Add seed to kmeans. Log seeds.

experiments/kmeans.py

@@ -29,6 +29,7 @@ def parse_args():
         help="distance metric",
     )
     parser.add_argument("--rank", type=int, default=1, help="decomposition rank")
+    parser.add_argument("--seed", type=int, default=0, help="seed for random")
 
     return parser.parse_args()
 
@@ -165,6 +166,7 @@ def compute_inter_distance(centroids):
         rank=args.rank,
         num_clusters=num_clusters,
         dist=args.distance,
+        seed=args.seed,
     )
 
     # Print the inertia list for each iteration

experiments/run.py

@@ -5,6 +5,7 @@
 from kmeans import custom_kmeans, decompose
 import wandb
 import random
+import os
 
 
 def parse_args():
@@ -54,6 +55,7 @@ def parse_args():
     parser.add_argument(
         "--initialization", type=int, default=100, help="number of initialization"
     )
+    parser.add_argument("--output", type=str, default="./", help="output directory")
 
     return parser.parse_args()
 
@@ -102,8 +104,9 @@ def run(data_combined, data_processed, method, rank):
         if inertia < inertia_min:
             inertia_min = inertia
             ARI = adjusted_rand_score(ground_truth, labels)
+            seed = i
 
-    return ARI, inertia_min
+    return ARI, inertia_min, seed
 
 
 def compute_std(x):
@@ -134,13 +137,16 @@ def topk_positions(x, y, k=10):
 def main():
     name = f"distance = {args.distance} rank = {args.rank} clusters-std = [{args.clusters_std_min} {args.clusters_std_max}] noise-std = [{args.noise_std_min} {args.noise_std_max}]"
     wandb.init(name=name, project=f"CORING_CustomKmeans", config=vars(args))
+    if not os.path.isdir(args.output):
+        os.makedirs(args.outputI)
 
     ARIs_tensor = []
-    inertias_tensor = []
     ARIs_matrix = []
-    inertias_matrix = []
+    seeds_tensor = []
+    seeds_matrix = []
+    i = 0
 
-    for i in tqdm(range(args.runs)):
+    while i < args.runs:
         # Create dataset
         clusters_std = random.uniform(args.clusters_std_min, args.clusters_std_max)
         noise_std = random.uniform(args.noise_std_min, args.noise_std_max)
@@ -153,7 +159,8 @@ def main():
         closely_similar_filters = add_noise_to_centroids(
             initial_filters, args.satellites, noise_std
         )
-        np.save(f"{i}.npy", [initial_filters, closely_similar_filters])
+        path = os.path.join(args.output, f"{i}.npy")
+        np.save(path, [initial_filters, closely_similar_filters])
 
         # Combine both initial_filters and closely_similar_filters
         data_combined = np.vstack((initial_filters, closely_similar_filters))
@@ -166,13 +173,13 @@ def main():
 
         # Apply K-means
         try:
-            ARI_tensor, inertia_tensor = run(
+            ARI_tensor, inertia_tensor, seed_tensor = run(
                 data_combined=data_combined,
                 data_processed=data_processed_tensor,
                 method="tensor",
                 rank=args.rank,
             )
-            ARI_matrix, inertia_matrix = run(
+            ARI_matrix, inertia_matrix, seed_matrix = run(
                 data_combined=data_combined,
                 data_processed=data_processed_matrix,
                 method="matrix",
@@ -185,13 +192,17 @@ def main():
                     "inertia_tensor": inertia_tensor,
                     "ARI_matrix": ARI_matrix,
                     "inertia_matrix": inertia_matrix,
+                    "seed_tensor": seed_tensor,
+                    "seed_matrix": seed_matrix,
                 }
             )
 
             ARIs_tensor.append(ARI_tensor)
-            inertias_tensor.append(inertia_tensor)
             ARIs_matrix.append(ARI_matrix)
-            inertias_matrix.append(inertia_matrix)
+            seeds_tensor.append(seed_tensor)
+            seeds_matrix.append(seed_matrix)
+
+            i += 1
 
         except Exception as error:
             print(error)
@@ -206,7 +217,7 @@ def main():
     top_positions = topk_positions(ARIs_tensor, ARIs_matrix)
     for i, position in enumerate(top_positions):
         print(
-            f"Position {i + 1}: ARIs_tensor[{position}] = {ARIs_tensor[position]}, ARIs_matrix[{position}] = {ARIs_matrix[position]}"
+            f"Position {i + 1}: ARIs_tensor[{position}] = {ARIs_tensor[position]}, ARIs_matrix[{position}] = {ARIs_matrix[position]}, seed_tensor = {seed_tensor}, seed_matrix = {seed_matrix}"
         )