Resize images

README.md

@@ -5,14 +5,16 @@ gflownet is library built upon [PyTorch](https://pytorch.org/) to easily train a
 In a nutshell, GFlowNets can be regarded as a generative model designed to sample objects $x \in \mathcal{X}$ proportionally to a reward function $R(x)$. This results in the potential of sampling diverse objects with high rewards. For example, given the reward landscape depicted below, defined over a two-dimensional space, a well-trained GFlowNet will be able to sample from the four high-reward corners with high probability.
 
 <p align="center">
-  <img width="50" src="docs/images/reward_landscape.png" />
+  <img width="400" src="docs/images/reward_landscape.png" />
 </p>
 
 GFlowNets rely on the principle of **compositionality** to generate samples. A meaningful decomposition of samples $x$ into multiple intermediate states $s_0\rightarrow s_1 \rightarrow \dots \rightarrow x$ can yield generalisable patterns. These patterns can then be learned by neural networks trained to model the value of transitions $F_{\theta}(s_t \rightarrow s_{t+1})$.
 
 Consider the problem of generating [Tetris](https://en.wikipedia.org/wiki/Tetris)-like boards. A natural decomposition of the sample generation process would be to add one piece at a time, starting from an empty board. For any state representing a board with pieces, we could identify its valid parents and children, as illustrated in the figure below.
 
-![Tetris Environment](docs/images/tetris_flows.png)
+<p align="center">
+  <img width="400" src="docs/images/tetris_flows.png" />
+</p>
 
 We could define a reward function $R(x)$ as the number of cells occupied by pieces, for instance. The goal of training a GFlowNet on this task would be to discover (sample) diverse solutions (boards with pieces) with high rewards. This represents an intuitive yet complex problem where GFlowNets can be used. Many problems in scientific discoveries, such as the inverse design of proteins, molecules, or crystals share similarties with this intuitive task.