formatter_class=argparse.ArgumentDefaultsHelpFormatter,
)
-parser.add_argument("--task", type=str, default="picoclvr")
+parser.add_argument(
+ "--task", type=str, default="picoclvr", help="picoclvr, mnist, maze, snake"
+)
-parser.add_argument("--log_filename", type=str, default="train.log")
+parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
parser.add_argument("--result_dir", type=str, default="results_default")
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=None)
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": {
- "batch_size": 20,
+ "nb_epochs": 5,
+ "batch_size": 25,
},
}
######################################################################
+# 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 tqdm.tqdm(
- zip(input.split(batch_size), ar_mask.split(batch_size)),
- dynamic_ncols=True,
- desc="autoregression",
- total=input.size(0) // 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 compute_error(self, model, split="train", nb_to_use=-1):
nb_total, nb_correct = 0, 0
+ count = torch.zeros(
+ self.width * self.height, self.width * self.height, device=self.device
+ )
for input in task.batches(split, nb_to_use):
result = input.clone()
ar_mask = result.new_zeros(result.size())
model, self.batch_size, result, ar_mask, device=self.device
)
mazes, paths = self.seq2map(result)
- nb_correct += maze.path_correctness(mazes, paths).long().sum()
+ path_correctness = maze.path_correctness(mazes, paths)
+ nb_correct += path_correctness.long().sum()
nb_total += mazes.size(0)
- return nb_total, nb_correct
+ optimal_path_lengths = (
+ (input[:, self.height * self.width :] == maze.v_path).long().sum(1)
+ )
+ predicted_path_lengths = (
+ (result[:, self.height * self.width :] == maze.v_path).long().sum(1)
+ )
+ optimal_path_lengths = optimal_path_lengths[path_correctness]
+ predicted_path_lengths = predicted_path_lengths[path_correctness]
+ count[optimal_path_lengths, predicted_path_lengths] += 1
+
+ if count.max() == 0:
+ count = None
+ else:
+ count = count[
+ : count.sum(1).nonzero().max() + 1, : count.sum(0).nonzero().max() + 1
+ ]
+
+ return nb_total, nb_correct, count
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(
+ train_nb_total, train_nb_correct, count = 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(
+ test_nb_total, test_nb_correct, count = 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}%"
)
+ if count is not None:
+ with open(
+ os.path.join(args.result_dir, f"maze_result_{n_epoch:04d}.txt"), "w"
+ ) as f:
+ for i in range(count.size(0)):
+ for j in range(count.size(1)):
+ eol = " " if j < count.size(1) - 1 else "\n"
+ f.write(f"{count[i,j]}{eol}")
+
input = self.test_input[:48]
result = input.clone()
ar_mask = result.new_zeros(result.size())
######################################################################
-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=}')
+import snake
class TaskSnake(Task):
self.device = device
self.prompt_length = prompt_length
- self.train_input, self.train_prior_visits = generate_snake_sequences(
+ self.train_input, self.train_prior_visits, _, _ = snake.generate_sequences(
nb_train_samples,
height,
width,
prompt_length,
self.device,
)
- self.test_input, self.test_prior_visits = generate_snake_sequences(
+ self.test_input, self.test_prior_visits, _, _ = snake.generate_sequences(
nb_test_samples,
height,
width,
)
result *= 1 - ar_mask
- # snake_solver(result,ar_mask)
+ # snake.solver(result,ar_mask)
masked_inplace_autoregression(
model, self.batch_size, result, ar_mask, device=self.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)