make h5merger more robust

h5_merger.py

@@ -2943,8 +2943,16 @@ def parse_input():
     elif ' ' in args.h5_tables:
         args.h5_tables = args.h5_tables[0].split()
 
-    if type(args.h5_tables) == str:
-        args.h5_tables = glob(args.h5_tables)
+
+    if type(args.h5_tables) == str or len(args.h5_tables) == 1:
+        if len(args.h5_tables) == 1:
+            args.h5_tables = args.h5_tables[0]
+        if ' ' not in args.h5_tables:
+            args.h5_tables = glob(args.h5_tables)
+        else:
+            args.h5_tables = args.h5_tables.split()
+        if len(args.h5_tables) == 0:
+            sys.exit(f"ERROR: parsing of input h5s failed")
     elif type(args.h5_tables) == list and len(args.h5_tables) == 1:
         args.h5_tables = glob(args.h5_tables[0])
     elif type(args.h5_tables) == list:

neural_networks/plots/confusion_matrix_hacky.py

@@ -0,0 +1,77 @@
+from pathlib import Path
+import sys
+import os
+
+SCRIPT_DIR = Path(os.path.dirname(os.path.abspath(__file__)))
+sys.path.append(os.path.dirname(SCRIPT_DIR))
+import matplotlib.pyplot as plt
+import numpy as np
+import torch
+from sklearn.metrics import confusion_matrix, ConfusionMatrixDisplay
+
+
+
+@torch.no_grad()
+def get_confusion_matrix(predictor, dataloader, mean, std, thresholds):
+    conf_matrix_dict = {thres: np.zeros((2, 2)) for thres in thresholds}
+    thresholds_tensor = torch.tensor(thresholds)
+    for i, (img, label) in enumerate(dataloader):
+        data = predictor.prepare_batch(img, mean=mean, std=std)
+        pred = torch.sigmoid(predictor.model(data)).to("cpu")
+        preds_thres = pred >= thresholds_tensor
+        for i, thres in enumerate(thresholds):
+            conf_matrix_dict[thres] += confusion_matrix(
+                label, preds_thres[:, i], labels=[0, 1]
+            )
+
+    return conf_matrix_dict
+
+
+def plot_conf_matrices(savedir, conf_matrix_dict):
+    os.makedirs(savedir, exist_ok=True)
+    for thres, conf_matrix in conf_matrix_dict.items():
+
+        disp = ConfusionMatrixDisplay(
+            conf_matrix / np.sum(conf_matrix, axis=1, keepdims=True),  # Normalization
+            display_labels=["stop", "continue"]
+        )
+
+        fig, ax = plt.subplots()
+        disp.plot(cmap='Blues', ax=ax)
+        im = ax.images[0]
+        im.set_clim(0, 1)
+
+        ax.set_xlabel("Predicted Label", fontsize=16)
+        ax.set_ylabel("True Label", fontsize=16)
+
+        ax.tick_params(axis='both', which='major', labelsize=16)
+
+        for text in disp.text_.ravel():
+            text.set_fontsize(16)
+
+        colorbar = im.colorbar
+        colorbar.set_label("Accuracy", fontsize=16)
+        colorbar.ax.tick_params(labelsize=16)
+
+        fig.tight_layout(pad=2.0)
+
+        plt.savefig(f"{savedir}/confusion_thres_{thres:.3f}.png", dpi=150)
+        plt.close(fig)
+
+
+
+if __name__ == "__main__":
+    # Latest model
+    model_name = "surf/dinov2_09739_rotations"
+    TESTING = True
+    architecture_name = "surf/TransferLearning"
+    DEVICE = "cuda"
+    thresholds = [0.2, 0.3, 0.4, 0.5]
+    dict_fname = f"plots/dinov2_09739_rotations/conf_matrix.npydict"
+    print(Path(dict_fname).exists(), dict_fname)
+    if TESTING:
+        conf_matrix_dict = np.load(dict_fname, allow_pickle=True)[()]
+
+    print(conf_matrix_dict)
+
+    plot_conf_matrices('/home/jurjen/Documents/ELAIS/paperplots', conf_matrix_dict, thresholds)