlas, DeCur and DOFA are included in TerraTorch.
• Powerful and state-of-the-art vision transformers to experiment with modern neural network architectures
• TerraTorch-Iterate build on top of PyTorch, Optuna, MLFlow and Ray Tune for Hyperparameter Optimization (HPO), Neural Architecture Search (NAS) and Foundation Model Benchmarking (GeoBench), where TerraTorch became the reference implementation

The fine-tuning and inference process is completely described in a single YAML config file. There, the architectural building blocks of the model (backbone, neck, decoder, head) are defined. The Model Factory assembles the model using the build-in and custom registries. In addition, the Optimizer and Data Modules are created as defined in the config. Finally, everything is passed to the Lightning Trainer, who executes the task.

With PyTorch’s flexibility, we were able to prototype quickly, iterate on model architectures, and deploy pipelines for a range of geospatial applications — from flood and biomass detection to increasing resolution of climate data, where some of our our work became part of the IBM Granite Geospatial Model Family.

Architecture of the Prithvi-EO-2.0-600M foundation model which IBM Research developed together with NASA

Solving AI Challenges with PyTorch

PyTorch helped us to tackle three major challenges:

• Ease of experimentation: Dynamic computation graphs, automatic differentiation, full abstraction of CUDA and rich visualization tools made it simple to test different models and training strategies.
• Scalability: With DDP, FSDP, PyTorch Lightning and TorchGeo, we could train models on large-scale datasets without worrying about infrastructure.
• Community support: PyTorch – the de-facto standard in AI research – with its active community and excellent documentation made it easy to overcome hurdles and stay up to date with the latest advancements in AI research.

A Word from IBM Research

“PyTorch gave me the power to turn complex linear algebra and optimization problems into accessible, shareable solutions for the community. It feels empowering that we’re building and fine-tuning models for anyone curious about understanding our planet through AI.”

— Romeo Kienzler, AI Research Engineer at IBM Research Zurich, Rueschlikon

The Benefits of Using PyTorch

Using PyTorch allowed us to:

• Build a reproducible, open-source framework for fine-tuning geospatial foundation models
• Share our work with the community through easy-to-follow notebooks, TerraTorch configuration files, tutorials and model checkpoints on HuggingFace
• Rapidly iterate over foundation model architectures and deploy fine-tuned models for inference, from research to real-world client products

Learn More

For more information about this project and to explore the code, visit:

• GitHub Repository
• IBM Research: Simplifying Geospatial AI with TerraTorch 1.0
• TerraTorch PrithviEOv2 example notebooks
• TerraMind example notebooks
• Run TerraMind using TerraTorch on Colab

Read More

Earth Observation-based analytics are becoming essential for understanding our planet — from monitoring deforestation to tracking urban development and analyzing the impacts of climate change. However, the coding and deep learning skills for applying AI models to satellite imagery and earth observation data has traditionally been a major barrier for many practitioners.

By IBM Research’s launch of TerraTorch 1.0, a PyTorch domain library for fine-tuning of Geospatial Computer Vision Foundation Models, we make geospatial AI not only more accessible but also more practical for the wider PyTorch community. Our goal: simplify the process so that any data scientist, researcher, or enthusiast can build powerful geospatial models with ease and low GPU and data processing requirements.

The power of foundation models, even with 75-95% of the input data removed, the models do a fantastic job in reconstruction of the input data – therefore learning the underlying physics of our planet in a deep, latent space

The Business Challenge

Our goal was to remove the technical barriers that prevent people from working with satellite imagery, weather and climate data at scale. Together with NASA, we’ve developed the Prithvi family of foundation models. Integrating the latest innovations of AI research using the clean API PyTorch provides has facilitated the job.

We wanted to create a framework that anyone can use to go from raw data to inference ready models in just a few steps.

How a weather and climate foundation model created and fine-tuned on PyTorch is used for weather forecasts

How IBM Research Used PyTorch

We’ve built TerraTorch on top of PyTorch, leveraging its dynamic ecosystem to integrate:

• PyTorch Lightning for clean, scalable training loops
• TorchGeo for geospatial data handling and transformations (PyTorch transforms)
• For foundation models like the leading generative multimodal foundation model ‘Terramind’, co-developed by IBM and ESA, and the ‘Prithvi’ family, co-developed by IBM and NASA, TerraTorch has been used to fine-tune all of the downstream geospatial models for satellite imagery, weather and climate data. It includes the family of fine-tuned models that IBM has released as part of Granite. In addition, other interesting foundation models and ecosystem components like Clay, SatMAE, Satlas, DeCur and DOFA are included in TerraTorch.
• Powerful and state-of-the-art vision transformers to experiment with modern neural network architectures
• TerraTorch-Iterate build on top of PyTorch, Optuna, MLFlow and Ray Tune for Hyperparameter Optimization (HPO), Neural Architecture Search (NAS) and Foundation Model Benchmarking (GeoBench), where TerraTorch became the reference implementation

The fine-tuning and inference process is completely described in a single YAML config file. There, the architectural building blocks of the model (backbone, neck, decoder, head) are defined. The Model Factory assembles the model using the build-in and custom registries. In addition, the Optimizer and Data Modules are created as defined in the config. Finally, everything is passed to the Lightning Trainer, who executes the task.

With PyTorch’s flexibility, we were able to prototype quickly, iterate on model architectures, and deploy pipelines for a range of geospatial applications — from flood and biomass detection to increasing resolution of climate data, where some of our our work became part of the IBM Granite Geospatial Model Family.

Architecture of the Prithvi-EO-2.0-600M foundation model which IBM Research developed together with NASA

Solving AI Challenges with PyTorch

PyTorch helped us to tackle three major challenges:

• Ease of experimentation: Dynamic computation graphs, automatic differentiation, full abstraction of CUDA and rich visualization tools made it simple to test different models and training strategies.
• Scalability: With DDP, FSDP, PyTorch Lightning and TorchGeo, we could train models on large-scale datasets without worrying about infrastructure.
• Community support: PyTorch – the de-facto standard in AI research – with its active community and excellent documentation made it easy to overcome hurdles and stay up to date with the latest advancements in AI research.

A Word from IBM Research

“PyTorch gave me the power to turn complex linear algebra and optimization problems into accessible, shareable solutions for the community. It feels empowering that we’re building and fine-tuning models for anyone curious about understanding our planet through AI.”

— Romeo Kienzler, AI Research Engineer at IBM Research Zurich, Rueschlikon

The Benefits of Using PyTorch

Using PyTorch allowed us to:

• Build a reproducible, open-source framework for fine-tuning geospatial foundation models
• Share our work with the community through easy-to-follow notebooks, TerraTorch configuration files, tutorials and model checkpoints on HuggingFace
• Rapidly iterate over foundation model architectures and deploy fine-tuned models for inference, from research to real-world client products

Learn More

For more information about this project and to explore the code, visit:

• GitHub Repository
• IBM Research: Simplifying Geospatial AI with TerraTorch 1.0
• TerraTorch PrithviEOv2 example notebooks
• TerraMind example notebooks
• Run TerraMind using TerraTorch on Colab

Read More

We’re thrilled to announce the 2025 PyTorch Docathon! This is a hackathon-style event aimed at enhancing PyTorch documentation with the support of the community. Documentation is a vital component of any technology, and by refining it, we can simplify the onboarding process for new users, help them effectively utilize PyTorch’s features, and ultimately speed up the transition from research to production in machine learning.

WHY PARTICIPATE

Low Barrier to Entry

Unlike many open-source projects that require deep knowledge of the codebase and previous contributions to join hackathon events, the Docathon is tailored for newcomers. While we expect participants to be familiar with Python, and have basic knowledge of PyTorch and machine learning, there are tasks related to website issues that don’t even require that level of expertise.

Tangible Results

A major advantage of the Docathon is witnessing the immediate impact of your contributions. Enhancing documentation significantly boosts a project’s usability and accessibility, and you’ll be able to observe these improvements directly. Seeing tangible outcomes can also be a strong motivator to continue contributing.

Collaborative Environment

The Docathon fosters a collaborative atmosphere, offering you the chance to work alongside other contributors and PyTorch maintainers to improve the documentation. This is a fantastic opportunity to learn from peers, exchange ideas, and build connections.

Learning Opportunities

Even if you’re not a PyTorch expert, the Docathon offers a valuable learning experience. You’ll have the chance to delve into PyTorch modules, test tutorials on your machine, and explore them in the CI environment.

WHO SHOULD PARTICIPATE

Whether you’re a seasoned documentation expert or just starting out, we invite everyone to join in the PyTorch docathon to contribute and develop your skills and knowledge to help improve the documentation for everyone! We will have issues labelled by skill level, and the PyTorch Discord will be available for collaboration and help.

EVENT DETAILS

• June 3: Kick-off 10 AM PT
• June 4 – June 15: Submissions and Feedback
• June 16 – June 17: Final Reviews
• June 18: Winner Announcements

Make sure to RSVP to the event so you receive all the notifications and instructions on how to participate.

Further details about the Docathon will be shared during the Kick-off call on June 3.

Don’t forget to register for this year’s event: RSVP now

Read More

We’re thrilled to announce the 2025 PyTorch Docathon! This is a hackathon-style event aimed at enhancing PyTorch documentation with the support of the community. Documentation is a vital component of any technology, and by refining it, we can simplify the onboarding process for new users, help them effectively utilize PyTorch’s features, and ultimately speed up the transition from research to production in machine learning.

WHY PARTICIPATE

Low Barrier to Entry

Unlike many open-source projects that require deep knowledge of the codebase and previous contributions to join hackathon events, the Docathon is tailored for newcomers. While we expect participants to be familiar with Python, and have basic knowledge of PyTorch and machine learning, there are tasks related to website issues that don’t even require that level of expertise.

Tangible Results

A major advantage of the Docathon is witnessing the immediate impact of your contributions. Enhancing documentation significantly boosts a project’s usability and accessibility, and you’ll be able to observe these improvements directly. Seeing tangible outcomes can also be a strong motivator to continue contributing.

Collaborative Environment

The Docathon fosters a collaborative atmosphere, offering you the chance to work alongside other contributors and PyTorch maintainers to improve the documentation. This is a fantastic opportunity to learn from peers, exchange ideas, and build connections.

Learning Opportunities

Even if you’re not a PyTorch expert, the Docathon offers a valuable learning experience. You’ll have the chance to delve into PyTorch modules, test tutorials on your machine, and explore them in the CI environment.

WHO SHOULD PARTICIPATE

Whether you’re a seasoned documentation expert or just starting out, we invite everyone to join in the PyTorch docathon to contribute and develop your skills and knowledge to help improve the documentation for everyone! We will have issues labelled by skill level, and the PyTorch Discord will be available for collaboration and help.

EVENT DETAILS

• June 3: Kick-off 10 AM PT
• June 4 – June 15: Submissions and Feedback
• June 16 – June 17: Final Reviews
• June 18: Winner Announcements

Make sure to RSVP to the event so you receive all the notifications and instructions on how to participate.

Further details about the Docathon will be shared during the Kick-off call on June 3.

Don’t forget to register for this year’s event: RSVP now

Read More

Earth Observation-based analytics are becoming essential for understanding our planet — from monitoring deforestation to tracking urban development and analyzing the impacts of climate change. However, the coding and deep learning skills for applying AI models to satellite imagery and earth observation data has traditionally been a major barrier for many practitioners.

By IBM Research’s launch of TerraTorch 1.0, a PyTorch domain library for fine-tuning of Geospatial Computer Vision Foundation Models, we make geospatial AI not only more accessible but also more practical for the wider PyTorch community. Our goal: simplify the process so that any data scientist, researcher, or enthusiast can build powerful geospatial models with ease and low GPU and data processing requirements.

The power of foundation models, even with 75-95% of the input data removed, the models do a fantastic job in reconstruction of the input data – therefore learning the underlying physics of our planet in a deep, latent space

The Business Challenge

Our goal was to remove the technical barriers that prevent people from working with satellite imagery, weather and climate data at scale. Together with NASA, we’ve developed the Prithvi family of foundation models. Integrating the latest innovations of AI research using the clean API PyTorch provides has facilitated the job.

We wanted to create a framework that anyone can use to go from raw data to inference ready models in just a few steps.

How a weather and climate foundation model created and fine-tuned on PyTorch is used for weather forecasts

How IBM Research Used PyTorch

We’ve built TerraTorch on top of PyTorch, leveraging its dynamic ecosystem to integrate:

• PyTorch Lightning for clean, scalable training loops
• TorchGeo for geospatial data handling and transformations (PyTorch transforms)
• For foundation models like the leading generative multimodal foundation model ‘Terramind’, co-developed by IBM and ESA, and the ‘Prithvi’ family, co-developed by IBM and NASA, TerraTorch has been used to fine-tune all of the downstream geospatial models for satellite imagery, weather and climate data. It includes the family of fine-tuned models that IBM has released as part of Granite. In addition, other interesting foundation models and ecosystem components like Clay, SatMAE, Satlas, DeCur and DOFA are included in TerraTorch.
• Powerful and state-of-the-art vision transformers to experiment with modern neural network architectures
• TerraTorch-Iterate build on top of PyTorch, Optuna, MLFlow and Ray Tune for Hyperparameter Optimization (HPO), Neural Architecture Search (NAS) and Foundation Model Benchmarking (GeoBench), where TerraTorch became the reference implementation

The fine-tuning and inference process is completely described in a single YAML config file. There, the architectural building blocks of the model (backbone, neck, decoder, head) are defined. The Model Factory assembles the model using the build-in and custom registries. In addition, the Optimizer and Data Modules are created as defined in the config. Finally, everything is passed to the Lightning Trainer, who executes the task.

With PyTorch’s flexibility, we were able to prototype quickly, iterate on model architectures, and deploy pipelines for a range of geospatial applications — from flood and biomass detection to increasing resolution of climate data, where some of our our work became part of the IBM Granite Geospatial Model Family.

Architecture of the Prithvi-EO-2.0-600M foundation model which IBM Research developed together with NASA

Solving AI Challenges with PyTorch

PyTorch helped us to tackle three major challenges:

• Ease of experimentation: Dynamic computation graphs, automatic differentiation, full abstraction of CUDA and rich visualization tools made it simple to test different models and training strategies.
• Scalability: With DDP, FSDP, PyTorch Lightning and TorchGeo, we could train models on large-scale datasets without worrying about infrastructure.
• Community support: PyTorch – the de-facto standard in AI research – with its active community and excellent documentation made it easy to overcome hurdles and stay up to date with the latest advancements in AI research.

A Word from IBM Research

“PyTorch gave me the power to turn complex linear algebra and optimization problems into accessible, shareable solutions for the community. It feels empowering that we’re building and fine-tuning models for anyone curious about understanding our planet through AI.”

— Romeo Kienzler, AI Research Engineer at IBM Research Zurich, Rueschlikon

The Benefits of Using PyTorch

Using PyTorch allowed us to:

• Build a reproducible, open-source framework for fine-tuning geospatial foundation models
• Share our work with the community through easy-to-follow notebooks, TerraTorch configuration files, tutorials and model checkpoints on HuggingFace
• Rapidly iterate over foundation model architectures and deploy fine-tuned models for inference, from research to real-world client products

Learn More

For more information about this project and to explore the code, visit:

• GitHub Repository
• IBM Research: Simplifying Geospatial AI with TerraTorch 1.0
• TerraTorch PrithviEOv2 example notebooks
• TerraMind example notebooks
• Run TerraMind using TerraTorch on Colab

Read More