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

tune hyperparams and fix forward_step shift, re-enable training

Browse Source 
- Switch TEXT to "HALLO MAUSI! SUPER HASE!", H_SIZE to 64, lr to 0.1
- Extract v2char() helper from vc2char()
- Fix shift_left to use vocab_size() step instead of 1
- Re-enable model.train() and model.save() in main
- Clean up dead prediction loop code

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
jens
2026-06-05 17:47:30 +02:00
co-authored by Claude Sonnet 4.6
parent f7ed8563d7
commit 5ac4608ba8
1 changed files with 15 additions and 24 deletions
Download Patch File Download Diff File Expand all files Collapse all files
JayRnn.py
+15 -24
View File
@@ -24,16 +24,16 @@ from rbm.train import train
from rbm.status import Status
# ── Hyper-parameters ──────────────────────────────────────────────────────────
TEXT = "0123456789" # the character sequence to learn (matches diagram example)
TEXT = "HALLO MAUSI! SUPER HASE!" # the character sequence to learn (matches diagram example)
WIN = 3 # sliding-window width (= N in the diagram)
STRIDE = 1 # sliding-window step size
UNITS = 3
H_SIZE = 32 # hidden units per RBM cell
H_SIZE = 64 # hidden units per RBM cell
NUM_EPOCHS = 1000
NUM_ITERATIONS = 1
TRAIN_PARAMS = TrainingParams(
learning_rate = 0.04,
learning_rate = 0.1,
momentum = 0.5,
num_epochs = NUM_EPOCHS,
do_rao_blackwell = True,
@@ -76,6 +76,9 @@ def batch_delay(_batch: Mat, delay=0):
def vc2char(_vc: Mat):
_v, _ = split(_vc.reshape(1, WIN*vocab_size()+H_SIZE), H_SIZE, axis=1)
return v2char(_v)
def v2char(_v: Mat):
return vec2str(_v.reshape([WIN, vocab_size()]), axis=1)
class RnnModel(Model):
@@ -101,13 +104,14 @@ class RnnModel(Model):
def forward_step(self, _vc: Mat):
for unit in self.units:
# print(f"forward_step in : {unit.name}:{vc2char(_vc)}")
_c = unit.forward(_vc)
_vc = unit.reconstruct(_c)
print(f"forward_step out: {unit.name}:{vc2char(_vc)}")
_v, _ = split(_vc, H_SIZE, axis=1)
_v_cl = clamp(_v.reshape([WIN, vocab_size()]), axis=1).reshape([1, WIN * vocab_size()])
_v = shift_left(_v_cl, 1)
_v_cl = clamp(_v.reshape([WIN, vocab_size()]), axis=1)
_v = shift_left(_v_cl.reshape([1, WIN*vocab_size()]), vocab_size())
_vc = concat(_v, _c, axis=1)
print(f"forward_step: {unit.name}:{vc2char(_vc)}")
return _vc
def forward(self, x: Mat):
@@ -139,27 +143,14 @@ if __name__ == "__main__":
# context will be updated after training
c_train = np.zeros([batch.shape[0], H_SIZE])
vc_train = concat(batch, c_train, axis=1)
# model.train(vc_train, status=Status())
# model.save()
model.train(vc_train, status=Status())
model.save()
# test the model
seed_str = '1'
seed_padded = '^'*(WIN-len(seed_str)) + seed_str
seed_str = 'HA'
seed_padded = seed_str + '^'*(WIN-len(seed_str))
v_test = str2vec(seed_padded).reshape([1, WIN*vocab_size()])
c_test = np.zeros([1, H_SIZE])
vc_test = concat(v_test, c_test, axis=1)
for i in range(len(TEXT) + 5):
for i in range(len(TEXT)):
vc_test = model.forward_step(vc_test)
if 0:
for i in range(len(TEXT)+5):
v_mat = v_forward.reshape([WIN, vocab_size()])
print(f"Forward {i:02d}: {vec2str(v_mat, axis=1)}")
v_predict, context = model.forward_step(v_mat, context)
v_predict_mat = v_predict.reshape([WIN, vocab_size()])
v_predict_clamp = clamp(v_predict_mat, axis=1)
v_predict_str = vec2str(v_predict_clamp, axis=1)
text_predict += v_predict_str[WIN-1]
print(f"Predict {i:02d}: {v_predict_str}")
v_forward = shift_left(v_predict_clamp.reshape([1, WIN*vocab_size()]), vocab_size())
print(seed_padded[0:WIN-1] + text_predict)