Merge branch 'main' of github.com:milancurcic/pfml-iccs-talk

slides/fortran-lang.md

@@ -2,9 +2,9 @@
 
 ## Fortran continues to improve in many ways..
 
-* MPI+Fortran keeps getting better (Thanks, Jeff)
-* Native parallel Fortran continues to improve (Thanks, Kate, Damian & CLaSS team)
-* Fortran 2028 will bring type-safe generics (Thanks, Tom & The Generics team)
+* Latest release is Fortran 2023
+* Native and MPI-based parallelism continue to improve
+* Fortran 2028 will bring powerful, type-safe, generics
 * The modern ecosystem of Fortran tools and libraries is growing rapidly
 </section>
 
@@ -19,11 +19,21 @@
   and the US Fortran Standards Committee
 * Supported by the Google Summer of Code Program 4 years in a row (19 student projects to date)
 * Grants from the Sovereign Tech Fund in 2022 and 2023
-* And now a [NumFOCUS](https://numfocus.org) fiscally sponsored project
+* A [NumFOCUS](https://numfocus.org/project/fortran-lang) fiscally sponsored project
 
 </section>
 
 
+<section>
+
+## Fortran's popularity according to TIOBE
+
+![TIOBE](assets/tiobe_fortran.png)
+
+<div class="reference">Image credit: Jacob Williams</div>
+</section>
+
+
 <section>
 
 ## fpm: Fortran Package Manager

slides/machine-learning.md

@@ -18,12 +18,12 @@
 
 ## ...and so it eats the Numerical Weather Prediction (NWP) as well
 
-* We've entered a new revolution in NWP.
-* AI NWP models are:
+* We've entered a new era of NWP.
+* Data-based (AI) NWP models are:
   - orders of magnitude faster;
   - match or beat the accuracy of traditional models on _some_ metrics.
 * New AI models appear rapidly: FourCastNet, GraphCast, PanguWeather, FuXi, NeuralGCM, AtmoRep, ClimaX, STORMER, ACE, StormCast, etc.
-* Though extremely useful for prediction, they cannot (yet) replace the traditional physics-based models.
+* Though extremely useful for prediction, they can't (yet) replace the physics-based models.
 </section>
 
 
@@ -60,10 +60,11 @@ $\rightarrow$ Domain scientist asks: How do I run ML within my Fortran applicati
 
 ## Multiple approaches to ML in Fortran
 
+* Two-language disk-based interop, e.g. [Fortran Keras Bridge](https://doi.org/10.1155/2020/8888811)
 * PyTorch bindings: [pytorch-fortran](https://github.com/alexeedm/pytorch-fortran), [ftorch](https://github.com/Cambridge-ICCS/FTorch), [torchfort](https://github.com/NVIDIA/TorchFort)
 * TensorFlow bindings: [fortran-tf-lib](https://github.com/Cambridge-ICCS/fortran-tf-lib)
 * _Pure Fortran ML_:
-  - Naturally great fit for ML: First-class multi-dimensional arrays, `matmul`, `dot_product`, `sum`, `pack`, `unpack`, `reshape` etc.
+  - Natural fit for ML: First-class multi-dimensional arrays, `matmul`, `dot_product`, `sum`, `pack`, `unpack`, `reshape` etc.
   - Fast linear algebra libraries
   - No complexity associated with multi-language applications
   - Many Fortran programmers would love to use pure Fortran ML.

slides/neural-fortran.md

@@ -41,10 +41,10 @@
 
 ## Design principles
 
-1. **Easy to use**: nice, concise, high-level API;
-2. **Easy to extend**: modular design;
-3. **Be hardware-agnostic**: let the compiler optimize when possible;
-4. **Be self-contained**: no external dependencies when possible;
+1. **Easy to use**: nice, concise, high-level API
+2. **Easy to extend**: modular design
+3. **Hardware-agnostic**: let the compiler optimize & accelerate
+4. **Self-contained**: no external dependencies when possible
 </section>
 
 
@@ -91,7 +91,7 @@ end program simple
 
 <section>
 
-## Creating, training, and running a CNN with neural-fortran
+## Convolutional network training and inference
 
 ```fortran
 use nf
@@ -125,7 +125,37 @@ print *, net % predict(test_images)
 
 <section>
 
-## More Fortran ML goodness
+## nf-keras-hdf5: Load Keras models saved in HDF5
+
+* [https://github.com/neural-fortran/nf-keras-hdf5](https://github.com/neural-fortran/nf-keras-hdf5)
+* Optional plugin for neural-fortran
+* Instead of defining the network, give it the HDF5 path string
+
+```fortran
+program cnn_from_keras
+  use nf, only: network
+
+  type(network) :: net
+
+  net = network("keras_cnn_mnist.h5")
+
+  ! run inference
+```
+</section>
+
+
+<section>
+
+## Single-core CPU performance in training
+
+![benchmark_dense_mnist](assets/benchmark_dense_mnist.png)
+
+</section>
+
+
+<section>
+
+## Other Fortran ML libraries
 
 * [inference-engine](https://github.com/BerkeleyLab/inference-engine) - A deep learning library for HPC applications
 * [athena](https://github.com/nedtaylor/athena) - Another ML framework, inspired by neural-fortran

slides/takeaways.md

@@ -1,20 +1,19 @@
 <section>
 
-## Takeaways and some personal outlooks
+## Takeaways
 
-* NWP has entered a new revolution with AI models
+* Machine Learning is transforming the the NWP and Earth System Modeling enterprise
 * Phenomenal *forecasting* tools that won't replace traditional models for *research*
-* Pure Fortran ML has its place in Earth System Modeling
-* What does the future of Earth System Modeling look like?
-  - A wide, diverse, spectrum of traditional, AI, and hybrid models
-  - ML implemented in many languages for respective needs
-  - Hardware will continue to evolve and remain ahead of software
-* Fortran will continue to rock and improve.
+* Pure Fortran ML approach may appeal to:
+  - Domain scientist who mostly work with Fortran-based models
+  - Fortran enthusiasts
+* Fortran the language and its ecosystem of compilers, tools, and libraries
+  continues to improve
+* If you haven't tried modern Fortran before, please do!
 </section>
 
 
 <section>
 
-## Thank you and get off my porch!
-😉
+## Thank you!
 </section>

slides/title.md

@@ -23,7 +23,7 @@ _September 10, 2024_
 * I study ocean waves and their role in the Earth system
 * Mix of theory, field and laboratory measurements, and numerical modeling
 * I wrote a popular book for Fortran beginners
-* I'm not an ML practitioner, but I love learning new things
+* I'm not an ML practitioner yet, but I enjoy making useful tools
 </section>
 
 
@@ -69,9 +69,6 @@ _September 10, 2024_
     introducing an integration complexity when used in Fortran applications.
 * **Pure Fortran** ML
   - Fortran itself provides all the necessary building blocks for ML,
-    making a Pure Fortran approach one candidate to integrating ML in physics-based
+    making a Pure Fortran approach a candidate to applying ML in physics-based
     Fortran models.
-* **Pure Fortran ML**
-  - We will look at one candidate solution to this problem and the challenges in
-    the design and implementation.
 </section>