refactored forward_step

JayRnn.py

@@ -24,13 +24,13 @@ from rbm.train import train
 from rbm.status import Status
 
 # ── Hyper-parameters ──────────────────────────────────────────────────────────
-TEXT    = "  HALLO SUPER JENS. SUPPE!    " # the character sequence to learn (matches diagram example)
+TEXT    = "  HALLO SUPER JENS!" # the character sequence to learn (matches diagram example)
 WIN    = 3         # sliding-window width  (= N in the diagram)
 STRIDE = 1     	   # sliding-window step size
 UNITS  = 1
 H_SIZE = 128         # hidden units per RBM cell
-NUM_EPOCHS = 100
-NUM_ITERATIONS = 100
+NUM_EPOCHS = 1000
+NUM_ITERATIONS = 1
 
 TRAIN_PARAMS = TrainingParams(
 	learning_rate    = 0.25,
@@ -84,11 +84,12 @@ class RnnModel(Model):
 			_c = unit.forward(vc)
 			vc[1:, WIN*vocab_size():] = _c[0:-1,:]
 
-	def forward_step(self, _vc: Mat):
+	def forward_step(self, _v: Mat, _c: Mat) -> tuple[Mat, Mat]:
+		_vc = concat(_v.reshape([1, WIN*vocab_size()]), _c.reshape([1, H_SIZE]), axis=1)
 		for unit in self.units:
 			_c = unit.forward(_vc)
 			_vc = unit.reconstruct(_c)
-		return _vc
+		return split(_vc.flatten(), H_SIZE)
 
 	def forward(self, x: Mat):
 		pass
@@ -128,10 +129,9 @@ if __name__ == "__main__":
 	v_forward = seed.flatten()
 	context = c_train[0]
 	for i in range(len(TEXT)+5):
-		print(f"Forward {i:02d}: {vec2str(v_forward.reshape([WIN, vocab_size()]), axis=1)}")
-		vc_forward = concat(v_forward.reshape([1, WIN*vocab_size()]), context.reshape([1, H_SIZE]), axis=1)
-		vc_predict = model.forward_step(vc_forward)
-		v_predict, context = split(vc_predict.flatten(), H_SIZE)
+		v_mat = v_forward.reshape([WIN, vocab_size()])
+		print(f"Forward {i:02d}: {vec2str(v_mat, axis=1)}")
+		v_predict, context = model.forward_step(v_mat, context)
 		v_predict_clamp = clamp(v_predict.reshape([WIN, vocab_size()]), axis=1)
 		print(f"Predict {i:02d}: {vec2str(v_predict_clamp, axis=1)}")
 		v_forward = shift_left(v_predict_clamp.reshape([1, WIN*vocab_size()]), vocab_size())