projects
/
picoclvr.git
/ blobdiff
commit
grep
author
committer
pickaxe
?
search:
re
summary
|
shortlog
|
log
|
commit
|
commitdiff
|
tree
raw
|
inline
| side by side
Update.
[picoclvr.git]
/
world.py
diff --git
a/world.py
b/world.py
index
64c7434
..
12c6553
100755
(executable)
--- a/
world.py
+++ b/
world.py
@@
-61,6
+61,19
@@
class SignSTE(nn.Module):
else:
return s
else:
return s
+class DiscreteSampler2d(nn.Module):
+ def __init__(self):
+ super().__init__()
+
+ def forward(self, x):
+ s = (x >= x.max(-3,keepdim=True).values).float()
+
+ if self.training:
+ u = x.softmax(dim=-3)
+ return s + u - u.detach()
+ else:
+ return s
+
def loss_H(binary_logits, h_threshold=1):
p = binary_logits.sigmoid().mean(0)
def loss_H(binary_logits, h_threshold=1):
p = binary_logits.sigmoid().mean(0)
@@
-72,9
+85,9
@@
def loss_H(binary_logits, h_threshold=1):
def train_encoder(
train_input,
test_input,
def train_encoder(
train_input,
test_input,
- depth=2,
+ depth,
+ nb_bits_per_token,
dim_hidden=48,
dim_hidden=48,
- nb_bits_per_token=8,
lambda_entropy=0.0,
lr_start=1e-3,
lr_end=1e-4,
lambda_entropy=0.0,
lr_start=1e-3,
lr_end=1e-4,
@@
-83,8
+96,6
@@
def train_encoder(
logger=None,
device=torch.device("cpu"),
):
logger=None,
device=torch.device("cpu"),
):
- if logger is None:
- logger = lambda s: print(s)
mu, std = train_input.float().mean(), train_input.float().std()
mu, std = train_input.float().mean(), train_input.float().std()
@@
-144,7
+155,7
@@
def train_encoder(
nb_parameters = sum(p.numel() for p in model.parameters())
nb_parameters = sum(p.numel() for p in model.parameters())
- logger(f"nb_parameters {nb_parameters}")
+ logger(f"
vqae
nb_parameters {nb_parameters}")
model.to(device)
model.to(device)
@@
-159,7
+170,7
@@
def train_encoder(
for input in tqdm.tqdm(train_input.split(batch_size), desc="vqae-train"):
input = input.to(device)
z = encoder(input)
for input in tqdm.tqdm(train_input.split(batch_size), desc="vqae-train"):
input = input.to(device)
z = encoder(input)
- zq =
z if k < 2 else
quantizer(z)
+ zq = quantizer(z)
output = decoder(zq)
output = output.reshape(
output = decoder(zq)
output = output.reshape(
@@
-182,7
+193,7
@@
def train_encoder(
for input in tqdm.tqdm(test_input.split(batch_size), desc="vqae-test"):
input = input.to(device)
z = encoder(input)
for input in tqdm.tqdm(test_input.split(batch_size), desc="vqae-test"):
input = input.to(device)
z = encoder(input)
- zq =
z if k < 1 else
quantizer(z)
+ zq = quantizer(z)
output = decoder(zq)
output = output.reshape(
output = decoder(zq)
output = output.reshape(
@@
-196,7
+207,7
@@
def train_encoder(
train_loss = acc_train_loss / train_input.size(0)
test_loss = acc_test_loss / test_input.size(0)
train_loss = acc_train_loss / train_input.size(0)
test_loss = acc_test_loss / test_input.size(0)
- logger(f"
train_ae
{k} lr {lr} train_loss {train_loss} test_loss {test_loss}")
+ logger(f"
vqae train
{k} lr {lr} train_loss {train_loss} test_loss {test_loss}")
sys.stdout.flush()
return encoder, quantizer, decoder
sys.stdout.flush()
return encoder, quantizer, decoder
@@
-353,6
+364,8
@@
def create_data_and_processors(
nb_test_samples,
mode,
nb_steps,
nb_test_samples,
mode,
nb_steps,
+ depth=3,
+ nb_bits_per_token=8,
nb_epochs=10,
device=torch.device("cpu"),
device_storage=torch.device("cpu"),
nb_epochs=10,
device=torch.device("cpu"),
device_storage=torch.device("cpu"),
@@
-363,6
+376,9
@@
def create_data_and_processors(
if mode == "first_last":
steps = [True] + [False] * (nb_steps + 1) + [True]
if mode == "first_last":
steps = [True] + [False] * (nb_steps + 1) + [True]
+ if logger is None:
+ logger = lambda s: print(s)
+
train_input, train_actions = generate_episodes(nb_train_samples, steps)
train_input, train_actions = train_input.to(device_storage), train_actions.to(
device_storage
train_input, train_actions = generate_episodes(nb_train_samples, steps)
train_input, train_actions = train_input.to(device_storage), train_actions.to(
device_storage
@@
-375,6
+391,8
@@
def create_data_and_processors(
encoder, quantizer, decoder = train_encoder(
train_input,
test_input,
encoder, quantizer, decoder = train_encoder(
train_input,
test_input,
+ depth=depth,
+ nb_bits_per_token=nb_bits_per_token,
lambda_entropy=1.0,
nb_epochs=nb_epochs,
logger=logger,
lambda_entropy=1.0,
nb_epochs=nb_epochs,
logger=logger,
@@
-388,6
+406,8
@@
def create_data_and_processors(
pow2 = (2 ** torch.arange(z.size(1), device=device))[None, None, :]
z_h, z_w = z.size(2), z.size(3)
pow2 = (2 ** torch.arange(z.size(1), device=device))[None, None, :]
z_h, z_w = z.size(2), z.size(3)
+ logger(f"vqae input {train_input[0].size()} output {z[0].size()}")
+
def frame2seq(input, batch_size=25):
seq = []
p = pow2.to(device)
def frame2seq(input, batch_size=25):
seq = []
p = pow2.to(device)
@@
-440,7
+460,7
@@
if __name__ == "__main__":
seq2frame,
) = create_data_and_processors(
25000, 1000,
seq2frame,
) = create_data_and_processors(
25000, 1000,
- nb_epochs=
10
,
+ nb_epochs=
5
,
mode="first_last",
nb_steps=20,
)
mode="first_last",
nb_steps=20,
)