X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;ds=inline;f=reasoning.py;h=c545e97f0b6517d7bbec6d163a55b447c2e0a989;hb=cb64cb72d5fff242cd90c13e4d8b9ca43e6e8837;hp=b8d39ee0a08d8563946901e032709de65d1e1160;hpb=1fb70f7c2d14ad7a1f600cdf472d179d84c438bf;p=culture.git diff --git a/reasoning.py b/reasoning.py index b8d39ee..c545e97 100755 --- a/reasoning.py +++ b/reasoning.py @@ -27,9 +27,9 @@ class Reasoning(problem.Problem): ("cyan", [0, 255, 255]), ("violet", [255, 0, 255]), ("lightgreen", [192, 255, 192]), - ("pink", [255, 192, 192]), + ("brown", [165, 42, 42]), ("lightblue", [192, 192, 255]), - ("gray", [192, 192, 192]), + ("gray", [128, 128, 128]), ] def __init__(self, device=torch.device("cpu")): @@ -42,6 +42,31 @@ class Reasoning(problem.Problem): ###################################################################### def frame2img(self, x, scale=15): + x = x.reshape(x.size(0), self.height, -1) + m = torch.logical_and(x >= 0, x < self.nb_token_values()).long() + x = self.colors[x * m].permute(0, 3, 1, 2) + s = x.shape + x = x[:, :, :, None, :, None].expand(-1, -1, -1, scale, -1, scale) + x = x.reshape(s[0], s[1], s[2] * scale, s[3] * scale) + + x[:, :, :, torch.arange(0, x.size(3), scale)] = 0 + x[:, :, torch.arange(0, x.size(2), scale), :] = 0 + x = x[:, :, 1:, 1:] + + for n in range(m.size(0)): + for i in range(m.size(1)): + for j in range(m.size(2)): + if m[n, i, j] == 0: + for k in range(2, scale - 2): + for l in [0, 1]: + x[n, :, i * scale + k, j * scale + k - l] = 0 + x[ + n, :, i * scale + scale - 1 - k, j * scale + k - l + ] = 0 + + return x + + def frame2img_(self, x, scale=15): x = x.reshape(x.size(0), self.height, -1) x = self.colors[x].permute(0, 3, 1, 2) s = x.shape @@ -169,14 +194,17 @@ class Reasoning(problem.Problem): def nb_token_values(self): return len(self.colors) - def rec_coo(self, x, n, min_height=3, min_width=3): - K = 3 - N = 4000 + # That's quite a tensorial spaghetti mess to sample + # non-overlapping rectangles quickly, but made the generation of + # 100k samples go from 1h50 with a lame pure python code to 3min30s + # with this one. + def rec_coo(self, nb_rec, min_height=3, min_width=3): + nb_trials = 200 while True: v = ( ( - torch.rand(N * K, self.height + 1, device=self.device) + torch.rand(nb_trials * nb_rec, self.height + 1, device=self.device) .sort(dim=-1) .indices < 2 @@ -188,7 +216,7 @@ class Reasoning(problem.Problem): h = ( ( - torch.rand(N * K, self.width + 1, device=self.device) + torch.rand(nb_trials * nb_rec, self.width + 1, device=self.device) .sort(dim=-1) .indices < 2 @@ -203,10 +231,10 @@ class Reasoning(problem.Problem): ) v, h = v[i], h[i] - v = v[: v.size(0) - v.size(0) % K] - h = h[: h.size(0) - h.size(0) % K] - v = v.reshape(v.size(0) // K, K, -1) - h = h.reshape(h.size(0) // K, K, -1) + v = v[: v.size(0) - v.size(0) % nb_rec] + h = h[: h.size(0) - h.size(0) % nb_rec] + v = v.reshape(v.size(0) // nb_rec, nb_rec, -1) + h = h.reshape(h.size(0) // nb_rec, nb_rec, -1) r = v[:, :, :, None] * h[:, :, None, :] @@ -256,23 +284,23 @@ class Reasoning(problem.Problem): ###################################################################### def task_replace_color(self, A, f_A, B, f_B): - N = 3 - c = torch.randperm(len(self.colors) - 1)[: N + 1] + 1 + nb_rec = 3 + c = torch.randperm(len(self.colors) - 1)[: nb_rec + 1] + 1 for X, f_X in [(A, f_A), (B, f_B)]: - r = self.rec_coo(X, N) - for n in range(N): + r = self.rec_coo(nb_rec) + for n in range(nb_rec): i1, j1, i2, j2 = r[n] X[i1:i2, j1:j2] = c[n] f_X[i1:i2, j1:j2] = c[n if n > 0 else -1] - def task_move(self, A, f_A, B, f_B): - di, dj = torch.randint(2, (2,)) * 2 - 1 - N = 3 - c = torch.randperm(len(self.colors) - 1)[:N] + 1 + def task_translate(self, A, f_A, B, f_B): + di, dj = torch.randint(3, (2,)) - 1 + nb_rec = 3 + c = torch.randperm(len(self.colors) - 1)[:nb_rec] + 1 for X, f_X in [(A, f_A), (B, f_B)]: while True: - r = self.rec_coo(X, N) - i1, j1, i2, j2 = r[N - 1] + r = self.rec_coo(nb_rec) + i1, j1, i2, j2 = r[nb_rec - 1] if ( i1 + di >= 0 and i2 + di < X.size(0) @@ -281,29 +309,29 @@ class Reasoning(problem.Problem): ): break - for n in range(N): + for n in range(nb_rec): i1, j1, i2, j2 = r[n] X[i1:i2, j1:j2] = c[n] - if n == N - 1: + if n == nb_rec - 1: f_X[i1 + di : i2 + di, j1 + dj : j2 + dj] = c[n] else: f_X[i1:i2, j1:j2] = c[n] def task_grow(self, A, f_A, B, f_B): di, dj = torch.randint(2, (2,)) * 2 - 1 - N = 3 - c = torch.randperm(len(self.colors) - 1)[:N] + 1 + nb_rec = 3 + c = torch.randperm(len(self.colors) - 1)[:nb_rec] + 1 direction = torch.randint(2, (1,)) for X, f_X in [(A, f_A), (B, f_B)]: while True: - r = self.rec_coo(X, N) - i1, j1, i2, j2 = r[N - 1] + r = self.rec_coo(nb_rec) + i1, j1, i2, j2 = r[nb_rec - 1] if i1 + 3 < i2 and j1 + 3 < j2: break - for n in range(N): + for n in range(nb_rec): i1, j1, i2, j2 = r[n] - if n == N - 1: + if n == nb_rec - 1: if direction == 0: X[i1 + 1 : i2 - 1, j1 + 1 : j2 - 1] = c[n] f_X[i1:i2, j1:j2] = c[n] @@ -316,61 +344,233 @@ class Reasoning(problem.Problem): def task_color_grow(self, A, f_A, B, f_B): di, dj = torch.randint(2, (2,)) * 2 - 1 - N = 3 - c = torch.randperm(len(self.colors) - 1)[: 2 * N] + 1 - direction = torch.randint(2, (1,)) + nb_rec = 3 + c = torch.randperm(len(self.colors) - 1)[: 2 * nb_rec] + 1 + direction = torch.randint(4, (1,)) for X, f_X in [(A, f_A), (B, f_B)]: - r = self.rec_coo(X, N) - for n in range(N): + r = self.rec_coo(nb_rec) + for n in range(nb_rec): i1, j1, i2, j2 = r[n] - i = (i1 + i2) // 2 X[i1:i2, j1:j2] = c[2 * n] - X[i : i + 1, j1:j2] = c[2 * n + 1] f_X[i1:i2, j1:j2] = c[2 * n] - if n == N - 1: - f_X[i:i2, j1:j2] = c[2 * n + 1] - else: - f_X[i : i + 1, j1:j2] = c[2 * n + 1] + # Not my proudest moment + if direction == 0: + i = (i1 + i2) // 2 + X[i : i + 1, j1:j2] = c[2 * n + 1] + if n == nb_rec - 1: + f_X[i:i2, j1:j2] = c[2 * n + 1] + else: + f_X[i : i + 1, j1:j2] = c[2 * n + 1] + elif direction == 1: + i = (i1 + i2 - 1) // 2 + X[i : i + 1, j1:j2] = c[2 * n + 1] + if n == nb_rec - 1: + f_X[i1 : i + 1, j1:j2] = c[2 * n + 1] + else: + f_X[i : i + 1, j1:j2] = c[2 * n + 1] + elif direction == 2: + j = (j1 + j2) // 2 + X[i1:i2, j : j + 1] = c[2 * n + 1] + if n == nb_rec - 1: + f_X[i1:i2, j:j2] = c[2 * n + 1] + else: + f_X[i1:i2, j : j + 1] = c[2 * n + 1] + elif direction == 3: + j = (j1 + j2 - 1) // 2 + X[i1:i2, j : j + 1] = c[2 * n + 1] + if n == nb_rec - 1: + f_X[i1:i2, j1 : j + 1] = c[2 * n + 1] + else: + f_X[i1:i2, j : j + 1] = c[2 * n + 1] def task_frame(self, A, f_A, B, f_B): - N = 3 - c = torch.randperm(len(self.colors) - 1)[: N + 1] + 1 + nb_rec = 3 + c = torch.randperm(len(self.colors) - 1)[: nb_rec + 1] + 1 for X, f_X in [(A, f_A), (B, f_B)]: - r = self.rec_coo(X, N) - for n in range(N): + r = self.rec_coo(nb_rec) + for n in range(nb_rec): i1, j1, i2, j2 = r[n] X[i1:i2, j1:j2] = c[n] f_X[i1:i2, j1:j2] = c[n] - if n == N - 1: + if n == nb_rec - 1: f_X[i1 + 1 : i2 - 1, j1 + 1 : j2 - 1] = 0 def task_detect(self, A, f_A, B, f_B): - N = 3 - c = torch.randperm(len(self.colors) - 1)[: N + 1] + 1 + nb_rec = 3 + c = torch.randperm(len(self.colors) - 1)[: nb_rec + 1] + 1 for X, f_X in [(A, f_A), (B, f_B)]: - r = self.rec_coo(X, N) - for n in range(N): + r = self.rec_coo(nb_rec) + for n in range(nb_rec): i1, j1, i2, j2 = r[n] X[i1:i2, j1:j2] = c[n] - f_X[i1, j1] = c[-1] + if n < nb_rec - 1: + f_X[i1, j1] = c[-1] + + def task_count(self, A, f_A, B, f_B): + N = torch.randint(4, (1,)) + 2 + c = torch.randperm(len(self.colors) - 1)[:N] + 1 + + for X, f_X in [(A, f_A), (B, f_B)]: + + def contact(i, j, q): + nq, nq_diag = 0, 0 + no = 0 + + for ii, jj in [ + (i - 1, j - 1), + (i - 1, j), + (i - 1, j + 1), + (i, j - 1), + (i, j + 1), + (i + 1, j - 1), + (i + 1, j), + (i + 1, j + 1), + ]: + if ii >= 0 and ii < self.height and jj >= 0 and jj < self.width: + if X[ii, jj] != 0 and X[ii, jj] != q: + no += 1 + + for ii, jj in [ + (i - 1, j - 1), + (i - 1, j + 1), + (i + 1, j - 1), + (i + 1, j + 1), + ]: + if ii >= 0 and ii < self.height and jj >= 0 and jj < self.width: + if X[ii, jj] == q and X[i, jj] != q and X[ii, j] != q: + nq_diag += 1 + + for ii, jj in [(i - 1, j), (i, j - 1), (i, j + 1), (i + 1, j)]: + if ii >= 0 and ii < self.height and jj >= 0 and jj < self.width: + if X[ii, jj] == q: + nq += 1 + + return no, nq, nq_diag + + nb = torch.zeros(N, dtype=torch.int64) + q = torch.randint(N, (self.height * self.width,)) + k = torch.randperm(self.height * self.width) + for p in range(self.height * self.width): + i, j = k[p] % self.height, k[p] // self.height + no, nq, nq_diag = contact(i, j, c[q[p]]) + if no == 0 and nq_diag == 0: + if nq == 0: + if nb[q[p]] < self.width: + X[i, j] = c[q[p]] + nb[q[p]] += 1 + if nq == 1: + X[i, j] = c[q[p]] + + for n in range(N): + for j in range(nb[n]): + f_X[n, j] = c[n] + + def task_trajectory(self, A, f_A, B, f_B): + c = torch.randperm(len(self.colors) - 1)[:2] + 1 + for X, f_X in [(A, f_A), (B, f_B)]: + while True: + di, dj = torch.randint(7, (2,)) - 3 + i, j = torch.randint(self.height, (1,)), torch.randint(self.width, (1,)) + if ( + abs(di) + abs(dj) > 0 + and i + 2 * di >= 0 + and i + 2 * di < self.height + and j + 2 * dj >= 0 + and j + 2 * dj < self.width + ): + break + + k = 0 + while ( + i + k * di >= 0 + and i + k * di < self.height + and j + k * dj >= 0 + and j + k * dj < self.width + ): + if k < 2: + X[i + k * di, j + k * dj] = c[k] + f_X[i + k * di, j + k * dj] = c[min(k, 1)] + k += 1 + + def task_bounce(self, A, f_A, B, f_B): + c = torch.randperm(len(self.colors) - 1)[:3] + 1 + for X, f_X in [(A, f_A), (B, f_B)]: + + def free(i, j): + return ( + i >= 0 + and i < self.height + and j >= 0 + and j < self.width + and f_X[i, j] == 0 + ) + + while True: + f_X[...] = 0 + X[...] = 0 + + for _ in range((self.height * self.width) // 10): + i, j = torch.randint(self.height, (1,)), torch.randint( + self.width, (1,) + ) + X[i, j] = c[0] + f_X[i, j] = c[0] + + while True: + di, dj = torch.randint(7, (2,)) - 3 + if abs(di) + abs(dj) == 1: + break + + i, j = torch.randint(self.height, (1,)), torch.randint(self.width, (1,)) + + X[i, j] = c[1] + f_X[i, j] = c[1] + l = 0 + + while True: + l += 1 + if free(i + di, j + dj): + pass + elif free(i - dj, j + di): + di, dj = -dj, di + if free(i + dj, j - di): + if torch.rand(1) < 0.5: + di, dj = -di, -dj + elif free(i + dj, j - di): + di, dj = dj, -di + else: + break + + i, j = i + di, j + dj + f_X[i, j] = c[2] + if l <= 1: + X[i, j] = c[2] + + if l >= self.width: + break + + f_X[i, j] = c[1] + X[i, j] = c[1] + + if l > 3: + break ###################################################################### def generate_prompts_and_answers(self, nb, device="cpu"): tasks = [ self.task_replace_color, - self.task_move, + self.task_translate, self.task_grow, self.task_color_grow, self.task_frame, self.task_detect, + self.task_count, + self.task_trajectory, + self.task_bounce, ] - prompts = torch.zeros( - nb, self.height, self.width * 3, dtype=torch.int64, device=self.device - ) - answers = torch.zeros( - nb, self.height, self.width, dtype=torch.int64, device=self.device - ) + prompts = torch.zeros(nb, self.height, self.width * 3, dtype=torch.int64) + answers = torch.zeros(nb, self.height, self.width, dtype=torch.int64) w = self.width for prompt, answer in tqdm.tqdm( @@ -385,6 +585,7 @@ class Reasoning(problem.Problem): f_B = answer task = tasks[torch.randint(len(tasks), (1,))] task(A, f_A, B, f_B) + return prompts.flatten(1), answers.flatten(1) def save_quizzes( @@ -418,14 +619,15 @@ if __name__ == "__main__": delay = time.perf_counter() - start_time print(f"{prompts.size(0)/delay:02f} seq/s") - # predicted_prompts = torch.rand(prompts.size(0)) < 0.5 - # predicted_answers = torch.logical_not(predicted_prompts) + predicted_prompts = torch.rand(prompts.size(0)) < 0.5 + predicted_answers = torch.logical_not(predicted_prompts) reasoning.save_quizzes( "/tmp", "test", - prompts[:36], - answers[:36], + prompts[:64], + answers[:64], # You can add a bool to put a frame around the predicted parts - # predicted_prompts, predicted_answers + # predicted_prompts[:64], + # predicted_answers[:64], )