Wrinkle/09_standalone_model/generate_samples.py

#!/usr/bin/env python3
"""Generate text samples from the trained WrinkleBrane model."""
import sys
sys.path.insert(0, 'src')
import torch
from wrinklebrane.standalone_model import WrinkleBraneModel
from wrinklebrane.data import encode_bytes, decode_tokens, BOS_ID

# Load best checkpoint
print("Loading best checkpoint...")
ckpt = torch.load('checkpoints/best_model.pt', weights_only=False, map_location='cpu')
config = ckpt['config']
model = WrinkleBraneModel(config)
model.load_state_dict(ckpt['model_state_dict'])
model.eval()

print(f"Model: {config.d_model}d, {config.n_layers}L, {config.n_heads}H")
print(f"Checkpoint step: {ckpt.get('step')}, val_loss: {ckpt.get('val_loss'):.4f}")
print()

def generate(prompt, max_tokens=200, temperature=0.8):
    tokens = [BOS_ID] + encode_bytes(prompt)
    input_ids = torch.tensor([tokens], dtype=torch.long)

    with torch.no_grad():
        logits, states = model.forward_sequential(input_ids)

    generated = list(tokens)
    current = generated[-1]

    for _ in range(max_tokens):
        inp = torch.tensor([[current]], dtype=torch.long)
        with torch.no_grad():
            logits, states = model.forward_sequential(inp, states)
        probs = torch.softmax(logits[0, -1] / temperature, dim=-1)
        current = torch.multinomial(probs, 1).item()
        generated.append(current)
        if current == 2:  # EOS
            break

    return decode_tokens(generated)

# Generate samples
prompts = [
    "Once upon a time",
    "The cat sat on",
    "2 + 3 =",
    "def hello",
    "The little bear",
]

for prompt in prompts:
    print(f"=== Prompt: '{prompt}' ===")
    sample = generate(prompt, max_tokens=150)
    print(sample[:400])
    print()