SAM 3 路 LiteRT (.tflite) (fp32 / fp16 / int8 / w8 / q4 / w4)

Conversi贸n de facebook/sam3 con litert-torch y cuantizaci贸n con litert-quantizer.

Archivos

variante vision_encoder text_encoder prompt_decoder
fp32 1.7 GB 1.3 GB 128.7 MB
fp16 869.1 MB 675.0 MB 83.4 MB
int8 441.5 MB 341.5 MB 61.2 MB
w8 441.6 MB 341.6 MB 61.3 MB
q4 241.6 MB 173.2 MB 50.0 MB
w4 241.7 MB 173.4 MB 49.9 MB
sufijo qu茅 es
fp16 pesos fp16, c贸mputo fp32
int8 din谩mica int8 (pesos por canal + activaciones cuantizadas en ejecuci贸n)
w8 solo pesos int8 (c贸mputo float, m谩s fiel)
q4 din谩mica int4 por canal
w4 solo pesos int4 por canal

Entradas / salidas

grafo entradas salidas
vision_encoder pixel_values [B,3,1008,1008] float32 fpn_hidden_state_0 [B,256,288,288], _1 [B,256,144,144], _2 [B,256,72,72]
text_encoder input_ids, attention_mask [B,32] text_features [B,32,256]
prompt_decoder fpn_hidden_state_0/1/2, text_features, attention_mask pred_masks [B,200,288,288] (logits), pred_boxes [B,200,4] xyxy normalizado, pred_logits [B,200], presence_logits [B,1], semantic_seg [B,1,288,288]

Score de cada consulta = sigmoid(pred_logits) * sigmoid(presence_logits). El preprocesado (resize 1008, normalizaci贸n) y el tokenizer (CLIP, max_length=32, padding="max_length") vienen en este repo v铆a Sam3Processor.

En LiteRT input_ids y attention_mask son int32, y las salidas se llaman output_0鈥 en el mismo orden que la tabla (ver manifest.json).

Validaci贸n (imagen COCO, prompt "cat")

variante objetos (ref) IoU m谩scara vs PyTorch 螖 score m谩x
fp32 1 (1) 1.0 0.0
fp16 1 (1) 1.0 0.0
int8 1 (1) 0.9981 0.0016
w8 1 (1) 0.9995 0.0003
q4 1 (1) 0.9923 0.0287
w4 1 (1) 0.9927 0.0304

Uso (Python, ai-edge-litert)

import numpy as np
from huggingface_hub import hf_hub_download
from ai_edge_litert.interpreter import Interpreter
from transformers import Sam3Processor
from PIL import Image

repo, v = "tepirale/sam3-LiteRT", "_q4"
proc = Sam3Processor.from_pretrained(repo)
R = lambda n: Interpreter(model_path=hf_hub_download(repo, f"{n}{v}.tflite")).get_signature_runner()
o = lambda d: [d[k] for k in sorted(d, key=lambda s: int(s.split("_")[-1]))]

image = Image.open("foto.jpg").convert("RGB")
pv  = proc(images=image, return_tensors="np")["pixel_values"].astype(np.float32)
tok = proc.tokenizer("cat", padding="max_length", max_length=32, truncation=True, return_tensors="np")
ids, am = tok["input_ids"].astype(np.int32), tok["attention_mask"].astype(np.int32)
fpn = o(R("vision_encoder")(pixel_values=pv))
tf  = o(R("text_encoder")(input_ids=ids, attention_mask=am))[0]
pred_masks, pred_boxes, pred_logits, presence_logits, semantic_seg = o(R("prompt_decoder")(
    fpn_hidden_state_0=fpn[0], fpn_hidden_state_1=fpn[1], fpn_hidden_state_2=fpn[2], text_features=tf, attention_mask=am))

En Android/iOS usa CompiledModel / Interpreter de LiteRT con los mismos nombres de firma.

Licencia

Pesos derivados de facebook/sam3, publicados por Meta bajo la SAM License (no Apache-2.0). Revisa y acepta sus t茅rminos antes de usar o redistribuir.

Downloads last month
92
Inference Providers NEW
This model isn't deployed by any Inference Provider. 馃檵 Ask for provider support

Model tree for tepirale/sam3-LiteRT

Base model

facebook/sam3
Finetuned
(44)
this model