X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=tasks.py;h=11879fd44866094711a76b5c7bdebebdda28fa41;hb=0b8185b90014369f0d39892e128ad04a7d9ae872;hp=29f1e5a783024a3948867d2ffc0dae8c71b09954;hpb=2be22c9825d8aebe8d184e9501355a31318abf2b;p=picoclvr.git diff --git a/tasks.py b/tasks.py index 29f1e5a..11879fd 100755 --- a/tasks.py +++ b/tasks.py @@ -5,7 +5,7 @@ # Written by Francois Fleuret -import math, os, tqdm +import math, os, tqdm, warnings import torch, torchvision @@ -1893,13 +1893,11 @@ class Escape(Task): states, actions, rewards = escape.generate_episodes( nb_train_samples + nb_test_samples, height, width, T, nb_walls ) - seq = escape.episodes2seq(states, actions, rewards, lookahead_delta=T) + seq = escape.episodes2seq(states, actions, rewards) # seq = seq[:, seq.size(1) // 3 : 2 * seq.size(1) // 3] self.train_input = seq[:nb_train_samples].to(self.device) self.test_input = seq[nb_train_samples:].to(self.device) - self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1 - def batches(self, split="train", nb_to_use=-1, desc=None): assert split in {"train", "test"} input = self.train_input if split == "train" else self.test_input @@ -1913,7 +1911,7 @@ class Escape(Task): yield batch def vocabulary_size(self): - return self.nb_codes + return escape.nb_codes def thinking_autoregression( self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000 @@ -1922,10 +1920,15 @@ class Escape(Task): t = torch.arange(result.size(1), device=result.device)[None, :] state_len = self.height * self.width - index_action = state_len - index_reward = state_len + 1 - index_lookahead_reward = state_len + 2 - it_len = state_len + 3 # state / action / reward / lookahead_reward + index_lookahead_reward = 0 + index_states = 1 + index_action = state_len + 1 + index_reward = state_len + 2 + it_len = state_len + 3 # lookahead_reward / state / action / reward + + result[:, it_len:] = -1 + + snapshots = [] def ar(result, ar_mask, logit_biases=None): ar_mask = ar_mask.expand_as(result) @@ -1940,42 +1943,59 @@ class Escape(Task): device=self.device, progress_bar_desc=None, ) + warnings.warn("keeping thinking snapshots", RuntimeWarning) + snapshots.append(result[:10].detach().clone()) # Generate iteration after iteration - optimistic_bias = result.new_zeros(self.nb_codes, device=result.device) + optimistic_bias = result.new_zeros(escape.nb_codes, device=result.device) optimistic_bias[escape.lookahead_reward2code(-1)] = -math.log(1e1) optimistic_bias[escape.lookahead_reward2code(1)] = math.log(1e1) for u in tqdm.tqdm( range(it_len, result.size(1) - it_len + 1, it_len), desc="thinking" ): - # Re-generate the lookahead_reward pessimistically in the - # previous iterations - ar_mask = (t < u).long() * (t % it_len == index_lookahead_reward).long() - ar(result, ar_mask, logit_biases=-optimistic_bias) + lr, _, _, _ = escape.seq2episodes(result[:, :u], self.height, self.width) - # Generate the state - ar_mask = (t >= u).long() * (t < u + state_len).long() + # Generate the lookahead_reward and state + ar_mask = (t % it_len == index_lookahead_reward).long() * ( + t <= u + index_lookahead_reward + ).long() ar(result, ar_mask) - # Re-generate the lookahead_reward optimistically in the - # previous iterations - ar_mask = (t < u).long() * (t % it_len == index_lookahead_reward).long() + # Generate the lookahead_reward and state + ar_mask = (t >= u + index_states).long() * ( + t < u + index_states + state_len + ).long() + ar(result, ar_mask) + + # Re-generate the lookahead_reward + ar_mask = (t % it_len == index_lookahead_reward).long() * ( + t <= u + index_lookahead_reward + ).long() ar(result, ar_mask, logit_biases=optimistic_bias) # Generate the action and reward ar_mask = (t >= u + index_action).long() * (t <= u + index_reward).long() ar(result, ar_mask) + filename = os.path.join(result_dir, f"test_thinking_compute_{n_epoch:04d}.txt") + with open(filename, "w") as f: + for n in range(10): + for s in snapshots: + lr, s, a, r = escape.seq2episodes( + s[n : n + 1], self.height, self.width + ) + str = escape.episodes2str( + lr, s, a, r, unicode=True, ansi_colors=True + ) + f.write(str) + f.write("\n\n") + # Saving the generated sequences - s, a, r, lr = escape.seq2episodes( - result, self.height, self.width, lookahead=True - ) - str = escape.episodes2str( - s, a, r, lookahead_rewards=lr, unicode=True, ansi_colors=True - ) + lr, s, a, r = escape.seq2episodes(result, self.height, self.width) + str = escape.episodes2str(lr, s, a, r, unicode=True, ansi_colors=True) filename = os.path.join(result_dir, f"test_thinking_seq_{n_epoch:04d}.txt") with open(filename, "w") as f: @@ -1989,12 +2009,12 @@ class Escape(Task): # Saving the ground truth - s, a, r, lr = escape.seq2episodes( - result, self.height, self.width, lookahead=True - ) - str = escape.episodes2str( - s, a, r, lookahead_rewards=lr, unicode=True, ansi_colors=True + lr, s, a, r = escape.seq2episodes( + result, + self.height, + self.width, ) + str = escape.episodes2str(lr, s, a, r, unicode=True, ansi_colors=True) filename = os.path.join(result_dir, f"test_true_seq_{n_epoch:04d}.txt") with open(filename, "w") as f: @@ -2021,12 +2041,12 @@ class Escape(Task): # Saving the generated sequences - s, a, r, lr = escape.seq2episodes( - result, self.height, self.width, lookahead=True - ) - str = escape.episodes2str( - s, a, r, lookahead_rewards=lr, unicode=True, ansi_colors=True + lr, s, a, r = escape.seq2episodes( + result, + self.height, + self.width, ) + str = escape.episodes2str(lr, s, a, r, unicode=True, ansi_colors=True) filename = os.path.join(result_dir, f"test_seq_{n_epoch:04d}.txt") with open(filename, "w") as f: