X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=beaver.py;h=86008f6717ef964a17d51c759614d75221ac855e;hb=24d40cc2693dddf190fce9d43e458a86b685a7d3;hp=6ec0fb290e2109077b6aefe1a2ae63d032e755b2;hpb=2cd3f15987d2bf9050f737cd13506740ad3e90cb;p=beaver.git diff --git a/beaver.py b/beaver.py index 6ec0fb2..86008f6 100755 --- a/beaver.py +++ b/beaver.py @@ -64,6 +64,8 @@ parser.add_argument("--dropout", type=float, default=0.1) parser.add_argument("--deterministic_synthesis", action="store_true", default=False) +parser.add_argument("--random_regression_order", action="store_true", default=False) + parser.add_argument("--no_checkpoint", action="store_true", default=False) parser.add_argument("--overwrite_results", action="store_true", default=False) @@ -129,29 +131,45 @@ for n in vars(args): ###################################################################### -def random_order(result, fixed_len): - order = torch.rand(result.size(), device=result.device) - order[:, :fixed_len] = torch.linspace(-2, -1, fixed_len, device=order.device) - return order.sort(1).indices +def generation_order(x, fixed_len): + if args.random_regression_order: + order = torch.rand(x.size(), device=x.device) + order[:, :fixed_len] = torch.linspace(-2, -1, fixed_len, device=x.device) + order = order.sort(1).indices + else: + order = ( + torch.arange(x.size(1), device=x.device).unsqueeze(0).expand(x.size(0), -1) + ) + return order -def shuffle(x, order, reorder=False): - if x.dim() == 3: - order = order.unsqueeze(-1).expand(-1, -1, x.size(-1)) - if reorder: - y = x.new(x.size()) - y.scatter_(1, order, x) - return y +def reorder(x, order, reverse=False): # x is NxTxD1x...xDk, order is NxT' + u = x.reshape(x.size()[:2] + (-1,)) + order = order.unsqueeze(-1).expand(-1, -1, u.size(-1)) + if reverse: + v = u.new(u.size()) + v.scatter_(1, order, u) else: - return x.gather(1, order) + v = u.gather(1, order) + v = v.reshape(v.size()[:2] + x.size()[2:]) + return v + +def shuffle(x, fixed_len): + order = generation_order(x, fixed_len) + return reorder(x, order), order + + +###################################################################### # ar_mask is a Boolean matrix of same shape as input, with 1s on the # tokens that should be generated def masked_inplace_autoregression(model, batch_size, input, ar_mask, order=None): - for input, ar_mask in zip(input.split(batch_size), ar_mask.split(batch_size)): + for input, ar_mask, order in zip( + input.split(batch_size), ar_mask.split(batch_size), order.split(batch_size) + ): i = (ar_mask.sum(0) > 0).nonzero() if i.min() > 0: # Needed to initialize the model's cache @@ -170,7 +188,7 @@ def masked_inplace_autoregression(model, batch_size, input, ar_mask, order=None) ###################################################################### -def compute_perplexity(model, split="train"): +def compute_perplexity(model, task, fixed_len, split="train"): with torch.autograd.no_grad(): t = model.training model.eval() @@ -179,9 +197,9 @@ def compute_perplexity(model, split="train"): for input in task.batches(split=split): input = input.to(device) - order = random_order(input, task.height * task.width) - input = shuffle(input, order) - output = model(mygpt.BracketedSequence(input), order=order).x + x, order = shuffle(input, fixed_len) + x = model(mygpt.BracketedSequence(x), order=order).x + output = reorder(x, order, reverse=True) loss = F.cross_entropy(output.transpose(1, 2), input) acc_loss += loss.item() * input.size(0) nb_samples += input.size(0) @@ -245,10 +263,9 @@ def oneshot(gpt, task): acc_train_loss, nb_train_samples = 0, 0 for mazes, policies in task.policy_batches(split="train"): - order = random_order(input, task.height * task.width) - x = shuffle(mazes, order) + x, order = shuffle(mazes, task.height * task.width) x = gpt(mygpt.BracketedSequence(x), mode=args.oneshot_input, order=order).x - output_gpt = shuffle(x, order, reorder=True) + output_gpt = reorder(x, order, reverse=True) output = model(output_gpt) loss = compute_loss(mazes, output, policies, task.height, task.width) @@ -261,10 +278,9 @@ def oneshot(gpt, task): acc_test_loss, nb_test_samples = 0, 0 for mazes, policies in task.policy_batches(split="test"): - order = random_order(input, task.height * task.width) - x = shuffle(mazes, order) + x, order = shuffle(mazes, task.height * task.width) x = gpt(mygpt.BracketedSequence(x), mode=args.oneshot_input, order=order).x - output_gpt = shuffle(x, order, reorder=True) + output_gpt = reorder(x, order, reverse=True) output = model(output_gpt) loss = compute_loss(mazes, output, policies, task.height, task.width) acc_test_loss += loss.item() * mazes.size(0) @@ -277,10 +293,9 @@ def oneshot(gpt, task): # ------------------- mazes = task.test_input[:32, : task.height * task.width] policies = task.test_policies[:32] - order = random_order(input, task.height * task.width) - x = shuffle(mazes, order) + x, order = shuffle(mazes, task.height * task.width) x = gpt(mygpt.BracketedSequence(x), mode=args.oneshot_input, order=order).x - output_gpt = shuffle(x, order, reorder=True) + output_gpt = reorder(x, order, reverse=True) output = model(output_gpt) if args.oneshot_output == "policy": targets = policies.permute(0, 2, 1) @@ -299,15 +314,17 @@ def oneshot(gpt, task): scores = scores.reshape(-1, task.height, task.width) mazes = mazes.reshape(-1, task.height, task.width) targets = targets.reshape(-1, task.height, task.width) + filename = ( + f"oneshot_{args.oneshot_input}_{args.oneshot_output}_{n_epoch:04d}.png" + ) maze.save_image( - os.path.join( - args.result_dir, - f"oneshot_{args.oneshot_input}_{args.oneshot_output}_{n_epoch:04d}.png", - ), + os.path.join(args.result_dir, filename), mazes=mazes, score_paths=scores, score_truth=targets, ) + log_string(f"wrote {filename}") + # ------------------- gpt.train(t) @@ -419,11 +436,11 @@ class TaskMaze(Task): ar_mask = result.new_zeros(result.size()) ar_mask[:, self.height * self.width :] = 1 result *= 1 - ar_mask - order = random_order(result, self.height * self.width) + x, order = shuffle(result, self.height * self.width) masked_inplace_autoregression( - model, self.batch_size, result, ar_mask, order=order + model, self.batch_size, x, ar_mask, order=order ) - result = shuffle(result, order, reorder=True) + result = reorder(x, order, reverse=True) mazes, paths = self.seq2map(result) nb_correct += maze.path_correctness(mazes, paths).long().sum() nb_total += mazes.size(0) @@ -454,17 +471,23 @@ class TaskMaze(Task): ar_mask = result.new_zeros(result.size()) ar_mask[:, self.height * self.width :] = 1 result *= 1 - ar_mask - masked_inplace_autoregression(model, self.batch_size, result, ar_mask) + x, order = shuffle(result, self.height * self.width) + masked_inplace_autoregression( + model, self.batch_size, x, ar_mask, order=order + ) + result = reorder(x, order, reverse=True) mazes, paths = self.seq2map(input) _, predicted_paths = self.seq2map(result) + filename = f"result_{n_epoch:04d}.png" maze.save_image( - os.path.join(args.result_dir, f"result_{n_epoch:04d}.png"), + os.path.join(args.result_dir, filename), mazes=mazes, target_paths=paths, predicted_paths=predicted_paths, path_correct=maze.path_correctness(mazes, predicted_paths), ) + log_string(f"wrote {filename}") model.train(t) @@ -570,8 +593,12 @@ log_string(f"learning_rate_schedule {learning_rate_schedule}") if nb_epochs_finished >= args.nb_epochs: n_epoch = nb_epochs_finished - train_perplexity = compute_perplexity(model, split="train") - test_perplexity = compute_perplexity(model, split="test") + train_perplexity = compute_perplexity( + model, task, fixed_len=task.height * task.width, split="train" + ) + test_perplexity = compute_perplexity( + model, task, fixed_len=task.height * task.width, split="test" + ) log_string( f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}" @@ -579,8 +606,6 @@ if nb_epochs_finished >= args.nb_epochs: task.produce_results(n_epoch, model) - exit(0) - ############################## for n_epoch in range(nb_epochs_finished, args.nb_epochs): @@ -603,9 +628,9 @@ for n_epoch in range(nb_epochs_finished, args.nb_epochs): for input in task.batches(split="train"): input = input.to(device) - order = random_order(input, task.height * task.width) - input = shuffle(input, order) - output = model(mygpt.BracketedSequence(input), order=order).x + x, order = shuffle(input, task.height * task.width) + x = model(mygpt.BracketedSequence(x), order=order).x + output = reorder(x, order, reverse=True) loss = F.cross_entropy(output.transpose(1, 2), input) acc_train_loss += loss.item() * input.size(0) nb_train_samples += input.size(0) @@ -615,7 +640,9 @@ for n_epoch in range(nb_epochs_finished, args.nb_epochs): optimizer.step() train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples)) - test_perplexity = compute_perplexity(model, split="test") + test_perplexity = compute_perplexity( + model, task, fixed_len=task.height * task.width, split="test" + ) log_string( f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"