# 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("--snake_width", type=int, default=8)
-parser.add_argument("--snake_nb_colors", type=int, default=3)
+parser.add_argument("--snake_nb_colors", type=int, default=5)
-parser.add_argument("--snake_length", type=int, default=400)
+parser.add_argument("--snake_length", type=int, default=200)
######################################################################
######################################################################
+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": 5,
+ "batch_size": 25,
+ },
+}
+
+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())
######################################################################
+# ra_mask is boolean, with 1s on the values to generate
+
+
def masked_inplace_autoregression(
- model, batch_size, input, ar_mask, forbidden_tokens=None, device=torch.device("cpu")
+ model,
+ batch_size,
+ input,
+ ar_mask,
+ forbidden_tokens=None,
+ progress_bar_desc="autoregression",
+ device=torch.device("cpu"),
):
- for input, ar_mask in zip(input.split(batch_size), ar_mask.split(batch_size)):
+ batches = zip(input.split(batch_size), ar_mask.split(batch_size))
+ if progress_bar_desc is not None:
+ tqdm.tqdm(
+ batches,
+ dynamic_ncols=True,
+ desc=progress_bar_desc,
+ total=input.size(0) // batch_size,
+ )
+ for input, ar_mask in batches:
i = (ar_mask.sum(0) > 0).nonzero()
if i.min() > 0:
model(
input,
ar_masks,
forbidden_tokens,
+ progress_bar_desc=None,
device=self.device,
)
model.train(t)
######################################################################
-def generate_snake_sequences(
- nb, height, width, nb_colors, length, device=torch.device("cpu")
-):
- worlds = torch.randint(nb_colors, (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)
- 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 = (
- torch.rand_like(val) * val * torch.tensor([[1.0, 4.0, 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[:, 2 * l + 1] = snake_direction
-
- # nb x 2
- snake_position = snake_next_position[i, j]
-
- return sequences, worlds
-
-
-# generate_snake_sequences(nb=1, height=4, width=6, nb_colors=3, length=20)
-# exit(0)
+import snake
class TaskSnake(Task):
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_worlds = generate_snake_sequences(
- nb_train_samples, height, width, nb_colors, length, self.device
+ 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_worlds = generate_snake_sequences(
- nb_test_samples, height, width, nb_colors, 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.test_input.max()) + 1
t = model.training
model.eval()
- def compute_nb_correct(input):
+ def compute_nb_correct(input, prior_visits):
result = input.clone()
- i = torch.arange(result.size(1), device=result.device)
- ar_mask = torch.logical_and(i >= i.size(0) // 2, i % 2 == 0)[
- None, :
- ].long()
+ 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 = ar_mask.sum() * input.size(0)
- nb_correct = ((result == input).long() * ar_mask).sum()
+ 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)
+ # 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}%"
- )
+ # 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)
+ 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}%"
width=args.snake_width,
nb_colors=args.snake_nb_colors,
length=args.snake_length,
+ prompt_length=args.snake_length // 2,
device=device,
)
for input in task.batches(split="test"):
input = input.to(device)
- # 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)