def generate_inf_cases(train,
seed,
board_size=25,
min_dens=0.01,
max_dens=0.99,
warm_up=5,
min_delta=1,
max_delta=5,
dtype=np.int,
return_one_but_last=False):
rs = np.random.RandomState(seed)
zer = np.zeros(shape=(board_size, board_size), dtype=dtype)
while True:
density = rs.uniform(min_dens, max_dens)
start = rs.choice([1, 0],
size=(board_size, board_size),
p=[density, 1.0 - density])
for _ in range(warm_up):
start = life_step(start)
delta = rs.randint(min_delta, max_delta + 1)
stop = start.copy()
one_but_last = None
for _ in range(delta):
one_but_last = stop
stop = life_step(stop)
if not (stop == zer).all():
if return_one_but_last:
yield delta, one_but_last, stop
elif train:
yield delta, start, stop
else:
yield delta, stop
def test_sim_toad():
toad_1 = np.array([[0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0],
[0, 0, 1, 1, 1, 0], [0, 1, 1, 1, 0, 0],
[0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0]])
toad_2 = np.array([[0, 0, 0, 0, 0, 0], [0, 0, 0, 1, 0, 0],
[0, 1, 0, 0, 1, 0], [0, 1, 0, 0, 1, 0],
[0, 0, 1, 0, 0, 0], [0, 0, 0, 0, 0, 0]])
assert (life_step(toad_1) == toad_2).all()
assert (life_step(toad_2) == toad_1).all()
def test_generate_train_set():
xs = list(generate_train_set(10, 234))
assert len(xs) == 10
for delta, start, stop in xs:
for i in range(delta):
start = life_step(start)
assert (start == stop).all()
def test_sim_block():
block = np.array([
[0, 0, 0, 0],
[0, 1, 1, 0],
[0, 1, 1, 0],
[0, 0, 0, 0],
])
assert (life_step(block) == block).all()
def test_step_back_all_3x3(alg_class):
alg = alg_class(tile_graph)
for Aorig in tqdm(generate_all(3, 3)):
X = np.copy(Aorig)
X = life_step(X)
A = alg.step_back(X)
sc = score(1, A, X)
assert sc == 1.0
def train(self, delta, start, stop):
X = start
m = X.shape[0]
n = X.shape[1]
for s in range(delta):
Y = life_step(X) if s < delta - 1 else stop
for i in range(m):
for j in range(n):
a_id = self.__get_tile_id(X, i, j)
b_id = self.__get_tile_id(Y, i, j)
self.trans[a_id][b_id] += 1
def preprocess():
# Tiles - all possible titles 3x3
T = list(generate_all(3, 3))
assert (len(T) == 512)
# Backward possibilities
# dict: central bit -> list of possible prev tiles 3x3
B = [[], []]
for i, tile in enumerate(T):
next_tile = life_step(tile)
B[int(next_tile[1][1])].append(i)
# Fun fact:
# >>> len(B[0])
# 372
# >>> len(B[1])
# 140
# Matrix horiz[i,j] - true if tile j can be put horizontally on the right side of tile i.
# 0 0 1 0 1 0 0 0 1 0
# 0[1]0 + 1[0]0 = 0[1|0]0
# 0 0 0 0 0 1 0 0 0 1
horiz = np.zeros((len(T), len(T)), dtype=np.bool)
# Matrix verti[i,j] - true if tile j can be put vertically under tile i.
verti = np.zeros((len(T), len(T)), dtype=np.bool)
# Diagonal relationships -- they don't seem to be needed.
#diago_se = np.zeros((len(T), len(T)), dtype=np.bool)
#diago_sw = np.zeros((len(T), len(T)), dtype=np.bool)
for i, x in enumerate(T):
for j, y in enumerate(T):
# Check if left tile is compatible with the right tile.
# I.e., left tile's right side is equal to right tile's left side.
horiz[i, j] = (x[:, (1, 2)] == y[:, (0, 1)]).all()
# Check if upper tile is compatible with the lower tile.
# I.e., upper tile's lower side is equal to lower tile's upper side.
verti[i, j] = (x[(1, 2), :] == y[(0, 1), :]).all()
# Diagonal relationships -- they don't seem to be needed.
#diago_se[i, j] = (x[(1, 2), (1, 2)] == y[(0, 1), (0, 1)]).all()
#diago_sw[i, j] = (x[(1, 2), (0, 1)] == y[(0, 1), (1, 2)]).all()
return T, B, horiz, verti
def test_step_back_simple(alg_class):
alg = alg_class(tile_graph)
block = np.array([
[0, 0, 0, 0],
[0, 1, 1, 0],
[0, 1, 1, 0],
[0, 0, 0, 0],
])
A = alg.step_back(block)
B = life_step(A)
assert (B == block).all()
toad_1 = np.array([[0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0],
[0, 0, 1, 1, 1, 0], [0, 1, 1, 1, 0, 0],
[0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0]])
toad_2 = np.array([[0, 0, 0, 0, 0, 0], [0, 0, 0, 1, 0, 0],
[0, 1, 0, 0, 1, 0], [0, 1, 0, 0, 1, 0],
[0, 0, 1, 0, 0, 0], [0, 0, 0, 0, 0, 0]])
A = alg.step_back(toad_2)
assert (A == toad_1).all()
def score(delta, start, stop):
X = start
for i in range(delta):
X = life_step(X)
return np.count_nonzero(X == stop) / stop.size
def likely_starts(delta, stop):
# Just test a couple of "likely" starting boards and pick the best one.
starts = [np.zeros_like(stop), stop, life_step(stop)]
scores = [score(delta, start, stop) for start in starts]
idx = np.argmax(scores)
return starts[idx]
def sim_density(delta, stop):
# If stop is a still board, then return that. Otherwise, return zeros.
n = life_step(stop)
return stop if (stop == n).all() else np.zeros_like(stop)
def train(
gen_arch,
fwd_arch,
device,
writer,
batchSize,
niter,
lr,
beta1,
dry_run,
outf,
workers=1,
start_iter=0,
gen_path=None,
fwd_path=None,
ngpu=1,
nz=100,
epoch_samples=64*100,
learn_forward=True,
sigmoid=True):
os.makedirs(outf, exist_ok=True)
gen_path = gen_path if start_iter == 0 else os.path.join(outf, f'netG_epoch_{start_iter-1}.pth')
fwd_path = fwd_path if start_iter == 0 else os.path.join(outf, f'netF_epoch_{start_iter-1}.pth')
# Prediction threshold
pred_th = 0.5 if sigmoid else 0.0
dataset = DataGenerator(823131 + start_iter, sigmoid)
dataloader = torch.utils.data.DataLoader(dataset, batch_size=batchSize,
shuffle=False, num_workers=int(workers))
val_size = 1024
val_set = bitmap.generate_test_set(set_size=val_size, seed=9568382)
deltas_val, stops_val = cnnify_batch(zip(*val_set))
ones_val = np.ones_like(deltas_val)
noise_val = torch.randn(val_size, nz, 1, 1, device=device)
netG = get_generator_net(gen_arch).to(device)
init_model(netG, gen_path)
netF = get_forward_net(fwd_arch).to(device)
init_model(netF, fwd_path)
criterion = nn.BCELoss()
fixed_noise = torch.randn(batchSize, nz, 1, 1, device=device)
fixed_ones = np.ones((batchSize,), dtype=np.int)
# setup optimizer
optimizerD = optim.Adam(netF.parameters(), lr=lr, betas=(beta1, 0.999))
optimizerG = optim.Adam(netG.parameters(), lr=lr, betas=(beta1, 0.999))
scores = []
for i in range(5):
noise = torch.randn(val_size, nz, 1, 1, device=device)
one_step_pred_batch = (netG(torch.Tensor(stops_val).to(device), noise) > pred_th).cpu()
model_scores = scoring.score_batch(ones_val, np.array(one_step_pred_batch, dtype=np.bool), stops_val)
scores.append(model_scores)
zeros = np.zeros_like(one_step_pred_batch, dtype=np.bool)
zeros_scores = scoring.score_batch(ones_val, zeros, stops_val)
scores.append(zeros_scores)
best_scores = np.max(scores, axis=0)
print(
f'Mean error one step: model {1 - np.mean(model_scores)}, zeros {1 - np.mean(zeros_scores)}, ensemble {1 - np.mean(best_scores)}')
#for epoch in range(opt.niter):
epoch = start_iter
samples_in_epoch = 0
samples_before = start_iter * epoch_samples
for j, data in enumerate(dataloader, 0):
i = start_iter * epoch_samples // batchSize + j
############################
# (1) Update F (forward) network -- in the original GAN, it's a "D" network (discriminator)
# Original comment: Update D network: maximize log(D(x)) + log(1 - D(G(z)))
###########################
# train with real starting board -- data set provides ground truth
netF.zero_grad()
start_real_cpu = data[0].to(device)
stop_real_cpu = data[1].to(device)
batch_size = start_real_cpu.size(0)
output = netF(start_real_cpu)
errD_real = criterion(output, stop_real_cpu)
if learn_forward:
errD_real.backward()
D_x = (output.round().eq(stop_real_cpu)).sum().item() / output.numel()
# train with fake -- use simulator (life_step) to generate ground truth
# TODO: replace with fixed forward model (should be faster, in batches and on GPU)
noise = torch.randn(batch_size, nz, 1, 1, device=device)
fake = netG(stop_real_cpu, noise)
fake_np = (fake > pred_th).detach().cpu().numpy()
fake_next_np = life_step(fake_np)
fake_next = torch.tensor(fake_next_np, dtype=torch.float32).to(device)
output = netF(fake.detach())
errD_fake = criterion(output, fake_next)
if learn_forward:
errD_fake.backward()
D_G_z1 = (output.round().eq(fake_next)).sum().item() / output.numel()
errD = errD_real + errD_fake
if learn_forward:
optimizerD.step()
# just for reporting...
true_stop_np = (stop_real_cpu > pred_th).detach().cpu().numpy()
fake_scores = scoring.score_batch(fixed_ones, fake_np, true_stop_np, show_progress=False)
fake_mae = 1 - fake_scores.mean()
fake_density = fake_np.mean()
############################
# (2) Update G network: maximize log(D(G(z)))
###########################
netG.zero_grad()
output = netF(fake)
errG = criterion(output, stop_real_cpu)
errG.backward()
D_G_z2 = (output.round().eq(fake_next)).sum().item() / output.numel()
optimizerG.step()
samples_in_epoch += batch_size
s = samples_before + samples_in_epoch
writer.add_scalar('Loss/forward', errD.item(), i)
writer.add_scalar('Loss/gen', errG.item(), i)
writer.add_scalar('MAE/train', fake_mae.item(), i)
print('[%d/%d][%d] Loss_F: %.4f Loss_G: %.4f fwd acc(real): %.2f fwd acc(fake): %.2f / %.2f, fake dens: %.2f, MAE: %.4f'
% (epoch, start_iter+niter, i,
errD.item(), errG.item(), D_x, D_G_z1, D_G_z2, fake_density, fake_mae))
if samples_in_epoch >= epoch_samples:
"""
multi_step_pred_batch = predict(netG, deltas_val, stops_val, fixed_noise)
multi_step_mean_err = 1 - np.mean(scoring.score_batch(deltas_val, np.array(multi_step_pred_batch, dtype=np.bool), stops_val))
"""
one_step_pred_batch = (netG(torch.Tensor(stops_val).to(device), noise_val) > pred_th).detach().cpu().numpy()
one_step_mean_err = 1 - np.mean(scoring.score_batch(ones_val, np.array(one_step_pred_batch, dtype=np.bool), stops_val))
print(f'Mean error: one step {one_step_mean_err}')
writer.add_scalar('MAE/val', one_step_mean_err, epoch)
vutils.save_image(start_real_cpu,
'%s/real_samples.png' % outf,
normalize=True)
fake = netG(stop_real_cpu, fixed_noise).detach()
vutils.save_image(fake,
'%s/fake_samples_epoch_%03d.png' % (outf, epoch),
normalize=True)
grid = vutils.make_grid(start_real_cpu)
writer.add_image('real', grid, epoch)
grid = vutils.make_grid(fake)
writer.add_image('fake', grid, epoch)
# do checkpointing
torch.save(netG.state_dict(), '%s/netG_epoch_%d.pth' % (outf, epoch))
torch.save(netF.state_dict(), '%s/netF_epoch_%d.pth' % (outf, epoch))
epoch += 1
samples_in_epoch = 0
if epoch - start_iter >= niter:
break
if dry_run:
break
return one_step_mean_err
