is_reusable

This commit is contained in:
Brett Kuprel 2022-06-30 11:25:24 -04:00
parent 3e64e868ef
commit f951424e38
4 changed files with 64 additions and 167 deletions

View File

@ -44,9 +44,9 @@ def generate_image(
image_path: str,
token_count: int
):
is_expendable = True
is_reusable = False
if is_torch:
image_generator = MinDalleTorch(is_mega, is_expendable, token_count)
image_generator = MinDalleTorch(is_mega, is_reusable, token_count)
if token_count < image_generator.config['image_length']:
image_tokens = image_generator.generate_image_tokens(text, seed)
@ -56,7 +56,7 @@ def generate_image(
image = image_generator.generate_image(text, seed)
else:
image_generator = MinDalleFlax(is_mega, is_expendable=True)
image_generator = MinDalleFlax(is_mega, is_reusable)
image = image_generator.generate_image(text, seed)
save_image(image, image_path)

187
min_dalle.ipynb vendored

File diff suppressed because one or more lines are too long

View File

@ -9,11 +9,11 @@ from .models.dalle_bart_decoder_flax import DalleBartDecoderFlax
class MinDalleFlax(MinDalleBase):
def __init__(self, is_mega: bool, is_expendable: bool = False):
def __init__(self, is_mega: bool, is_reusable: bool = True):
super().__init__(is_mega)
self.is_expendable = is_expendable
self.is_reusable = is_reusable
print("initializing MinDalleFlax")
if not is_expendable:
if is_reusable:
self.init_encoder()
self.init_decoder()
self.init_detokenizer()
@ -48,12 +48,12 @@ class MinDalleFlax(MinDalleBase):
def generate_image(self, text: str, seed: int) -> Image.Image:
text_tokens = self.tokenize_text(text)
if self.is_expendable: self.init_encoder()
if not self.is_reusable: self.init_encoder()
print("encoding text tokens")
encoder_state = self.encoder(text_tokens)
if self.is_expendable: del self.encoder
if not self.is_reusable: del self.encoder
if self.is_expendable:
if not self.is_reusable:
self.init_decoder()
params = self.model_params.pop('decoder')
else:
@ -65,13 +65,13 @@ class MinDalleFlax(MinDalleBase):
jax.random.PRNGKey(seed),
params
)
if self.is_expendable: del self.decoder
if not self.is_reusable: del self.decoder
image_tokens = torch.tensor(numpy.array(image_tokens))
if self.is_expendable: self.init_detokenizer()
if not self.is_reusable: self.init_detokenizer()
print("detokenizing image")
image = self.detokenizer.forward(image_tokens).to(torch.uint8)
if self.is_expendable: del self.detokenizer
if not self.is_reusable: del self.detokenizer
image = Image.fromarray(image.to('cpu').detach().numpy())
return image

View File

@ -16,14 +16,14 @@ class MinDalleTorch(MinDalleBase):
def __init__(
self,
is_mega: bool,
is_expendable: bool = False,
is_reusable: bool = True,
token_count: int = 256
):
super().__init__(is_mega)
self.is_expendable = is_expendable
self.is_reusable = is_reusable
self.token_count = token_count
print("initializing MinDalleTorch")
if not is_expendable:
if is_reusable:
self.init_encoder()
self.init_decoder()
self.init_detokenizer()
@ -84,24 +84,24 @@ class MinDalleTorch(MinDalleBase):
text_tokens = torch.tensor(text_tokens).to(torch.long)
if torch.cuda.is_available(): text_tokens = text_tokens.cuda()
if self.is_expendable: self.init_encoder()
if not self.is_reusable: self.init_encoder()
print("encoding text tokens")
encoder_state = self.encoder.forward(text_tokens)
if self.is_expendable: del self.encoder
if not self.is_reusable: del self.encoder
if self.is_expendable: self.init_decoder()
if not self.is_reusable: self.init_decoder()
print("sampling image tokens")
torch.manual_seed(seed)
image_tokens = self.decoder.forward(text_tokens, encoder_state)
if self.is_expendable: del self.decoder
if not self.is_reusable: del self.decoder
return image_tokens
def generate_image(self, text: str, seed: int) -> Image.Image:
image_tokens = self.generate_image_tokens(text, seed)
if self.is_expendable: self.init_detokenizer()
if not self.is_reusable: self.init_detokenizer()
print("detokenizing image")
image = self.detokenizer.forward(image_tokens).to(torch.uint8)
if self.is_expendable: del self.detokenizer
if not self.is_reusable: del self.detokenizer
image = Image.fromarray(image.to('cpu').detach().numpy())
return image