Add object detection, tracking, mobile notifications, and search to any security camera.
less bad func names
roryclear0c3e7cd4dcba
diff --git a/clip.py b/clip.pyindex c1bfb71..de63fac 100644--- a/clip.py+++ b/clip.py@@ -74,8 +74,8 @@ class CachedCLIPSearch:# for BEAMif prewarm:- precompute_embeddings(self.model, Tensor.rand((1, 3, 224, 224), dtype=dtypes.float32))- precompute_embeddings(self.model, Tensor.rand((16, 3, 224, 224), dtype=dtypes.float32))+ precompute_embeddings_jit(self.model, Tensor.rand((1, 3, 224, 224), dtype=dtypes.float32))+ precompute_embeddings_jit(self.model, Tensor.rand((16, 3, 224, 224), dtype=dtypes.float32))def find_object_folders(self, base_path="data/cameras"):object_folders = []@@ -138,11 +138,11 @@ class CachedCLIPSearch:batch_np = np.stack(batch_np)if len(batch_np) == batch_size:- embeddings = precompute_embeddings(self.model, Tensor(batch_np)).numpy()+ embeddings = precompute_embeddings_jit(self.model, Tensor(batch_np)).numpy()else:embeddings = []for j in range(len(batch_np)):- emb = precompute_embedding_bs1(self.model, Tensor(batch_np[j:j+1])).numpy()+ emb = precompute_embedding_jit_bs1(self.model, Tensor(batch_np[j:j+1])).numpy()embeddings.append(emb)emb_ret.extend(embeddings)path_ret.extend(batch_paths)@@ -154,6 +154,7 @@ class CachedCLIPSearch:os.makedirs(os.path.dirname(cache_file), exist_ok=True)with open(cache_file, "wb") as f: pickle.dump({"embeddings": folder_embeddings, "paths": folder_paths}, f)return emb_ret, path_ret+def precompute_embedding_bs1_np(self, img):if type(img) == bytes: # todo, another level up?img = cv2.imdecode(np.frombuffer(img, np.uint8), cv2.IMREAD_COLOR)@@ -162,14 +163,14 @@ class CachedCLIPSearch:img = cv2.imread(img)img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)img = [preprocess(img)]- ret = precompute_embedding_bs1(self.model, Tensor(img)).numpy()+ ret = precompute_embedding_jit_bs1(self.model, Tensor(img)).numpy()return ret@TinyJit-def precompute_embeddings(model, x): return precompute_embedding(model, x)+def precompute_embeddings_jit(model, x): return precompute_embedding(model, x)@TinyJit-def precompute_embedding_bs1(model, x): return precompute_embedding(model, x)+def precompute_embedding_jit_bs1(model, x): return precompute_embedding(model, x)def precompute_embedding(model, x):x = model.visual_conv1(x)@@ -217,22 +218,4 @@ def preprocess(img):img = img.astype(np.float32) / 255.0img = (img - 0.5) / 0.5img = np.transpose(img, (2, 0, 1))- return img--'''-if __name__ == "__main__":- searcher = CachedCLIPSearch(- model_name="ViT-L-14",- pretrained_name="laion2b_s32b_b82k"- )-- while True:- object_folders = searcher.find_object_folders("data/cameras")- print(f"{datetime.now()}: Found {len(object_folders)} object folders")-- for folder in object_folders:- searcher.precompute_embeddings(folder)-- print(f"{datetime.now()}: Sleeping for 15 mins...")- time.sleep(900)-'''\ No newline at end of file+ return img\ No newline at end of file