######################################################################
-parser = argparse.ArgumentParser(
- description="An implementation of GPT with cache to solve a toy geometric reasoning task."
-)
+parser = argparse.ArgumentParser(description="A maze shortest path solving with a GPT.")
parser.add_argument("--log_filename", type=str, default="train.log")
parser.add_argument("--nb_epochs", type=int, default=25)
-parser.add_argument("--batch_size", type=int, default=100)
+parser.add_argument("--nb_train_samples", type=int, default=200000)
+
+parser.add_argument("--nb_test_samples", type=int, default=50000)
-parser.add_argument("--data_size", type=int, default=-1)
+parser.add_argument("--batch_size", type=int, default=25)
parser.add_argument("--optim", type=str, default="adam")
parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
+parser.add_argument("--random_regression_order", action="store_true", default=False)
+
+parser.add_argument("--noncausal_prompt", 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)
parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
##############################
-# picoclvr options
+# maze options
+
+parser.add_argument("--maze_height", type=int, default=13)
+
+parser.add_argument("--maze_width", type=int, default=21)
+
+parser.add_argument("--maze_nb_walls", type=int, default=15)
+
+##############################
+# one-shot prediction
-parser.add_argument("--world_height", type=int, default=23)
+parser.add_argument("--oneshot", action="store_true", default=False)
-parser.add_argument("--world_width", type=int, default=31)
+parser.add_argument("--oneshot_input", type=str, default="head")
-parser.add_argument("--world_nb_walls", type=int, default=15)
+parser.add_argument("--oneshot_output", type=str, default="trace")
######################################################################
args = parser.parse_args()
-assert args.prune_properties in {"none", "train+eval", "eval"}
-
try:
os.mkdir(args.result_dir)
except FileExistsError:
######################################################################
-def masked_inplace_autoregression(
- model, batch_size, input, ar_mask, forbidden_tokens=None, device=torch.device("cpu")
-):
+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()).scatter_(1, order, u)
+ else:
+ v = u.gather(1, order)
+ v = v.reshape(v.size()[:2] + x.size()[2:])
+ return v
- for input, ar_mask in zip(input.split(batch_size), ar_mask.split(batch_size)):
+
+def shuffle(x, prompt_len):
+ if args.random_regression_order:
+ order = torch.rand(x.size(), device=x.device)
+ order[:, :prompt_len] = torch.arange(-prompt_len, 0, 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 reorder(x, order), order
+
+
+def eval_mygpt(model, input, mode="standard", prompt_len=0):
+ x, order = shuffle(input, prompt_len)
+ x = model(mygpt.BracketedSequence(x), mode=mode, order=order).x
+ return reorder(x, order, reverse=True)
+
+
+######################################################################
+
+# 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, 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:
- model(
- mygpt.BracketedSequence(input, 0, i.min())
- ) # Needed to initialize the model's cache
+ # Needed to initialize the model's cache
+ model(mygpt.BracketedSequence(input, 0, i.min()), order=order)
for s in range(i.min(), i.max() + 1):
- output = model(mygpt.BracketedSequence(input, s, 1)).x
+ output = model(mygpt.BracketedSequence(input, s, 1), order=order).x
logits = output[:, s]
- if forbidden_tokens is not None:
- logits = logits.masked_fill(forbidden_tokens, float("-inf"))
if args.deterministic_synthesis:
t_next = logits.argmax(1)
else:
######################################################################
+def compute_perplexity(model, task, prompt_len, split="train"):
+ with torch.autograd.no_grad():
+ t = model.training
+ model.eval()
+
+ nb_samples, acc_loss = 0, 0.0
+
+ for input in task.batches(split=split):
+ input = input.to(device)
+ output = eval_mygpt(model, input, prompt_len=prompt_len)
+ if args.noncausal_prompt:
+ d = input.size(1) // 2
+ loss = F.cross_entropy(output[:, d:].transpose(1, 2), input[:, d:])
+ else:
+ loss = F.cross_entropy(output.transpose(1, 2), input)
+ acc_loss += loss.item() * input.size(0)
+ nb_samples += input.size(0)
+
+ model.train(t)
+
+ return math.exp(min(100, acc_loss / nb_samples))
+
+
+######################################################################
+
+
+def oneshot_policy_loss(mazes, output, policies, height, width):
+ masks = (mazes == maze.v_empty).unsqueeze(-1)
+ targets = policies.permute(0, 2, 1) * masks
+ output = output * masks
+ return -(output.log_softmax(-1) * targets).sum() / masks.sum()
+
+
+def oneshot_trace_loss(mazes, output, policies, height, width):
+ masks = mazes == maze.v_empty
+ targets = maze.stationary_densities(
+ mazes.view(-1, height, width), policies.view(-1, 4, height, width)
+ ).flatten(-2)
+ targets = targets * masks
+ output = output.squeeze(-1) * masks
+ return (output - targets).abs().sum() / masks.sum()
+
+
+def oneshot(gpt, learning_rate_scheduler, task):
+ t = gpt.training
+ gpt.eval()
+
+ if args.oneshot_input == "head":
+ dim_in = args.dim_model
+ elif args.oneshot_input == "deep":
+ dim_in = args.dim_model * args.nb_blocks * 2
+ else:
+ raise ValueError(f"{args.oneshot_input=}")
+
+ if args.oneshot_output == "policy":
+ dim_out = 4
+ compute_loss = oneshot_policy_loss
+ elif args.oneshot_output == "trace":
+ dim_out = 1
+ compute_loss = oneshot_trace_loss
+ else:
+ raise ValueError(f"{args.oneshot_output=}")
+
+ model = nn.Sequential(
+ nn.Linear(dim_in, args.dim_model),
+ nn.ReLU(),
+ nn.Linear(args.dim_model, args.dim_model),
+ nn.ReLU(),
+ nn.Linear(args.dim_model, dim_out),
+ ).to(device)
+
+ learning_rate_scheduler.reset()
+
+ for n_epoch in range(args.nb_epochs):
+ learning_rate = learning_rate_scheduler.get_learning_rate()
+ optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
+
+ acc_train_loss, nb_train_samples = 0, 0
+ for mazes, policies in task.policy_batches(split="train"):
+ output_gpt = eval_mygpt(
+ gpt, mazes, mode=args.oneshot_input, prompt_len=task.height * task.width
+ )
+ output = model(output_gpt)
+
+ loss = compute_loss(mazes, output, policies, task.height, task.width)
+ acc_train_loss += loss.item() * mazes.size(0)
+ nb_train_samples += mazes.size(0)
+
+ optimizer.zero_grad()
+ loss.backward()
+ optimizer.step()
+
+ learning_rate_scheduler.update(n_epoch + 1, acc_train_loss)
+
+ acc_test_loss, nb_test_samples = 0, 0
+ for mazes, policies in task.policy_batches(split="test"):
+ output_gpt = eval_mygpt(
+ gpt, mazes, mode=args.oneshot_input, prompt_len=task.height * task.width
+ )
+ output = model(output_gpt)
+ loss = compute_loss(mazes, output, policies, task.height, task.width)
+ acc_test_loss += loss.item() * mazes.size(0)
+ nb_test_samples += mazes.size(0)
+
+ log_string(
+ f"diff_ce {n_epoch} train {acc_train_loss/nb_train_samples} test {acc_test_loss/nb_test_samples}"
+ )
+
+ # -------------------
+ mazes = task.test_input[:32, : task.height * task.width]
+ policies = task.test_policies[:32]
+ output_gpt = eval_mygpt(
+ gpt, mazes, mode=args.oneshot_input, prompt_len=task.height * task.width
+ )
+ output = model(output_gpt)
+ if args.oneshot_output == "policy":
+ targets = policies.permute(0, 2, 1)
+ scores = (
+ (F.one_hot(output.argmax(-1), num_classes=4) * targets).sum(-1) == 0
+ ).float()
+ elif args.oneshot_output == "trace":
+ targets = maze.stationary_densities(
+ mazes.view(-1, task.height, task.width),
+ policies.view(-1, 4, task.height, task.width),
+ ).flatten(-2)
+ scores = output
+ else:
+ raise ValueError(f"{args.oneshot_output=}")
+
+ 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, filename),
+ mazes=mazes,
+ score_paths=scores,
+ score_truth=targets,
+ )
+ log_string(f"wrote {filename}")
+
+ # -------------------
+
+ gpt.train(t)
+
+
+######################################################################
+
+
+class LearningRateScheduler:
+ def get_learning_rate(self):
+ pass
+
+ def update(self, nb_finished_epochs, loss):
+ pass
+
+ def reset(self):
+ pass
+
+ def get_state(self):
+ return vars(self)
+
+ def set_state(self, state):
+ print(f"{state=}")
+ for k, v in state.items():
+ setattr(self, k, v)
+
+
+class StepWiseScheduler(LearningRateScheduler):
+ def __init__(self, schedule):
+ self.nb_finished_epochs = 0
+ self.schedule = schedule
+
+ def get_learning_rate(self):
+ return self.schedule[self.nb_finished_epochs]
+
+ def update(self, nb_finished_epochs, loss):
+ self.nb_finished_epochs = nb_finished_epochs
+
+ def reset(self):
+ self.nb_finished_epochs = 0
+
+ def get_state(self):
+ return {"nb_finished_epochs": self.nb_finished_epochs}
+
+
+class AutoScheduler(LearningRateScheduler):
+ def __init__(self, learning_rate_init, growth=1.0, degrowth=0.2):
+ self.learning_rate_init = learning_rate_init
+ self.learning_rate = learning_rate_init
+ self.growth = growth
+ self.degrowth = degrowth
+ self.pred_loss = None
+
+ def get_learning_rate(self):
+ return self.learning_rate
+
+ def update(self, nb_finished_epochs, loss):
+ if self.pred_loss is not None:
+ if loss >= self.pred_loss:
+ self.learning_rate *= self.degrowth
+ else:
+ self.learning_rate *= self.growth
+ self.pred_loss = loss
+
+ def reset(self):
+ self.learning_rate = self.learning_rate_init
+
+ def get_state(self):
+ return {
+ "learning_rate_init": self.learning_rate_init,
+ "pred_loss": self.pred_loss,
+ }
+
+
+######################################################################
+
+
class Task:
- def batches(self, split="train"):
+ def batches(self, split="train", nb_to_use=-1, desc=None):
pass
def vocabulary_size(self):
######################################################################
-import picoclvr
-
-
-class TaskPicoCLVR(Task):
-
- # Make a tensor from a list of strings
- def tensorize(self, descr):
- token_descr = [s.strip().split(" ") for s in descr]
- l = max([len(s) for s in token_descr])
- token_descr = [s + ["<nul>"] * (l - len(s)) for s in token_descr]
- id_descr = [[self.token2id[u] for u in s] for s in token_descr]
- return torch.tensor(id_descr, device=self.device)
-
- # Make a list of strings from a tensor
- def detensorize(self, x):
- return [" ".join([self.id2token[t.item()] for t in r]) for r in x]
-
- # trim all the tensors in the tuple z to remove as much token from
- # left and right in the first tensor. If z is a tuple, all its
- # elements are trimed according to the triming for the first
- def trim(self, z, token="<nul>"):
- n = self.token2id[token]
- if type(z) == tuple:
- x = z[0]
- i = (1 - (F.pad(x, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
- a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
- return tuple([t[:, a:b] for t in z])
- else:
- i = (1 - (F.pad(z, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
- a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
- return z[:, a:b]
-
- ######################
- # Not the cleanest part of the code
-
- # Extract the last image of each sequence, from the last <img>
- # included, and set to <nul> all the tokens from the beginning of
- # that image to the end
- def excise_last_image(self, input):
- t_img, t_nul = self.token2id["<img>"], self.token2id["<nul>"]
- nb_img_tokens = self.height * self.width + 1
-
- input = input.clone()
- t = (input == t_img).long()
- tail_masks = (t.cumsum(dim=1) == t.sum(dim=1, keepdim=True)).long()
- i = (t * tail_masks).nonzero(as_tuple=True)
- j = (
- i[0][:, None],
- i[1][:, None] + torch.arange(nb_img_tokens, device=input.device)[None, :],
- )
- images = self.trim(input[j])
- input[j] = t_nul
- loss_masks = 1 - tail_masks
- input, loss_masks = self.trim((input, loss_masks))
- return input, loss_masks, images
-
- def add_true_image(self, input, images, loss_masks):
- t_nul = self.token2id["<nul>"]
- nb_img_tokens = self.height * self.width + 1
- input = F.pad(input, (0, nb_img_tokens), value=t_nul)
- loss_masks = F.pad(loss_masks, (0, nb_img_tokens), value=0)
- t = (input == t_nul).long()
- i = (t.cumsum(dim=1) == 1).nonzero(as_tuple=True)
- j = (
- i[0][:, None],
- i[1][:, None] + torch.arange(nb_img_tokens, device=input.device)[None, :],
- )
- input[j] = images
- loss_masks[j] = 1
- input, loss_masks = self.trim((input, loss_masks))
- return input, loss_masks
-
- def add_generated_image(self, input, loss_masks, model):
- t_img, t_nul = self.token2id["<img>"], self.token2id["<nul>"]
- nb_img_tokens = self.height * self.width + 1
-
- input = F.pad(input, (0, nb_img_tokens), value=t_nul)
- loss_masks = F.pad(loss_masks, (0, nb_img_tokens), value=0)
- t = (input == t_nul).long()
- i = (t.cumsum(dim=1) == 1).nonzero(as_tuple=True)
- input[i] = t_img
-
- j = (
- i[0][:, None],
- i[1][:, None]
- + 1
- + torch.arange(nb_img_tokens - 1, device=input.device)[None, :],
- )
- ar_masks = input.new_zeros(input.size(), dtype=torch.int64)
- ar_masks[j] = 1
- forbidden_tokens = (
- torch.arange(self.vocabulary_size(), device=input.device) == t_nul
- )
- with torch.autograd.no_grad():
- t = model.training
- model.eval()
- masked_inplace_autoregression(
- model,
- self.batch_size,
- input,
- ar_masks,
- forbidden_tokens,
- device=self.device,
- )
- model.train(t)
+import maze
- input, loss_masks = self.trim((input, loss_masks))
- return input, loss_masks
+class TaskMaze(Task):
+ def map2seq(self, *m):
+ return torch.cat([x.flatten(1) for x in m], 1)
- ######################
+ def seq2map(self, s):
+ s = s.reshape(s.size(0), -1, self.height, self.width)
+ return (s[:, k] for k in range(s.size(1)))
def __init__(
self,
+ nb_train_samples,
+ nb_test_samples,
batch_size,
height,
width,
- nb_colors=5,
+ nb_walls,
device=torch.device("cpu"),
- pruner_train=None,
- pruner_eval=None,
):
- def generate_descr(nb, cache_suffix, pruner):
- return picoclvr.generate(
- nb,
- height=self.height,
- width=self.width,
- nb_colors=nb_colors,
- pruner=pruner,
- )
-
+ self.batch_size = batch_size
self.height = height
self.width = width
- self.batch_size = batch_size
self.device = device
- nb = args.data_size if args.data_size > 0 else 250000
- self.pruner_train = pruner_train
- self.pruner_eval = pruner_eval
-
- param = {
- "nb": nb,
- "height": height,
- "width": width,
- "nb_colors": nb_colors,
- "batch_size": batch_size,
- "rng_state": list(torch.get_rng_state()),
- }
- log_string(f"generating {nb} samples (can take some time)")
- self.train_descr = generate_descr(
- (nb * 4) // 5, "train", pruner=self.pruner_train
+ train_mazes, train_paths, train_policies = maze.create_maze_data(
+ nb_train_samples,
+ height=height,
+ width=width,
+ nb_walls=nb_walls,
+ progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
+ )
+ self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
+ self.train_policies = train_policies.flatten(-2).to(device)
+
+ test_mazes, test_paths, test_policies = maze.create_maze_data(
+ nb_test_samples,
+ height=height,
+ width=width,
+ nb_walls=nb_walls,
+ progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
)
- self.test_descr = generate_descr((nb * 1) // 5, "test", pruner=None)
-
- # Build the tokenizer
- tokens = {"<nul>", "<img>"}
- for d in [self.train_descr, self.test_descr]:
- for s in d:
- for t in s.strip().split(" "):
- tokens.add(t)
- # make this set a sorted list to get the same tensors given
- # the same descr
- tokens = list(tokens)
- tokens.sort()
- self.token2id = dict([(t, n) for n, t in enumerate(tokens)])
- self.id2token = dict([(n, t) for n, t in enumerate(tokens)])
-
- # Tokenize the train and test sets
- self.train_input = self.tensorize(self.train_descr)
- self.test_input = self.tensorize(self.test_descr)
-
- def batches(self, split="train"):
+ self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
+ self.test_policies = test_policies.flatten(-2).to(device)
+
+ self.nb_codes = self.train_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
+ if nb_to_use > 0:
+ input = input[:nb_to_use]
+ if desc is None:
+ desc = f"epoch-{split}"
for batch in tqdm.tqdm(
- input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
+ input.split(self.batch_size), dynamic_ncols=True, desc=desc
):
- yield self.trim(batch)
-
- def vocabulary_size(self):
- return len(self.token2id)
+ yield batch
- def compute_missing_properties(self, n_epoch, model, pruner=None):
+ def policy_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
+ policies = self.train_policies if split == "train" else self.test_policies
+ input = input[:, : self.height * self.width]
+ policies = policies * (input != maze.v_wall)[:, None]
- acc_nb_requested_properties = []
- acc_nb_missing_properties = []
- acc_nb_results = 0
+ if nb_to_use > 0:
+ input = input[:nb_to_use]
+ policies = policies[:nb_to_use]
- for input in tqdm.tqdm(
- self.test_input.split(self.batch_size),
+ if desc is None:
+ desc = f"epoch-{split}"
+ for batch in tqdm.tqdm(
+ zip(input.split(self.batch_size), policies.split(self.batch_size)),
dynamic_ncols=True,
- desc=f"test-properties",
+ desc=desc,
):
- tape, loss_masks, _ = self.excise_last_image(input)
- tape, loss_masks = self.add_generated_image(tape, loss_masks, model)
- result_descr = self.detensorize(tape)
- np = picoclvr.nb_properties(
- result_descr,
- height=self.height,
- width=self.width,
- pruner=pruner,
- )
- nb_requested_properties, _, nb_missing_properties = zip(*np)
- acc_nb_requested_properties += nb_requested_properties
- acc_nb_missing_properties += nb_missing_properties
- acc_nb_results += len(result_descr)
-
- nb_requested_properties = sum(acc_nb_requested_properties)
- nb_missing_properties = sum(acc_nb_missing_properties)
+ yield batch
- prefix = "" if pruner is None else "pruned_"
- log_string(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
- log_string(
- f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
- )
- log_string(
- f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
- )
+ def vocabulary_size(self):
+ return self.nb_codes
+
+ def compute_error(self, model, split="train", nb_to_use=-1):
+ nb_total, nb_correct = 0, 0
+ for input in task.batches(split, nb_to_use):
+ result = input.clone()
+ ar_mask = result.new_zeros(result.size())
+ ar_mask[:, self.height * self.width :] = 1
+ result *= 1 - 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(result)
+ nb_correct += maze.path_correctness(mazes, paths).long().sum()
+ nb_total += mazes.size(0)
- ######################################################################
+ return nb_total, nb_correct
def produce_results(self, n_epoch, model):
+ with torch.autograd.no_grad():
+ t = model.training
+ model.eval()
- self.compute_missing_properties(n_epoch, model)
-
- if self.pruner_eval is not None:
- self.compute_missing_properties(n_epoch, model, self.pruner_eval)
-
- nb_tokens_to_generate = self.height * self.width + 3
- result_descr = []
- nb_per_primer = 8
- primer = []
-
- for primer_descr in [
- "red above green <sep> green top <sep> blue right of red",
- "there is red <sep> there is yellow <sep> there is blue",
- "red below yellow <sep> yellow below green <sep> green below blue <sep> red right <sep> yellow left <sep> green right <sep> blue left",
- "green bottom <sep> yellow bottom <sep> green left of blue <sep> yellow right of blue <sep> blue top",
- ]:
- primer += [primer_descr] * nb_per_primer
-
- tape = self.tensorize(primer)
- loss_masks = 1 - (tape == self.token2id["<nul>"]).long()
- tape, loss_masks = self.add_generated_image(tape, loss_masks, model)
- result_descr = self.detensorize(tape)
-
- np = picoclvr.nb_properties(result_descr, height=self.height, width=self.width)
-
- acc_nb_requested_properties, _, acc_nb_missing_properties = zip(*np)
- acc_nb_results = len(result_descr)
-
- nb_requested_properties = sum(acc_nb_requested_properties)
- nb_missing_properties = sum(acc_nb_missing_properties)
+ train_nb_total, train_nb_correct = self.compute_error(
+ model, "train", nb_to_use=1000
+ )
+ log_string(
+ f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
+ )
- prefix = "demo_"
- log_string(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
- log_string(
- f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
- )
- log_string(
- f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
- )
+ test_nb_total, test_nb_correct = self.compute_error(
+ model, "test", nb_to_use=1000
+ )
+ log_string(
+ f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
+ )
- img = picoclvr.descr2img(result_descr, height=self.height, width=self.width)
+ input = self.test_input[:32]
+ result = input.clone()
+ ar_mask = result.new_zeros(result.size())
+ ar_mask[:, self.height * self.width :] = 1
+ result *= 1 - 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, filename),
+ mazes=mazes,
+ target_paths=paths,
+ predicted_paths=predicted_paths,
+ path_correct=maze.path_correctness(mazes, predicted_paths),
+ )
+ log_string(f"wrote {filename}")
- if img.dim() == 5:
- if img.size(1) == 1:
- img = F.pad(img.squeeze(1), pad=(1, 1, 1, 1), value=64)
- else:
- img = torch.cat(
- [
- torchvision.utils.make_grid(x, padding=1, pad_value=64)[None]
- for x in img
- ],
- 0,
- )
-
- image_name = os.path.join(args.result_dir, f"result_{n_epoch:04d}.png")
- torchvision.utils.save_image(
- img / 255.0, image_name, nrow=nb_per_primer, padding=1, pad_value=1.0
- )
- log_string(f"wrote {image_name}")
+ model.train(t)
######################################################################
log_string(f"device {device}")
-def pruner_horizontal_green(p):
- return not ("green" in p and ("left" in p or "right" in p))
-
-
-task = TaskPicoCLVR(
+task = TaskMaze(
+ nb_train_samples=args.nb_train_samples,
+ nb_test_samples=args.nb_test_samples,
batch_size=args.batch_size,
- height=args.height,
- width=args.width,
- nb_colors=args.nb_colors,
+ height=args.maze_height,
+ width=args.maze_width,
+ nb_walls=args.maze_nb_walls,
device=device,
- pruner_train=pruner_horizontal_green
- if args.prune_properties in {"train+eval"}
- else None,
- pruner_eval=(lambda p: not pruner_horizontal_green(p))
- if args.prune_properties in {"train+eval", "eval"}
- else None,
)
+
vocabulary_size = task.vocabulary_size()
log_string(f"vocabulary_size {vocabulary_size}")
##############################
+
+def noncausal_prompt_amm_generator(d):
+ q = torch.arange(d)[:, None]
+ k = torch.arange(d)[None, :]
+ s = args.maze_height * args.maze_width
+ # return torch.logical_and(q < k, torch.logical_or(q >= s, k >= s))
+ return q < k
+
+
+amm_generator = None
+
+if args.noncausal_prompt:
+ amm_generator = noncausal_prompt_amm_generator
+
model = mygpt.MyGPT(
vocabulary_size=vocabulary_size,
dim_model=args.dim_model,
nb_blocks=args.nb_blocks,
causal=True,
dropout=args.dropout,
+ amm_generator=amm_generator,
)
model.to(device)
######################################################################
+if args.learning_rate_schedule == "auto":
+ learning_rate_scheduler = AutoScheduler(args.learning_rate)
+
+elif args.learning_rate_schedule == "cos":
+ schedule = {}
+ for n_epoch in range(args.nb_epochs):
+ u = n_epoch / args.nb_epochs * math.pi
+ schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
+ learning_rate_scheduler = StepWiseScheduler(schedule)
+ log_string(f"learning_rate_schedule {schedule}")
+
+else:
+ u = {
+ int(k): float(v)
+ for k, v in [
+ tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
+ ]
+ }
+
+ schedule = {}
+ learning_rate = args.learning_rate
+ for n_epoch in range(args.nb_epochs):
+ if n_epoch in u:
+ learning_rate = u[n_epoch]
+ schedule[n_epoch] = learning_rate
+ learning_rate_scheduler = StepWiseScheduler(schedule)
+ log_string(f"learning_rate_schedule {schedule}")
+
+######################################################################
+
nb_epochs_finished = 0
if args.no_checkpoint:
checkpoint = torch.load(checkpoint_name)
nb_epochs_finished = checkpoint["nb_epochs_finished"]
model.load_state_dict(checkpoint["model_state"])
+ learning_rate_scheduler.set_state(checkpoint["learning_rate_scheduler_state"])
torch.set_rng_state(checkpoint["rng_state"])
if torch.cuda.is_available():
torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
except FileNotFoundError:
log_string("starting from scratch.")
- except:
- log_string("error when loading the checkpoint.")
- exit(1)
+ # except:
+ # log_string("error when loading the checkpoint.")
+ # exit(1)
######################################################################
-nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
-
token_count = 0
for input in task.batches(split="train"):
token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
##############################
-if args.learning_rate_schedule == "cos":
- learning_rate_schedule = {}
- for n_epoch in range(args.nb_epochs):
- u = n_epoch / args.nb_epochs * math.pi
- learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
-else:
- u = {
- int(k): float(v)
- for k, v in [
- tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
- ]
- }
+if nb_epochs_finished >= args.nb_epochs:
+ n_epoch = nb_epochs_finished
+ train_perplexity = compute_perplexity(
+ model, task, prompt_len=task.height * task.width, split="train"
+ )
+ test_perplexity = compute_perplexity(
+ model, task, prompt_len=task.height * task.width, split="test"
+ )
- learning_rate_schedule = {}
- learning_rate = args.learning_rate
- for n_epoch in range(args.nb_epochs):
- if n_epoch in u:
- learning_rate = u[n_epoch]
- learning_rate_schedule[n_epoch] = learning_rate
+ log_string(
+ f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
+ )
-log_string(f"learning_rate_schedule {learning_rate_schedule}")
+ task.produce_results(n_epoch, model)
##############################
-nb_samples_seen = 0
-
-if nb_epochs_finished >= nb_epochs:
- task.produce_results(nb_epochs_finished, model)
-
-for n_epoch in range(nb_epochs_finished, nb_epochs):
+learning_rate_scheduler.reset()
- learning_rate = learning_rate_schedule[n_epoch]
+for n_epoch in range(nb_epochs_finished, args.nb_epochs):
+ learning_rate = learning_rate_scheduler.get_learning_rate()
log_string(f"learning_rate {learning_rate}")
elif args.optim == "adamw":
optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
else:
- raise ValueError(f"Unknown optimizer {args.optim}.")
+ raise ValueError(f"{args.optim=}")
model.train()
for input in task.batches(split="train"):
input = input.to(device)
- output = model(mygpt.BracketedSequence(input)).x
- loss = F.cross_entropy(output.transpose(1, 2), input)
+ output = eval_mygpt(model, input, prompt_len=task.height * task.width)
+ if args.noncausal_prompt:
+ d = input.size(1) // 2
+ loss = F.cross_entropy(output[:, d:].transpose(1, 2), input[:, d:])
+ else:
+ loss = F.cross_entropy(output.transpose(1, 2), input)
acc_train_loss += loss.item() * input.size(0)
nb_train_samples += input.size(0)
- nb_samples_seen += input.size(0)
optimizer.zero_grad()
loss.backward()
optimizer.step()
- with torch.autograd.no_grad():
-
- model.eval()
+ learning_rate_scheduler.update(n_epoch + 1, acc_train_loss)
- nb_test_samples, acc_test_loss = 0, 0.0
+ train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
+ test_perplexity = compute_perplexity(
+ model, task, prompt_len=task.height * task.width, split="test"
+ )
- for input in task.batches(split="test"):
- input = input.to(device)
+ log_string(
+ f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
+ )
- # input, loss_masks, true_images = task.excise_last_image(input)
- # input, loss_masks = task.add_true_image(input, true_images, loss_masks)
-
- output = model(mygpt.BracketedSequence(input)).x
- loss = F.cross_entropy(output.transpose(1, 2), input)
- acc_test_loss += loss.item() * input.size(0)
- nb_test_samples += input.size(0)
-
- train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
- test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
-
- log_string(
- f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
- )
-
- task.produce_results(n_epoch, model)
+ task.produce_results(n_epoch, model)
checkpoint = {
"nb_epochs_finished": n_epoch + 1,
"model_state": model.state_dict(),
+ "learning_rate_scheduler_state": learning_rate_scheduler.get_state(),
"rng_state": torch.get_rng_state(),
}
log_string(f"saved checkpoint {checkpoint_name}")
######################################################################
+
+if args.oneshot:
+ oneshot(model, learning_rate_scheduler, task)
+
+######################################################################