X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=escape.py;h=7596beaec0e6b99ed4684675b161c9112802f117;hb=290c261e54a98cdea6115e2a0ee91ce92257d13b;hp=da3e4959cc402dffa6254abf611ce4bf3aefa0a1;hpb=690470307a7995cc3117eb54545f921eedcecba5;p=picoclvr.git diff --git a/escape.py b/escape.py index da3e495..7596bea 100755 --- a/escape.py +++ b/escape.py @@ -11,13 +11,13 @@ from torch.nn import functional as F ###################################################################### -nb_state_codes = 4 +nb_states_codes = 5 nb_actions_codes = 5 nb_rewards_codes = 3 -nb_lookahead_rewards_codes = 3 +nb_lookahead_rewards_codes = 4 # stands for -1, 0, +1, and UNKNOWN -first_state_code = 0 -first_actions_code = first_state_code + nb_state_codes +first_states_code = 0 +first_actions_code = first_states_code + nb_states_codes first_rewards_code = first_actions_code + nb_actions_codes first_lookahead_rewards_code = first_rewards_code + nb_rewards_codes nb_codes = first_lookahead_rewards_code + nb_lookahead_rewards_codes @@ -25,21 +25,68 @@ nb_codes = first_lookahead_rewards_code + nb_lookahead_rewards_codes ###################################################################### -def generate_episodes(nb, height=6, width=6, T=10): +def state2code(r): + return r + first_states_code + + +def code2state(r): + return r - first_states_code + + +def action2code(r): + return r + first_actions_code + + +def code2action(r): + return r - first_actions_code + + +def reward2code(r): + return r + 1 + first_rewards_code + + +def code2reward(r): + return r - first_rewards_code - 1 + + +def lookahead_reward2code(r): + # -1, 0, +1 or 2 for UNKNOWN + return r + 1 + first_lookahead_rewards_code + + +def code2lookahead_reward(r): + return r - first_lookahead_rewards_code - 1 + + +###################################################################### + + +def generate_episodes(nb, height=6, width=6, T=10, nb_walls=3, nb_coins=2): rnd = torch.rand(nb, height, width) rnd[:, 0, :] = 0 rnd[:, -1, :] = 0 rnd[:, :, 0] = 0 rnd[:, :, -1] = 0 wall = 0 - - for k in range(3): + for k in range(nb_walls): wall = wall + ( rnd.flatten(1).argmax(dim=1)[:, None] == torch.arange(rnd.flatten(1).size(1))[None, :] ).long().reshape(rnd.size()) + rnd = rnd * (1 - wall.clamp(max=1)) + rnd = torch.rand(nb, height, width) + coins = torch.zeros(nb, T, height, width, dtype=torch.int64) + rnd = rnd * (1 - wall.clamp(max=1)) + for k in range(nb_coins): + coins[:, 0] = coins[:, 0] + ( + rnd.flatten(1).argmax(dim=1)[:, None] + == torch.arange(rnd.flatten(1).size(1))[None, :] + ).long().reshape(rnd.size()) + + rnd = rnd * (1 - coins[:, 0].clamp(max=1)) + states = wall[:, None, :, :].expand(-1, T, -1, -1).clone() agent = torch.zeros(states.size(), dtype=torch.int64) @@ -93,11 +140,14 @@ def generate_episodes(nb, height=6, width=6, T=10): ) hit = (hit > 0).long() - assert hit.min() == 0 and hit.max() <= 1 + # assert hit.min() == 0 and hit.max() <= 1 + + got_coin = (agent[:, t + 1] * coins[:, t]).flatten(1).sum(dim=1) + coins[:, t + 1] = coins[:, t] * (1 - agent[:, t + 1]) - rewards[:, t + 1] = -hit + (1 - hit) * agent[:, t + 1, -1, -1] + rewards[:, t + 1] = -hit + (1 - hit) * got_coin - states += 2 * agent + 3 * monster + states = states + 2 * agent + 3 * monster + 4 * coins return states, agent_actions, rewards @@ -105,79 +155,39 @@ def generate_episodes(nb, height=6, width=6, T=10): ###################################################################### -def episodes2seq(states, actions, rewards, lookahead_delta=None): - states = states.flatten(2) + first_state_code - actions = actions[:, :, None] + first_actions_code - - if lookahead_delta is not None: - # r = rewards - # u = F.pad(r, (0, lookahead_delta - 1)).as_strided( - # (r.size(0), r.size(1), lookahead_delta), - # (r.size(1) + lookahead_delta - 1, 1, 1), - # ) - # a = u[:, :, 1:].min(dim=-1).values - # b = u[:, :, 1:].max(dim=-1).values - # s = (a < 0).long() * a + (a >= 0).long() * b - # lookahead_rewards = (1 + s[:, :, None]) + first_lookahead_rewards_code - - # a[n,t]=min_s>t r[n,s] - a = rewards.new_zeros(rewards.size()) - b = rewards.new_zeros(rewards.size()) - for t in range(a.size(1) - 1): - a[:, t] = rewards[:, t + 1 :].min(dim=-1).values - b[:, t] = rewards[:, t + 1 :].max(dim=-1).values - s = (a < 0).long() * a + (a >= 0).long() * b - lookahead_rewards = (1 + s[:, :, None]) + first_lookahead_rewards_code - - r = rewards[:, :, None] - rewards = (r + 1) + first_rewards_code - - assert ( - states.min() >= first_state_code - and states.max() < first_state_code + nb_state_codes - ) - assert ( - actions.min() >= first_actions_code - and actions.max() < first_actions_code + nb_actions_codes - ) - assert ( - rewards.min() >= first_rewards_code - and rewards.max() < first_rewards_code + nb_rewards_codes - ) - - if lookahead_delta is None: - return torch.cat([states, actions, rewards], dim=2).flatten(1) - else: - assert ( - lookahead_rewards.min() >= first_lookahead_rewards_code - and lookahead_rewards.max() - < first_lookahead_rewards_code + nb_lookahead_rewards_codes - ) - return torch.cat([states, actions, rewards, lookahead_rewards], dim=2).flatten( - 1 - ) - - -def seq2episodes(seq, height, width, lookahead=False): - seq = seq.reshape(seq.size(0), -1, height * width + (3 if lookahead else 2)) - states = seq[:, :, : height * width] - first_state_code +def episodes2seq(states, actions, rewards): + neg = rewards.new_zeros(rewards.size()) + pos = rewards.new_zeros(rewards.size()) + for t in range(neg.size(1) - 1): + neg[:, t] = rewards[:, t:].min(dim=-1).values + pos[:, t] = rewards[:, t:].max(dim=-1).values + s = (neg < 0).long() * neg + (neg >= 0).long() * pos + + return torch.cat( + [ + lookahead_reward2code(s[:, :, None]), + state2code(states.flatten(2)), + action2code(actions[:, :, None]), + reward2code(rewards[:, :, None]), + ], + dim=2, + ).flatten(1) + + +def seq2episodes(seq, height, width): + seq = seq.reshape(seq.size(0), -1, height * width + 3) + lookahead_rewards = code2lookahead_reward(seq[:, :, 0]) + states = code2state(seq[:, :, 1 : height * width + 1]) states = states.reshape(states.size(0), states.size(1), height, width) - actions = seq[:, :, height * width] - first_actions_code - rewards = seq[:, :, height * width + 1] - first_rewards_code - 1 - - if lookahead: - lookahead_rewards = ( - seq[:, :, height * width + 2] - first_lookahead_rewards_code - 1 - ) - return states, actions, rewards, lookahead_rewards - else: - return states, actions, rewards + actions = code2action(seq[:, :, height * width + 1]) + rewards = code2reward(seq[:, :, height * width + 2]) + return lookahead_rewards, states, actions, rewards def seq2str(seq): def token2str(t): - if t >= first_state_code and t < first_state_code + nb_state_codes: - return " #@$"[t - first_state_code] + if t >= first_states_code and t < first_states_code + nb_states_codes: + return " #@T$"[t - first_states_code] elif t >= first_actions_code and t < first_actions_code + nb_actions_codes: return "ISNEW"[t - first_actions_code] elif t >= first_rewards_code and t < first_rewards_code + nb_rewards_codes: @@ -186,7 +196,7 @@ def seq2str(seq): t >= first_lookahead_rewards_code and t < first_lookahead_rewards_code + nb_lookahead_rewards_codes ): - return "n.p"[t - first_lookahead_rewards_code] + return "n.pU"[t - first_lookahead_rewards_code] else: return "?" @@ -197,14 +207,14 @@ def seq2str(seq): def episodes2str( - states, actions, rewards, lookahead_rewards=None, unicode=False, ansi_colors=False + lookahead_rewards, states, actions, rewards, unicode=False, ansi_colors=False ): if unicode: - symbols = " █@$" + symbols = "·█@T$" # vert, hori, cross, thin_hori = "║", "═", "╬", "─" vert, hori, cross, thin_vert, thin_hori = "┃", "━", "╋", "│", "─" else: - symbols = " #@$" + symbols = " #@T$" vert, hori, cross, thin_vert, thin_hori = "|", "-", "+", "|", "-" hline = (cross + hori * states.size(-1)) * states.size(1) + cross + "\n" @@ -227,7 +237,7 @@ def episodes2str( + "\n" ) - result += (vert + thin_hori * states.size(-1)) * states.size(1) + vert + "\n" + # result += (vert + thin_hori * states.size(-1)) * states.size(1) + vert + "\n" def status_bar(a, r, lr=None): a, r = a.item(), r.item() @@ -246,27 +256,17 @@ def episodes2str( + sb_lr ) - if lookahead_rewards is None: - result += ( - vert - + vert.join([status_bar(a, r) for a, r in zip(actions[n], rewards[n])]) - + vert - + "\n" - ) - else: - result += ( - vert - + vert.join( - [ - status_bar(a, r, lr) - for a, r, lr in zip( - actions[n], rewards[n], lookahead_rewards[n] - ) - ] - ) - + vert - + "\n" + result += ( + vert + + vert.join( + [ + status_bar(a, r, lr) + for a, r, lr in zip(actions[n], rewards[n], lookahead_rewards[n]) + ] ) + + vert + + "\n" + ) result += hline @@ -280,11 +280,11 @@ def episodes2str( ###################################################################### if __name__ == "__main__": - nb, height, width, T = 1000, 4, 6, 20 - states, actions, rewards = generate_episodes(nb, height, width, T) - seq = episodes2seq(states, actions, rewards, lookahead_delta=T) - s, a, r, lr = seq2episodes(seq, height, width, lookahead=True) - print(episodes2str(s, a, r, lookahead_rewards=lr, unicode=True, ansi_colors=True)) + nb, height, width, T, nb_walls = 5, 5, 7, 10, 5 + states, actions, rewards = generate_episodes(nb, height, width, T, nb_walls) + seq = episodes2seq(states, actions, rewards) + lr, s, a, r = seq2episodes(seq, height, width) + print(episodes2str(lr, s, a, r, unicode=True, ansi_colors=True)) # print() # for s in seq2str(seq): # print(s)