getter setter for weights

nnet_gpu/network.py

@@ -64,14 +64,14 @@ def predict(self, X_inp):
 	def train_on_batch(self, X_inp, labels):
 		X_inp = self.forward(X_inp.astype(self.dtype, copy=False))
 		grads = self.del_loss(X_inp, labels.astype(self.dtype, copy=False))
-		self.backprop(grads, do_d_inp=False)		# The gradients with input layer will NOT be calculated.
+		self.backprop(grads, do_d_inp=False)  # The gradients with input layer will NOT be calculated.
 		self.optimizer(self.sequence, self.learning_rate, self.beta)
 		return X_inp
 
 	def not_train_on_batch(self, X_inp, labels):
 		X_inp = self.forward(X_inp.astype(self.dtype, copy=False))
 		grads = self.del_loss(X_inp, labels.astype(self.dtype, copy=False))
-		grads = self.backprop(grads, do_d_inp=True)	# Gradients with input layer will be calculated.
+		grads = self.backprop(grads, do_d_inp=True)  # Gradients with input layer will be calculated.
 		return X_inp, grads
 
 	def fit(self, X_inp=None, labels=None, iterator=None, batch_size=1, epochs=1, validation_data=None, shuffle=True, accuracy_metric=True,
@@ -156,14 +156,33 @@ def validate(self, validation_data, batch_size, info_beta=0.2):
 		end = time.time()
 		print(f"\rValidation Accuracy: {(vacc / lnvx).get():.4f} - val_loss: {vloss.get():.4f} - Time: {end - start:.3f}s")
 
-	def save_weights(self, path):						# TODO - make a proper saving mechanism.
+	@property
+	def weights(self):
 		sv_me = []
 		for obj in self.sequence:
 			if obj.param > 0:
 				if isinstance(obj, layers.BatchNormalization):
 					sv_me.append((obj.weights, obj.biases, obj.moving_mean, obj.moving_var))
 				else:
 					sv_me.append((obj.weights, obj.biases))  # ,obj.w_m,obj.w_v,obj.b_m,obj.b_v))
+		return sv_me
+
+	@weights.setter
+	def weights(self, sv_me):
+		idx = 0
+		for obj in self.sequence:
+			if obj.param > 0:
+				if isinstance(obj, layers.BatchNormalization):
+					obj.kernels, obj.biases, obj.moving_mean, obj.moving_var = sv_me[idx]
+				else:
+					obj.kernels, obj.biases = sv_me[idx]
+					if isinstance(obj, layers.Conv2D):  # TODO - Verify isinstance works.
+						obj.init_back()
+				obj.weights = obj.kernels
+				idx += 1
+
+	def save_weights(self, path):  # TODO - make a proper saving mechanism.
+		sv_me = self.weights
 		if isinstance(path, str):
 			with open(path, 'wb') as f:
 				pickle.dump(sv_me, f)
@@ -176,19 +195,9 @@ def load_weights(self, path):
 				sv_me = pickle.load(f)
 		else:
 			sv_me = pickle.load(path)
-		idx = 0
-		for obj in self.sequence:
-			if obj.param > 0:
-				if isinstance(obj, layers.BatchNormalization):
-					obj.kernels, obj.biases, obj.moving_mean, obj.moving_var = sv_me[idx]
-				else:
-					obj.kernels, obj.biases = sv_me[idx]
-					if isinstance(obj, layers.Conv2D):		# TODO - Verify isinstance works.
-						obj.init_back()
-				obj.weights = obj.kernels
-				idx += 1
+		self.weights = sv_me
 
-	def summary(self):					# TODO - Show connections.
+	def summary(self):  # TODO - Show connections.
 		ipl = layers.InputLayer(self.sequence[0].input_shape)
 		reps = 90
 		print(chr(9149) * reps)