mamba

[anhhuyalex]

downstream evals
try scaling, e.g. reconstruct full volume

[anhhuyalex]

to do: try varying the amount of input patches fed (from 25% of full to close to 100%) because we want the encoder to be robust to the number of patches fed as input

[anhhuyalex]

/weka/home-alexnguyen/mamba_fmri/ckpts/mar24_encoder32_decoder1_mamba64_lr3e-4/last.pth failed

[anhhuyalex]

• encoder->decoder (1M input embeddings are too much)
• unfreeze mae model during downstream

[anhhuyalex]

working runs:
apr22_encoder32_decoder32_equalseqlen_learnableposemb_encoder_outdim_512_lr4e-4

random tube mask:
apr19_encoder32_decoder1_randomtubemask_fixdecodemask_learnableposemb_encoder_outdim_512_lr4e-6