1 files changed, 45 insertions, 0 deletions

diff --git a/util/files.py b/util/files.py
new file mode 100644
index 0000000..2712525
--- /dev/null
+++ b/util/files.py
@@ -0,0 +1,45 @@
+from pathlib import Path
+import json
+
+from models.clip.embeddings import ManagedCLIPTextEmbeddings
+from models.clip.tokenizer import MultiCLIPTokenizer
+
+from safetensors import safe_open
+
+
+def load_config(filename):
+    with open(filename, 'rt') as f:
+        config = json.load(f)
+
+    args = config["args"]
+
+    if "base" in config:
+        args = load_config(Path(filename).parent / config["base"]) | args
+
+    return args
+
+
+def load_embeddings_from_dir(tokenizer: MultiCLIPTokenizer, embeddings: ManagedCLIPTextEmbeddings, embeddings_dir: Path):
+    if not embeddings_dir.exists() or not embeddings_dir.is_dir():
+        return []
+
+    filenames = [filename for filename in embeddings_dir.iterdir() if filename.is_file()]
+    tokens = [filename.stem for filename in filenames]
+
+    new_ids: list[list[int]] = []
+    new_embeds = []
+
+    for filename in filenames:
+        with safe_open(filename, framework="pt", device="cpu") as file:
+            embed = file.get_tensor("embed")
+
+        added = tokenizer.add_multi_tokens(filename.stem, embed.shape[0])
+        new_ids.append(added)
+        new_embeds.append(embed)
+
+    embeddings.resize(len(tokenizer))
+
+    for (new_id, embeds) in zip(new_ids, new_embeds):
+        embeddings.add_embed(new_id, embeds)
+
+    return tokens, new_ids