深度学习中遇到的知识点

1.grounded_sam_demo.py

点击查看代码

import argparse
import os
import sys

import numpy as np
import json
import torch
from PIL import Image

sys.path.append(os.path.join(os.getcwd(), "./GroundingDINO"))
sys.path.append(os.path.join(os.getcwd(), "./segment_anything"))


# Grounding DINO
import GroundingDINO.groundingdino.datasets.transforms as T
from GroundingDINO.groundingdino.models import build_model
from GroundingDINO.groundingdino.util.slconfig import SLConfig
from GroundingDINO.groundingdino.util.utils import clean_state_dict, get_phrases_from_posmap


# segment anything
from segment_anything import (
    sam_model_registry,
    sam_hq_model_registry,
    SamPredictor
)
import cv2
import numpy as np
import matplotlib.pyplot as plt


def load_image(image_path):
    # load image
    image_pil = Image.open(image_path).convert("RGB")  # load image

    transform = T.Compose(
        [
            T.RandomResize([800], max_size=1333),
            T.ToTensor(),
            T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
        ]
    )
    image, _ = transform(image_pil, None)  # 3, h, w
    return image_pil, image


def load_model(model_config_path, model_checkpoint_path, bert_base_uncased_path, device):
    args = SLConfig.fromfile(model_config_path)
    args.device = device
    args.bert_base_uncased_path = bert_base_uncased_path
    model = build_model(args)
    checkpoint = torch.load(model_checkpoint_path, map_location="cpu")
    load_res = model.load_state_dict(clean_state_dict(checkpoint["model"]), strict=False)
    print(load_res)
    _ = model.eval() # .eval()作用是将模型切换到推理模式（evaluation mode）；_ 是“占位符”，表示你不关心返回值
    return model


def get_grounding_output(model, image, caption, box_threshold, text_threshold, with_logits=True, device="cpu"):
    caption = caption.lower()
    caption = caption.strip()
    if not caption.endswith("."):
        caption = caption + "."
    model = model.to(device)
    image = image.to(device)
    with torch.no_grad():
        outputs = model(image[None], captions=[caption])
    logits = outputs["pred_logits"].cpu().sigmoid()[0]  # (nq, 256)
    boxes = outputs["pred_boxes"].cpu()[0]  # (nq, 4)
    logits.shape[0]

    # filter output
    logits_filt = logits.clone()
    boxes_filt = boxes.clone()
    filt_mask = logits_filt.max(dim=1)[0] > box_threshold
    logits_filt = logits_filt[filt_mask]  # num_filt, 256
    boxes_filt = boxes_filt[filt_mask]  # num_filt, 4
    logits_filt.shape[0]

    # get phrase
    tokenlizer = model.tokenizer
    tokenized = tokenlizer(caption)
    # build pred
    pred_phrases = []
    for logit, box in zip(logits_filt, boxes_filt):
        pred_phrase = get_phrases_from_posmap(logit > text_threshold, tokenized, tokenlizer)
        if with_logits:
            pred_phrases.append(pred_phrase + f"({str(logit.max().item())[:4]})")
        else:
            pred_phrases.append(pred_phrase)

    return boxes_filt, pred_phrases

def show_mask(mask, ax, random_color=False):
    if random_color:
        color = np.concatenate([np.random.random(3), np.array([0.6])], axis=0)
    else:
        color = np.array([30/255, 144/255, 255/255, 0.6])
    h, w = mask.shape[-2:]
    mask_image = mask.reshape(h, w, 1) * color.reshape(1, 1, -1)
    ax.imshow(mask_image)


def show_box(box, ax, label):
    x0, y0 = box[0], box[1]
    w, h = box[2] - box[0], box[3] - box[1]
    ax.add_patch(plt.Rectangle((x0, y0), w, h, edgecolor='green', facecolor=(0,0,0,0), lw=2))
    ax.text(x0, y0, label)


def save_mask_data(output_dir, mask_list, box_list, label_list):
    value = 0  # 0 for background

    mask_img = torch.zeros(mask_list.shape[-2:])
    for idx, mask in enumerate(mask_list):
        mask_img[mask.cpu().numpy()[0] == True] = value + idx + 1
    plt.figure(figsize=(10, 10))
    plt.imshow(mask_img.numpy())
    plt.axis('off')
    plt.savefig(os.path.join(output_dir, 'mask.jpg'), bbox_inches="tight", dpi=300, pad_inches=0.0)

    json_data = [{
        'value': value,
        'label': 'background'
    }]
    for label, box in zip(label_list, box_list):
        value += 1
        name, logit = label.split('(')
        logit = logit[:-1] # the last is ')'
        json_data.append({
            'value': value,
            'label': name,
            'logit': float(logit),
            'box': box.numpy().tolist(),
        })
    with open(os.path.join(output_dir, 'mask.json'), 'w') as f:
        json.dump(json_data, f)


if __name__ == "__main__":

    parser = argparse.ArgumentParser("Grounded-Segment-Anything Demo", add_help=True)
    parser.add_argument("--config", type=str, required=True, help="path to config file")
    parser.add_argument(
        "--grounded_checkpoint", type=str, required=True, help="path to checkpoint file"
    )
    parser.add_argument(
        "--sam_version", type=str, default="vit_h", required=False, help="SAM ViT version: vit_b / vit_l / vit_h"
    )
    parser.add_argument(
        "--sam_checkpoint", type=str, required=False, help="path to sam checkpoint file"
    )
    parser.add_argument(
        "--sam_hq_checkpoint", type=str, default=None, help="path to sam-hq checkpoint file"
    )
    parser.add_argument(
        "--use_sam_hq", action="store_true", help="using sam-hq for prediction"
    )
    parser.add_argument("--input_image", type=str, required=True, help="path to image file")
    parser.add_argument("--text_prompt", type=str, required=True, help="text prompt")
    parser.add_argument(
        "--output_dir", "-o", type=str, default="outputs", required=True, help="output directory"
    )

    parser.add_argument("--box_threshold", type=float, default=0.3, help="box threshold")
    parser.add_argument("--text_threshold", type=float, default=0.25, help="text threshold")

    parser.add_argument("--device", type=str, default="cpu", help="running on cpu only!, default=False")
    parser.add_argument("--bert_base_uncased_path", type=str, required=False, help="bert_base_uncased model path, default=False")
    args = parser.parse_args()
    print(args)
    # cfg
    config_file = "./GroundingDINO/groundingdino/config/GroundingDINO_SwinT_OGC.py"  # change the path of the model config file
    grounded_checkpoint = "groundingdino_swint_ogc.pth"  # change the path of the model
    sam_version = "vit_h"
    sam_checkpoint = "sam_vit_h_4b8939.pth"
    sam_hq_checkpoint = ""
    use_sam_hq = ""
    image_path ="./assets/demo7.jpg"
    text_prompt = "Horse. Clouds. Grasses. Sky. Hill."
    output_dir = "outputs"
    box_threshold = 0.3
    text_threshold = 0.25
    device = "cpu"
    bert_base_uncased_path = ""

    # make dir
    os.makedirs(output_dir, exist_ok=True)
    # load image
    image_pil, image = load_image(image_path)
    # load model
    model = load_model(config_file, grounded_checkpoint, bert_base_uncased_path, device=device)

    # visualize raw image
    image_pil.save(os.path.join(output_dir, "raw_image.jpg"))

    # run grounding dino model
    boxes_filt, pred_phrases = get_grounding_output(
        model, image, text_prompt, box_threshold, text_threshold, device=device
    )

    # initialize SAM
    if use_sam_hq:
        predictor = SamPredictor(sam_hq_model_registry[sam_version](checkpoint=sam_hq_checkpoint).to(device))
    else:
        predictor = SamPredictor(sam_model_registry[sam_version](checkpoint=sam_checkpoint).to(device))
    image = cv2.imread(image_path)
    image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
    predictor.set_image(image)

    size = image_pil.size
    H, W = size[1], size[0]
    for i in range(boxes_filt.size(0)):
        boxes_filt[i] = boxes_filt[i] * torch.Tensor([W, H, W, H])
        boxes_filt[i][:2] -= boxes_filt[i][2:] / 2
        boxes_filt[i][2:] += boxes_filt[i][:2]

    boxes_filt = boxes_filt.cpu()
    transformed_boxes = predictor.transform.apply_boxes_torch(boxes_filt, image.shape[:2]).to(device)

    masks, _, _ = predictor.predict_torch(
        point_coords = None,
        point_labels = None,
        boxes = transformed_boxes.to(device),
        multimask_output = False,
    )

    # draw output image
    plt.figure(figsize=(10, 10))
    plt.imshow(image)
    for mask in masks:
        show_mask(mask.cpu().numpy(), plt.gca(), random_color=True)
    for box, label in zip(boxes_filt, pred_phrases):
        show_box(box.numpy(), plt.gca(), label)

    plt.axis('off')
    plt.savefig(
        os.path.join(output_dir, "grounded_sam_output.jpg"),
        bbox_inches="tight", dpi=300, pad_inches=0.0
    )

    save_mask_data(output_dir, masks, boxes_filt, pred_phrases)

1.1

def load_model(model_config_path, model_checkpoint_path, bert_base_uncased_path, device):
    args = SLConfig.fromfile(model_config_path)
    args.device = device
    args.bert_base_uncased_path = bert_base_uncased_path
    model = build_model(args)
    checkpoint = torch.load(model_checkpoint_path, map_location="cpu")
    load_res = model.load_state_dict(clean_state_dict(checkpoint["model"]), strict=False)
    print(load_res)
    _ = model.eval() 

✅ checkpoint = torch.load(model_checkpoint_path, map_location="cpu")

🔍 作用（What it does）：
从磁盘加载一个 训练好的模型权重文件（通常是 .pth 或 .pt 文件），比如 groundingdino_swint_ogc.pth。

这个文件里保存了模型在训练过程中学到的所有参数（权重、偏置等）。

📥 参数说明：
model_checkpoint_path：字符串，表示模型权重文件的路径
例如："groundingdino_swint_ogc.pth" 或 "./weights/dino.pth"
map_location="cpu"：告诉 PyTorch 把模型参数加载到 CPU 上，如果你用 GPU 训练但想在 CPU 上运行（推理），必须加这个
如果你想用 GPU，可以写 map_location="cuda" 或 "cuda:0"
💡 小知识：.pth 文件本质是一个 Python 字典（dict），可以用 torch.load() 读出来。

✅ checkpoint["model"]

checkpoint 是一个字典，它通常包含多个字段

✅ clean_state_dict(...)

clean_state_dict 是一个自定义函数（在 GroundingDINO 项目中定义）
作用：清理模型权重名称，比如去掉多余的前缀（如 module.）

原来的 key: module.backbone.conv1.weight
清理后 key: backbone.conv1.weight

✅ 返回值 load_res

load_res 是一个命名元组（NamedTuple），包含两个字段：

• missing_keys：模型中找不到对应权重的层
• unexpected_keys：权重中有但模型中没有的层

打印它可以看到哪些层没加载成功，方便调试

✅ _ = model.eval()

🔍 作用：
将模型切换到 推理模式（evaluation mode）
_ = model.eval() 中的 _ 是“占位符”，表示你不关心返回值

1.2

def get_grounding_output(model, image, caption, box_threshold, text_threshold, with_logits=True, device="cpu"):
    caption = caption.lower()
    caption = caption.strip()
    if not caption.endswith("."):
        caption = caption + "."
    model = model.to(device)
    image = image.to(device)
    with torch.no_grad():
        outputs = model(image[None], captions=[caption])
    logits = outputs["pred_logits"].cpu().sigmoid()[0]  # (nq, 256)
    boxes = outputs["pred_boxes"].cpu()[0]  # (nq, 4)
    logits.shape[0]

    # filter output
    logits_filt = logits.clone()
    boxes_filt = boxes.clone()
    filt_mask = logits_filt.max(dim=1)[0] > box_threshold
    logits_filt = logits_filt[filt_mask]  # num_filt, 256
    boxes_filt = boxes_filt[filt_mask]  # num_filt, 4
    logits_filt.shape[0]

    # get phrase
    tokenlizer = model.tokenizer
    tokenized = tokenlizer(caption)
    # build pred
    pred_phrases = []
    for logit, box in zip(logits_filt, boxes_filt):
        pred_phrase = get_phrases_from_posmap(logit > text_threshold, tokenized, tokenlizer)
        if with_logits:
            pred_phrases.append(pred_phrase + f"({str(logit.max().item())[:4]})")
        else:
            pred_phrases.append(pred_phrase)

    return boxes_filt, pred_phrases

流程

输入：图像 + 文本提示
     ↓
文本预处理：转小写、去空格、加句号
     ↓
模型推理：model(image, caption) → 得到 raw outputs
     ↓
提取 logits 和 boxes（去掉 batch 维度）
     ↓
按置信度过滤：max(logits) > box_threshold
     ↓
使用 tokenizer 和 get_phrases_from_posmap
     ↓
生成人类可读的短语 + 可选置信度
     ↓
输出：[boxes], ["phrase1(score)", "phrase2(score)", ...]

函数签名

def get_grounding_output(
    model,           # 加载好的 Grounding DINO 模型
    image,           # 预处理后的图像张量 (3, H, W)
    caption,         # 文本提示，例如 "dog", "a cat on the table"
    box_threshold,   # 置信度阈值（过滤低置信度的框）
    text_threshold,  # 文本匹配阈值
    with_logits=True,# 是否在输出标签中包含置信度分数
    device="cpu"     # 运行设备
):

image[None]

给图像增加一个 batch 维度：

• 原 shape: (3, H, W) → 表示单张图像
• 加 None 后: (1, 3, H, W) → 表示 batch_size=1 的图像批次

提取并处理模型输出

logits = outputs["pred_logits"].cpu().sigmoid()[0]  # (nq, 256)
boxes = outputs["pred_boxes"].cpu()[0]              # (nq, 4)

表达式	说明
outputs["pred_logits"]	模型原始输出，形状 (1, nq, 256)，其中 nq 是查询数量（如 900）
.cpu()	把结果从 GPU 拿回 CPU（便于后续处理）
.sigmoid()	将 logits 转换为 [0,1] 区间的概率值（越接近 1 表示匹配越好）
[0]	去掉 batch 维度，得到 (nq, 256)

根据阈值过滤低质量预测

# 复制一份用于过滤
logits_filt = logits.clone()
boxes_filt = boxes.clone()

# 创建过滤掩码：只保留最大得分 > box_threshold 的框
filt_mask = logits_filt.max(dim=1)[0] > box_threshold

# 应用掩码
logits_filt = logits_filt[filt_mask]  # (num_filt, 256)
boxes_filt = boxes_filt[filt_mask]    # (num_filt, 4)

从模型输出中提取人类可读的文本短语

tokenlizer = model.tokenizer        # 获取分词器
tokenized = tokenlizer(caption)     # 对输入文本进行编码

• tokenizer：把文本拆成单词/子词（tokens），并转换为数字 ID
• tokenized：包含 input_ids, attention_mask 等信息

构建最终的预测短语列表

核心函数：get_phrases_from_posmap(...)
它做的事情是：
logit > text_threshold：得到一个布尔向量，标记哪些 token 匹配成功。
根据这些 token 的位置，反推出原始文本中的词语合并连续的 token 成完整短语
✅ 示例：
输入 "a black dog."
模型发现第2、3个 token 匹配成功 → 输出 "black dog"

1.3

def show_mask(mask, ax, random_color=False):
    if random_color:
        color = np.concatenate([np.random.random(3), np.array([0.6])], axis=0)
    else:
        color = np.array([30/255, 144/255, 255/255, 0.6])
    h, w = mask.shape[-2:]
    mask_image = mask.reshape(h, w, 1) * color.reshape(1, 1, -1)
    ax.imshow(mask_image)

mask.shape[-2:]

获取掩码的高度和宽度。这里假设输入的掩码至少是二维的，.shape[-2:] 确保即使掩码有额外的维度（例如批次维度），也能正确获取其高度和宽度。

广播乘法

当执行 mask.reshape(h, w, 1) * color.reshape(1, 1, -1) 操作时，NumPy会自动应用广播规则来对这两个不同形状的数组进行元素乘法。

• 掩码数组的形状为 (H, W, 1)。
• 颜色数组的形状为 (1, 1, 4)。
• 在乘法过程中，这些数组会被“扩展”到共同的形状 (H, W, 4)，其中掩码中的每个非零元素都会被替换为颜色数组中的相应颜色，而零元素则保持不变（实际上，由于乘法操作，它们会变成透明的，因为颜色数组的第四个通道代表透明度）。