fix: code copy button (#20)

_posts/2023-05-16-functionality-checks.md

@@ -14,6 +14,7 @@ subclass: 'post'
 ---
 
 # Available Functionalities
+<div markdown="block" class="with-new-line">
 
 ```python
 
@@ -78,6 +79,7 @@ trainer = pl.Trainer(gpus=4, num_nodes=8, precision=16, limit_train_batches=0.5)
 trainer.fit(model, train_loader, val_loader)
 
 ```
+</div>
 
 # Latex Formulas
 

_posts/2023-05-17-welcome.md

@@ -35,8 +35,8 @@ Picture this: Within a mere five minutes, you'll have your first agent trained a
 
 <div class="two-columns-container code-flex" style="gap: 15px; margin:15px">
 	<div>
-		<div class="code-block" style="opacity: 1">
-			<code class="language-bash with-new-line" data-lang="bash" style="opacity: 1; color: #f8f8f2; height: 250px">
+		<div class="code-block with-new-line" style="opacity: 1">
+			<code class="language-bash" data-lang="bash" style="opacity: 1; color: #f8f8f2; height: 250px">
 git clone https://github.com/Eclectic-Sheep/sheeprl.git
 cd sheeprl
 python3.10 -m venv .venv

_posts/2023-07-06-dreamer_v2.md

@@ -93,6 +93,8 @@ Our PyTorch implementation aims to be a simple, scalable and well-documented rep
 
 As an example, the implementation of the *KL balancing* directly follows the equation above:
 
+<div markdown="block" class="with-new-line">
+
 ```python
 from torch.distributions import Independent, OneHotCategoricalStraightThrough
 
@@ -108,6 +110,8 @@ rhs = kl_divergence(
 kl_loss = alpha * lhs + (1 - alpha) * rhs
 ```
 
+</div>
+
 Do you want to know more about how we implemented Dreamer-V2? Check out [our implementation](https://github.com/Eclectic-Sheep/sheeprl/tree/main/sheeprl/algos/dreamer_v2){:target="_blank"}.
 
 ### References

_posts/2023-08-10-dreamer_v3.md

@@ -68,6 +68,8 @@ $$
 #### Uniform Mix
 To prevent spikes in the KL loss, the categorical distributions (the one for discrete actions and the one for the posteriors/priors) are parametrized as mixtures of $1\%$ uniform and $99\%$ neural network output. This avoid the distributions to become near deterministic. To implement the *uniform mix*, we applied the *uniform mix* function to the logits returned by the neural networks.
 
+<div markdown="block" class="with-new-line">>
+
 ```python
 import torch
 from torch import Tensor
@@ -86,6 +88,8 @@ def uniform_mix(self, logits: Tensor, unimix: float = 0.01) -> Tensor:
     return logits
 ```
 
+</div>
+
 #### Return regularizer for the policy
 The main difficulty in Dreamer-V2 *actor learning phase* is the choosing of the entropy regularizer, which heavily depends on the scale and the frequency of the rewards. To have a single entropy coefficient, it is necessary to normalize the returns using moving statistics. In particular, they found out that it is more convenient to scale down large rewards and not scale up small rewards, to avoid adding noise.
 

assets/js/main.js

@@ -49,7 +49,8 @@ $(document).ready(function () {
     });
 
     // Document Ctrl + C
-    const sources = document.querySelectorAll("code:not(.with-new-line)");
+    const sources = document.querySelectorAll(":not(.with-new-line) code");
+    const sources_new_line = document.querySelectorAll(".with-new-line code");
 
     sources.forEach(source => {
         source.addEventListener("copy", (event) => {
@@ -58,4 +59,12 @@ $(document).ready(function () {
             event.preventDefault();
         });
     });
+
+    sources_new_line.forEach(source => {
+        source.addEventListener("copy", (event) => {
+            const selection = document.getSelection();
+            event.clipboardData.setData("text/plain", selection.toString());
+            event.preventDefault();
+        });
+    });
 });