patchfm 2.2.5__tar.gz → 3.0.0__tar.gz

{patchfm-2.2.5/src/patchfm.egg-info → patchfm-3.0.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: patchfm
-Version: 2.2.5
+Version: 3.0.0
 Summary: a Foundation Model for Univariate Time Series Forecasting
 Author-email: Samy-Melwan Vilhes <samy-melwan.vilhes@insa-rouen.fr>
 License: MIT License

{patchfm-2.2.5 → patchfm-3.0.0}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "patchfm"
-version = "2.2.5"
+version = "3.0.0"
 authors = [
   { name="Samy-Melwan Vilhes", email="samy-melwan.vilhes@insa-rouen.fr" },
 ]

{patchfm-2.2.5 → patchfm-3.0.0}/src/patchfm/configs/model_config.py

@@ -8,7 +8,6 @@ class PatchFMConfig:
     d_model: int = 2048
     n_heads: int = 32
     n_layers_encoder: int = 6
-    use_xsa: bool = True
     quantiles: list[float] = field(default_factory=lambda: [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9])
     
     compile: bool = True

{patchfm-2.2.5 → patchfm-3.0.0}/src/patchfm/inference/forecaster.py

@@ -18,17 +18,12 @@ class Forecaster(nn.Module):
         self.d_model = config["d_model"]
         self.n_heads = config["n_heads"]
         self.n_layers_encoder = config["n_layers_encoder"]
-        self.use_xsa = config["use_xsa"]
         self.quantiles = config["quantiles"]
         self.n_quantiles = len(self.quantiles)
         self.max_patches = self.max_seq_len // self.patch_len
 
-        if not self.use_xsa:
-            print("Loading base model from HuggingFace Hub...")
-            base_model = PatchFM.from_pretrained("vilhess/PatchFM")
-        else:
-            print("Loading XSA model from HuggingFace Hub...")
-            base_model = PatchFM.from_pretrained("vilhess/PatchFM-XSA")
+        print("Loading base model from HuggingFace Hub...")
+        base_model = PatchFM.from_pretrained("vilhess/PatchFM")
 
         self._init_from_base(base_model)
 
@@ -53,7 +48,6 @@ class Forecaster(nn.Module):
             d_model=self.d_model, 
             n_heads=self.n_heads, 
             n_layers=self.n_layers_encoder,
-            use_xsa=self.use_xsa
         )
         self.proj_output = ResidualBlock(
             in_dim=self.d_model, 
@@ -167,7 +161,7 @@ class Forecaster(nn.Module):
         bs, ws = x.size()
 
         if flip_equivariance:
-            print("Flip equivariance enabled: forecast = (f(x) - f(-x)) / 2. This requires multiplying by 2 the batch size (Reverso: Efficient Time Series Foundation Models for Zero-shot Forecasting).")
+
             bs = context.size(0)
             x_flipped = -x
             concat_x = torch.cat([x, x_flipped], dim=0)

{patchfm-2.2.5 → patchfm-3.0.0}/src/patchfm/inference/modules.py

@@ -185,7 +185,7 @@ class ResidualBlock(nn.Module):
         return out
 
 class MultiHeadAttention(nn.Module):
-    def __init__(self, d_model, n_heads, last=False, use_xsa=False):
+    def __init__(self, d_model, n_heads, last=False):
         super().__init__()
         assert d_model%n_heads==0, f"d_model ({d_model}) must be divisible by n_heads ({n_heads})"
 
@@ -204,7 +204,6 @@ class MultiHeadAttention(nn.Module):
         self.v_cache = None
 
         self.last = last
-        self.use_xsa = use_xsa
 
     def forward(self, q):
         bs, context, dim = q.size()
@@ -240,13 +239,6 @@ class MultiHeadAttention(nn.Module):
 
         values = nn.functional.scaled_dot_product_attention(q, k, v, is_causal=is_causal)
 
-        if self.use_xsa:
-            if v.size(2) != values.size(2):
-                v = v[:, :, -values.size(2) :, :]
-                
-            vn = torch.nn.functional.normalize(v, dim=-1)
-            values = values - (values * vn).sum(dim=-1, keepdim=True) * vn
-
         values = values.transpose(1, 2).reshape(bs, -1, dim)
         values = self.out_proj(values)
         return values
@@ -275,10 +267,10 @@ class FeedForward(nn.Module):
 
 
 class TransformerEncoderLayer(nn.Module):
-    def __init__(self, d_model, n_heads, last=False, use_xsa=False):
+    def __init__(self, d_model, n_heads, last=False):
         super().__init__()
         self.ln1 = nn.LayerNorm(d_model)
-        self.attn = MultiHeadAttention(d_model=d_model, n_heads=n_heads, last=last, use_xsa=use_xsa)
+        self.attn = MultiHeadAttention(d_model=d_model, n_heads=n_heads, last=last)
         self.ln2 = nn.LayerNorm(d_model)
         self.ff = FeedForward(d_model=d_model)
     
@@ -289,15 +281,15 @@ class TransformerEncoderLayer(nn.Module):
         return out
     
 class TransformerEncoder(nn.Module):
-    def __init__(self, d_model, n_heads, n_layers, use_xsa=False):
+    def __init__(self, d_model, n_heads, n_layers):
         super().__init__()
         self.layers = nn.ModuleList(
             [
-                TransformerEncoderLayer(d_model=d_model, n_heads=n_heads, use_xsa=use_xsa)
+                TransformerEncoderLayer(d_model=d_model, n_heads=n_heads)
                 for _ in range(n_layers-1)
             ]
         )
-        self.layers.append(TransformerEncoderLayer(d_model=d_model, n_heads=n_heads, last=True, use_xsa=use_xsa))
+        self.layers.append(TransformerEncoderLayer(d_model=d_model, n_heads=n_heads, last=True))
         self.norm = nn.LayerNorm(d_model)
 
     def forward(self, x):
@@ -314,7 +306,6 @@ class PatchFM(nn.Module, PyTorchModelHubMixin):
         self.d_model = config["d_model"]
         self.n_heads = config["n_heads"]
         self.n_layers_encoder = config["n_layers_encoder"]
-        self.use_xsa = config["use_xsa"]
         self.quantiles = config["quantiles"]
         self.n_quantiles = len(self.quantiles)
 
@@ -329,7 +320,6 @@ class PatchFM(nn.Module, PyTorchModelHubMixin):
             d_model=self.d_model, 
             n_heads=self.n_heads, 
             n_layers=self.n_layers_encoder, 
-            use_xsa=self.use_xsa
         )
         self.proj_output = ResidualBlock(
             in_dim=self.d_model, 

{patchfm-2.2.5 → patchfm-3.0.0/src/patchfm.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: patchfm
-Version: 2.2.5
+Version: 3.0.0
 Summary: a Foundation Model for Univariate Time Series Forecasting
 Author-email: Samy-Melwan Vilhes <samy-melwan.vilhes@insa-rouen.fr>
 License: MIT License