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)
######################################################################
-def generation_order(x, fixed_len):
+def generation_order(x, prompt_len=0):
if args.random_regression_order:
order = torch.rand(x.size(), device=x.device)
- order[:, :fixed_len] = torch.linspace(-2, -1, fixed_len, device=order.device)
+ order[:, :prompt_len] = torch.arange(-prompt_len, 0, device=x.device)
order = order.sort(1).indices
else:
order = (
return order
-def reorder(x, order, back=False): # x is NxTxD1x...xDk, order is NxT'
+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 back:
- v = u.new(u.size())
- v.scatter_(1, order, u)
+ 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
-def shuffle(x, fixed_len):
- order = generation_order(x, fixed_len)
+def shuffle(x, prompt_len):
+ order = generation_order(x, prompt_len)
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
def masked_inplace_autoregression(model, batch_size, input, ar_mask, order=None):
- for input, ar_mask in zip(input.split(batch_size), ar_mask.split(batch_size)):
+ 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:
# Needed to initialize the model's cache
######################################################################
-def compute_perplexity(model, split="train"):
+def compute_perplexity(model, task, prompt_len, split="train"):
with torch.autograd.no_grad():
t = model.training
model.eval()
for input in task.batches(split=split):
input = input.to(device)
- input, order = shuffle(input, task.height * task.width)
- output = model(mygpt.BracketedSequence(input), order=order).x
- loss = F.cross_entropy(output.transpose(1, 2), input)
+ 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)
acc_train_loss, nb_train_samples = 0, 0
for mazes, policies in task.policy_batches(split="train"):
- x, order = shuffle(mazes, task.height * task.width)
- x = gpt(mygpt.BracketedSequence(x), mode=args.oneshot_input, order=order).x
- output_gpt = reorder(x, order, back=True)
+ 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, nb_test_samples = 0, 0
for mazes, policies in task.policy_batches(split="test"):
- x, order = shuffle(mazes, task.height * task.width)
- x = gpt(mygpt.BracketedSequence(x), mode=args.oneshot_input, order=order).x
- output_gpt = reorder(x, order, back=True)
+ 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)
# -------------------
mazes = task.test_input[:32, : task.height * task.width]
policies = task.test_policies[:32]
- x, order = shuffle(mazes, task.height * task.width)
- x = gpt(mygpt.BracketedSequence(x), mode=args.oneshot_input, order=order).x
- output_gpt = reorder(x, order, back=True)
+ 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 = 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,
- f"oneshot_{args.oneshot_input}_{args.oneshot_output}_{n_epoch:04d}.png",
- ),
+ os.path.join(args.result_dir, filename),
mazes=mazes,
score_paths=scores,
score_truth=targets,
)
+ log_string(f"wrote {filename}")
+
# -------------------
gpt.train(t)
ar_mask = result.new_zeros(result.size())
ar_mask[:, self.height * self.width :] = 1
result *= 1 - ar_mask
- result, order = shuffle(result, self.height * self.width)
+ x, order = shuffle(result, self.height * self.width)
masked_inplace_autoregression(
- model, self.batch_size, result, ar_mask, order=order
+ model, self.batch_size, x, ar_mask, order=order
)
- result = reorder(result, order, back=True)
+ 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)
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)
+ 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, f"result_{n_epoch:04d}.png"),
+ 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}")
model.train(t)
##############################
+
+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 nb_epochs_finished >= args.nb_epochs:
n_epoch = nb_epochs_finished
- train_perplexity = compute_perplexity(model, split="train")
- test_perplexity = compute_perplexity(model, split="test")
+ 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"
+ )
log_string(
f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
for input in task.batches(split="train"):
input = input.to(device)
- input, order = shuffle(input, task.height * task.width)
- output = model(mygpt.BracketedSequence(input), order=order).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)
optimizer.step()
train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
- test_perplexity = compute_perplexity(model, split="test")
+ test_perplexity = compute_perplexity(
+ model, task, prompt_len=task.height * task.width, split="test"
+ )
log_string(
f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"