Upload modeling_move.py with huggingface_hub

modeling_move.py

@@ -0,0 +1,75 @@
+"""Transformers-compatible wrapper for harpertoken/move.
+
+The model classifies KTH actions from 16 frames with max-pooling over time,
+not a transformer. This subclasses PreTrainedModel so the weights serialize
+honestly as config.json plus model.safetensors and load through AutoModel.
+Importing this module registers the architecture.
+"""
+
+import torch
+import torch.nn as nn
+from transformers import AutoConfig, AutoModel, PretrainedConfig, PreTrainedModel
+
+
+class MoveConfig(PretrainedConfig):
+    model_type = "move_temporal"
+
+    def __init__(self, num_labels=6, dropout=0.25, **kwargs):
+        super().__init__(**kwargs)
+        self.num_labels = num_labels
+        self.dropout = dropout
+
+
+LABELS = ["walking", "jogging", "running", "boxing", "handwaving", "handclapping"]
+
+
+class MoveTemporal(PreTrainedModel):
+    config_class = MoveConfig
+    base_model_prefix = "move"
+
+    # Transformers 5 reads this mapping while finalising a load. Nothing here
+    # is tied, but the attribute has to exist or from_pretrained raises
+    # AttributeError before the weights are placed.
+    all_tied_weights_keys = {}
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.cnn = nn.ModuleDict(
+            {
+                "conv1": nn.Conv2d(1, 32, 3, padding=1),
+                "conv2": nn.Conv2d(32, 64, 3, padding=1),
+                "fc1": nn.Linear(64 * 8 * 8, 128),
+            }
+        )
+        self.pool = nn.MaxPool2d(2, 2)
+        self.adapt = nn.AdaptiveAvgPool2d((8, 8))
+        self.head = nn.Linear(128, config.num_labels)
+        self.drop = nn.Dropout(config.dropout)
+
+    def forward(self, clip):
+        b, t, _, h, w = clip.shape
+        x = clip.view(b * t, 1, h, w)
+        x = self.pool(torch.relu(self.cnn["conv1"](x)))
+        x = self.pool(torch.relu(self.cnn["conv2"](x)))
+        x = self.adapt(x)
+        feats = torch.relu(self.cnn["fc1"](x.view(b * t, -1))).view(b, t, -1)
+        return self.head(self.drop(feats.amax(1)))
+
+    def predict_label(self, clip):
+        """clip: 16x64x64 float array in [0, 1]. Returns the action."""
+        self.eval()
+        x = torch.as_tensor(clip, dtype=torch.float32)
+        if x.dim() == 3:
+            x = x.unsqueeze(0).unsqueeze(0)
+        elif x.dim() == 4:
+            x = x.unsqueeze(0)
+        with torch.no_grad():
+            return LABELS[int(self.forward(x).argmax(1)[0])]
+
+
+def register():
+    AutoConfig.register(MoveConfig.model_type, MoveConfig, exist_ok=True)
+    AutoModel.register(MoveConfig, MoveTemporal, exist_ok=True)
+
+
+register()