pyRBM/JayRnn.py

#!/usr/bin/env python3
"""Character-level RNN-RBM (unrolled) — see docs/Rnn.drawio.png.

Diagram recap
    t=0:  v[0]={0}  v[1:N]=[' ',' ','J']  W[0]    → h
    t=1:  v[0]=h₀   v[1:N]=[' ','J','A']  W[1]    → h
    ...
    t=M-1: v[0]=h_{M-2}  v[1:N]=['J','A','Y']  W[M-1]  → h

v[0]   = recurrent context (previous hidden state, or zeros at t=0)
v[1:N] = N_WIN one-hot-encoded characters concatenated (the sliding window)
W[t]   = RBM weight matrix for time step t  (one per step → unrolled mode)
"""
import sys
import os

sys.path.insert(0, os.path.join(os.path.dirname(__file__), 'src'))

from rbm.entity import EntityParams, TrainingParams, Entity
from rbm.matrix import np, Mat
from model.model import Model
from stack.rnn_helper import vocab_size, shift_left, clamp, concat, split, vec2idx, idx2ch, ch2idx, idx2vec
from rbm.train import train
from rbm.status import Status

# ── Hyper-parameters ──────────────────────────────────────────────────────────
TEXT    = "0123456789" # the character sequence to learn (matches diagram example)
WIN    = 3         # sliding-window width  (= N in the diagram)
STRIDE = 1     	   # sliding-window step size
UNITS  = 3
H_SIZE = 32         # hidden units per RBM cell
NUM_EPOCHS = 1000
NUM_ITERATIONS = 1

TRAIN_PARAMS = TrainingParams(
	learning_rate    = 0.04,
	momentum         = 0.5,
	num_epochs       = NUM_EPOCHS,
	do_rao_blackwell = True,
	num_gibbs_samples= 1
)
#WIN=3
#STRIDE=1
#UNIT=3
#WIN |   |
#U0: "JAY IS COOL"
#U1: "AY IS COOL "
#U2: "Y IS COOL  "

def vec2str(mat: Mat, axis=0):
	return idx2ch(vec2idx(mat, axis))

def str2vec(ch_str: str) -> Mat:
	return idx2vec(ch2idx(ch_str))

def to_window(text: str):
	win_text = []
	text_padded = text + ' '*WIN
	for i in range(len(TEXT)):
		win_text.append(text_padded[i:i+WIN])

	return win_text

def to_batch(_win_str: list[str]):
	_batch = np.zeros([len(TEXT), WIN*vocab_size()])
	for i, win in enumerate(_win_str):
		_batch[i, :] = str2vec(win).flatten()
	return _batch

def batch_delay(_batch: Mat, delay=0):
	result = _batch
	if delay > 0:
		result[0:-delay] = _batch[delay:]
		result[-delay:] = np.zeros([delay, _batch.shape[1]])
	return result

def vc2char(_vc: Mat):
	_v, _ = split(_vc.reshape(1, WIN*vocab_size()+H_SIZE), H_SIZE, axis=1)
	return vec2str(_v.reshape([WIN, vocab_size()]), axis=1)

class RnnModel(Model):
	def __init__(self, name: str, work_dir: str = '.'):
		super().__init__(name, work_dir)

		self.units: list[Entity] = []
		for index in range(UNITS):
			unit = Entity((WIN*vocab_size() + H_SIZE, H_SIZE), EntityParams(do_gaussian_visible=False, do_gaussian_hidden=False), training_params=TRAIN_PARAMS, index=index)
			self.units.append(unit)

	def train(self, vc: Mat, status: Status = None):
		_v, _c = split(vc, H_SIZE, axis=1)
		for delay, unit in enumerate(self.units):
			_vd = batch_delay(_v, delay)
			_vc = concat(_vd, _c, axis=1)
			for i in range(vc.shape[0]):
				print(f"train: {unit.name}:{vc2char(_vc[i,:])}")
			train(unit, _vc, status)
			# For the next unit: Update context portion of vc
			_c = unit.forward(_vc)


	def forward_step(self, _vc: Mat):
		for unit in self.units:
			_c = unit.forward(_vc)
			_vc = unit.reconstruct(_c)
			_v, _ = split(_vc, H_SIZE, axis=1)
			_v_cl = clamp(_v.reshape([WIN, vocab_size()]), axis=1).reshape([1, WIN * vocab_size()])
			_v = shift_left(_v_cl, 1)
			_vc = concat(_v, _c, axis=1)
			print(f"forward_step: {unit.name}:{vc2char(_vc)}")
		return _vc

	def forward(self, x: Mat):
		pass


if __name__ == "__main__":
	if 1:
		# Test of helper function
		# ToDo: move to tests
		vec = idx2vec(ch2idx("JENS"))
		vec = clamp(vec, axis=1)
		indices = vec2idx(vec, axis=1)
		ch_str = idx2ch(indices)
		print(ch_str)

		win_text = to_window(TEXT)
		print(win_text)

		batch = to_batch(win_text)
		print(batch)


	model = RnnModel(name='JayRnn', work_dir='results')
	model.init(0.01)
	model.load()

	# vc contains vis + context
	# context will be updated after training
	c_train = np.zeros([batch.shape[0], H_SIZE])
	vc_train = concat(batch, c_train, axis=1)
#	model.train(vc_train, status=Status())
#	model.save()

	# test the model
	seed_str = '1'
	seed_padded = '^'*(WIN-len(seed_str)) + seed_str
	v_test = str2vec(seed_padded).reshape([1, WIN*vocab_size()])
	c_test = np.zeros([1, H_SIZE])
	vc_test = concat(v_test, c_test, axis=1)
	for i in range(len(TEXT) + 5):
		vc_test = model.forward_step(vc_test)
	if 0:
		for i in range(len(TEXT)+5):
			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_mat = v_predict.reshape([WIN, vocab_size()])
			v_predict_clamp = clamp(v_predict_mat, axis=1)
			v_predict_str = vec2str(v_predict_clamp, axis=1)
			text_predict += v_predict_str[WIN-1]
			print(f"Predict {i:02d}: {v_predict_str}")
			v_forward = shift_left(v_predict_clamp.reshape([1, WIN*vocab_size()]), vocab_size())

		print(seed_padded[0:WIN-1] + text_predict)