tanner/min-dalle-test
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

add temperature parameter

Browse Source
This commit is contained in:
Brett Kuprel
2022-07-11 12:28:20 -04:00
parent 2feabd7847
commit d64e957731
5 changed files with 74 additions and 55 deletions
Download Patch File Download Diff File Expand all files Collapse all files

86
min_dalle/min_dalle.py
View File

@@ -177,8 +177,9 @@ class MinDalle:
seed: int,
image_count: int,
log2_mid_count: int,
log2_k: int = 6,
log2_supercondition_factor: int = 3,
temperature: float = 1,
top_k: int = 256,
supercondition_factor: int = 16,
is_verbose: bool = False
) -> Iterator[FloatTensor]:
assert(log2_mid_count in range(5))
@@ -206,10 +207,10 @@ class MinDalle:
with torch.cuda.amp.autocast(dtype=self.dtype):
encoder_state, attention_mask, attention_state, image_tokens = (
self.decoder.decode_initial(
seed,
image_count,
text_tokens,
encoder_state
seed=seed,
image_count=image_count,
text_tokens=text_tokens,
encoder_state=encoder_state
)
)
@@ -220,12 +221,13 @@ class MinDalle:
with torch.cuda.amp.autocast(dtype=self.dtype):
attention_state, image_tokens = self.decoder.decode_row(
row_index,
log2_k,
log2_supercondition_factor,
encoder_state,
attention_mask,
attention_state,
image_tokens
temperature=temperature,
top_k=top_k,
supercondition_factor=supercondition_factor,
encoder_state=encoder_state,
attention_mask=attention_mask,
attention_state=attention_state,
image_tokens_sequence=image_tokens
)
with torch.cuda.amp.autocast(dtype=torch.float32):
if ((row_index + 1) * (2 ** log2_mid_count)) % row_count == 0:
@@ -240,18 +242,20 @@ class MinDalle:
seed: int,
grid_size: int,
log2_mid_count: int,
log2_k: int = 6,
log2_supercondition_factor: int = 3,
temperature: float = 1,
top_k: int = 256,
supercondition_factor: int = 16,
is_verbose: bool = False
) -> Iterator[Image.Image]:
images_stream = self.generate_images_stream(
text,
seed,
grid_size ** 2,
log2_mid_count,
log2_k,
log2_supercondition_factor,
is_verbose
text=text,
seed=seed,
image_count=grid_size ** 2,
log2_mid_count=log2_mid_count,
temperature=temperature,
top_k=top_k,
supercondition_factor=supercondition_factor,
is_verbose=is_verbose
)
for images in images_stream:
yield self.grid_from_images(images)
@@ -262,19 +266,21 @@ class MinDalle:
text: str,
seed: int = -1,
image_count: int = 1,
log2_k: int = 6,
log2_supercondition_factor: int = 3,
temperature: float = 1,
top_k: int = 1024,
supercondition_factor: int = 16,
is_verbose: bool = False
) -> FloatTensor:
log2_mid_count = 0
images_stream = self.generate_images_stream(
text,
seed,
image_count,
log2_mid_count,
log2_k,
log2_supercondition_factor,
is_verbose
text=text,
seed=seed,
image_count=image_count,
temperature=temperature,
log2_mid_count=log2_mid_count,
top_k=top_k,
supercondition_factor=supercondition_factor,
is_verbose=is_verbose
)
return next(images_stream)
@@ -284,18 +290,20 @@ class MinDalle:
text: str,
seed: int = -1,
grid_size: int = 1,
log2_k: int = 6,
log2_supercondition_factor: int = 3,
temperature: float = 1,
top_k: int = 1024,
supercondition_factor: int = 16,
is_verbose: bool = False
) -> Image.Image:
log2_mid_count = 0
image_stream = self.generate_image_stream(
text,
seed,
grid_size,
log2_mid_count,
log2_k,
log2_supercondition_factor,
is_verbose
text=text,
seed=seed,
grid_size=grid_size,
log2_mid_count=log2_mid_count,
temperature=temperature,
top_k=top_k,
supercondition_factor=supercondition_factor,
is_verbose=is_verbose
)
return next(image_stream)

28
min_dalle/models/dalle_bart_decoder.py
View File

@@ -140,8 +140,9 @@ class DalleBartDecoder(nn.Module):
def decode_step(
self,
log2_k: int,
log2_supercondition_factor: int,
temperature: float,
top_k: int,
supercondition_factor: int,
attention_mask: BoolTensor,
encoder_state: FloatTensor,
attention_state: FloatTensor,
@@ -166,18 +167,17 @@ class DalleBartDecoder(nn.Module):
)
decoder_state = self.final_ln(decoder_state)
logits = self.lm_head(decoder_state)
a = 2 ** log2_supercondition_factor
a = supercondition_factor
logits: FloatTensor = (
logits[:image_count, -1] * (1 - a) +
logits[image_count:, -1] * a
)
top_logits, _ = logits.topk(2 ** log2_k, dim=-1)
probs = torch.where(
logits < top_logits[:, [-1]],
self.zero_prob,
torch.exp(logits - top_logits[:, [0]])
)
top_logits, _ = logits.topk(top_k, dim=-1)
is_kept = logits >= top_logits[:, [-1]]
logits -= top_logits[:, [0]]
logits /= max(temperature, 1e-6)
probs = torch.where(is_kept, torch.exp(logits), self.zero_prob)
probs[:, 2 ** 14:] = 0 # vqgan vocab_count is only 2 ** 14
return probs, attention_state
@@ -185,8 +185,9 @@ class DalleBartDecoder(nn.Module):
def decode_row(
self,
row_index: int,
log2_k: int,
log2_supercondition_factor: int,
temperature: float,
top_k: int,
supercondition_factor: int,
encoder_state: FloatTensor,
attention_mask: BoolTensor,
attention_state: FloatTensor,
@@ -195,8 +196,9 @@ class DalleBartDecoder(nn.Module):
for col_index in range(16):
i = 16 * row_index + col_index
probs, attention_state = self.decode_step(
log2_k = log2_k,
log2_supercondition_factor = log2_supercondition_factor,
temperature = temperature,
top_k = top_k,
supercondition_factor = supercondition_factor,
attention_mask = attention_mask,
encoder_state = encoder_state,
attention_state = attention_state,