Update README.md
Browse files
README.md
CHANGED
@@ -18,24 +18,24 @@ tags:
|
|
18 |
|
19 |
# Europe Reanalysis Super Resolution
|
20 |
|
21 |
-
The aim of the project is to create a Machine learning (ML) model that can generate high-resolution regional reanalysis data (similar to the one produced by CERRA) by
|
|
|
22 |
|
|
|
|
|
|
|
23 |
|
24 |
-
|
|
|
|
|
25 |
|
|
|
|
|
26 |
|
27 |
-
|
|
|
28 |
|
29 |
-
|
30 |
-
Moreover, tools for interpretability of DL models can be used to understand the inner workings and decision-making processes of these complex structures by analyzing the activations of different neurons and the importance of different features in the input data.
|
31 |
-
|
32 |
-
This work is funded by [Code for Earth 2023](https://codeforearth.ecmwf.int/) initiative. The model **ConvSwin2SR** is released in Apache 2.0, making it usable without restrictions anywhere.
|
33 |
-
|
34 |
-
|
35 |
-
|
36 |
-
|
37 |
-
|
38 |
-
# Table of Contents
|
39 |
|
40 |
- [Model Card for Europe Reanalysis Super Resolution](#model-card-for--model_id-)
|
41 |
- [Table of Contents](#table-of-contents)
|
@@ -57,8 +57,10 @@ This work is funded by [Code for Earth 2023](https://codeforearth.ecmwf.int/) in
|
|
57 |
- [Metrics](#metrics)
|
58 |
- [Results](#results)
|
59 |
- [Technical Specifications](#technical-specifications-optional)
|
60 |
-
- [Model Architecture
|
61 |
-
|
|
|
|
|
62 |
- [Computing Infrastructure](#computing-infrastructure)
|
63 |
- [Hardware](#hardware)
|
64 |
- [Software](#software)
|
@@ -185,12 +187,6 @@ across different temporal segments.
|
|
185 |
The testing data samples correspond to the three-year period from 2018 to 2020, inclusive. This segment is crucial for assessing the model's real-world applicability and
|
186 |
its performance on the most recent data points, ensuring its relevance and reliability in current and future scenarios.
|
187 |
|
188 |
-
|
189 |
-
### Factors
|
190 |
-
|
191 |
-
<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
|
192 |
-
|
193 |
-
|
194 |
## Results
|
195 |
|
196 |
In our evaluation, the proposed model displayed a significant enhancement over the established baseline, which employs bicubic interpolation for the same task.
|
@@ -210,23 +206,29 @@ In comparison to the bicubic interpolation baseline, our model's superior predic
|
|
210 |
![rmse](metric_global_map_diff_var-rmse.png)
|
211 |
|
212 |
|
213 |
-
|
214 |
# Technical Specifications
|
215 |
|
216 |
-
## Model Architecture
|
|
|
|
|
|
|
217 |
|
218 |
-
|
219 |
|
220 |
-
|
221 |
|
222 |
-
|
223 |
-
|
|
|
|
|
|
|
224 |
|
225 |
-
|
226 |
|
227 |
-
|
|
|
228 |
|
229 |
-
|
230 |
|
231 |
The Swin2 transformer optimizes its parameters using a composite loss function that aggregates multiple \( \mathcal{L}_1 \) loss terms to enhance its predictive
|
232 |
accuracy across different resolutions and representations:
|
|
|
18 |
|
19 |
# Europe Reanalysis Super Resolution
|
20 |
|
21 |
+
The aim of the project is to create a Machine learning (ML) model that can generate high-resolution regional reanalysis data (similar to the one produced by CERRA) by
|
22 |
+
downscaling global reanalysis data from ERA5.
|
23 |
|
24 |
+
This will be accomplished by using state-of-the-art Deep Learning (DL) techniques like U-Net, conditional GAN, and diffusion models (among others). Additionally,
|
25 |
+
an ingestion module will be implemented to assess the possible benefit of using CERRA pseudo-observations as extra predictors. Once the model is designed and trained,
|
26 |
+
a detailed validation framework takes the place.
|
27 |
|
28 |
+
It combines classical deterministic error metrics with in-depth validations, including time series, maps, spatio-temporal correlations, and computer vision metrics,
|
29 |
+
disaggregated by months, seasons, and geographical regions, to evaluate the effectiveness of the model in reducing errors and representing physical processes.
|
30 |
+
This level of granularity allows for a more comprehensive and accurate assessment, which is critical for ensuring that the model is effective in practice.
|
31 |
|
32 |
+
Moreover, tools for interpretability of DL models can be used to understand the inner workings and decision-making processes of these complex structures by analyzing
|
33 |
+
the activations of different neurons and the importance of different features in the input data.
|
34 |
|
35 |
+
This work is funded by [Code for Earth 2023](https://codeforearth.ecmwf.int/) initiative. The model **ConvSwin2SR** is released in Apache 2.0, making it usable without
|
36 |
+
restrictions anywhere.
|
37 |
|
38 |
+
# Table of Contents
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
39 |
|
40 |
- [Model Card for Europe Reanalysis Super Resolution](#model-card-for--model_id-)
|
41 |
- [Table of Contents](#table-of-contents)
|
|
|
57 |
- [Metrics](#metrics)
|
58 |
- [Results](#results)
|
59 |
- [Technical Specifications](#technical-specifications-optional)
|
60 |
+
- [Model Architecture](#model-architecture)
|
61 |
+
- [Components](#components)
|
62 |
+
- [Configuration details](#configuration-details)
|
63 |
+
- [Loss function](#loss-function)
|
64 |
- [Computing Infrastructure](#computing-infrastructure)
|
65 |
- [Hardware](#hardware)
|
66 |
- [Software](#software)
|
|
|
187 |
The testing data samples correspond to the three-year period from 2018 to 2020, inclusive. This segment is crucial for assessing the model's real-world applicability and
|
188 |
its performance on the most recent data points, ensuring its relevance and reliability in current and future scenarios.
|
189 |
|
|
|
|
|
|
|
|
|
|
|
|
|
190 |
## Results
|
191 |
|
192 |
In our evaluation, the proposed model displayed a significant enhancement over the established baseline, which employs bicubic interpolation for the same task.
|
|
|
206 |
![rmse](metric_global_map_diff_var-rmse.png)
|
207 |
|
208 |
|
|
|
209 |
# Technical Specifications
|
210 |
|
211 |
+
## Model Architecture
|
212 |
+
|
213 |
+
Our model's design is deeply rooted in the Swin2 architecture, specifically tailored for Super Resolution (SR) tasks.
|
214 |
+
We've harnessed the [transformers library](https://github.com/huggingface/transformers) to streamline and simplify the model's design.
|
215 |
|
216 |
+
![Model Architecture](architecture.png)
|
217 |
|
218 |
+
### Components
|
219 |
|
220 |
+
- **Transformers Component**: Central to our model is the [transformers.Swin2SRModel](https://huggingface.co/docs/transformers/model_doc/swin2sr#transformers.Swin2SRModel). This component amplifies the spatial resolution of its inputs by a factor of 8. Notably, Swin2SR exclusively supports upscaling ratios that are powers of 2.
|
221 |
+
- **Convolutional Neural Network (CNN) Component**: Given that our actual upscale ratio is approximately 5 and the designated output shape is (160, 240),
|
222 |
+
we've integrated a CNN. This serves as a preprocessing unit, transforming inputs into (20, 30) feature maps suitable for the Swin2SRModel.
|
223 |
+
|
224 |
+
The underlying objective of this network is to master the residuals stemming from bicubic interpolation.
|
225 |
|
226 |
+
### Configuration Details
|
227 |
|
228 |
+
For those inclined towards the intricacies of the model, the specific parameters governing its behavior are meticulously detailed in the
|
229 |
+
[config.json](https://huggingface.co/predictia/convswin2sr_mediterranean/blob/main/config.json).
|
230 |
|
231 |
+
### Loss function
|
232 |
|
233 |
The Swin2 transformer optimizes its parameters using a composite loss function that aggregates multiple \( \mathcal{L}_1 \) loss terms to enhance its predictive
|
234 |
accuracy across different resolutions and representations:
|