Skip to content

Latest commit

 

History

History
104 lines (76 loc) · 4.27 KB

useful_info.md

File metadata and controls

104 lines (76 loc) · 4.27 KB

GPU

nvidia-smi
ps aux | grep <process number>
kill -9 

import psutil
memory_info = psutil.virtual_memory()
print(f"Available memory: {memory_info.available / 1024 / 1024 / 1024} GB")
import sys
print(f"Model size: {sys.getsizeof(trained_modelstate)} bytes")


import torch
torch.cuda.set_per_process_memory_fraction(0.6, 0)
#torch.cuda.list_gpu_processes(0)
# torch.cuda.memory_allocated(0) / 1024 / 1024 / 1024
# total_memory = torch.cuda.get_device_properties(0).total_memory / 1024 / 1024 / 1024
# total_memory
total_memory = torch.cuda.get_device_properties(0).total_memory / 1024 / 1024 / 1024
total_memory*0.6
torch.cuda.list_gpu_processes(0)
torch.cuda.empty_cache()

import importlib importlib.reload(modelviz_utils)

Change env and parameters inside configs.py whenever doing training or inference - we're importing it in many files inside

Codes

  • Comparison of movement policies
  • predictions over entire plot - in 2D env
  • some atoms seem to converge predictions to single line or centre in the maze
  1. comparison between previous SR and DSM - 1 gamma model vs ensemble of gamma (train with config distributional False) vs DSM
  • changing a basis feature - turning off a PC (change activity to 0) - for all sources and generate rate map
  • 2% of world is changed - subtle changes to input can cause some PCs to change bt not others - previous SR models may not show that effect -
  • models break atoms - predict - variations in atom responses may mimic the brain
  • To test how much of the responses f an atom is a reslt of a change we made - keep same latents

  1. Generalisation changes train with a quarter of the maze not explored - and test with sources in unexplored env

  2. DONE ?? : how to evaluate - to get the directional responses of atoms - 1 source - pmake dsm produce 1 sample - get head direction this way - rayleigh vector thing Cell directionality Head direction cells signal the direction in which an animal's head is pointing- part of the brain's internal compass system. The Rayleigh vector length is a measure used in circular statistics- of the concentration of a set of directions.

HIGH Rayleigh vector length indicates that the directions are concentrated around a certain value, while a LOW Rayleigh vector length indicates that the directions are spread out.

  • Directionality of cells - collect all the recorder activities of a neuron and all the head directions then bin them. Strictly you also want to normalise each bin count by the number of times you sampled a particular head direction.

  1. DONE: Checkpoints - euclidean similarity with original PCs - normalize cuz outputs different range - - learning profiles - checkpoints

  2. Code: models trained on different sampling regimes - comparison random vs high Th

  • measure symmetry of rate maps - skew = width/height ------ do animals with high Th have skewed PC rate maps ?
  • isometric vs concentration of field - entropy of states - check Slack -
  • average of all PCs as a population measure
  • similarity to original PCs - spatial correlation - colum vectorn - pearson corr diff CHECK WHAT WAS SENT ON SLACK - SPARSITY FOR PLACE FIELD DEFINITION
  • rate map outputs not constrained 0-1 in thesis: Atoms as respresenting different parts of world in real brain? in dsm - unified training but change in env - compare mixture of how PCs change
  • relu instead of leaky_relu - models.py

Errors (In development of another dataset with different number of place cells)

ScopeParamShapeError: Initializer expected to generate shape (, 32) but got shape (, 32)

  • In train.py: change num_state_dims or -- replace math.prod(observation_spec.shape) everywhere with desired shape in train.py

wsl