add details for ttp

Author: likyoo

configs/_base_/models/ttp_vit-sam-l.py

@@ -1,6 +1,5 @@
 # model settings
 norm_cfg = dict(type='SyncBN', requires_grad=True)
-bit_norm_cfg = dict(type='LN', requires_grad=True)
 fpn_norm_cfg = dict(type='mmpretrain.LN2d', requires_grad=True)
 data_preprocessor = dict(
     type='DualInputSegDataPreProcessor',

configs/ttp/README.md

@@ -1,6 +1,6 @@
 # TTP
 
-Time Travelling Pixels: Bitemporal Features Integration with Foundation Model for Remote Sensing Image Change Detection
+[Time Travelling Pixels: Bitemporal Features Integration with Foundation Model for Remote Sensing Image Change Detection](https://arxiv.org/abs/2312.16202)
 
 ## Introduction
 
@@ -42,7 +42,7 @@ pip install peft
 
 | Method | Backbone  | Crop Size | Lr schd | Precision | Recall | F1-Score | IoU  |                            config                            |
 | :----: | --------- | :-------: | :-----: | :-------: | :----: | :------: | :--: | :----------------------------------------------------------: |
-|  TTP   | ViT-SAM-L |  512x512  |  300e   |           |        |          |      | [config](https://github.com/likyoo/open-cd/blob/main/configs/ttp/ttp_vit-sam-l_512x512_300e_levircd.py) |
+|  TTP   | ViT-SAM-L |  512x512  |  300e   |   93.0    |  91.7  |   92.1   | 85.6 | [config](https://github.com/likyoo/open-cd/blob/main/configs/ttp/ttp_vit-sam-l_512x512_300e_levircd.py) |
 
 
 - All metrics are based on the category "change".

configs/ttp/ttp_vit-sam-l_512x512_300e_levircd.py

@@ -53,4 +53,5 @@
     ),
 ]
 
-train_cfg = dict(type='EpochBasedTrainLoop', max_epochs=max_epochs, val_interval=5)
+train_cfg = dict(type='EpochBasedTrainLoop', max_epochs=max_epochs, val_interval=5)
+default_hooks = dict(checkpoint=dict(interval=5))

opencd/models/utils/ttp_layer.py

@@ -54,7 +54,7 @@ def __init__(self,
             dropout_layer=dict(type='DropPath', drop_prob=drop_path_rate),
             act_cfg=act_cfg)
 
-        if self.window_size > 0:
+        if self.window_size > 0: # TODO: Maybe it should be ``self.window_size == 0`` here.
             in_channels = embed_dims * 2
             self.down_channel = torch.nn.Conv2d(in_channels, 1, kernel_size=1, stride=1, bias=False)
             self.down_channel.weight.data.fill_(1.0 / in_channels)
@@ -89,7 +89,7 @@ def forward(self, x):
 
         x = self.ffn(self.ln2(x), identity=x)
         # time phase fusion
-        if self.window_size > 0:
+        if self.window_size > 0: # TODO: Maybe it should be ``self.window_size == 0`` here.
             x = einops.rearrange(x, 'b h w d -> b d h w')  # 2B, C, H, W
             x0 = x[:x.size(0)//2]
             x1 = x[x.size(0)//2:]  # B, C, H, W