- fixed armadillo data read

- added shape to layer
- changed construction of layer from Stack factory
- fixed rms_error_accu scaling
- RBM: added image display using opencv
- improved Status print

pyproject.toml

@@ -12,7 +12,9 @@ authors = [
     { name = "Jens Ahrensfeld" }
 ]
 dependencies = [
-    "numpy"
+    "numpy",
+    "opencv-python",
+    "opencv-contrib-python"
 ]
 readme = "README.md"
 

src/rbm/helper.py

@@ -21,6 +21,6 @@ def rms_error(d_err: np.ndarray):
 
 def rms_error_accu(d_err: np.ndarray):
 	d_err_squared = d_err * d_err
-	s =  np.sum(d_err_squared) / len(d_err)
+	s =  np.sum(d_err_squared) / d_err.size
 	return s
 

src/rbm/layer.py

@@ -7,9 +7,10 @@ from status import Status
 from cd_train import cd_jens
 
 class Layer:
-	def __init__(self, name: str, dim: tuple[int, int], params: RbmParams):
+	def __init__(self, name: str, shape: tuple[int, int, int, int], params: RbmParams):
 		self.name = name
-		self.state = RbmState.from_layer_params(dim)
+		self.shape = shape
+		self.state = RbmState.from_layer_params((shape[0]*shape[1]+shape[2], shape[3]))
 		self.params = params
 		self.state_filename = f"{self.name}_state.npz"
 

src/rbm/rbm.py

@@ -1,29 +1,7 @@
 import os.path
-
 import numpy as np
-from layer import Layer
+import cv2 as cv
-from stack import Stack, StackType, StackException
 from stack_factory import StackFactory
-from params import RbmParams
-
-def test():
-	params = RbmParams()
-	stack = Stack(StackType.Deep, "Stack")
-	dims = [(2, 3), (3, 4), (4, 5)]
-
-	for n, dim in enumerate(dims):
-		layer = Layer(f"Layer-{n}", dim, params)
-		stack.append(layer)
-
-	stack.init(std=0.1)
-	stack.state_save()
-
-	lay0 = stack.layers[0]
-	lay1 = stack.from_name("Layer-1")
-	lay2 = stack.from_index(2)
-	stack.remove(lay2)
-	stack.remove(lay1)
-	stack.remove(lay0)
 
 def read_armadillo(filename: str) -> np.ndarray:
 	result = None
@@ -37,14 +15,18 @@ def read_armadillo(filename: str) -> np.ndarray:
 		print(f"shape: {shape}")
 
 		result = np.zeros(shape=shape, dtype=np.float64)
-		line = fp.readline().replace("\n", '').split('  ')
-		line = line[1:]
-		data = [float(s) for s in line]
 		for row in range(shape[0]):
+			line = fp.readline().replace("\n", '').split('  ')
+			line = line[1:]
+			data = [float(s) for s in line]
 			result[row, :] = data
 
 	return result
 
+def cv_show(vec: np.array, shape):
+	img = cv.Mat(np.resize(vec, shape))
+	img_n = cv.normalize(src=img, dst=None, alpha=255, beta=0, norm_type=cv.NORM_MINMAX, dtype=cv.CV_8U)
+	cv.imshow(f"training samples", img_n)
 
 def main(prj_name: str = "test"):
 	work_dir = "../../results"
@@ -63,12 +45,21 @@ def main(prj_name: str = "test"):
 	training_data_path = os.path.join(prj_root, f"{prj_name}.training.dat")
 	batch = read_armadillo(training_data_path)
 
+	# Shape of training vector
+	shape = stack.from_index(0).shape[0:2] + (1,)
+
 	# Train
 	stack.train(batch)
 
 	# Save state
 	stack.state_save()
 
+	for n in range(batch.shape[0]):
+		cv_show(batch[n,:], shape)
+		cv.waitKeyEx(200)
+
 if __name__ == "__main__":
 	main("norb_small_16h")
+	cv.destroyAllWindows()
+
 	print("Test: [passed]")

src/rbm/stack_factory.py

@@ -48,7 +48,7 @@ class StackFactory:
 			params = RbmParams.from_dict(layer_params, params_version)
 
 			# Create layer
-			layer_obj = Layer(f"{layer_name}-{layer_id}", (num_visible_x*num_visible_y+num_context, num_hidden), params)
+			layer_obj = Layer(f"{layer_name}-{layer_id}", (num_visible_x, num_visible_y, num_context, num_hidden), params)
 
 			# Add layer to stack
 			obj.append(layer_obj)

src/rbm/status.py

@@ -14,5 +14,8 @@ class Status:
 				for key in status.keys():
 					value = status[key]['value']
 					unit = status[key]['unit']
-					print(f"{key}              : {value}{unit}")
+					if isinstance(value, float):
+						print(f"{key}              : {value:0.6f}{unit}")
+					else:
+						print(f"{key}              : {value}{unit}")
 		return True