Update README.md
Browse files
README.md
CHANGED
@@ -190,14 +190,6 @@ its performance on the most recent data points, ensuring its relevance and relia
|
|
190 |
<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
|
191 |
|
192 |
|
193 |
-
|
194 |
-
|
195 |
-
### Metrics
|
196 |
-
|
197 |
-
<!-- These are the evaluation metrics being used, ideally with a description of why. -->
|
198 |
-
|
199 |
-
More information needed
|
200 |
-
|
201 |
## Results
|
202 |
|
203 |
In our evaluation, the proposed model displayed a significant enhancement over the established baseline, which employs bicubic interpolation for the same task.
|
@@ -233,25 +225,53 @@ This network is trained to learn the residuals of the bicubic interpolation.
|
|
233 |
|
234 |
The specific parameters of this network are available in [config.json](https://huggingface.co/predictia/convswin2sr_mediterranean/blob/main/config.json).
|
235 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
236 |
|
237 |
-
|
|
|
238 |
|
239 |
-
|
240 |
|
241 |
-
|
|
|
242 |
|
243 |
-
|
|
|
244 |
|
245 |
-
- **
|
|
|
|
|
246 |
|
247 |
-
|
|
|
|
|
248 |
|
249 |
-
|
250 |
|
|
|
|
|
|
|
251 |
|
252 |
-
### Software
|
253 |
|
254 |
-
|
|
|
255 |
|
256 |
|
257 |
### Authors
|
|
|
190 |
<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
|
191 |
|
192 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
193 |
## Results
|
194 |
|
195 |
In our evaluation, the proposed model displayed a significant enhancement over the established baseline, which employs bicubic interpolation for the same task.
|
|
|
225 |
|
226 |
The specific parameters of this network are available in [config.json](https://huggingface.co/predictia/convswin2sr_mediterranean/blob/main/config.json).
|
227 |
|
228 |
+
## Loss function
|
229 |
+
|
230 |
+
The Swin2 transformer optimizes its parameters using a composite loss function that aggregates multiple \( \mathcal{L}_1 \) loss terms to enhance its predictive
|
231 |
+
accuracy across different resolutions and representations:
|
232 |
+
|
233 |
+
1. **Primary Predictions Loss**:
|
234 |
+
- This term computes the \( \mathcal{L}_1 \) loss between the primary model predictions and the reference values. It ensures that the transformer's outputs
|
235 |
+
- closely match the ground truth across the primary spatial resolution.
|
236 |
+
|
237 |
+
2. **Downsampled Predictions Loss**:
|
238 |
+
- Recognizing the importance of accuracy across varying resolutions, this term calculates the \( \mathcal{L}_1 \) loss between the downsampled versions of the
|
239 |
+
- predictions and the reference values. By incorporating this term, the model is incentivized to preserve critical information even when the data is represented
|
240 |
+
- at a coarser scale.
|
241 |
+
|
242 |
+
3. **Blurred Predictions Loss**:
|
243 |
+
- To ensure the model's robustness against small perturbations and noise, this term evaluates the \( \mathcal{L}_1 \) loss between blurred versions of the
|
244 |
+
- predictions and the references. This encourages the model to produce predictions that maintain accuracy even under slight modifications in the data representation.
|
245 |
|
246 |
+
By combining these loss terms, the Swin2 transformer is trained to produce accurate predictions across different resolutions and under various data transformations,
|
247 |
+
ensuring its versatility and robustness in diverse scenarios.
|
248 |
|
249 |
+
## Technical Infrastructure
|
250 |
|
251 |
+
Leveraging GPUs in deep learning initiatives greatly amplifies the pace of model training and inference. This computational edge not only diminishes the total
|
252 |
+
computational duration but also equips us to proficiently navigate complex tasks and extensive datasets.
|
253 |
|
254 |
+
Our profound gratitude extends to our collaborative partners, whose invaluable contribution and support have been cornerstones in the fruition of this project.
|
255 |
+
Their substantial inputs have immensely propelled our research and developmental strides.
|
256 |
|
257 |
+
- **AI4EOSC**: Representing "Artificial Intelligence for the European Open Science Cloud," AI4EOSC functions under the aegis of the European Open Science Cloud (EOSC).
|
258 |
+
Initiated by the European Union, EOSC endeavors to orchestrate a cohesive platform for research data and services. AI4EOSC, a distinct arm within EOSC, concentrates
|
259 |
+
on embedding and leveraging artificial intelligence (AI) techniques within the open science domain.
|
260 |
|
261 |
+
- **European Weather Cloud**: Serving as a cloud-centric hub, this platform catalyzes collective efforts in meteorological application design and operations
|
262 |
+
throughout Europe. Its offerings are manifold, ranging from disseminating weather forecast data to proffering computational prowess, expert counsel, and
|
263 |
+
consistent support.
|
264 |
|
265 |
+
### Hardware Specifications
|
266 |
|
267 |
+
Our endeavor harnesses the capabilities of two virtual machines (VMs), each embedded with a dedicated GPU. One VM is bolstered with a 16GB GPU, while its counterpart
|
268 |
+
is equipped with an even potent 20GB GPU. This strategic hardware alignment proficiently caters to diverse computational needs, spanning data orchestration to model
|
269 |
+
fine-tuning and evaluation, ensuring the seamless flow and success of our project.
|
270 |
|
271 |
+
### Software Resources
|
272 |
|
273 |
+
For enthusiasts and researchers inclined towards a deeper probe, our model's training and evaluation code is transparently accessible.
|
274 |
+
Navigate to our GitHub Repository [ECMWFCode4Earth/DeepR](https://github.com/ECMWFCode4Earth/DeepR) under the ECWMF Code 4 Earth consortium.
|
275 |
|
276 |
|
277 |
### Authors
|