jens/pyRBM
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

[bugfix] - remove dead cd_jens; fix cd_gaussian_binary dbv; fix cd_binary_gaussian Gibbs sampling

Browse Source 
- Remove cd_jens: dead code never called, contained multiple bugs.
- cd_gaussian_binary: dbv formula simplified — b_v terms cancelled between
  positive and negative phases, making the subtraction redundant.
- cd_binary_gaussian: Gibbs loop (CD-k > 1) now properly samples both visible
  (sample(prob(...))) and hidden (+ sample_gaussian) states when
  do_rao_blackwell=False. Previously used means throughout, giving biased
  gradient estimates for k > 1.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
jens
2026-05-30 20:54:09 +02:00
co-authored by Claude Sonnet 4.6
parent a6bfdf8eaf
commit b5104ebc97
1 changed files with 7 additions and 44 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/rbm/train.py
+7 -44
View File
@@ -2,46 +2,6 @@ from .matrix import sample, sample_gaussian, prob, rms_error_accu, Mat, np
from .entity import Entity
from .status import Status
def cd_jens(entity: Entity, v_states: Mat):
params = entity.training_params
v_probs = prob(v_states)
h_states = entity.forward(v_states)
h_probs = h_states
if entity.params.do_gaussian_hidden:
h_states += sample_gaussian(h_states)
else:
h_probs = entity.forward(v_states)
if params.do_rao_blackwell:
h_states = h_probs
else:
h_states = sample(h_probs)
# Update weights (positive phase)
dw = np.dot(np.transpose(v_states), h_states)
dbv = np.sum(v_states, 0)
dbh = np.sum(h_states, 0)
# Gibbs sampling with training params
for i in range(params.num_gibbs_samples):
if params.do_gibbs_sample_hidden:
v_probs = entity.reconstruct(sample(h_probs))
else:
v_probs = entity.reconstruct(h_probs)
# Create hidden representation given v
if params.do_gibbs_sample_visible:
h_probs = entity.forward(sample(v_probs))
else:
h_probs = entity.forward(v_probs)
# Update weights (negative phase)
dw -= np.dot(np.transpose(v_probs), h_probs)
dbv -= np.sum(v_probs, 0)
dbh -= np.sum(h_probs, 0)
return dw, dbv, dbh
# %%%%%%%%% START POSITIVE PHASE %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
# data = batchdata(:,:,batch);
# poshidprobs = 1./(1 + exp(-data*vishid - repmat(hidbiases,numcases,1)));
@@ -99,7 +59,7 @@ def cd_gaussian_binary(entity: Entity, data_pos: Mat):
# Update weights (positive phase)
dw = np.dot(np.transpose(data_pos), h_probs_pos)
dbh = np.sum(h_probs_pos, 0)
dbv = np.sum(data_pos-entity.state.b_v, 0)
dbv = np.sum(data_pos, 0)
# Negative phase — sample binary hidden states to get a sharper fantasy particle
data_neg = entity.v_given_h(sample(h_probs_pos))
@@ -108,7 +68,7 @@ def cd_gaussian_binary(entity: Entity, data_pos: Mat):
# Update weights (negative phase)
dw -= np.dot(np.transpose(data_neg), h_probs_neg)
dbh -= np.sum(h_probs_neg, 0)
dbv -= np.sum(data_neg-entity.state.b_v, 0)
dbv -= np.sum(data_neg, 0)
return dw, dbv, dbh
@@ -169,10 +129,13 @@ def cd_binary_gaussian(entity: Entity, data_pos: Mat):
data_neg = prob(entity.v_given_h(h_probs_pos))
h_probs_neg = entity.h_given_v(data_neg)
# Gibbs sampling
# Gibbs sampling — sample both visible and hidden to continue the Markov chain
for _ in range(params.num_gibbs_samples-1):
data_neg = prob(entity.v_given_h(h_probs_neg))
v_probs = prob(entity.v_given_h(h_probs_neg))
data_neg = v_probs if params.do_rao_blackwell else sample(v_probs)
h_probs_neg = entity.h_given_v(data_neg)
if not params.do_rao_blackwell:
h_probs_neg = h_probs_neg + sample_gaussian(h_probs_neg)
# Update weights (negative phase)
dw -= np.dot(np.transpose(data_neg), h_probs_neg)