#!/usr/bin/env python3
"""Candidate 008, frozen five-offset perturbation diagnostic (no model fitting)."""
from ceb_behavior_pilot import ROOT,CACHE,Interpreter,resample_poly,sf,np,pd,math
import json

def main():
    sample=pd.read_csv(ROOT/'results/003_pilot_sample.csv')
    sample=sample[sample.event_duration<=1.2]
    labels=(CACHE/'v1_5_1_BirdNET_GLOBAL_6K_V2.4_Labels_en_uk.txt').read_text().splitlines()
    mapping={s.split('_')[0]:i for i,s in enumerate(labels)}
    model=Interpreter(model_path=str(CACHE/'v1_5_1_BirdNET_GLOBAL_6K_V2.4_Model_FP32.tflite'),num_threads=1)
    model.allocate_tensors();inp=model.get_input_details()[0];out=model.get_output_details()[0]
    rows=[]
    for file,group in sample.groupby('filepath'):
        audio,rate=sf.read(CACHE/'ceb_test_audio'/file,dtype='float32',always_2d=True)
        audio=audio.mean(axis=1);gcd=math.gcd(rate,48000)
        audio=resample_poly(audio,48000//gcd,rate//gcd).astype(np.float32)
        for e in group.to_dict('records'):
            start=round(e['start']*48000);original=audio[start:start+144000]
            for mode in ['circular','source']:
                for shift in [-.6,-.3,0.,.3,.6]:
                    offset=round(shift*48000)
                    chunk=np.roll(original,offset) if mode=='circular' else audio[start+offset:start+offset+144000]
                    assert len(chunk)==144000
                    model.set_tensor(inp['index'],chunk[None,:]);model.invoke()
                    logit=float(model.get_tensor(out['index'])[0,mapping[e['species']]])
                    rows.append(dict(event_id=e['event_id'],filepath=file,species=e['species'],voc=e['voc'],duration=e['event_duration'],mode=mode,shift=shift,score=float(1/(1+np.exp(-np.clip(logit,-15,15))))))
    df=pd.DataFrame(rows);df.to_csv(ROOT/'results/008_alignment_predictions.csv',index=False)
    summaries=[]
    for mode,g in df.groupby('mode'):
        limits=g.groupby('event_id').score.agg(['min','max'])
        summaries.append(dict(mode=mode,events=len(limits),species=g.species.nunique(),files=g.filepath.nunique(),
            crosses_01=int(((limits['min']<.1)&(limits['max']>=.1)).sum()),
            crosses_05=int(((limits['min']<.5)&(limits['max']>=.5)).sum()),
            median_score_range=float((limits['max']-limits['min']).median())))
    (ROOT/'results/008_alignment_summary.json').write_text(json.dumps(summaries,indent=2)+'\n')
    print(json.dumps(summaries,indent=2))

if __name__=='__main__':main()