add script to convert all models to torchscript

scripts/convert_to_torchscript.py

@@ -0,0 +1,78 @@
+"""Convert wsinfer models to torchscript."""
+
+import hashlib
+from pathlib import Path
+
+import torch
+from wsinfer import get_model_weights, list_all_models_and_weights
+
+
+def sha256sum(path) -> str:
+    """Calculate MD5 hash of a file."""
+    sha256 = hashlib.sha256()
+    with open(path, "rb") as f:
+        while True:
+            data = f.read(1024 * 64)  # 64 kb
+            if not data:
+                break
+            sha256.update(data)
+    return sha256.hexdigest()
+
+
+def get_torchscript_frozen(arch: str, name: str):
+    """Convert a wsinfer pytorch model to a frozen TorchScript model."""
+
+    print("Loading model...")
+    weights_obj = get_model_weights(architecture=arch, name=name)
+    del arch, name
+
+    width = weights_obj.patch_size_pixels
+    example_input = torch.ones(1, 3, width, width)
+
+    model = weights_obj.load_model()
+    model.eval()
+
+    print("Converting model to TorchScript...")
+    try:
+        model_jit = torch.jit.script(model, example_inputs=[(example_input,)])
+    except RuntimeError:
+        print("!!!!!!")
+        print("torch.jit.script failed!")
+    try:
+        model_jit = torch.jit.trace(model, example_input)
+    except Exception:
+        return None, None
+    model_jit_frozen = torch.jit.freeze(model_jit)
+
+    print("Validating that jit model gives same output as original...")
+    with torch.no_grad():
+        output: torch.Tensor = model(example_input)
+        output_jit: torch.Tensor = model_jit(example_input)
+        assert torch.allclose(output, output_jit)
+        output_jit_frozen: torch.Tensor = model_jit_frozen(example_input)
+        assert torch.allclose(output, output_jit_frozen)
+
+    return model_jit_frozen, weights_obj
+
+
+if __name__ == "__main__":
+    outdir = Path("jit-frozen-models")
+    outdir.mkdir(exist_ok=True)
+
+    for model, weights in list_all_models_and_weights():
+        print(f"{model} with {weights} weights")
+        frozen_model, w = get_torchscript_frozen(model, weights)
+        if frozen_model is None:
+            print("Failed to convert.")
+            print("-" * 40)
+            continue
+        p = Path(w.url_file_name)
+        p = outdir / f"{p.stem}-jit-frozen.pt"
+        print(f"Saving to {p}")
+        torch.jit.save(frozen_model, p)
+        print("Calculating sha256 hash")
+        sha = sha256sum(p)
+        newname = p.parent / f"{p.stem}-{sha[:8]}.pt"
+        print(f"Renaming to {newname}")
+        p.rename(newname)
+        print("-" * 40)