The code for fMRI-Image Reconstruction.

[weipipione]

          Hi, we don't have plans to release the code for fMRI-Image Reconstruction. A quick way to implement this is to replace the brain encoder in MindEye with ours. The output dimensions should be modified accordingly, and the training loss should be changed to InfoNCE, as stated in the [paper](https://arxiv.org/pdf/2404.07202). For details, please refer to Sec 4.5 Brain Retrieval and Sec 4.6 Visual Decoding. Both experiments are intended to demonstrate the advantages of the UMBRAE architecture, not as our primary focus.

Originally posted by @weihaox in #6 (comment)

[weipipione]

Hi, i replace the brain encoder in MindEye with Umbare for fMRI-Image Reconstruction, but i can't reprduce the result. The training loss in MindEye is MSE and BiMixCo, when i use Umbare, i just use InfoNCE as the training loss ?

[zhibotian]

Hello@weihaox , In MindEye, the embedding length is 768 for both Brain Retrieval and Visual Decoding tasks. Does UMBRAE also use an embedding length of 768 for these tasks?