import torch import torch.nn as nn from torch.utils.data import DataLoader import argparse import math import os import json import matplotlib.pyplot as plt import seaborn as sns from tqdm import tqdm from .model import DecoderOnlyTransformer from .dataset import prepare_data, train_or_load_tokenizer, Translation_Dataset from .train import evaluate, compute_translation_bleu def evaluate_language_usage(model, pairs, tokenizer, seq_len = 128, device = "cuda"): model.reset_moe_usage() model.set_moe_tracking(True) dataset = Translation_Dataset(pairs, tokenizer, max_seq_len = seq_len) loader = DataLoader(dataset, batch_size = 32, shuffle = False) model.eval() with torch.no_grad(): for batch in loader: input_ids = batch["input_ids"].to(device) _ = model(input_ids) raw_counts = model.get_moe_usage() total_per_layer = raw_counts.sum(dim = -1, keepdim = True).float() normalized = (raw_counts.float() / (total_per_layer + 1e-9)).cpu().numpy() model.set_moe_tracking(False) model.reset_moe_usage() return normalized def plot_language_heatmap(vi_usage, ja_usage, variant, save_path): num_layers, num_experts = vi_usage.shape y_labels = [f"Layer {i + 1}" for i in range(num_layers)] if variant == 4: x_labels = [f"Routed {i}" for i in range(num_experts - 1)] + ["Shared"] else: x_labels = [f"Expert {i}" for i in range(num_experts)] vmax = max(vi_usage.max(), ja_usage.max()) fig, axes = plt.subplots(1, 2, figsize = (12,5), sharey = True) sns.heatmap(vi_usage, ax = axes[0], annot = True, fmt = ".2f", cmap = "Blues", vmin = 0.0, vmax = vmax, xticklabels = x_labels, yticklabels = y_labels, cbar = False) axes[0].set_title(f"Variant {variant}: Vietnamese Activations", fontsize = 12, fontweight = "bold") axes[0].set_xlabel("Experts") axes[0].set_ylabel("Transformer Layers") sns.heatmap(ja_usage, ax = axes[1], annot = True, fmt = ".2f", cmap = "Blues", vmin = 0.0, vmax = vmax, xticklabels = x_labels, yticklabels = y_labels, cbar = True, cbar_kws = {"label": "Activation Proportion"}) axes[1].set_title(f"Variant {variant}: Japanese Activations", fontsize = 12, fontweight = "bold") axes[1].set_xlabel("Experts") plt.suptitle(f"Language Specialization Heatmap (Variant {variant})", fontsize = 14, y = 1.02) plt.tight_layout() os.makedirs(os.path.dirname(save_path), exist_ok = True) plt.savefig(save_path, dpi = 300, bbox_inches = "tight") plt.close() print(f"Saved heatmap to: {save_path}") def run_evaluation(args): device = "cuda" if torch.cuda.is_available() else "cpu" os.makedirs(args.output_dir, exist_ok = True) print("Loading test dataset and tokenizer...") _, test_pairs, raw_ds = prepare_data(num_samples = args.num_samples, test_size = args.test_size) tokenizer = train_or_load_tokenizer(raw_ds, vocab_size = args.vocab_size) vi_test = [p for p in test_pairs if p[0] == "vi"] ja_test = [p for p in test_pairs if p[0] == "ja"] print(f"Test split: Total = {len(test_pairs)} | vi = {len(vi_test)} | ja = {len(ja_test)}") test_ds = Translation_Dataset(test_pairs, tokenizer, max_seq_len = args.seq_len) test_loader = DataLoader(test_ds, batch_size = args.batch_size, shuffle = False) results = {} for v in range(1, 6): checkpoint_path = os.path.join(args.save_dir, f"best_variant_{v}.pt") if not os.path.exists(checkpoint_path): print(f"Skipping variant {v}: {checkpoint_path} not found") continue print(f"\n{'=' * 55}\n Evaluating Variant {v}({checkpoint_path})\n{'=' * 55}") model = DecoderOnlyTransformer(vocab_size = len(tokenizer), d_model = args.d_model, n_heads = args.n_heads, d_ff = args.d_ff, n_layers = args.n_layers, variant = v, max_seq_len = args.seq_len).to(device) state_dict = torch.load(checkpoint_path, map_location = device) model.load_state_dict(state_dict) model.eval() print("Computing test loss and perplexity...") test_loss, test_ppl = evaluate(model, test_loader, device = device) print("Computing BLEU scores...") bleu_overall = compute_translation_bleu(model, test_pairs, tokenizer, max_samples = args.bleu_samples, device = device) bleu_vi = compute_translation_bleu(model, vi_test, tokenizer, max_samples = args.bleu_samples // 2, device = device) bleu_ja = compute_translation_bleu(model, ja_test, tokenizer, max_samples = args.bleu_samples // 2, device = device) print(f"Results -> Loss: {test_loss:.4f} | PPL: {test_ppl:.2f} | BLEU Total: {bleu_overall:.2f} (Vi: {bleu_vi:.2f}, Ja: {bleu_ja:.2f})") if v > 1: print("Analyzing expert routing patterns per language...") vi_usage = evaluate_language_usage(model, vi_test, tokenizer, seq_len = args.seq_len, device = device) ja_usage = evaluate_language_usage(model, ja_test, tokenizer, seq_len = args.seq_len, device = device) heatmap_path = os.path.join(args.output_dir, f"heatmap_variant_{v}.png") plot_language_heatmap(vi_usage, ja_usage, variant = v, save_path = heatmap_path) results[f"variant_{v}"] = { "test_loss": round(float(test_loss), 4), "test_perplexity": round(float(test_ppl), 2), "bleu_overall": round(float(bleu_overall), 2), "bleu_vi_en": round(float(bleu_vi), 2), "bleu_ja_en": round(float(bleu_ja), 2), } results_file = os.path.join(args.output_dir, "task1_results.json") with open(results_file, "w") as f: json.dump(results, f, indent = 4) print(f"\nAll results saved to: {results_file}") print("\n" + "=" * 65) print(f"{'Variant':<12} | {'PPL':<8} | {'BLEU (All)':<12} | {'BLEU (Vi)':<10} | {'BLEU (Ja)':<10}") print("-" * 65) for v_name, metrics in results.items(): print(f"{v_name:<12} | {metrics['test_perplexity']:<8.2f} | {metrics['bleu_overall']:<12.2f} | {metrics['bleu_vi_en']:<10.2f} | {metrics['bleu_ja_en']:<10.2f}") print("=" * 65) if __name__ == "__main__": parser = argparse.ArgumentParser() parser.add_argument("--save_dir", type=str, default="checkpoints_part1") parser.add_argument("--output_dir", type=str, default="assets/part1_results") parser.add_argument("--d_model", type=int, default=384) parser.add_argument("--n_heads", type=int, default=6) parser.add_argument("--d_ff", type=int, default=1536) parser.add_argument("--n_layers", type=int, default=6) parser.add_argument("--vocab_size", type=int, default=8192) parser.add_argument("--seq_len", type=int, default=128) parser.add_argument("--batch_size", type=int, default=32) parser.add_argument("--num_samples", type=int, default=50000) parser.add_argument("--test_size", type=int, default=2000) parser.add_argument("--bleu_samples", type=int, default=100) args = parser.parse_args() run_evaluation(args)