scxfxt 1.0.0.0__tar.gz

scxfxt-1.0.0.0/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2024 [scxfxt]
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

scxfxt-1.0.0.0/MANIFEST.in

@@ -0,0 +1,2 @@
+include LICENSE
+include README.md

scxfxt-1.0.0.0/PKG-INFO

@@ -0,0 +1,26 @@
+Metadata-Version: 2.1
+Name: scxfxt
+Version: 1.0.0.0
+Summary: scxfxt
+Home-page: https://gitee.com/scxfxt
+Author: scxfxt_scxfxt
+Author-email: shanhe12@163.com
+Classifier: Programming Language :: Python :: 3
+Classifier: License :: OSI Approved :: MIT License
+Classifier: Operating System :: OS Independent
+Requires-Python: >=3.6
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: tqdm
+Requires-Dist: pandas
+Requires-Dist: matplotlib
+
+# Welecome to scxfxt
+
+A simple Python package that provides a timer decorator to measure the execution time of functions.
+
+## Installation
+ 
+pip install scxfxt
+ 
+pip install scxfxt -i https://mirrors.cloud.tencent.com/pypi/simple

scxfxt-1.0.0.0/README.md

@@ -0,0 +1,9 @@
+# Welecome to scxfxt
+
+A simple Python package that provides a timer decorator to measure the execution time of functions.
+
+## Installation
+ 
+pip install scxfxt
+ 
+pip install scxfxt -i https://mirrors.cloud.tencent.com/pypi/simple

scxfxt-1.0.0.0/scxfxt/3D/__init__.py

@@ -0,0 +1,25 @@
+import torch
+import numpy as np
+import matplotlib.pyplot as plt
+# from mpl_toolkits.mplot3d import Axes3D 
+from shancxn import crDir
+def plot3DJU(x,path="./3D_test1.png"):
+    data = x.cpu().numpy()  if x.is_cuda else x[:,:,:].numpy()    
+    data = data[:, ::5, ::5]
+    x1 = np.arange(data.shape[0])  # x轴的范围，取决于data的第一个维度
+    y1 = np.arange(data.shape[1])  # y轴的范围，取决于data的第二个维度
+    z1 = np.arange(data.shape[2])  # z轴的范围，取决于data的第三个维度
+    x1, y1, z1 = np.meshgrid(x1, y1, z1)
+    x1_flat = x1.flatten()
+    y1_flat = y1.flatten()
+    z1_flat = z1.flatten()
+    colors = data.flatten()
+    fig = plt.figure(figsize=(10, 7))
+    ax = fig.add_subplot(111, projection='3d')
+    scatter = ax.scatter(x1_flat, y1_flat, z1_flat, c=colors, cmap='viridis')
+    ax.set_title("3D Scatter Plot of Shape (400, 640, 400)")
+    plt.colorbar(scatter)
+    outpath = path
+    crDir(outpath)
+    plt.savefig(outpath)
+    plt.close()

scxfxt-1.0.0.0/scxfxt/Algo/Class.py

@@ -0,0 +1,11 @@
+import torch
+import copy
+def classify1h(pre0):
+    pre = copy.deepcopy(pre0)
+    pre[torch.logical_and(pre0 >= 0.1, pre0 <= 2.5)] = 1
+    pre[torch.logical_and(pre0 > 2.5, pre0 <= 8)] = 2
+    pre[torch.logical_and(pre0 > 8, pre0 <= 16)] = 3
+    pre[torch.logical_and(pre0 > 16, pre0 <= 300)] = 4
+    pre[pre0 > 300] = -1
+    pre[torch.isnan(pre0)] = -1
+    return pre

scxfxt-1.0.0.0/scxfxt/Algo/CudaPrefetcher1.py

@@ -0,0 +1,112 @@
+
+from torch.utils.data import Dataset, DataLoader
+import torch
+class CUDAPrefetcher1:
+    """
+    Use the CUDA side to accelerate data reading.
+
+    Args:
+        dataloader (DataLoader): Data loader. Combines a dataset and a sampler, and provides an iterable over the given dataset.
+        device (torch.device): Specify running device.
+    """
+
+    def __init__(self, dataloader: DataLoader, device: torch.device):
+        self.batch_data = None
+        self.original_dataloader = dataloader
+        self.device = device
+
+        self.data = iter(dataloader)
+        self.stream = torch.cuda.Stream()
+        self.preload()
+
+    def preload(self):
+        """
+        Load the next batch of data and move it to the specified device asynchronously.
+        """
+        try:
+            self.batch_data = next(self.data)
+        except StopIteration:
+            self.batch_data = None
+            return
+
+        # Asynchronously move the data to the GPU
+        with torch.cuda.stream(self.stream):
+            if isinstance(self.batch_data, dict):
+                self.batch_data = {
+                    k: v.to(self.device, non_blocking=True)
+                    if torch.is_tensor(v) else v
+                    for k, v in self.batch_data.items()
+                }
+            elif isinstance(self.batch_data, (list, tuple)):
+                self.batch_data = [
+                    x.to(self.device, non_blocking=True)
+                    if torch.is_tensor(x) else x
+                    for x in self.batch_data
+                ]
+            elif torch.is_tensor(self.batch_data):
+                self.batch_data = self.batch_data.to(self.device, non_blocking=True)
+            else:
+                raise TypeError(
+                    f"Unsupported data type {type(self.batch_data)}. Ensure the dataloader outputs tensors, dicts, or lists."
+                )
+
+    def __iter__(self):
+        """
+        Make the object iterable by resetting the dataloader and returning self.
+        """
+        self.reset()
+        return self
+
+    def __next__(self):
+        """
+        Wait for the current stream, return the current batch, and preload the next batch.
+        """
+        if self.batch_data is None:
+            raise StopIteration
+
+        torch.cuda.current_stream().wait_stream(self.stream)
+        batch_data = self.batch_data
+        self.preload()
+        return batch_data
+
+    def reset(self):
+        """
+        Reset the dataloader iterator and preload the first batch.
+        """
+        self.data = iter(self.original_dataloader)
+        self.preload()
+
+    def __len__(self):
+        """
+        Return the length of the dataloader.
+        """
+        return len(self.original_dataloader)
+
+#  train_data_prefetcher = CUDAPrefetcher1(degenerated_train_dataloader, device)
+
+import torch
+from torch.utils.data import DataLoader
+import torch.nn.functional as F
+def custom_collate_fn(batch):
+    """
+    Custom collate function to handle batches with varying tensor sizes by padding them to the same size.
+    """
+    if isinstance(batch[0], dict):
+        collated = {key: custom_collate_fn([d[key] for d in batch]) for key in batch[0]}
+        return collated
+    elif isinstance(batch[0], torch.Tensor):
+        # Find max dimensions
+        max_height = max(item.shape[1] for item in batch)
+        max_width = max(item.shape[2] for item in batch)
+
+        # Pad each tensor to the same size
+        padded_batch = []
+        for item in batch:
+            padding = (0, max_width - item.shape[2], 0, max_height - item.shape[1])
+            padded_item = F.pad(item, padding, mode="constant", value=0)
+            padded_batch.append(padded_item)
+        return torch.stack(padded_batch)
+    else:
+        return batch
+# device = torch.device("cuda:0")
+# paired_test_dataloader = DataLoader(datasetdata,device)

scxfxt-1.0.0.0/scxfxt/Algo/Fake_image.py

@@ -0,0 +1,24 @@
+# 加载最佳模型权重
+import torch
+import matplotlib.pyplot as plt
+from shancxn import crDir
+def Fakeimage(generator,device,torchvision.utils.make_grid,):
+    generator.load_state_dict(torch.load('/home/scx/train_Gan/best_generator_weights.pth'))
+    # 设置模型为评估模式
+    generator.eval()
+    with torch.no_grad():
+        fixed_noise = torch.randn(64, 100, device=device)  # 将噪声张量直接创建在 GPU 上
+        fake_images = generator(fixed_noise)    
+        # 检查生成图像的形状
+        print(f"生成的图像形状: {fake_images.shape}")    
+        # 反归一化处理
+        fake_images = (fake_images + 1) / 2  # 将图像像素值从[-1, 1]映射到[0, 1]  
+        print(f"图像最小值: {fake_images.min()}, 最大值: {fake_images.max()}")
+        grid = torchvision.utils.make_grid(fake_images, nrow=8, normalize=True)  # 设置normalize=True以确保正确的缩放
+        grid_cpu = grid.cpu().detach().numpy()
+        plt.imshow(grid_cpu.transpose(1, 2, 0))  # 调整图像通道顺序为(H, W, C)
+        plt.axis('off')  # 关闭坐标轴显示
+        outPath = './fake_images/best_weights_fake_images.png'
+        crDir(outPath)
+        plt.savefig(outPath)
+        plt.close()

scxfxt-1.0.0.0/scxfxt/Algo/Hsml.py

@@ -0,0 +1,391 @@
+
+import torchvision.models as models
+from torch import nn
+import torch
+from torchvision.models import vgg19
+# Define VGG Loss
+class VGGLoss(nn.Module):
+    def __init__(self, weights_path=None,device =None):
+        super().__init__()
+        self.vgg = models.vgg19(pretrained=False).features[:35].eval().to(device)
+        if weights_path:
+            pretrained_weights = torch.load(weights_path)
+            self.vgg.load_state_dict(pretrained_weights, strict=False)
+        self.vgg[0] = nn.Conv2d(1, 64, kernel_size=3, stride=1, padding=1).to(device)
+        for param in self.vgg.parameters():
+            param.requires_grad = False
+        self.loss = nn.MSELoss()
+    def forward(self, input, target):
+        input_features = self.vgg(input)
+        target_features = self.vgg(target)
+        return self.loss(input_features, target_features)
+"""
+vgg_loss = VGGLoss(weights_path="/mnt/wtx_weather_forecast/scx/stat/sat/sat2radar/vgg19-dcbb9e9d.pth").to(device)
+""" 
+ 
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from segment_anything import sam_model_registry
+
+class SAMLoss(nn.Module):
+    def __init__(self, model_type='vit_b', checkpoint_path=None, input_size=1024):
+        """
+        SAM-based perceptual loss with resolution handling
+        
+        Args:
+            model_type (str): SAM model type (vit_b, vit_l, vit_h)
+            checkpoint_path (str): Path to SAM checkpoint weights
+            input_size (int): Target input size for SAM (default 1024)
+        """
+        super().__init__()
+        self.input_size = input_size        
+        # Initialize SAM model
+        self.sam = sam_model_registry[model_type](checkpoint=None)        
+        # Load pretrained weights if provided
+        if checkpoint_path:
+            state_dict = torch.load(checkpoint_path)
+            self.sam.load_state_dict(state_dict)        
+        # Use image encoder only and freeze parameters
+        self.image_encoder = self.sam.image_encoder.eval()
+        for param in self.image_encoder.parameters():
+            param.requires_grad = False            
+        # Define loss function
+        self.loss = nn.MSELoss()        
+        # Normalization parameters
+        self.register_buffer('mean', torch.tensor([0.485, 0.456, 0.406]).view(1, 3, 1, 1))
+        self.register_buffer('std', torch.tensor([0.229, 0.224, 0.225]).view(1, 3, 1, 1))
+        
+    def preprocess(self, x):
+        """
+        Preprocess input to match SAM requirements:
+        1. Convert to 3-channel if needed
+        2. Normalize using ImageNet stats
+        3. Resize to target size
+        """
+        if x.shape[1] == 1:
+            x = x.expand(-1, 3, -1, -1)  # More memory efficient than repeat        
+        # Normalize
+        x = (x - self.mean) / self.std        
+        # Resize
+        if x.shape[-2:] != (self.input_size, self.input_size):
+            x = F.interpolate(x, size=(self.input_size, self.input_size), 
+                             mode='bilinear', align_corners=False)        
+        return x        
+    def forward(self, input, target):
+        # Preprocess
+        input = self.preprocess(input)
+        target = self.preprocess(target)        
+        # Process in batches if needed
+        batch_size = 4  # Adjust based on your GPU memory
+        input_features = []
+        target_features = []        
+        with torch.no_grad():
+            for i in range(0, input.size(0), batch_size):
+                input_batch = input[i:i+batch_size]
+                target_batch = target[i:i+batch_size]                
+                input_features.append(self.image_encoder(input_batch))
+                target_features.append(self.image_encoder(target_batch))        
+        return self.loss(torch.cat(input_features), torch.cat(target_features))
+
+ 
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from segment_anything import sam_model_registry
+ 
+
+class SAMLoss(nn.Module):
+    def __init__(self, model_type='vit_b', checkpoint_path=None):
+        super().__init__()
+        # 初始化 SAM 并加载预训练权重
+        self.sam = sam_model_registry[model_type](checkpoint=None)
+        if checkpoint_path:
+            state_dict = torch.load(checkpoint_path)
+            self.sam.load_state_dict(state_dict)
+        
+        # 提取 image_encoder 并冻结
+        self.image_encoder = self.sam.image_encoder.eval()
+        for param in self.image_encoder.parameters():
+            param.requires_grad = False
+        
+        # 获取 patch_size（通常为16）
+        self.patch_size = self.image_encoder.patch_embed.proj.kernel_size[0]
+        
+        # 保存原始 pos_embed 供动态调整
+        self.original_pos_embed = self.image_encoder.pos_embed
+        
+        # 归一化参数
+        self.register_buffer('mean', torch.tensor([0.485, 0.456, 0.406]).view(1, 3, 1, 1))
+        self.register_buffer('std', torch.tensor([0.229, 0.224, 0.225]).view(1, 3, 1, 1))
+        
+        self.loss = nn.MSELoss()
+
+    def adjust_pos_embed(self, x):
+        """
+        动态调整 pos_embed 以匹配输入尺寸
+        Args:
+            x: 输入张量 (B, C, H, W)
+        """
+        B, _, H, W = x.shape
+        # 计算当前特征图的分辨率（H/patch_size, W/patch_size）
+        h, w = H // self.patch_size, W // self.patch_size
+        
+        # 如果 pos_embed 尺寸不匹配，则进行插值
+        if (h, w) != self.original_pos_embed.shape[1:3]:
+            # 插值 pos_embed 到目标尺寸 (1, h, w, C) -> (1, C, h, w) -> 插值 -> 恢复形状
+            pos_embed = self.original_pos_embed.permute(0, 3, 1, 2)  # (1, C, H_orig, W_orig)
+            pos_embed = F.interpolate(
+                pos_embed,
+                size=(h, w),
+                mode='bilinear',
+                align_corners=False
+            ).permute(0, 2, 3, 1)  # (1, h, w, C)
+            
+            # 用 nn.Parameter 包装调整后的 pos_embed
+            self.image_encoder.pos_embed = nn.Parameter(pos_embed, requires_grad=False)
+        else:
+            self.image_encoder.pos_embed = self.original_pos_embed
+    def preprocess(self, x):
+        """预处理：通道扩展 + 归一化 + 尺寸调整"""
+        if x.shape[1] == 1:
+            x = x.expand(-1, 3, -1, -1)  # 单通道→三通        
+        # 归一化
+        x = (x - self.mean) / self.std        
+        # 确保尺寸是 patch_size 的整数倍
+        B, C, H, W = x.shape
+        H_new = (H // self.patch_size) * self.patch_size
+        W_new = (W // self.patch_size) * self.patch_size
+        if H != H_new or W != W_new:
+            x = F.interpolate(
+                x,
+                size=(H_new, W_new),
+                mode='bilinear',
+                align_corners=False
+            )
+        return x
+    def forward(self, input, target):
+        # 预处理
+        input = self.preprocess(input)
+        target = self.preprocess(target)
+        
+        # 动态调整 pos_embed
+        self.adjust_pos_embed(input)
+        
+        # 计算特征
+        with torch.no_grad():
+            input_feat = self.image_encoder(input)
+            target_feat = self.image_encoder(target)
+        
+        return self.loss(input_feat, target_feat)
+
+# saml = SAMLoss(checkpoint_path = "/path/to/sam_vit_b_01ec64.pth.1").to(device)
+
+
+import torch
+import torch.nn as nn
+from torchvision import models
+class WeightedVGGLoss(nn.Module):
+    def __init__(self, weights_path=None, device=None, apply_weighting=True):
+        super().__init__()
+        self.device = device
+        self.apply_weighting = apply_weighting
+        self.vgg = models.vgg19(pretrained=False).features[:35].eval().to(device)        
+        if weights_path:
+            pretrained_weights = torch.load(weights_path)
+            self.vgg.load_state_dict(pretrained_weights, strict=False)
+        self.vgg[0] = nn.Conv2d(1, 64, kernel_size=3, stride=1, padding=1).to(device)
+        for param in self.vgg.parameters():
+            param.requires_grad = False
+        self.mse_loss = nn.MSELoss()
+    def custom_weight_4d(self, target):
+        EPS = 1e-6
+        segments = [
+            [10, 20, 0.5, 0.1], [20, 30, 0.4, 0.12],
+            [30, 40, 0.3, 0.14], [40, 50, 0.2, 0.16],
+            [50, 60, 0.1, 0.18], [60, 70, 0.05, 0.2]
+        ]
+        weights = torch.full_like(target, 0.1)
+        for start, end, steepness, base_w in segments:
+            mask = (target >= start) & (target < end)
+            x = ((target - start) / (end - start + EPS)).clamp(0, 1)
+            delta_w = (1.0 - base_w) * (0.7 - steepness)
+            seg_weights = base_w + delta_w * (x / (x + steepness*(1 - x) + EPS))
+            weights = torch.where(mask, seg_weights, weights)
+        weights[target >= 70] = 1.0
+        return weights
+    def forward(self, input, target):
+        input_features = self.vgg(input)
+        target_features = self.vgg(target)
+        if self.apply_weighting:
+            weights = self.custom_weight_4d(target_features.detach())
+            loss = torch.mean(weights * (input_features - target_features)**2)
+        else:
+            loss = self.mse_loss(input_features, target_features)            
+        return loss
+# vgg_loss = WeightedVGGLoss(weights_path="/mnt/wtx_weather_forecast/scx/stat/sat/sat2radar/vgg19-dcbb9e9d.pth",device=device).to(device)
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+class WeightedL1Loss(nn.Module):
+    def __init__(self, segment_ranges=None, device=None):
+        super().__init__()
+        self.device = device
+        self.segment_ranges = segment_ranges or [
+            (10, 20, 0.5, 0.1), (20, 30, 0.4, 0.12),
+            (30, 40, 0.3, 0.14), (40, 50, 0.2, 0.16),
+            (50, 60, 0.1, 0.18), (60, 70, 0.05, 0.2)
+        ]
+    def compute_weights(self, target):
+        EPS = 1e-6
+        weights = torch.full_like(target, 0.1)  # 默认权重0.1        
+        for start, end, steepness, base_w in self.segment_ranges:
+            mask = (target >= start) & (target < end)
+            x = ((target - start) / (end - start + EPS)).clamp(0, 1)
+            delta_w = (1.0 - base_w) * (0.7 - steepness)
+            seg_weights = base_w + delta_w * (x / (x + steepness*(1 - x) + EPS))
+            weights = torch.where(mask, seg_weights, weights)        
+        weights[target >= 70] = 1.0
+        return weights
+    def forward(self, output, target):
+        """
+        非降维(reduction='none')计算流程:
+        1. 计算逐元素L1损失
+        2. 生成动态权重矩阵
+        3. 返回加权损失张量（保持输入维度）
+        """
+        # 保持维度的L1损失 (B,C,H,W)
+        l1_loss = F.l1_loss(output, target, reduction='none')        
+        weights = self.compute_weights(target.detach())  # 切断梯度  
+        # weighted_loss = torch.mean(weights * l1_loss)   
+        weighted_loss =  weights * l1_loss  
+        return weighted_loss
+# L1loss = WeightedL1Loss(device= device)
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from segment_anything import sam_model_registry
+class SAMLoss1(nn.Module):
+    def __init__(self, model_type='vit_b', checkpoint_path=None):
+        super().__init__()
+        self.sam = sam_model_registry[model_type](checkpoint=None)
+        if checkpoint_path:
+            state_dict = torch.load(checkpoint_path)
+            self.sam.load_state_dict(state_dict)        
+        self.image_encoder = self.sam.image_encoder.eval()
+        for param in self.image_encoder.parameters():
+            param.requires_grad = False        
+        self.patch_size = self.image_encoder.patch_embed.proj.kernel_size[0]        
+        self.original_pos_embed = self.image_encoder.pos_embed        
+        self.register_buffer('mean', torch.tensor([0.485, 0.456, 0.406]).view(1, 3, 1, 1))
+        self.register_buffer('std', torch.tensor([0.229, 0.224, 0.225]).view(1, 3, 1, 1))        
+        self.loss = nn.MSELoss()
+    def adjust_pos_embed(self, x):
+        """
+        动态调整 pos_embed 以匹配输入尺寸
+        Args:
+            x: 输入张量 (B, C, H, W)
+        """
+        B, _, H, W = x.shape
+        h, w = H // self.patch_size, W // self.patch_size        
+        if (h, w) != self.original_pos_embed.shape[1:3]:
+            pos_embed = self.original_pos_embed.permute(0, 3, 1, 2)  # (1, C, H_orig, W_orig)
+            pos_embed = F.interpolate(
+                pos_embed,
+                size=(h, w),
+                mode='bilinear',
+                align_corners=False
+            ).permute(0, 2, 3, 1)  # (1, h, w, C)            
+            self.image_encoder.pos_embed = nn.Parameter(pos_embed, requires_grad=False)
+        else:
+            self.image_encoder.pos_embed = self.original_pos_embed
+    def preprocess(self, x):
+        """预处理：通道扩展 + 归一化 + 尺寸调整"""
+        if x.shape[1] == 1:
+            x = x.expand(-1, 3, -1, -1)  # 单通道→三通道
+        x = (x - self.mean) / self.std        
+        B, C, H, W = x.shape
+        H_new = (H // self.patch_size) * self.patch_size
+        W_new = (W // self.patch_size) * self.patch_size
+        if H != H_new or W != W_new:
+            x = F.interpolate(
+                x,
+                size=(H_new, W_new),
+                mode='bilinear',
+                align_corners=False
+            )
+        return x
+    def forward(self, input, target):
+        input = self.preprocess(input)
+        target = self.preprocess(target)        
+        self.adjust_pos_embed(input)        
+        with torch.no_grad():
+            input_feat = self.image_encoder(input)
+            target_feat = self.image_encoder(target)        
+        return self.loss(input_feat, target_feat)   
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from segment_anything import sam_model_registry
+class SAMLoss2(nn.Module):
+    def __init__(self, model_type='vit_b', checkpoint_path=None):
+        super().__init__()
+        self.sam = sam_model_registry[model_type](checkpoint=None)
+        if checkpoint_path:
+            state_dict = torch.load(checkpoint_path)
+            self.sam.load_state_dict(state_dict)        
+        self.image_encoder = self.sam.image_encoder.eval()
+        for param in self.image_encoder.parameters():
+            param.requires_grad = False        
+        self.patch_size = self.image_encoder.patch_embed.proj.kernel_size[0]       
+        self.original_pos_embed = self.image_encoder.pos_embed        
+        self.loss = nn.MSELoss()
+    def adjust_pos_embed(self, x):
+        """
+        动态调整 pos_embed 以匹配输入尺寸
+        Args:
+            x: 输入张量 (B, C, H, W)
+        """
+        B, _, H, W = x.shape
+        h, w = H // self.patch_size, W // self.patch_size        
+        if (h, w) != self.original_pos_embed.shape[1:3]:
+            pos_embed = self.original_pos_embed.permute(0, 3, 1, 2)  # (1, C, H_orig, W_orig)
+            pos_embed = F.interpolate(
+                pos_embed,
+                size=(h, w),
+                mode='bilinear',
+                align_corners=False
+            ).permute(0, 2, 3, 1)  # (1, h, w, C)            
+            self.image_encoder.pos_embed = nn.Parameter(pos_embed, requires_grad=False)
+        else:
+            self.image_encoder.pos_embed = self.original_pos_embed
+    def preprocess(self, x):
+        """预处理：仅通道扩展 + 尺寸调整（跳过归一化）"""
+        if x.shape[1] == 1:
+            x = x.expand(-1, 3, -1, -1)  # 单通道→三通道        
+        B, C, H, W = x.shape
+        H_new = (H // self.patch_size) * self.patch_size
+        W_new = (W // self.patch_size) * self.patch_size
+        if H != H_new or W != W_new:
+            x = F.interpolate(
+                x,
+                size=(H_new, W_new),
+                mode='bilinear',
+                align_corners=False
+            )
+        return x
+    def forward(self, input, target):
+        # 预处理（input和target均不归一化）
+        input = self.preprocess(input)
+        target = self.preprocess(target)        
+        # 动态调整 pos_embed
+        self.adjust_pos_embed(input)        
+        # 计算特征
+        with torch.no_grad():
+            input_feat = self.image_encoder(input)
+            target_feat = self.image_encoder(target)        
+        return self.loss(input_feat, target_feat)
+#saml = SAMLoss(checkpoint_path = "/mnt/wtx_weather_forecast/scx/stat/sat/sat2radar/sam_vit_b_01ec64.pth.1").to(device)

scxfxt-1.0.0.0/scxfxt/Algo/L2Loss.py

@@ -0,0 +1,10 @@
+import torch
+def L2loss(model,loss,lambda_reg =0.01):    
+    l2_reg = 0.0
+    for param in model.parameters():
+        l2_reg += torch.norm(param)**2  # 模型参数的平方和
+        print(l2_reg)
+    loss += lambda_reg * l2_reg
+    return loss
+
+