"""Distillation training v3 — multi-corpus large-scale run (>=10x v2 data).

Cache: /tmp/work/teacher_cache_v3/shard_*.pt (precomputed in parallel on 2x T4).
Recipe: same losses/adapter as v2; 4000 AdamW steps (batch 8 -> ~32k samples seen).
Held-out quality gate stays on RAVDESS actors 19-24 (never in any training pool).
"""

import glob
import math
import os
import random
import sys
import time
import warnings

warnings.filterwarnings("ignore")
import logging

logging.disable(logging.CRITICAL)

import numpy as np
import torch
import torch.nn.functional as F

sys.path.insert(0, "/tmp/work/s2st_emotion")
from surgical_emotion import SurgicalConfig, SurgicalEmotionEncoder

FUNASR_DIR = "/tmp/work/emotion2vec_plus_base"
CACHE_DIR = "/tmp/work/teacher_cache_v3"
OUT_DIR = "/tmp/work/surgical_out"
BATCH = 8
LR = 3e-4
STEPS = 4000
WARMUP = 200
SEED = 42
VAL_FRAC = 0.05
FP16 = True
DEV = "cuda:0"


def load_cache():
    shards = sorted(glob.glob(os.path.join(CACHE_DIR, "shard_*.pt")))
    assert shards, "no teacher shards found"
    cache = []
    for s in shards:
        cache.extend(torch.load(s, map_location="cpu", weights_only=False))
    return cache


def make_batch(pool):
    items = random.sample(pool, min(BATCH, len(pool)))
    T = max(it["states"].size(0) for it in items)
    B = len(items)
    states = torch.zeros(B, T, 768)
    probs = torch.zeros(B, T, 9)
    inten = torch.zeros(B, T, 1)
    lengths = torch.zeros(B, dtype=torch.long)
    for j, it in enumerate(items):
        t = it["states"].size(0)
        states[j, :t] = it["states"].float()
        probs[j, :t] = it["probs"].float()
        inten[j, :t] = it["inten"].float()
        lengths[j] = t
    return states.to(DEV), probs.to(DEV), inten.to(DEV), lengths.to(DEV)


def compute_loss(enc, batch):
    states, tprobs, tinten, lengths = batch
    out = enc.adapter(states, lengths)
    embs, flen = out["frame_embs"], out["frame_lengths"]
    T = states.size(1)
    T_out = embs.size(1)

    up = F.interpolate(embs.transpose(1, 2), size=T, mode="nearest").transpose(1, 2)
    m = (torch.arange(T, device=states.device)[None, :] < lengths[:, None]).unsqueeze(-1)
    cos = F.cosine_similarity(up, states, dim=-1).unsqueeze(-1)
    l_frame = ((1 - cos) * m).sum() / m.sum().clamp(min=1)
    l_mse = F.mse_loss(F.layer_norm(up, (768,)), F.layer_norm(states, (768,)), reduction="none")
    l_mse = (l_mse * m).sum() / m.sum().clamp(min=1)

    om = (torch.arange(T_out, device=states.device)[None, :] < flen[:, None]).unsqueeze(-1)
    utt_s = (embs * om).sum(1) / flen.unsqueeze(-1).float()
    utt_probs_s = torch.softmax(enc.adapter.emotion_head(utt_s.float()), dim=-1)
    utt_t = (states * m).sum(1) / lengths.unsqueeze(-1).float()
    l_utt_cos = (1 - F.cosine_similarity(utt_s, utt_t, dim=-1)).mean()
    utt_probs_t = torch.softmax(enc.teacher.proj(utt_t), dim=-1)
    l_utt_kl = F.kl_div(torch.log(utt_probs_s.clamp(min=1e-8)), utt_probs_t, reduction="batchmean")

    tprobs_g = F.interpolate(tprobs.transpose(1, 2), size=T_out, mode="area").transpose(1, 2)
    tprobs_g = tprobs_g / tprobs_g.sum(-1, keepdim=True).clamp(min=1e-8)
    l_frm = F.kl_div(torch.log(out["frame_probs"].clamp(min=1e-8)), tprobs_g,
                     reduction="none").sum(-1)
    l_frm = (l_frm * om.squeeze(-1)).sum() / om.sum().clamp(min=1)

    tinten_g = F.interpolate(tinten.transpose(1, 2), size=T_out, mode="area").transpose(1, 2)
    l_int = (((out["frame_intensity"] - tinten_g) ** 2) * om).sum() / om.sum().clamp(min=1)

    loss = (1.0 * l_frame + 0.5 * l_mse + 0.5 * l_utt_cos
            + 1.0 * l_utt_kl + 1.0 * l_frm + 0.5 * l_int)
    return loss, {
        "frame": l_frame.item(), "mse": l_mse.item(), "utt_cos": l_utt_cos.item(),
        "utt_kl": l_utt_kl.item(), "frm_kl": l_frm.item(), "inten": l_int.item(),
    }


def main():
    random.seed(SEED); np.random.seed(SEED); torch.manual_seed(SEED)
    os.makedirs(OUT_DIR, exist_ok=True)
    log = open(os.path.join(OUT_DIR, "train_v3.log"), "a", buffering=1)

    def P(*a):
        print(*a, flush=True); print(*a, file=log)

    P(f"=== distill v3 start {time.strftime('%H:%M:%S')} device={DEV} ===")

    enc = SurgicalEmotionEncoder.from_pretrained(FUNASR_DIR, device=DEV)
    adapter = enc.adapter
    opt = torch.optim.AdamW(adapter.parameters(), lr=LR, weight_decay=0.01)
    sched = torch.optim.lr_scheduler.LambdaLR(
        opt, lambda s: min((s + 1) / WARMUP, 0.5 * (1 + math.cos(math.pi * s / STEPS))))
    scaler = torch.amp.GradScaler("cuda", enabled=FP16)

    cache = load_cache()
    P(f"teacher cache: {len(cache)} utts from {len(glob.glob(CACHE_DIR + '/shard_*.pt'))} shards")
    random.shuffle(cache)
    n_val = max(16, int(len(cache) * VAL_FRAC))
    val, train = cache[:n_val], cache[n_val:]
    P(f"train={len(train)}  val={len(val)}")

    best_val = float("inf")
    for step in range(STEPS):
        adapter.train()
        with torch.autocast("cuda", dtype=torch.float16, enabled=FP16):
            loss, parts = compute_loss(enc, make_batch(train))
        opt.zero_grad(set_to_none=True)
        scaler.scale(loss).backward()
        scaler.unscale_(opt)
        torch.nn.utils.clip_grad_norm_(adapter.parameters(), 1.0)
        scaler.step(opt); scaler.update(); sched.step()
        if step % 100 == 0:
            P(f"step {step:4d}  loss {loss.item():.4f}  " +
              " ".join(f"{k}={v:.3f}" for k, v in parts.items()))
        if step % 400 == 0 or step == STEPS - 1:
            adapter.eval()
            vlosses = []
            with torch.no_grad(), torch.autocast("cuda", dtype=torch.float16, enabled=FP16):
                for _ in range(4):
                    _, vp = compute_loss(enc, make_batch(val))
                    vlosses.append(sum(vp.values()))
            vloss = float(np.mean(vlosses))
            P(f"  >> val step {step}: {vloss:.4f}")
            if vloss < best_val:
                best_val = vloss
                enc.save_surgical(os.path.join(OUT_DIR, "surgical_emotion_module.pt"))
                P("  >> saved best")

    enc.save_surgical(os.path.join(OUT_DIR, "surgical_emotion_module_last.pt"))
    P(f"=== done, best_val={best_val:.4f} ===")


if __name__ == "__main__":
    main()