File size: 934 Bytes
df6cf36 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 | import torch
from core.ledger import VitalisLedger
from core.nse.sync_manager import TripleHeadSyncManager
class NSETrainer:
def __init__(self, d_model=256):
self.ledger = VitalisLedger()
self.model = TripleHeadSyncManager(d_model)
self.optimizer = torch.optim.Adam(self.model.parameters(), lr=0.001)
def train_step(self, input_data, fe_stats):
self.optimizer.zero_grad()
# Consensus pass
recon, lr_mult = self.model(input_data, fe_stats)
# Loss calculation (e.g., reconstruction error)
loss = torch.mean((recon - input_data) ** 2)
loss.backward()
self.optimizer.step()
# Immutable Ledger Log
self.ledger.write_entry("training_step", {
"loss": loss.item(),
"lr_multiplier": lr_mult.item(),
"status": "verified_update"
})
return loss.item()
|