contriever-mlx / common.py
masahiroid's picture
Publish validated original-checkpoint conversion with provenance and usage
618a741 verified
Raw History Blame Contribute Delete
4.55 kB
import json
from pathlib import Path
import numpy as np
import torch
from transformers import AutoModel, AutoModelForSequenceClassification, AutoTokenizer
ROOT=Path(__file__).resolve().parent
NAMES=['roberta-base-go_emotions','chronos-bolt-small','contriever','mdeberta-v3-base']
torch.set_num_threads(4)
def load(name):
path=ROOT/'sources'/name
if name=='chronos-bolt-small':
from chronos.chronos_bolt import ChronosBoltModelForForecasting
model=ChronosBoltModelForForecasting.from_pretrained(path).eval()
model.config.use_cache=False
model.decoder.config.use_cache=False
else:
cls=AutoModelForSequenceClassification if name=='roberta-base-go_emotions' else AutoModel
model=cls.from_pretrained(path,attn_implementation='eager').eval()
return model
class Wrapper(torch.nn.Module):
def __init__(self,model,name): super().__init__(); self.model=model; self.name=name
def forward(self,input_ids,attention_mask=None):
if self.name=='chronos-bolt-small': return self.model(context=input_ids).quantile_preds
out=self.model(input_ids=input_ids,attention_mask=attention_mask,return_dict=False)[0]
if self.name=='contriever':
m=attention_mask.unsqueeze(-1).to(out.dtype)
return (out*m).sum(1)/m.sum(1).clamp(min=1)
return out
class ExportWrapper(Wrapper):
def forward(self,input_ids,attention_mask=None):
if self.name!='chronos-bolt-small': return super().forward(input_ids,attention_mask)
m=self.model; context=input_ids
valid=context==context
count=valid.sum(-1,keepdim=True).clamp(min=1).float()
loc=torch.where(valid,context,0.0).sum(-1,keepdim=True)/count
scale=(torch.where(valid,(context-loc).square(),0.0).sum(-1,keepdim=True)/count).sqrt()
scale=torch.where(valid.any(-1,keepdim=True),scale,1.0)
scale=torch.where(scale==0,1e-5,scale)
x=torch.where(valid,(context-loc)/scale,0.0)
size=m.chronos_config.input_patch_size
patches=x.reshape(context.shape[0],-1,size)
mask=valid.float().reshape(context.shape[0],-1,size)
am=mask.sum(-1)>0
embeds=m.input_patch_embedding(torch.cat([patches,mask],-1))
if m.chronos_config.use_reg_token:
reg=m.shared.weight[1].reshape(1,1,-1).expand(context.shape[0],1,-1)
embeds=torch.cat([embeds,reg],1)
am=torch.cat([am,torch.ones((context.shape[0],1),dtype=torch.bool)],1)
hidden=m.encoder(inputs_embeds=embeds,attention_mask=am,return_dict=False)[0]
dec=m.shared.weight[0].reshape(1,1,-1).expand(context.shape[0],1,-1)
output=m.decoder(inputs_embeds=dec,encoder_hidden_states=hidden,encoder_attention_mask=am,use_cache=False,return_dict=False)[0]
return m.output_patch_embedding(output).reshape(context.shape[0],m.num_quantiles,m.chronos_config.prediction_length)*scale.unsqueeze(-1)+loc.unsqueeze(-1)
def cases(name,length=128):
if name=='chronos-bolt-small':
length=2048
t=np.arange(length,dtype=np.float32)
return [dict(context=np.array([np.sin(t/9)+t/100],dtype=np.float32)),
dict(context=np.full((1,length),7,dtype=np.float32)),
dict(context=np.where(np.arange(length)[None]%11==0,np.nan,np.cos(t/7)[None]).astype(np.float32))]
tok=AutoTokenizer.from_pretrained(ROOT/'sources'/name,use_fast=name!='mdeberta-v3-base')
texts=['Thank you so much! I love this wonderful surprise.', 'I am angry and disappointed. This is terrible.',
'The document describes information retrieval and machine learning.', '今日は良い天気です。検索モデルを検証します。',
'', 'A short sentence. '*80]
return [dict(tok(t,padding='max_length',truncation=True,max_length=length,return_tensors='np')) for t in texts]
def args_for(name,case):
if name=='chronos-bolt-small': return (torch.tensor(case['context']),)
return tuple(torch.tensor(case[k],dtype=torch.int32) for k in ['input_ids','attention_mask'])
def reference(name,case,wrapper):
with torch.no_grad(): return wrapper(*args_for(name,case)).float().numpy()
def metrics(ref,out):
ref=ref.astype(np.float32); out=out.astype(np.float32)
assert ref.shape==out.shape and np.isfinite(out).all(), (ref.shape,out.shape)
return dict(max_abs=float(np.max(np.abs(ref-out))),mean_abs=float(np.mean(np.abs(ref-out))),
cosine=float(np.dot(ref.ravel(),out.ravel())/(np.linalg.norm(ref)*np.linalg.norm(out)+1e-30)))