Update README.md

README.md

@@ -168,8 +168,8 @@ rectification, and grid interpolation. The methodology employed in each step is
 
 - Inference speed: The ConvSwin2SR model demonstrates a commendable inference speed, particularly when handling a substantial batch of samples.
   Specifically, when tasked with downscaling 248 samples, which is synonymous with processing data for an entire month at 3-hour intervals,
-  the model completes the operation in a mere 21 seconds. This level of efficiency is observed in a local computing environment outfitted with 16GB o
-  f RAM and 4GB of GPU memory. 
+  the model completes the operation in a mere 21 seconds. This level of efficiency is observed in a local computing environment outfitted with 16GB of
+  RAM and 4GB of GPU memory. 
  
 # Evaluation
 
@@ -233,17 +233,23 @@ The Swin2 transformer optimizes its parameters using a composite loss function t
 accuracy across different resolutions and representations:
 
 1. **Primary Predictions Loss**: 
-   - This term computes the L1 loss between the primary model predictions and the reference values. It ensures that the transformer's outputs
-     closely match the ground truth across the primary spatial resolution.
+   - This term computes the L1 loss between the primary model predictions and the reference values. It ensures that the transformer's
+     outputs closely match the ground truth.
 
 2. **Downsampled Predictions Loss**: 
-   - Recognizing the importance of accuracy across varying resolutions, this term calculates the L1 loss between the downsampled versions of the
-     predictions and the reference values. By incorporating this term, the model is incentivized to preserve critical information even when the data is represented
-     at a coarser scale.
+   - This term calculates the L1 loss between the downsampled versions of the predictions and the reference values. By incorporating this term,
+     the model is incentivized to preserve the underlying relations between both spatial resolutions. The references and predictions are upscaled
+     by average pooling by a factor of x5 to match the source resolution. Although  this loss term could be (technically) computed with respect
+     to the low-resolution sample, the upscaled reference values are considered, due to the fact that the average pooling used for upscaling does
+     not represent the true relationship between both datasets considered.
 
 3. **Blurred Predictions Loss**: 
    - To ensure the model's robustness against small perturbations and noise, this term evaluates the L1 loss between blurred versions of the
-     predictions and the references. This encourages the model to produce predictions that maintain accuracy even under slight modifications in the data representation.
+     predictions and the references. This encourages the model to produce predictions that maintain accuracy even under slight modifications
+     in the data representation. On the other hand, it can smooth the prediction field too much, so it is a term whose use should be studied
+     before including it in your model. To produce the blurred values, a gaussian kernel of size 5 is applied.
+
+By combining these loss terms, the ConvSwin2SR is trained to produce realistic predictions.
 
 ## Computing Infrastructure