ae / ae.py
ChadlieOak's picture
Upload 22 files
c9e8126 verified
Raw History Blame Contribute Delete
20.8 kB
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.utils.data import Dataset, DataLoader
from torch.optim.lr_scheduler import CosineAnnealingWarmRestarts
import math
import random
import numpy as np
# Set random seed for reproducibility
torch.manual_seed(42)
np.random.seed(42)
random.seed(42)
# =============================================================================
# UPGRADE 1: Dynamic Fractal-Based Attention Mechanism
# =============================================================================
class FractalAttentionalResonance(nn.Module):
"""
UPGRADED: The fractal resonance bias is now DYNAMICALLY generated based
on the input sequence, making it content-aware.
"""
def __init__(self, dim, num_heads=4):
super().__init__()
self.dim = dim
self.num_heads = num_heads
self.head_dim = dim // num_heads
self.q_proj = nn.Linear(dim, dim)
self.k_proj = nn.Linear(dim, dim)
self.v_proj = nn.Linear(dim, dim)
self.out_proj = nn.Linear(dim, dim)
# --- UPGRADE 1 CHANGE ---
# Replace static nn.Parameter with a network to generate the bias
self.bias_generator = nn.Sequential(
nn.Linear(dim, dim // 2),
nn.ReLU(),
nn.Linear(dim // 2, self.num_heads * self.head_dim)
)
# --- END UPGRADE 1 CHANGE ---
def forward(self, x):
B, T, D = x.shape
# --- UPGRADE 1 CHANGE ---
# Generate a sequence-specific dynamic bias
global_context = x.mean(dim=1) # (B, D)
dynamic_bias = self.bias_generator(global_context).view(B, self.num_heads, self.head_dim)
# --- END UPGRADE 1 CHANGE ---
Q = self.q_proj(x).view(B, T, self.num_heads, self.head_dim).transpose(1, 2)
K = self.k_proj(x).view(B, T, self.num_heads, self.head_dim).transpose(1, 2)
V = self.v_proj(x).view(B, T, self.num_heads, self.head_dim).transpose(1, 2)
scores = torch.matmul(Q, K.transpose(-2, -1)) / np.sqrt(self.head_dim)
# Apply the DYNAMIC Fractal Attentional Resonance
bias = dynamic_bias.unsqueeze(-1) # Shape: (B, num_heads, head_dim, 1)
fractal_resonance = torch.matmul(Q, bias) # Result shape: (B, num_heads, T, 1)
scores = scores + fractal_resonance # Broadcasting applies the bias
attn_weights = F.softmax(scores, dim=-1)
context = torch.matmul(attn_weights, V)
context = context.transpose(1, 2).contiguous().view(B, T, D)
return self.out_proj(context)
# All other modules from the original code remain the same...
class FAR_TransformerEncoderLayer(nn.Module):
def __init__(self, d_model, nhead, dim_feedforward, dropout=0.1, activation='gelu'):
super().__init__()
self.self_attn = FractalAttentionalResonance(d_model, nhead) # Uses the upgraded FAR
self.linear1 = nn.Linear(d_model, dim_feedforward)
self.dropout = nn.Dropout(dropout)
self.linear2 = nn.Linear(dim_feedforward, d_model)
self.norm1 = nn.LayerNorm(d_model)
self.norm2 = nn.LayerNorm(d_model)
self.dropout1 = nn.Dropout(dropout)
self.dropout2 = nn.Dropout(dropout)
self.activation = F.gelu if activation == 'gelu' else F.relu
def forward(self, src):
src2 = self.self_attn(src)
src = src + self.dropout1(src2)
src = self.norm1(src)
src2 = self.linear2(self.dropout(self.activation(self.linear1(src))))
src = src + self.dropout2(src2)
src = self.norm2(src)
return src
class CustomTransformerEncoder(nn.Module):
def __init__(self, encoder_layer, num_layers):
super().__init__()
self.layers = nn.ModuleList([encoder_layer for _ in range(num_layers)])
self.num_layers = num_layers
def forward(self, src):
output = src
for mod in self.layers:
output = mod(output)
return output
# All other helper modules (FractalEmbeddingLayer, DynamicGating, etc.) are identical
# to the original script and are included here for completeness.
class FractalEmbeddingLayer(nn.Module):
def __init__(self, dim, num_fractals=4):
super().__init__()
self.num_fractals = num_fractals
self.dim = dim
self.dims = nn.Parameter(torch.rand(num_fractals) * 2 + 1)
self.weight_generator = nn.Sequential(
nn.Linear(dim, dim // 2),
nn.ReLU(),
nn.Linear(dim // 2, num_fractals)
)
self.fractal_functions = [
lambda x: torch.sin(x * 2 * math.pi),
lambda x: x - torch.floor(x),
lambda x: 4 * x * (1 - x),
lambda x: torch.sigmoid(5 * (x - 0.5))
]
def forward(self, x):
x_safe = torch.sigmoid(x)
p_logits = self.weight_generator(x.mean(dim=1))
p_weights = F.softmax(p_logits, dim=-1).unsqueeze(1).unsqueeze(-1)
fractal_outputs = []
for i in range(self.num_fractals):
d = self.dims[i]
x_powered = torch.pow(x_safe, 1.0 / d)
f_i = self.fractal_functions[i]
output = f_i(x_powered)
fractal_outputs.append(output.unsqueeze(-2))
fractal_stack = torch.cat(fractal_outputs, dim=-2)
weighted_fractals = p_weights * fractal_stack
pfaf_embedding = torch.sum(weighted_fractals, dim=-2)
return x + pfaf_embedding
class DynamicGating(nn.Module):
def __init__(self, dim, num_experts=4):
super().__init__()
self.gate_network = nn.Sequential(nn.Linear(dim, num_experts), nn.Softmax(dim=-1))
self.expert_networks = nn.ModuleList([nn.Linear(dim, dim) for _ in range(num_experts)])
def forward(self, x):
gates = self.gate_network(x)
expert_outputs = [expert(x) for expert in self.expert_networks]
expert_stack = torch.stack(expert_outputs, dim=-1)
gates = gates.unsqueeze(-2)
return torch.sum(expert_stack * gates, dim=-1)
class ContinuousResonanceField(nn.Module):
def __init__(self, dim):
super().__init__()
self.field_mlp = nn.Sequential(nn.Linear(dim, dim), nn.GELU(), nn.LayerNorm(dim))
self.field_attention = nn.MultiheadAttention(dim, num_heads=4, batch_first=True)
def forward(self, query, context):
field_context = self.field_mlp(context)
attended_field, _ = self.field_attention(query, field_context, field_context)
return attended_field
class QuantumInspiredFusion(nn.Module):
def __init__(self, dim, num_qubits=4):
super().__init__()
self.num_qubits = num_qubits
self.state_prep_know = nn.Linear(dim, num_qubits * 2)
self.state_prep_exec = nn.Linear(dim, num_qubits * 2)
self.entanglement_gate = nn.Parameter(torch.randn(num_qubits, num_qubits))
self.measurement = nn.Linear(num_qubits * 2, dim)
self.norm = nn.LayerNorm(dim)
def forward(self, know, exec):
B, T, _ = know.shape
know_state = self.state_prep_know(know).view(B, T, self.num_qubits, 2)
exec_state = self.state_prep_exec(exec).view(B, T, self.num_qubits, 2)
k_real, k_imag = know_state[..., 0], know_state[..., 1]
e_real, e_imag = exec_state[..., 0], exec_state[..., 1]
ent_real = k_real * e_real - k_imag * e_imag
ent_imag = k_real * e_imag + k_imag * e_real
ent_real = torch.matmul(ent_real, self.entanglement_gate)
ent_imag = torch.matmul(ent_imag, self.entanglement_gate)
state = torch.stack([ent_real, ent_imag], dim=-1).reshape(B, T, -1)
return self.norm(self.measurement(state))
class HierarchicalMemorySystem(nn.Module):
def __init__(self, dim, scales=[32, 16, 8]):
super().__init__()
self.memory_banks = nn.ParameterList([nn.Parameter(torch.randn(1, s, dim)) for s in scales])
self.projs = nn.ModuleList([nn.Linear(dim, dim) for _ in range(len(scales) * 3)])
self.gates = nn.ModuleList([nn.Linear(dim * 2, dim) for _ in scales])
self.scale_attention = nn.MultiheadAttention(dim, num_heads=4, batch_first=True)
self.consolidation = nn.GRU(dim, dim, batch_first=True)
def forward(self, x):
B, T, D = x.shape
scale_outputs = []
for i, memory in enumerate(self.memory_banks):
mem = memory.expand(B, -1, -1)
query, key, value = self.projs[i*3](x), self.projs[i*3+1](mem), self.projs[i*3+2](mem)
scores = torch.bmm(query, key.transpose(1, 2)) / math.sqrt(D)
attended_mem = torch.bmm(F.softmax(scores, dim=-1), value)
gate = torch.sigmoid(self.gates[i](torch.cat([x, attended_mem], dim=-1)))
scale_outputs.append(x * (1 - gate) + attended_mem * gate)
fused_scales, _ = self.scale_attention(x, torch.cat(scale_outputs, dim=1), torch.cat(scale_outputs, dim=1))
consolidated, _ = self.consolidation(fused_scales)
return consolidated
class NeuralODE(nn.Module):
def __init__(self, dim, num_steps=4):
super().__init__()
self.num_steps = num_steps
self.ode_net = nn.Sequential(nn.Linear(dim, dim * 2), nn.Tanh(), nn.Linear(dim * 2, dim))
def forward(self, x):
state = x
dt = 1.0 / self.num_steps
for _ in range(self.num_steps):
state = state + dt * self.ode_net(state)
return state
# =============================================================================
# ZRIA-Ultimate Main Model with UPGRADES
# =============================================================================
class ZRIA_Ultimate_PFAF(nn.Module):
def __init__(self, dim, vocab_size, action_dim, max_seq_len=128):
super().__init__()
self.token_embedding = nn.Embedding(vocab_size, dim)
self.positional_embedding = nn.Parameter(torch.randn(1, max_seq_len, dim))
self.modality_embedding = nn.Parameter(torch.randn(2, dim))
self.fractal_embedding = FractalEmbeddingLayer(dim)
# Using the upgraded FAR encoder
far_encoder_layer = FAR_TransformerEncoderLayer(d_model=dim, nhead=4, dim_feedforward=dim * 2)
self.knowledge_encoder = CustomTransformerEncoder(far_encoder_layer, num_layers=2)
self.execution_encoder = CustomTransformerEncoder(far_encoder_layer, num_layers=2)
# --- UPGRADE 2 CHANGE ---
# Add cross-attention modules for path communication
self.know_to_exec_cross_attn = nn.MultiheadAttention(dim, num_heads=4, batch_first=True)
self.exec_to_know_cross_attn = nn.MultiheadAttention(dim, num_heads=4, batch_first=True)
self.cross_norm_k = nn.LayerNorm(dim)
self.cross_norm_e = nn.LayerNorm(dim)
# --- END UPGRADE 2 CHANGE ---
self.dynamic_gating = DynamicGating(dim)
self.resonance_field = ContinuousResonanceField(dim)
self.quantum_fusion = QuantumInspiredFusion(dim)
self.hierarchical_memory = HierarchicalMemorySystem(dim)
self.neural_ode = NeuralODE(dim)
self.fusion_norm_1 = nn.LayerNorm(dim)
self.fusion_norm_2 = nn.LayerNorm(dim)
self.output_projection = nn.Linear(dim, dim)
self.output_router = nn.Linear(dim, 2)
self.text_head = nn.Linear(dim, 3)
self.action_head = nn.Sequential(nn.Linear(dim, dim // 2), nn.GELU(), nn.Linear(dim // 2, action_dim))
def forward(self, input_ids, modality_idx, mode='auto'):
B, T = input_ids.shape
x = self.token_embedding(input_ids) + self.positional_embedding[:, :T, :] + self.modality_embedding[modality_idx].unsqueeze(1)
pfaf_x = self.fractal_embedding(x)
know_path = self.knowledge_encoder(pfaf_x)
exec_path = self.execution_encoder(pfaf_x)
# --- UPGRADE 2 CHANGE ---
# Allow paths to communicate via cross-attention
exec_from_know, _ = self.know_to_exec_cross_attn(query=exec_path, key=know_path, value=know_path)
exec_path = self.cross_norm_e(exec_path + exec_from_know)
know_from_exec, _ = self.exec_to_know_cross_attn(query=know_path, key=exec_path, value=exec_path)
know_path = self.cross_norm_k(know_path + know_from_exec)
# --- END UPGRADE 2 CHANGE ---
base = know_path + exec_path
fused = self.dynamic_gating(base) + base
fused = self.resonance_field(fused, base) + fused
fused = self.quantum_fusion(fused, base) + fused
fused = self.fusion_norm_1(fused)
mem = self.hierarchical_memory(fused) + fused
ode = self.neural_ode(mem) + mem
final_repr = self.output_projection(self.fusion_norm_2(ode))[:, -1, :]
if mode == 'auto':
route_probs = F.softmax(self.output_router(final_repr), dim=-1)
return {'text_logits': self.text_head(final_repr), 'action_output': self.action_head(final_repr), 'route_probs': route_probs}
elif mode == 'text':
return self.text_head(final_repr)
else:
return self.action_head(final_repr)
# =============================================================================
# Data Handling, Training, and Inference (Unchanged)
# =============================================================================
class AdvancedTokenizer:
def __init__(self, corpus):
special_tokens = ['<PAD>', '<UNK>', '<KNOW>', '<DO>']
all_chars = sorted(list(set("".join(corpus))))
self.char_to_idx = {t: i for i, t in enumerate(special_tokens)}
for char in all_chars:
if char not in self.char_to_idx: self.char_to_idx[char] = len(self.char_to_idx)
self.vocab_size = len(self.char_to_idx)
self.pad_idx = self.char_to_idx['<PAD>']
def encode(self, text, prefix):
tokens = [self.char_to_idx['<KNOW>' if prefix == 'KNOW' else '<DO>']]
for char in text.lower(): tokens.append(self.char_to_idx.get(char, self.char_to_idx['<UNK>']))
return tokens
class EnhancedDataset(Dataset):
def __init__(self, data, tokenizer, max_len=64):
self.data, self.tokenizer, self.max_len = data, tokenizer, max_len
def __len__(self): return len(self.data)
def __getitem__(self, idx):
prefix, text, label = self.data[idx]
tokens = self.tokenizer.encode(text, prefix)
tokens += [self.tokenizer.pad_idx] * (self.max_len - len(tokens))
return {'input_ids': torch.tensor(tokens[:self.max_len], dtype=torch.long),
'label': torch.tensor(label, dtype=torch.float32),
'mode': 'text' if prefix == 'KNOW' else 'action',
'modality_idx': torch.tensor(0 if prefix == 'KNOW' else 1, dtype=torch.long)}
def advanced_train_loop(model, dataloader, optimizer, scheduler, epochs=150, device='cpu'):
model.to(device)
model.train()
text_crit, action_crit, route_crit = nn.CrossEntropyLoss(), nn.MSELoss(), nn.CrossEntropyLoss()
print("=== Starting ZRIA Training (Upgraded Architecture) ===")
for epoch in range(epochs):
total_loss = 0
for batch in dataloader:
ids, labels, mod_idx = batch['input_ids'].to(device), batch['label'].to(device), batch['modality_idx'].to(device)
optimizer.zero_grad()
outputs = model(ids, mod_idx, mode='auto')
text_mask = torch.tensor([m == 'text' for m in batch['mode']], device=device)
action_mask = ~text_mask
loss = 0
if text_mask.any():
loss += text_crit(outputs['text_logits'][text_mask], labels[text_mask].long())
loss += 0.1 * route_crit(outputs['route_probs'][text_mask], torch.zeros(text_mask.sum(), dtype=torch.long, device=device))
if action_mask.any():
loss += 0.1 * action_crit(outputs['action_output'][action_mask].squeeze(-1), labels[action_mask])
loss += 0.1 * route_crit(outputs['route_probs'][action_mask], torch.ones(action_mask.sum(), dtype=torch.long, device=device))
if isinstance(loss, torch.Tensor):
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
optimizer.step()
total_loss += loss.item()
scheduler.step()
if (epoch + 1) % 10 == 0 or epoch == epochs - 1:
print(f"Epoch {epoch+1}/{epochs} | Avg Loss: {total_loss/len(dataloader):.4f} | LR: {scheduler.get_last_lr()[0]:.6f}")
print("=== Training Complete ===")
def enhanced_inference(model, tokenizer, test_cases, device='cpu'):
model.to(device).eval()
print("\n=== Enhanced Inference Results (Upgraded Architecture) ===")
with torch.no_grad():
for prefix, text, expected in test_cases:
tokens = tokenizer.encode(text, prefix)
tokens += [tokenizer.pad_idx] * (64 - len(tokens))
ids = torch.tensor([tokens], dtype=torch.long).to(device)
mod_idx = torch.tensor([0 if prefix == 'KNOW' else 1], dtype=torch.long).to(device)
outputs = model(ids, mod_idx, mode='auto')
probs = outputs['route_probs'][0]
print(f"\nInput: '{prefix}: {text}' | Expected: {expected}")
print(f"Routing -> Text: {probs[0]:.3f}, Action: {probs[1]:.3f}")
if probs[0] > probs[1]:
pred = torch.argmax(outputs['text_logits'][0]).item()
print(f"Prediction: Routed to KNOWLEDGE -> {['Negative', 'Neutral', 'Positive'][pred]}")
else:
pred = outputs['action_output'][0].item()
print(f"Prediction: Routed to EXECUTION -> Vowel Count: {pred:.2f}")
# =============================================================================
# Main Execution Block
# =============================================================================
if __name__ == '__main__':
device = 'cuda' if torch.cuda.is_available() else 'cpu'
print(f"Using device: {device}")
enhanced_data = [
('KNOW', 'i absolutely love this amazing product', 2), ('KNOW', 'this is the best thing ever created', 2),
('KNOW', 'wonderful experience with great service', 2), ('KNOW', 'a truly fantastic and brilliant idea', 2),
('KNOW', 'excellent work, very impressive', 2), ('KNOW', 'i am so happy with the results', 2),
('KNOW', 'i hate this terrible awful thing', 0), ('KNOW', 'horrible experience with bad service', 0),
('KNOW', 'a disappointing and frustrating product', 0), ('KNOW', 'this is the worst i have ever seen', 0),
('KNOW', 'i am very unhappy and angry', 0), ('KNOW', 'a complete failure and a waste of time', 0),
('KNOW', 'the meeting is scheduled for tomorrow', 1), ('KNOW', 'this is a factual statement', 1),
('KNOW', 'the document is on the table', 1), ('KNOW', 'standard operating procedure', 1),
('DO', 'how many vowels in this sentence', 8), ('DO', 'count all the vowels here please', 9),
('DO', 'the quick brown fox jumps over the lazy dog', 11), ('DO', 'artificial intelligence', 7),
('DO', 'rhythm and blues', 3), ('DO', 'fly by', 1), ('DO', 'aeiou', 5),
] * 10
random.shuffle(enhanced_data)
corpus = [text for _, text, _ in enhanced_data]
tokenizer = AdvancedTokenizer(corpus)
dataset = EnhancedDataset(enhanced_data, tokenizer, max_len=64)
dataloader = DataLoader(dataset, batch_size=16, shuffle=True)
model = ZRIA_Ultimate_PFAF(dim=128, vocab_size=tokenizer.vocab_size, action_dim=1, max_seq_len=64)
total_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
print(f"Model Architecture: ZRIA-Ultimate (Upgraded)")
print(f"Trainable parameters: {total_params:,} (~{total_params * 4 / 1e6:.1f}MB)")
optimizer = optim.AdamW(model.parameters(), lr=3e-4, weight_decay=0.01)
scheduler = CosineAnnealingWarmRestarts(optimizer, T_0=30, T_mult=2, eta_min=1e-6)
advanced_train_loop(model, dataloader, optimizer, scheduler, epochs=150, device=device)
test_cases = [
("KNOW", "this is a wonderful experience", "Positive"),
("KNOW", "i hate this terrible product", "Negative"),
("KNOW", "the report is due on friday", "Neutral"),
("DO", "count the vowels in this example", 8),
("DO", "a quick test", 3),
("DO", "why", 0),
]
enhanced_inference(model, tokenizer, test_cases, device=device)