# 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("--seed", type=int, default=0)
-parser.add_argument("--nb_epochs", type=int, default=25)
+parser.add_argument("--nb_epochs", type=int, default=None)
-parser.add_argument("--batch_size", type=int, default=25)
+parser.add_argument("--batch_size", type=int, default=None)
parser.add_argument("--nb_train_samples", type=int, default=250000)
parser.add_argument("--maze_nb_walls", type=int, default=15)
+##############################
+# Snake options
+
+parser.add_argument("--snake_height", type=int, default=6)
+
+parser.add_argument("--snake_width", type=int, default=8)
+
+parser.add_argument("--snake_nb_colors", type=int, default=5)
+
+parser.add_argument("--snake_length", type=int, default=400)
+
######################################################################
args = parser.parse_args()
######################################################################
+default_args = {
+ "picoclvr": {
+ "nb_epochs": 25,
+ "batch_size": 25,
+ },
+ "mnist": {
+ "nb_epochs": 25,
+ "batch_size": 10,
+ },
+ "maze": {
+ "nb_epochs": 25,
+ "batch_size": 25,
+ },
+ "snake": {
+ "nb_epochs": 25,
+ "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(
masked_inplace_autoregression(
model, self.batch_size, results, ar_mask, device=self.device
)
- image_name = os.path.join(args.result_dir, f"result_mnist_{n_epoch:04d}.png")
+ 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,
mazes, paths = self.seq2map(input)
_, predicted_paths = self.seq2map(result)
- filename = os.path.join(args.result_dir, f"result_{n_epoch:04d}.png")
+ filename = os.path.join(args.result_dir, f"maze_result_{n_epoch:04d}.png")
maze.save_image(
filename,
mazes=mazes,
######################################################################
+
+import snake
+
+
class TaskSnake(Task):
def __init__(
self,
batch_size,
height,
width,
- nb_walls,
+ 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.test_input =
+ self.train_input, self.train_prior_visits = snake.generate_sequences(
+ nb_train_samples,
+ height,
+ width,
+ nb_colors,
+ length,
+ prompt_length,
+ self.device,
+ )
+ self.test_input, self.test_prior_visits = snake.generate_sequences(
+ nb_test_samples,
+ height,
+ width,
+ nb_colors,
+ length,
+ prompt_length,
+ self.device,
+ )
- self.nb_codes = max(self.train_input.max(), self.train_input.max()) + 1
+ 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"}
):
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]
+ )
+
+ 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)
+
######################################################################
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}")