X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=beaver.py;h=f5f092bea8247a9dce308f29004a3e2fcf78d99e;hb=beb6caeb2ffc8c5185bddbc07c4486b3cb9c8495;hp=69116eac0ef5313bfc8429b3686b92ed5bd375f6;hpb=706569c57bfbdc7e9a0791cdb608236208012710;p=beaver.git diff --git a/beaver.py b/beaver.py index 69116ea..f5f092b 100755 --- a/beaver.py +++ b/beaver.py @@ -134,7 +134,7 @@ for n in vars(args): 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=order.device) + order[:, :fixed_len] = torch.linspace(-2, -1, fixed_len, device=x.device) order = order.sort(1).indices else: order = ( @@ -167,7 +167,9 @@ def shuffle(x, fixed_len): 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 @@ -186,7 +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() @@ -195,8 +197,9 @@ def compute_perplexity(model, split="train"): for input in task.batches(split=split): input = input.to(device) - input, order = shuffle(input, task.height * task.width) - 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, back=True) loss = F.cross_entropy(output.transpose(1, 2), input) acc_loss += loss.item() * input.size(0) nb_samples += input.size(0) @@ -311,15 +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) @@ -431,11 +436,11 @@ class TaskMaze(Task): ar_mask = result.new_zeros(result.size()) ar_mask[:, self.height * self.width :] = 1 result *= 1 - ar_mask - result, order = shuffle(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 = reorder(result, order, back=True) + result = reorder(x, order, back=True) mazes, paths = self.seq2map(result) nb_correct += maze.path_correctness(mazes, paths).long().sum() nb_total += mazes.size(0) @@ -466,17 +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, back=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) @@ -582,8 +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}" @@ -613,8 +628,9 @@ for n_epoch in range(nb_epochs_finished, args.nb_epochs): for input in task.batches(split="train"): input = input.to(device) - input, order = shuffle(input, task.height * task.width) - 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, back=True) loss = F.cross_entropy(output.transpose(1, 2), input) acc_train_loss += loss.item() * input.size(0) nb_train_samples += input.size(0) @@ -624,7 +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}"