From: François Fleuret Date: Sun, 9 Jul 2023 10:58:42 +0000 (+0200) Subject: Update. X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=commitdiff_plain;h=a8f039a9b491b1b4b47f6b9f8123c7261e758661;p=picoclvr.git Update. --- diff --git a/expr.py b/expr.py index 8690504..f294d68 100755 --- a/expr.py +++ b/expr.py @@ -45,10 +45,6 @@ def generate_program(nb_variables, length): s = "" variables = set() - # We take length itself half of the time, and uniform between 1 - # and length otherwise. The actual length can be slightly greater - - length = min(length, 1 + torch.randint(length * 2, (1,)).item()) while len(s) < length: v = random_var(nb_variables=nb_variables) s += v + "=" + random_expr(variables, budget=20) + ";" @@ -70,10 +66,15 @@ def generate_sequences(nb, nb_variables=5, length=20): assert nb_variables <= 26 sequences = [] result_max = 99 + for n in range(nb): + # We take length itself half of the time, and uniform between + # 1 and length otherwise. The actual length can be slightly + # greater + + l = min(length, 1 + torch.randint(length * 2, (1,)).item()) result = None while result == None or max(result.values()) > result_max: - l = length p, v = generate_program(nb_variables, l) v = ", ".join(['"' + v + '": ' + v for v in v]) ldict = {} diff --git a/tasks.py b/tasks.py index 04b8f84..4d7e90e 100755 --- a/tasks.py +++ b/tasks.py @@ -840,9 +840,8 @@ class Expr(Task): for batch in tqdm.tqdm( input.split(self.batch_size), dynamic_ncols=True, desc=desc ): - if split == "train": - last = (batch != self.filler).max(0).values.nonzero().max() + 3 - batch = batch[:, :last] + last = (batch != self.filler).max(0).values.nonzero().max() + 3 + batch = batch[:, :last] yield batch def vocabulary_size(self): @@ -866,7 +865,8 @@ class Expr(Task): def compute_nb_correct(input): result = input.clone() - ar_mask = (result == self.space).long().cumsum(dim=1).clamp(max=1) + s = (result == self.space).long() + ar_mask = (s.cumsum(dim=1) - s).clamp(min=0, max=1) result = (1 - ar_mask) * result + ar_mask * self.filler masked_inplace_autoregression( model, diff --git a/world.py b/world.py index 5ba0f36..fb5d5c7 100755 --- a/world.py +++ b/world.py @@ -67,7 +67,7 @@ def random_scene(): colors = [ (1.00, 0.00, 0.00), (0.00, 1.00, 0.00), - (0.00, 0.00, 1.00), + (0.60, 0.60, 1.00), (1.00, 1.00, 0.00), (0.75, 0.75, 0.75), ] @@ -100,8 +100,7 @@ def sequence(nb_steps=10, all_frames=False): ] while True: - - frames =[] + frames = [] scene = random_scene() xh, yh = tuple(x.item() for x in torch.rand(2)) @@ -150,8 +149,111 @@ def sequence(nb_steps=10, all_frames=False): return frames, actions +###################################################################### + + +# ||x_i - c_j||^2 = ||x_i||^2 + ||c_j||^2 - 2 +def sq2matrix(x, c): + nx = x.pow(2).sum(1) + nc = c.pow(2).sum(1) + return nx[:, None] + nc[None, :] - 2 * x @ c.t() + + +def update_centroids(x, c, nb_min=1): + _, b = sq2matrix(x, c).min(1) + b.squeeze_() + nb_resets = 0 + + for k in range(0, c.size(0)): + i = b.eq(k).nonzero(as_tuple=False).squeeze() + if i.numel() >= nb_min: + c[k] = x.index_select(0, i).mean(0) + else: + n = torch.randint(x.size(0), (1,)) + nb_resets += 1 + c[k] = x[n] + + return c, b, nb_resets + + +def kmeans(x, nb_centroids, nb_min=1): + if x.size(0) < nb_centroids * nb_min: + print("Not enough points!") + exit(1) + + c = x[torch.randperm(x.size(0))[:nb_centroids]] + t = torch.full((x.size(0),), -1) + n = 0 + + while True: + c, u, nb_resets = update_centroids(x, c, nb_min) + n = n + 1 + nb_changes = (u - t).sign().abs().sum() + nb_resets + t = u + if nb_changes == 0: + break + + return c, t + + +###################################################################### + + +def patchify(x, factor, invert_size=None): + if invert_size is None: + return ( + x.reshape( + x.size(0), #0 + x.size(1), #1 + factor, #2 + x.size(2) // factor,#3 + factor,#4 + x.size(3) // factor,#5 + ) + .permute(0, 2, 4, 1, 3, 5) + .reshape(-1, x.size(1), x.size(2) // factor, x.size(3) // factor) + ) + else: + return ( + x.reshape( + invert_size[0], #0 + factor, #1 + factor, #2 + invert_size[1], #3 + invert_size[2] // factor, #4 + invert_size[3] // factor, #5 + ) + .permute(0, 3, 1, 4, 2, 5) + .reshape(invert_size) + ) + + if __name__ == "__main__": - frames, actions = sequence(nb_steps=31,all_frames=True) - frames = torch.cat(frames,0) - print(f"{frames.size()=}") - torchvision.utils.save_image(frames, "seq.png", nrow=8) + import time + + all_frames = [] + nb = 1000 + start_time = time.perf_counter() + for n in range(nb): + frames, actions = sequence(nb_steps=31) + all_frames += frames + end_time = time.perf_counter() + print(f"{nb / (end_time - start_time):.02f} samples per second") + + input = torch.cat(all_frames, 0) + x = patchify(input, 8) + y = x.reshape(x.size(0), -1) + print(f"{x.size()=} {y.size()=}") + centroids, t = kmeans(y, 4096) + results = centroids[t] + results = results.reshape(x.size()) + results = patchify(results, 8, input.size()) + + print(f"{input.size()=} {results.size()=}") + + torchvision.utils.save_image(input[:64], "orig.png", nrow=8) + torchvision.utils.save_image(results[:64], "qtiz.png", nrow=8) + + # frames, actions = sequence(nb_steps=31, all_frames=True) + # frames = torch.cat(frames, 0) + # torchvision.utils.save_image(frames, "seq.png", nrow=8)