projects
/
beaver.git
/ blobdiff
commit
grep
author
committer
pickaxe
?
search:
re
summary
|
shortlog
|
log
|
commit
|
commitdiff
|
tree
raw
|
inline
| side by side
Update
[beaver.git]
/
beaver.py
diff --git
a/beaver.py
b/beaver.py
index
dca97cc
..
6a6343d
100755
(executable)
--- a/
beaver.py
+++ b/
beaver.py
@@
-131,10
+131,10
@@
for n in vars(args):
######################################################################
######################################################################
-def generation_order(x, fixed_len):
+def generation_order(x, fixed_len
=0
):
if args.random_regression_order:
order = torch.rand(x.size(), device=x.device)
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[:, :fixed_len] = torch.
arange(-fixed_len, 0, device=x
.device)
order = order.sort(1).indices
else:
order = (
order = order.sort(1).indices
else:
order = (
@@
-143,12
+143,11
@@
def generation_order(x, fixed_len):
return 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))
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:])
else:
v = u.gather(1, order)
v = v.reshape(v.size()[:2] + x.size()[2:])
@@
-160,6
+159,12
@@
def shuffle(x, fixed_len):
return reorder(x, order), order
return reorder(x, order), order
+def eval_mygpt(model, input, mode="standard", fixed_len=0):
+ x, order = shuffle(input, fixed_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
######################################################################
# ar_mask is a Boolean matrix of same shape as input, with 1s on the
@@
-167,7
+172,9
@@
def shuffle(x, fixed_len):
def masked_inplace_autoregression(model, batch_size, input, ar_mask, order=None):
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
i = (ar_mask.sum(0) > 0).nonzero()
if i.min() > 0:
# Needed to initialize the model's cache
@@
-186,7
+193,7
@@
def masked_inplace_autoregression(model, batch_size, input, ar_mask, order=None)
######################################################################
######################################################################
-def compute_perplexity(model, fixed_len, split="train"):
+def compute_perplexity(model,
task,
fixed_len, split="train"):
with torch.autograd.no_grad():
t = model.training
model.eval()
with torch.autograd.no_grad():
t = model.training
model.eval()
@@
-195,9
+202,7
@@
def compute_perplexity(model, fixed_len, split="train"):
for input in task.batches(split=split):
input = input.to(device)
for input in task.batches(split=split):
input = input.to(device)
- x, order = shuffle(input, fixed_len)
- x = model(mygpt.BracketedSequence(x), order=order).x
- output = reorder(x, order, back=True)
+ output = eval_mygpt(model, input, fixed_len=fixed_len)
loss = F.cross_entropy(output.transpose(1, 2), input)
acc_loss += loss.item() * input.size(0)
nb_samples += input.size(0)
loss = F.cross_entropy(output.transpose(1, 2), input)
acc_loss += loss.item() * input.size(0)
nb_samples += input.size(0)
@@
-261,9
+266,9
@@
def oneshot(gpt, task):
acc_train_loss, nb_train_samples = 0, 0
for mazes, policies in task.policy_batches(split="train"):
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, fixed_len=task.height * task.width
+ )
output = model(output_gpt)
loss = compute_loss(mazes, output, policies, task.height, task.width)
output = model(output_gpt)
loss = compute_loss(mazes, output, policies, task.height, task.width)
@@
-276,9
+281,9
@@
def oneshot(gpt, task):
acc_test_loss, nb_test_samples = 0, 0
for mazes, policies in task.policy_batches(split="test"):
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, fixed_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)
output = model(output_gpt)
loss = compute_loss(mazes, output, policies, task.height, task.width)
acc_test_loss += loss.item() * mazes.size(0)
@@
-291,9
+296,9
@@
def oneshot(gpt, task):
# -------------------
mazes = task.test_input[:32, : task.height * task.width]
policies = task.test_policies[:32]
# -------------------
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, fixed_len=task.height * task.width
+ )
output = model(output_gpt)
if args.oneshot_output == "policy":
targets = policies.permute(0, 2, 1)
output = model(output_gpt)
if args.oneshot_output == "policy":
targets = policies.permute(0, 2, 1)
@@
-312,15
+317,17
@@
def oneshot(gpt, task):
scores = scores.reshape(-1, task.height, task.width)
mazes = mazes.reshape(-1, task.height, task.width)
targets = targets.reshape(-1, task.height, task.width)
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(
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,
)
mazes=mazes,
score_paths=scores,
score_truth=targets,
)
+ log_string(f"wrote {filename}")
+
# -------------------
gpt.train(t)
# -------------------
gpt.train(t)
@@
-436,7
+443,7
@@
class TaskMaze(Task):
masked_inplace_autoregression(
model, self.batch_size, x, ar_mask, order=order
)
masked_inplace_autoregression(
model, self.batch_size, x, ar_mask, order=order
)
- result = reorder(x, 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)
mazes, paths = self.seq2map(result)
nb_correct += maze.path_correctness(mazes, paths).long().sum()
nb_total += mazes.size(0)
@@
-467,17
+474,23
@@
class TaskMaze(Task):
ar_mask = result.new_zeros(result.size())
ar_mask[:, self.height * self.width :] = 1
result *= 1 - ar_mask
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)
mazes, paths = self.seq2map(input)
_, predicted_paths = self.seq2map(result)
+ filename = f"result_{n_epoch:04d}.png"
maze.save_image(
maze.save_image(
- os.path.join(args.result_dir, f
"result_{n_epoch:04d}.png"
),
+ os.path.join(args.result_dir, f
ilename
),
mazes=mazes,
target_paths=paths,
predicted_paths=predicted_paths,
path_correct=maze.path_correctness(mazes, predicted_paths),
)
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)
model.train(t)
@@
-584,10
+597,10
@@
log_string(f"learning_rate_schedule {learning_rate_schedule}")
if nb_epochs_finished >= args.nb_epochs:
n_epoch = nb_epochs_finished
train_perplexity = compute_perplexity(
if nb_epochs_finished >= args.nb_epochs:
n_epoch = nb_epochs_finished
train_perplexity = compute_perplexity(
- model, fixed_len=task.height * task.width, split="train"
+ model,
task,
fixed_len=task.height * task.width, split="train"
)
test_perplexity = compute_perplexity(
)
test_perplexity = compute_perplexity(
- model, fixed_len=task.height * task.width, split="test"
+ model,
task,
fixed_len=task.height * task.width, split="test"
)
log_string(
)
log_string(
@@
-618,9
+631,9
@@
for n_epoch in range(nb_epochs_finished, args.nb_epochs):
for input in task.batches(split="train"):
input = input.to(device)
for input in task.batches(split="train"):
input = input.to(device)
- x, order = shuffle(input, task.height * task.width)
- x = model(mygpt.BracketedSequence(x), order=order).x
-
output = reorder(x, order, back=True
)
+ output = eval_mygpt(
+ model, input, mode=args.oneshot_input, fixed_len=task.height * task.width
+ )
loss = F.cross_entropy(output.transpose(1, 2), input)
acc_train_loss += loss.item() * input.size(0)
nb_train_samples += input.size(0)
loss = F.cross_entropy(output.transpose(1, 2), input)
acc_train_loss += loss.item() * input.size(0)
nb_train_samples += input.size(0)
@@
-631,7
+644,7
@@
for n_epoch in range(nb_epochs_finished, args.nb_epochs):
train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
test_perplexity = compute_perplexity(
train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
test_perplexity = compute_perplexity(
- model, fixed_len=task.height * task.width, split="test"
+ model,
task,
fixed_len=task.height * task.width, split="test"
)
log_string(
)
log_string(