"""Reusable exact two-detector/two-view scorer for regeneration candidates."""
import json
import os
from pathlib import Path
import subprocess
import numpy as np
import torch
from transformers import CLIPImageProcessor,CLIPVisionModelWithProjection

class Scorer:
    def __init__(self,device='cuda:0'):
        from check_minor_safety import MODELS,STATE,ROOT
        self.root=MODELS;self.device=device
        r=json.loads((MODELS/'clip-model.json').read_text())
        self.model=CLIPVisionModelWithProjection.from_pretrained(r['path'],local_files_only=True,dtype=torch.float32).to(device).eval()
        self.processor=CLIPImageProcessor.from_pretrained(r['path'],local_files_only=True)
        env=os.environ.copy();env.update(CUDA_VISIBLE_DEVICES='',TF_CPP_MIN_LOG_LEVEL='2')
        self.log=(STATE/'logs/regeneration-tensorflow.log').open('a')
        self.worker=subprocess.Popen([str(MODELS/'tf-venv/bin/python'),str(ROOT/'scripts/minor_safety_tf.py')],stdin=subprocess.PIPE,stdout=subprocess.PIPE,stderr=self.log,text=True,env=env)
        assert json.loads(self.worker.stdout.readline())['status']=='ready'
        self.input=MODELS/f'candidate-{os.getpid()}-input.npz';self.output=MODELS/f'candidate-{os.getpid()}-output.npz'

    def score(self,images):
        from check_minor_safety import foreground
        crops=[foreground(im) for im in images]
        pixels=self.processor(images=images+crops,return_tensors='pt')['pixel_values'].to(self.device)
        prior_tf32=torch.backends.cuda.matmul.allow_tf32
        torch.backends.cuda.matmul.allow_tf32=False
        try:
            with torch.inference_mode():
                e=self.model(pixel_values=pixels).image_embeds;e=e/e.norm(dim=-1,keepdim=True)
        finally:torch.backends.cuda.matmul.allow_tf32=prior_tf32
        np.savez(self.input,embeddings=e.cpu().numpy().astype('float64'),original=np.stack([np.asarray(im) for im in images]),foreground=np.stack([np.asarray(im) for im in crops]))
        self.worker.stdin.write(json.dumps(dict(input=str(self.input),output=str(self.output)))+'\n');self.worker.stdin.flush()
        response=self.worker.stdout.readline();assert response,'TensorFlow candidate scorer exited'
        assert json.loads(response)['status']=='scored';r=np.load(self.output);a=r['clip'];b=r['efficientnet']
        assert np.isfinite(a).all() and np.isfinite(b).all() and 0<=a.min()<=a.max()<=1
        results=[];names=['drawing','hentai','neutral','porn','sexy']
        for i in range(len(images)):
            views={name:dict(clip_nsfw=float(a[j]),efficientnet_nsfw=float(b[j,[1,3,4]].sum()),efficientnet_classes={k:float(v) for k,v in zip(names,b[j])}) for j,name in [(i,'original'),(i+len(images),'foreground')]}
            results.append(dict(views=views,max_score=max(v[k] for v in views.values() for k in ['clip_nsfw','efficientnet_nsfw'])))
        self.input.unlink();self.output.unlink();return results

    def close(self):
        self.worker.stdin.write('{"stop":true}\n');self.worker.stdin.flush();self.worker.wait(timeout=30);self.log.close()