feat: initialize repository and add project-level gitignore

app.py

@@ -41,7 +41,6 @@ from mambaeye.scan import generate_scan_positions
 from mambaeye.positional_encoding import sinusoidal_position_encoding_2d
 from mamba_ssm.utils.generation import InferenceParams
 
-TARGET_CANVAS_SIZE = 512
 PATCH_SIZE = 16
 CATEGORIES = ResNet50_Weights.IMAGENET1K_V1.meta["categories"]
 
@@ -149,26 +148,21 @@ def format_predictions(probs_np):
 def preprocess_image(image_arr):
     img = Image.fromarray(image_arr).convert("RGB")
     width, height = img.size
-    
-    ratio = min(TARGET_CANVAS_SIZE / width, TARGET_CANVAS_SIZE / height)
-    new_w = int(width * ratio)
-    new_h = int(height * ratio)
-    
-    img_resized = img.resize((new_w, new_h), Image.Resampling.LANCZOS)
     totensor = T.ToTensor()
-    img_tensor = totensor(img_resized)
-    
-    canvas = torch.zeros(3, TARGET_CANVAS_SIZE, TARGET_CANVAS_SIZE, dtype=torch.float32)
-    x_offset = (TARGET_CANVAS_SIZE - img_tensor.shape[1]) // 2
-    y_offset = (TARGET_CANVAS_SIZE - img_tensor.shape[2]) // 2
+    img_tensor = totensor(img)
+    canvas_size = max(width, height)
+    canvas = torch.zeros(3, canvas_size, canvas_size, dtype=torch.float32)
+    x_offset = (canvas_size - img_tensor.shape[1]) // 2
+    y_offset = (canvas_size - img_tensor.shape[2]) // 2
     
     canvas[:, x_offset : x_offset + img_tensor.shape[1], y_offset : y_offset + img_tensor.shape[2]] = img_tensor
     
-    return canvas, x_offset, y_offset, new_h, new_w
+    return canvas, x_offset, y_offset, height, width
 
 def extract_patch(canvas_tensor, px, py):
-    px = max(0, min(px, TARGET_CANVAS_SIZE - PATCH_SIZE))
-    py = max(0, min(py, TARGET_CANVAS_SIZE - PATCH_SIZE))
+    canvas_size = canvas_tensor.shape[1]
+    px = max(0, min(px, canvas_size - PATCH_SIZE))
+    py = max(0, min(py, canvas_size - PATCH_SIZE))
     patch = canvas_tensor[:, px : px + PATCH_SIZE, py : py + PATCH_SIZE]
     return patch.flatten()