# torch.backends.cuda.matmul.allow_tf23
# torch.autocast(torch.bfloat16)
-import math, sys, argparse, time, tqdm, itertools, os
+import math, sys, argparse, time, tqdm, os
import torch, torchvision
from torch import nn
######################################################################
parser = argparse.ArgumentParser(
- description="An implementation of GPT with cache to solve a toy geometric reasoning task."
+ description="An implementation of GPT with cache.",
+ formatter_class=argparse.ArgumentDefaultsHelpFormatter,
)
+parser.add_argument("--task", type=str, default="picoclvr")
+
parser.add_argument("--log_filename", type=str, default="train.log")
parser.add_argument("--result_dir", type=str, default="results_default")
parser.add_argument("--nb_epochs", type=int, default=25)
-parser.add_argument("--batch_size", type=int, default=100)
+parser.add_argument("--batch_size", type=int, default=None)
+
+parser.add_argument("--nb_train_samples", type=int, default=250000)
-parser.add_argument("--data_size", type=int, default=-1)
+parser.add_argument("--nb_test_samples", type=int, default=10000)
parser.add_argument("--optim", type=str, default="adam")
-parser.add_argument("--learning_rate", type=float, default=1e-3)
+parser.add_argument("--learning_rate", type=float, default=1e-4)
-parser.add_argument(
- "--learning_rate_schedule", type=str, default="10: 2e-4,20: 4e-5,30: 8e-6"
-)
+parser.add_argument("--learning_rate_schedule", type=str, default="10: 2e-5,30: 4e-6")
parser.add_argument("--dim_model", type=int, default=512)
##############################
# picoclvr options
-parser.add_argument("--nb_colors", type=int, default=5)
+parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
+
+parser.add_argument("--picoclvr_height", type=int, default=12)
+
+parser.add_argument("--picoclvr_width", type=int, default=16)
+
+parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
+
+##############################
+# 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)
-parser.add_argument("--height", type=int, default=12)
+##############################
+# Snake options
+
+parser.add_argument("--snake_height", type=int, default=6)
+
+parser.add_argument("--snake_width", type=int, default=8)
-parser.add_argument("--width", type=int, default=16)
+parser.add_argument("--snake_nb_colors", type=int, default=3)
-parser.add_argument("--prune_properties", type=str, default="none")
+parser.add_argument("--snake_length", type=int, default=400)
######################################################################
args = parser.parse_args()
-assert args.prune_properties in {"none", "train+eval", "eval"}
+assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
try:
os.mkdir(args.result_dir)
print(f"result directory {args.result_dir} already exists")
exit(1)
-log_file = open(os.path.join(args.result_dir, args.log_filename), "w")
+log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
if args.seed >= 0:
# torch.backends.cudnn.deterministic = True
######################################################################
+default_args = {
+ "picoclvr": {
+ "batch_size": 25,
+ },
+ "mnist": {
+ "batch_size": 10,
+ },
+ "maze": {
+ "batch_size": 25,
+ },
+ "snake": {
+ "batch_size": 20,
+ },
+}
+
+if args.task in default_args:
+ for k, v in default_args[args.task].items():
+ if getattr(args, k) is None:
+ setattr(args, k, v)
+
+######################################################################
+
def log_string(s):
t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
def masked_inplace_autoregression(
model, batch_size, input, ar_mask, forbidden_tokens=None, device=torch.device("cpu")
):
-
- for input, ar_mask in zip(input.split(batch_size), ar_mask.split(batch_size)):
+ for input, ar_mask in tqdm.tqdm(
+ zip(input.split(batch_size), ar_mask.split(batch_size)),
+ dynamic_ncols=True,
+ desc="autoregression",
+ total=input.size(0) // batch_size,
+ ):
i = (ar_mask.sum(0) > 0).nonzero()
if i.min() > 0:
model(
class TaskPicoCLVR(Task):
-
# Make a tensor from a list of strings
def tensorize(self, descr):
token_descr = [s.strip().split(" ") for s in descr]
def __init__(
self,
+ nb_train_samples,
+ nb_test_samples,
batch_size,
height,
width,
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,
+ "nb_train_samples": nb_train_samples,
+ "nb_test_samples": nb_test_samples,
"height": height,
"width": width,
"nb_colors": nb_colors,
"rng_state": list(torch.get_rng_state()),
}
- log_string(f"generating {nb} samples (can take some time)")
+ log_string(
+ f"generating {nb_train_samples+nb_test_samples} samples (can take some time)"
+ )
self.train_descr = generate_descr(
- (nb * 4) // 5, "train", pruner=self.pruner_train
+ nb_train_samples, "train", pruner=self.pruner_train
)
- self.test_descr = generate_descr((nb * 1) // 5, "test", pruner=None)
+ self.test_descr = generate_descr(nb_test_samples, "test", pruner=None)
# Build the tokenizer
tokens = {"<nul>", "<img>"}
return len(self.token2id)
def compute_missing_properties(self, n_epoch, model, pruner=None):
-
acc_nb_requested_properties = []
acc_nb_missing_properties = []
acc_nb_results = 0
######################################################################
def produce_results(self, n_epoch, model):
-
self.compute_missing_properties(n_epoch, model)
if self.pruner_eval is not None:
0,
)
- image_name = os.path.join(args.result_dir, f"result_{n_epoch:04d}.png")
+ image_name = os.path.join(args.result_dir, f"picoclvr_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"device {device}")
+class TaskMNIST(Task):
+ def __init__(self, batch_size, device=torch.device("cpu")):
+ self.device = device
+ self.batch_size = batch_size
+
+ def batches(self, split="train"):
+ assert split in {"train", "test"}
+ data_set = torchvision.datasets.MNIST(
+ root="./data", train=(split == "train"), download=True
+ )
+ data_input = data_set.data.view(-1, 28 * 28).long()
+ if args.nb_train_samples is not None:
+ data_input = data_input[: args.nb_train_samples]
+ for batch in tqdm.tqdm(
+ data_input.split(self.batch_size), desc=f"epoch-{split}"
+ ):
+ yield batch
+
+ def vocabulary_size(self):
+ return 256
+
+ def produce_results(self, n_epoch, model):
+ results = torch.empty(64, 28 * 28, device=self.device, dtype=torch.int64)
+ ar_mask = torch.full_like(results, 1)
+ masked_inplace_autoregression(
+ model, self.batch_size, results, ar_mask, device=self.device
+ )
+ image_name = os.path.join(args.result_dir, f"mnist_result_{n_epoch:04d}.png")
+ torchvision.utils.save_image(
+ 1 - results.reshape(-1, 1, 28, 28) / 255.0,
+ image_name,
+ nrow=16,
+ pad_value=0.8,
+ )
+ log_string(f"wrote {image_name}")
+
+
+######################################################################
+
+import maze
+
+
+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_walls,
+ device=torch.device("cpu"),
+ ):
+ self.batch_size = batch_size
+ self.height = height
+ self.width = width
+ self.device = device
+
+ train_mazes, train_paths, _ = 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))
+
+ test_mazes, test_paths, _ = 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_input = self.map2seq(test_mazes.to(device), test_paths.to(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
+ 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=desc
+ ):
+ yield batch
+
+ 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
+ masked_inplace_autoregression(
+ model, self.batch_size, result, ar_mask, device=self.device
+ )
+ 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()
+
+ 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}%"
+ )
+
+ 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}%"
+ )
+
+ input = self.test_input[:48]
+ result = input.clone()
+ 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, device=self.device
+ )
+
+ mazes, paths = self.seq2map(input)
+ _, predicted_paths = self.seq2map(result)
+
+ filename = os.path.join(args.result_dir, f"maze_result_{n_epoch:04d}.png")
+ maze.save_image(
+ 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)
+
+
+######################################################################
+
+
+def generate_snake_sequences(
+ nb, height, width, nb_colors, length, prompt_length, device=torch.device("cpu")
+):
+ worlds = torch.randint(nb_colors, (nb, height, width), device=device)
+ nb_prior_visits = torch.zeros(nb, height, width, device=device)
+
+ # nb x 2
+ snake_position = torch.cat(
+ (
+ torch.randint(height, (nb, 1), device=device),
+ torch.randint(width, (nb, 1), device=device),
+ ),
+ 1,
+ )
+ snake_direction = torch.randint(4, (nb,), device=device)
+ sequences = torch.empty(nb, 2 * length, device=device, dtype=torch.int64)
+ sequences_prior_visits = torch.zeros(
+ nb, 2 * length, device=device, dtype=torch.int64
+ )
+ i = torch.arange(nb, device=device) # [:,None]
+
+ for l in range(length):
+ # nb x 3
+ snake_next_direction = torch.cat(
+ (
+ (snake_direction[:, None] - 1) % 4,
+ snake_direction[:, None],
+ (snake_direction[:, None] + 1) % 4,
+ ),
+ 1,
+ )
+
+ # nb x 3
+ vh = (snake_next_direction + 1) % 2 * (snake_next_direction - 1)
+ vw = snake_next_direction % 2 * (snake_next_direction - 2)
+
+ # nb x 3 x 2
+ snake_next_speed = torch.cat((vh[:, :, None], vw[:, :, None]), 2)
+ snake_next_position = snake_position[:, None, :] + snake_next_speed
+
+ # nb x 3
+ val = torch.logical_and(
+ torch.logical_and(
+ snake_next_position[:, :, 0] >= 0, snake_next_position[:, :, 0] < height
+ ),
+ torch.logical_and(
+ snake_next_position[:, :, 1] >= 0, snake_next_position[:, :, 1] < width
+ ),
+ ).float()
+ val = (
+ # The multiplicative factors bias toward moving forward
+ torch.rand_like(val)
+ * val
+ * torch.tensor([[1.0, 2.0, 1.0]], device=device)
+ )
+
+ # nb
+ j = val.argmax(1)
+ snake_direction = snake_next_direction[i, j]
+
+ sequences[:, 2 * l] = worlds[i, snake_position[:, 0], snake_position[:, 1]] + 4
+ sequences_prior_visits[:, 2 * l] = nb_prior_visits[
+ i, snake_position[:, 0], snake_position[:, 1]
+ ]
+ if l < prompt_length:
+ nb_prior_visits[i, snake_position[:, 0], snake_position[:, 1]] += 1
+ sequences[:, 2 * l + 1] = snake_direction
+
+ # nb x 2
+ snake_position = snake_next_position[i, j]
+
+ return sequences, sequences_prior_visits
+
+
+# generate_snake_sequences(nb=1, height=4, width=6, nb_colors=3, length=20)
+# exit(0)
+
+
+def snake_solver(input, ar_mask):
+ for n in range(input.size(0)):
+ i, j, memory = 0, 0, {}
+ # print(input[n])
+ # print(ar_mask[n])
+ for l in range(input.size(1) // 2):
+ if ar_mask[n, 2 * l] == 1:
+ if memory.get((i, j)) is None:
+ input[n, 2 * l] = -1
+ else:
+ input[n, 2 * l] = memory[(i, j)]
+ else:
+ # print(f'@3 {memory=}')
+ if memory.get((i, j)) is None:
+ memory[(i, j)] = input[n, 2 * l]
+ else:
+ assert memory[(i, j)] == input[n, 2 * l], f"n={n} l={l}"
+ # print(f'@1 {i=} {j=}')
+ d = input[n, 2 * l + 1].item()
+ i += (d + 1) % 2 * (d - 1)
+ j += d % 2 * (d - 2)
+ # print(f'@2 {i=} {j=}')
+
+
+class TaskSnake(Task):
+ def __init__(
+ self,
+ nb_train_samples,
+ nb_test_samples,
+ batch_size,
+ height,
+ width,
+ nb_colors,
+ length,
+ prompt_length,
+ device=torch.device("cpu"),
+ ):
+ self.batch_size = batch_size
+ self.height = height
+ self.width = width
+ self.device = device
+ self.prompt_length = prompt_length
+
+ self.train_input, self.train_prior_visits = generate_snake_sequences(
+ nb_train_samples,
+ height,
+ width,
+ nb_colors,
+ length,
+ prompt_length,
+ self.device,
+ )
+ self.test_input, self.test_prior_visits = generate_snake_sequences(
+ nb_test_samples,
+ height,
+ width,
+ nb_colors,
+ length,
+ prompt_length,
+ 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
+ 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=desc
+ ):
+ yield batch
+
+ def vocabulary_size(self):
+ return self.nb_codes
+
+ def produce_results(self, n_epoch, model):
+ with torch.autograd.no_grad():
+ t = model.training
+ model.eval()
+
+ def compute_nb_correct(input, prior_visits):
+ result = input.clone()
+ i = torch.arange(result.size(1), device=result.device)[None, :]
+ ar_mask = (
+ torch.logical_and(i >= self.prompt_length * 2, i % 2 == 0)
+ .long()
+ .expand_as(result)
+ )
+ result *= 1 - ar_mask
+
+ # snake_solver(result,ar_mask)
+
+ masked_inplace_autoregression(
+ model, self.batch_size, result, ar_mask, device=self.device
+ )
+
+ nb_total = ((prior_visits > 0) * ar_mask).sum()
+
+ nb_correct = (
+ (result == input).long() * (prior_visits > 0) * ar_mask
+ ).sum()
+
+ # nb_total = result.size(0)
+ # nb_correct = ((result - input).abs().sum(1) == 0).sum()
+
+ return nb_total, nb_correct
+
+ # train_nb_total, train_nb_correct = compute_nb_correct(
+ # self.train_input, self.train_prior_visits
+ # )
+
+ # 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}%"
+ # )
+
+ test_nb_total, test_nb_correct = compute_nb_correct(
+ self.test_input[:1000], self.test_prior_visits[:1000]
+ )
-def pruner_horizontal_green(p):
+ 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}%"
+ )
+
+ model.train(t)
+
+
+######################################################################
+
+
+def picoclvr_pruner_horizontal_green(p):
return not ("green" in p and ("left" in p or "right" in p))
-task = TaskPicoCLVR(
- batch_size=args.batch_size,
- height=args.height,
- width=args.width,
- nb_colors=args.nb_colors,
- 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,
+picoclvr_pruner_train = (
+ picoclvr_pruner_horizontal_green
+ if args.picocvlr_prune_properties in {"train+eval"}
+ else None
)
+picoclvr_pruner_eval = (
+ (lambda p: not picoclvr_pruner_horizontal_green(p))
+ if args.picocvlr_prune_properties in {"train+eval", "eval"}
+ else None
+)
+
+######################################################################
+
+if args.task == "picoclvr":
+ task = TaskPicoCLVR(
+ nb_train_samples=args.nb_train_samples,
+ nb_test_samples=args.nb_test_samples,
+ batch_size=args.batch_size,
+ height=args.picoclvr_height,
+ width=args.picoclvr_width,
+ nb_colors=args.picoclvr_nb_colors,
+ device=device,
+ pruner_train=picoclvr_pruner_train,
+ pruner_eval=picoclvr_pruner_eval,
+ )
+
+elif args.task == "mnist":
+ task = TaskMNIST(
+ batch_size=args.batch_size,
+ device=device,
+ )
+
+elif args.task == "maze":
+ task = TaskMaze(
+ nb_train_samples=args.nb_train_samples,
+ nb_test_samples=args.nb_test_samples,
+ batch_size=args.batch_size,
+ height=args.maze_height,
+ width=args.maze_width,
+ nb_walls=args.maze_nb_walls,
+ device=device,
+ )
+
+elif args.task == "snake":
+ task = TaskSnake(
+ nb_train_samples=args.nb_train_samples,
+ nb_test_samples=args.nb_test_samples,
+ batch_size=args.batch_size,
+ height=args.snake_height,
+ width=args.snake_width,
+ nb_colors=args.snake_nb_colors,
+ length=args.snake_length,
+ prompt_length=args.snake_length // 2,
+ device=device,
+ )
+
+else:
+ raise ValueError(f"Unknown task {args.task}")
+
+######################################################################
+
+log_string(f"device {device}")
+
vocabulary_size = task.vocabulary_size()
log_string(f"vocabulary_size {vocabulary_size}")
task.produce_results(nb_epochs_finished, model)
for n_epoch in range(nb_epochs_finished, nb_epochs):
-
learning_rate = learning_rate_schedule[n_epoch]
log_string(f"learning_rate {learning_rate}")
optimizer.step()
with torch.autograd.no_grad():
-
model.eval()
nb_test_samples, acc_test_loss = 0, 0.0