add TTM model #20

frndtls · 2025-03-04T11:56:46Z

add TTM model (simplify some settings)
delete a print statement in dataloader.py

WenWeiTHU · 2025-03-04T12:00:31Z

layers/Mlp.py

+from layers.SelfAttention_Family import TTMGatedAttention
+
+
+class TTMmlp(nn.Module):


作为类名，MLP应该大写，文件名也要用大写MLP

WenWeiTHU · 2025-03-04T12:01:05Z

layers/Mlp.py

+class TTMmlp(nn.Module):
+    def __init__(self, in_features, out_features, e_factor, dropout):
+        super().__init__()
+        num_hidden = in_features * e_factor


e_factor用factor即可

WenWeiTHU · 2025-03-04T12:17:44Z

layers/Mlp.py

+from layers.SelfAttention_Family import TTMGatedAttention
+
+
+class TTMmlp(nn.Module):


可以加一些注释，例如特有的参数（e_factor）在论文中的含义等

WenWeiTHU · 2025-03-04T12:19:22Z

layers/Mlp.py

+        residual = x  # [B M N P]
+        x = self.norm(x)
+
+        assert self.mode in ["patch", "feature", "channel"]


关于mode的含义可提供一些注释

WenWeiTHU · 2025-03-04T12:19:48Z

layers/Mlp.py

+        elif self.mode == "channel":
+            x = x.permute(0, 3, 2, 1)  # [B M N P]
+        else:
+            pass


此处可rasie NotImplementedError

WenWeiTHU · 2025-03-04T12:21:28Z

run.py

@@ -81,6 +81,17 @@
    parser.add_argument('--kernel_size', type=int, default=25)
    parser.add_argument('--stride', type=int, default=8)

+    # TTM


关于这些ttm的参数可以给一些文中具体的--help解释

WenWeiTHU · 2025-03-04T12:22:55Z

models/ttm.py

+        return y_hat
+
+
+class TTMAPBlock(nn.Module):


可以加一些注释，特别是ttm特有的参数

WenWeiTHU · 2025-03-04T12:23:23Z

models/ttm.py

+
+
+class Model(nn.Module):
+    """


开源再标注一下预训练模型在huggingface的来源

WenWeiTHU · 2025-03-04T12:32:30Z

layers/SelfAttention_Family.py

@@ -140,3 +140,14 @@ def forward(self, queries, keys, values, attn_mask, n_vars=None, n_tokens=None,

        return self.out_projection(out), attn

+
+class TTMGatedAttention(nn.Module):


这个可以改名为TTMGatedLayer，forward流程与attention关联不大

WenWeiTHU · 2025-03-04T12:33:00Z

models/ttm.py

+    ):
+        super().__init__()
+        self.adapt_patch_level = adapt_patch_level
+        adaptive_patch_factor = 2**adapt_patch_level


这一块乘2的含义是啥，是否可以优化一下

run.py

WenWeiTHU · 2025-03-04T12:36:32Z

run.py

+    parser.add_argument("--e_factor", type=int, default=2)
+    parser.add_argument("--mode", type=str, default="mix_channel")
+    parser.add_argument("--AP_levels", type=int, default=0)
+    parser.add_argument("--head_dropout", type=float, default=0.2)


dropout如果影响不大，可以fix为一个默认参数到model的入参中，上面几个参数也类似

run.py

frndtls and others added 3 commits March 4, 2025 19:30

delete a print statement in data_loader.py

8988f2e

add TTM model

01de3df

Merge branch 'thuml:main' into main

10753cf

WenWeiTHU reviewed Mar 4, 2025

View reviewed changes

WenWeiTHU requested changes Mar 4, 2025

View reviewed changes

frndtls added 2 commits March 5, 2025 11:16

add doc info

c9fbb25

Merge branch 'main' of https://github.com/frndtls/OpenLTM

9e8e00b

frndtls closed this Mar 5, 2025

frndtls reopened this Mar 5, 2025

frndtls added 3 commits March 5, 2025 11:41

fix a argument error in run.py

27e5cc7

delete Mlp.py

8cbb7bd

add MLP.py

2cf2c83

WenWeiTHU requested changes Mar 5, 2025

View reviewed changes

run.py Show resolved Hide resolved

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

add TTM model #20

add TTM model #20

frndtls commented Mar 4, 2025

WenWeiTHU Mar 4, 2025

WenWeiTHU Mar 4, 2025

WenWeiTHU Mar 4, 2025

WenWeiTHU Mar 4, 2025

WenWeiTHU Mar 4, 2025

WenWeiTHU Mar 4, 2025

WenWeiTHU Mar 4, 2025

WenWeiTHU Mar 4, 2025

WenWeiTHU Mar 4, 2025

WenWeiTHU Mar 4, 2025

WenWeiTHU Mar 4, 2025

		from layers.SelfAttention_Family import TTMGatedAttention


		class TTMmlp(nn.Module):

		@@ -140,3 +140,14 @@ def forward(self, queries, keys, values, attn_mask, n_vars=None, n_tokens=None,

		return self.out_projection(out), attn


		class TTMGatedAttention(nn.Module):

add TTM model #20

Are you sure you want to change the base?

add TTM model #20

Conversation

frndtls commented Mar 4, 2025

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment