adding compute_baselines improvements

PlaceRec/Methods/hybridnet.py

@@ -186,7 +186,7 @@ def __call__(self, tensor):
 
 class HybridNet(SingleStageBaseModelWrapper):
     def __init__(self, pretrained: bool = True):
-        name = "alexnet_hybridnet"
+        name = "hybridnet"
         weight_path = join(config["weights_directory"], name + ".ckpt")
         self.model = HybridNetModel()
         if pretrained:
@@ -204,10 +204,4 @@ def __init__(self, pretrained: bool = True):
                 model=self.model, preprocess=preprocess, name=name, weight_path=None
             )
         # some layers not implemented on metal
-        self.set_device(self.device)
-
-    def set_device(self, device: str) -> None:
-        if "mps" in device:
-            device = "cpu"
-        self.device = device
-        self.model.to(device)
+        self.set_device("cpu")

compute_baselines.py

@@ -11,12 +11,11 @@
 from PlaceRec.Evaluate import Eval
 
 datasets = ["Pitts30k_Val"]
-methods = ["ResNet34_NetVLAD", "ResNet34_MixVPR", "ResNet34_GeM", "ResNet34_ConvAP"]
-
-columns = ["method_name", "model_memory"]
+methods = ["MixVPR", "DinoSalad", "EigenPlaces", "SFRS"]
 
 batch_size = 24
 
+columns = ["method_name", "model_memory"]
 for ds in datasets:
     columns += [f"{ds}_total_memory", f"{ds}_R1", f"{ds}_map_memory"]
 df = pd.DataFrame(columns=columns)

config.yaml

@@ -1,9 +1,9 @@
 # Desktop
-datasets_directory: "/media/oliver/9e8e8649-c1f2-4524-9025-f2c751d67f57/home/oliver/Documents/datasets/Datasets"
-weights_directory: "/media/oliver/9e8e8649-c1f2-4524-9025-f2c751d67f57/home/oliver/Documents/VisualLoc_Weights"
+#datasets_directory: "/media/oliver/9e8e8649-c1f2-4524-9025-f2c751d67f57/home/oliver/Documents/datasets/Datasets"
+#weights_directory: "/media/oliver/9e8e8649-c1f2-4524-9025-f2c751d67f57/home/oliver/Documents/VisualLoc_Weights"
 # Laptop
-#weights_directory: "/Users/olivergrainge/Documents/github/VisualLoc_Weights"
-#datasets_directory: "/Users/olivergrainge/Documents/github/Datasets"
+weights_directory: "/Users/olivergrainge/Documents/github/VisualLoc_Weights"
+datasets_directory: "/Users/olivergrainge/Documents/github/Datasets"
 #HPC
 #datasets_directory: "/scratch/oeg1n18/datasets"
 #weights_directory: "/scratch/oeg1n18/VisualLoc_Weights"
@@ -16,10 +16,10 @@ weights_directory: "/media/oliver/9e8e8649-c1f2-4524-9025-f2c751d67f57/home/oliv
 run: 
   datasets: ["Pitts30k"]
   methods: ["MixVPR"]
-  num_workers: 16
+  num_workers: 0
   pin_memory: False
-  batchsize: 198
-  device: "cuda"
+  batchsize: 32
+  device: "mps"
 
 eval: 
   datasets: ["pitts30k"]
@@ -39,7 +39,7 @@ train:
   eval_distance: "cosine"
   finetune: False
   checkpoint: False
-  num_workers: 16
+  num_workers: 0
 
   # optimization
   optimizer: "adam"

pruning_plots/memory/R1_total_memory.py

@@ -1,32 +1,35 @@
 import matplotlib.pyplot as plt
 import pandas as pd
-from brokenaxes import brokenaxes
-import matplotlib.gridspec as gridspec
 import numpy as np
 
-aggregation = 1.5
+# Define parameters
+aggregation = 0.75
 dataset = "Pitts30k_Val"
 
+# Read data from CSV files
 df = pd.read_csv("../results.csv")
 baselines = pd.read_csv("../baselines.csv")
 
 # Replace the method_name entries
 df["method_name"] = df["method_name"].replace("ResNet50_GeM", "GeM")
 
+# Filter the dataframe based on the specified aggregation
 df = df[["method_name", f"{dataset}_total_memory", f"{dataset}_R1", "agg_rate"]]
 df = df[df["agg_rate"] == aggregation]
 
-# Create a figure with broken x-axis
-fig = plt.figure(figsize=(7, 4))
-gs = gridspec.GridSpec(1, 1)
-bax = brokenaxes(xlims=((0, 400), (600, 700)), subplot_spec=gs[0], d=0.05)
+# Filter the baselines dataframe for the same dataset
+baselines = baselines[["method_name", f"{dataset}_total_memory", f"{dataset}_R1"]]
 
-# Manually plotting
+# Create a figure with normal axis
+fig, ax = plt.subplots(figsize=(7, 4))
+
+# Define color palette
 color_palette = plt.get_cmap("viridis")(np.linspace(0, 1, df["method_name"].nunique()))
 method_colors = dict(zip(df["method_name"].unique(), color_palette))
 
+# Plot data on the normal axis
 for name, group in df.groupby("method_name"):
-    bax.plot(
+    ax.plot(
         group[f"{dataset}_total_memory"],
         group[f"{dataset}_R1"],
         label=name,
@@ -35,21 +38,22 @@
         color=method_colors[name],
     )
 
-# Add baseline points
-for i in range(len(baselines)):
-    row = baselines.iloc[i]
-    memory = float(row[f"{dataset}_total_memory"])
-    name = str(row["method_name"])
-    if "dino" not in name.lower():
+# Plot baseline data points
+for name, group in baselines.groupby("method_name"):
+    if name == "CosPlace":
         continue
-    r1 = row[f"{dataset}_R1"]
-    bax.plot(memory, r1, "o", color="red")  # Red dot for baseline
-    bax.text(memory - 5, r1 - 0.02, name, ha="center", va="bottom", fontsize=11)
-
-bax.set_title(f"R@1 vs Memory Consumption for Pitts30k Val")
-bax.set_xlabel("Memory (Mb)", fontsize=13, labelpad=20)
-bax.set_ylabel("R@1", fontsize=13, labelpad=42)
-bax.grid(True)
-bax.legend(title="VPR Model", loc="center right")
-plt.savefig("total_mem.jpg", dpi=500, format="jpg")
+    ax.scatter(
+        group[f"{dataset}_total_memory"],
+        group[f"{dataset}_R1"],
+        label=f"{name} (baseline)",
+        marker="x",
+        color="black",
+    )
+
+# Add legend and labels
+ax.legend()
+ax.set_xlabel(f"{dataset}_total_memory")
+ax.set_ylabel(f"{dataset}_R1")
+
+# Show plot
 plt.show()

pruning_plots/preprocess_results.py

@@ -4,7 +4,18 @@
 
 df = pd.read_csv("../results.csv")
 
-print(df["weight_path"])
+# Function to extract sparsity value
+def extract_sparsity(path):
+    parts = path.split("_")
+    num = float(parts[5])
+    return num
+
+
+# Apply the function to the weight_path column
+df["sparsity"] = df["weight_path"].apply(extract_sparsity)
+
+# Convert the extracted values to numeric type
+df["sparsity"] = pd.to_numeric(df["sparsity"])
 df["agg_rate"] = df["weight_path"].str.extract("agg_([0-9]+\.?[0-9]*)")
 df["agg_rate"] = pd.to_numeric(df["agg_rate"])