127 lines
4.6 KiB
Python
127 lines
4.6 KiB
Python
import numpy as np
|
||
from sklearn.manifold import TSNE
|
||
import matplotlib.pyplot as plt
|
||
from opts import opts
|
||
from clip import clip
|
||
import torch
|
||
from utils.utils import prepare_directories, set_seed, get_dataset_loader, configure_clip_encoders, save_model, \
|
||
set_adapter_weights, get_text_feature, AverageMeter, accuracy, calculate_zeroshot_weights, gpt_clip_classifier,calculate_zeroshot_weights_GPT
|
||
import glob
|
||
import json
|
||
import os
|
||
import numpy as np
|
||
from PIL import Image
|
||
|
||
# 设置随机种子以确保可重复性
|
||
np.random.seed(0)
|
||
|
||
args = opts()
|
||
model, preprocess = clip.load(args.name)
|
||
model = model.cuda()
|
||
model.float()
|
||
CLIP_Text, Text_Encoder, CLIP_Image, Image_Encoder = configure_clip_encoders(args, model, 0, 1)
|
||
prepare_directories(args, CLIP_Text, CLIP_Image)
|
||
json_files = glob.glob('gpt_file/caltech_prompt.json')
|
||
for file_path in json_files:
|
||
# 打开并读取每个YAML文件
|
||
with open(file_path, 'r') as f:
|
||
gpt3_prompt = json.load(f)
|
||
|
||
|
||
dir = './weights/'
|
||
CLIP_Text=torch.load(dir+'CLIP_Text.pth')
|
||
CLIP_Image=torch.load(dir+'CLIP_Image.pth')
|
||
Image_Encoder=torch.load(dir+'Image_Encoder.pth')
|
||
Text_Encoder=torch.load(dir+'Text_Encoder.pth')
|
||
adapter=torch.load(dir+'adapter.pth')
|
||
|
||
image_folder = "./images"
|
||
image_folder2 = "./images_2"
|
||
image_folder3 = "./images_3"
|
||
# 读取图片并提取特征
|
||
features = []
|
||
for image_name in os.listdir(image_folder):
|
||
if image_name.lower().endswith(('.png', '.jpg', '.jpeg')):
|
||
# 读取图片
|
||
image_path = os.path.join(image_folder, image_name)
|
||
image = preprocess(Image.open(image_path)).unsqueeze(0).cuda()
|
||
with torch.no_grad():
|
||
input_target_temp = CLIP_Image(image)
|
||
input_target_add = Image_Encoder(input_target_temp)
|
||
|
||
|
||
# 存储提取的特征
|
||
# features.append(input_target_add.detach().numpy())
|
||
features.append(input_target_add.detach().cpu().numpy())
|
||
|
||
features2 = []
|
||
for image_name in os.listdir(image_folder2):
|
||
if image_name.lower().endswith(('.png', '.jpg', '.jpeg')):
|
||
# 读取图片
|
||
image_path = os.path.join(image_folder2, image_name)
|
||
image = preprocess(Image.open(image_path)).unsqueeze(0).cuda()
|
||
with torch.no_grad():
|
||
input_target_temp = CLIP_Image(image)
|
||
input_target_add = Image_Encoder(input_target_temp)
|
||
|
||
|
||
# 存储提取的特征
|
||
# features.append(input_target_add.detach().numpy())
|
||
features2.append(input_target_add.detach().cpu().numpy())
|
||
|
||
# 将特征列表转换为NumPy数组
|
||
features = np.vstack(features)
|
||
features2 = np.vstack(features2)
|
||
|
||
features3 = []
|
||
for image_name in os.listdir(image_folder3):
|
||
if image_name.lower().endswith(('.png', '.jpg', '.jpeg')):
|
||
# 读取图片
|
||
image_path = os.path.join(image_folder3, image_name)
|
||
image = preprocess(Image.open(image_path)).unsqueeze(0).cuda()
|
||
with torch.no_grad():
|
||
input_target_temp = CLIP_Image(image)
|
||
input_target_add = Image_Encoder(input_target_temp)
|
||
|
||
|
||
# 存储提取的特征
|
||
# features.append(input_target_add.detach().numpy())
|
||
features3.append(input_target_add.detach().cpu().numpy())
|
||
|
||
# 将特征列表转换为NumPy数组
|
||
features = np.vstack(features)
|
||
features2 = np.vstack(features2)
|
||
features3 = np.vstack(features3)
|
||
|
||
|
||
# 源域样本：一个二维特征
|
||
# source_domain_sample = np.array([[0.5, 0.5]]) # 示例特征
|
||
|
||
# 生成随机数据：目标域，20个样本，每个样本1024维
|
||
# target_domain_samples = np.random.rand(20, 1024)
|
||
target_domain_samples = features
|
||
source_domain_samples=features2
|
||
|
||
|
||
# 由于源域只有一个二维样本，我们不需要对它使用t-SNE
|
||
tsne = TSNE(n_components=2, random_state=0,perplexity=19)
|
||
source_domain_tsne_results = tsne.fit_transform(source_domain_samples)
|
||
|
||
# 直接对目标域样本应用t-SNE降维到2维
|
||
# tsne = TSNE(n_components=2, random_state=0,perplexity=7)
|
||
target_domain_tsne_results = tsne.fit_transform(target_domain_samples)
|
||
|
||
results3=tsne.fit_transform(features3)
|
||
|
||
# 可视化结果
|
||
plt.figure(figsize=(10, 6))
|
||
# plt.scatter(source_domain_tsne_results[:, 0], source_domain_tsne_results[:, 1], c='blue', label='Source Domain (Single Sample)', edgecolors='w', s=200)
|
||
plt.scatter(source_domain_tsne_results[:, 0], source_domain_tsne_results[:, 1], c='blue', label='Source Domain (Single Sample)')#, edgecolors='w')
|
||
plt.scatter(target_domain_tsne_results[:, 0], target_domain_tsne_results[:, 1], c='red', label='Target Domain')
|
||
plt.scatter(results3[:, 0], results3[:, 1], c='yellow', label='3')
|
||
plt.title('t-SNE visualization of feature representations')
|
||
plt.xlabel('Component 1')
|
||
plt.ylabel('Component 2')
|
||
plt.legend()
|
||
plt.show()
|