Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Batched CoCa Generation #498

Open
wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Batched CoCa Generation #498

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
136 changes: 79 additions & 57 deletions src/open_clip/coca_model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -167,6 +167,7 @@ def forward(self, image, text, embed_cls=True, image_latent=None, image_embs=Non
def generate(
self,
image,
device,
text=None,
seq_len=30,
max_seq_len=77,
Expand All @@ -182,12 +183,18 @@ def generate(
min_seq_len=5,
stopping_criteria=None,
repetition_penalty=1.0,
fixed_output_length=False # if True output.shape == (batch_size, seq_len)
fixed_output_length=False, # if True output.shape == (batch_size, seq_len)
batch_size=None,
):
# taking many ideas and components from HuggingFace GenerationMixin
# https://huggingface.co/docs/transformers/main/en/main_classes/text_generation
assert _has_transformers, "Please install transformers for generate functionality. `pip install transformers`."
assert seq_len > min_seq_len, "seq_len must be larger than min_seq_len"

if batch_size is None:
batch_size = image.shape[0]

outputs = []

with torch.no_grad():
sot_token_id = 49406 if sot_token_id is None else sot_token_id
Expand All @@ -207,26 +214,27 @@ def generate(
stopping_criteria
)

device = image.device

if generation_type == "beam_search":
output = self._generate_beamsearch(
image_inputs = image,
pad_token_id=pad_token_id,
eos_token_id=eos_token_id,
sot_token_id=sot_token_id,
num_beams=num_beams,
num_beam_groups=num_beam_groups,
min_seq_len=min_seq_len,
stopping_criteria=stopping_criteria,
logit_processor=logit_processor,
)
if fixed_output_length and output.shape[1] < seq_len:
return torch.cat(
(output, torch.ones(output.shape[0], seq_len-output.shape[1], device=device, dtype=output.dtype) * self.pad_id),
dim=1
for i in range(0, image.shape[0], batch_size):
output = self._generate_beamsearch(
image_inputs = image[i:i+batch_size].to(device),
pad_token_id=pad_token_id,
eos_token_id=eos_token_id,
sot_token_id=sot_token_id,
num_beams=num_beams,
num_beam_groups=num_beam_groups,
min_seq_len=min_seq_len,
stopping_criteria=stopping_criteria,
logit_processor=logit_processor,
)
return output
if fixed_output_length and output.shape[1] < seq_len:
output = torch.cat(
(output, torch.ones(output.shape[0], seq_len-output.shape[1], device=device, dtype=output.dtype) * self.pad_id),
dim=1
)
outputs += output.cpu()

return nn.utils.rnn.pad_sequence(outputs, batch_first=True, padding_value=eos_token_id)

elif generation_type == "top_p":
logit_warper = GENERATION_TYPES[generation_type](top_p)
Expand All @@ -238,54 +246,68 @@ def generate(
f"{'| ' + ' | '.join(list(GENERATION_TYPES.keys())) + ' |'}."
)

image_latent, image_embs = self._encode_image(image)

if text is None:
text = torch.ones((image.shape[0], 1), device=device, dtype=torch.long) * sot_token_id
for i in range(0, image.shape[0], batch_size):

was_training = self.training
num_dims = len(text.shape)
images = image[i:i+batch_size].to(device)

if num_dims == 1:
text = text[None, :]
image_latent, image_embs = self._encode_image(images)

cur_len = text.shape[1]
self.eval()
out = text

while True:
x = out[:, -max_seq_len:]
cur_len = x.shape[1]
logits = self(image, x, image_latent=image_latent, image_embs=image_embs, embed_cls=False)["logits"][:, -1]
mask = (out[:, -1] == eos_token_id) | (out[:, -1] == pad_token_id)
sample = torch.ones((out.shape[0], 1), device=device, dtype=torch.long) * pad_token_id

if mask.all():
if not fixed_output_length:
break
if text is None:
texts = torch.ones((images.shape[0], 1), device=device, dtype=torch.long) * sot_token_id
else:
logits = logits[~mask, :]
filtered_logits = logit_processor(x[~mask, :], logits)
filtered_logits = logit_warper(x[~mask, :], filtered_logits)
probs = F.softmax(filtered_logits / temperature, dim=-1)

if (cur_len + 1 == seq_len):
sample[~mask, :] = torch.ones((sum(~mask), 1), device=device, dtype=torch.long) * eos_token_id
texts = text[:]

was_training = self.training
num_dims = len(texts.shape)

if num_dims == 1:
texts = texts[None, :]
elif text is not None:
texts = texts[i:i+batch_size]

cur_len = texts.shape[1]
self.eval()
out = texts.to(device)

while True:
x = out[:, -max_seq_len:]
cur_len = x.shape[1]

logits = self(images, x, image_latent=image_latent, image_embs=image_embs, embed_cls=False)["logits"][:, -1]
mask = (out[:, -1] == eos_token_id) | (out[:, -1] == pad_token_id)
sample = torch.ones((images.shape[0], 1), device=device, dtype=torch.long) * pad_token_id

if mask.all():
if not fixed_output_length:
break
else:
sample[~mask, :] = torch.multinomial(probs, 1)
logits = logits[~mask, :]
filtered_logits = logit_processor(x[~mask, :], logits)
filtered_logits = logit_warper(x[~mask, :], filtered_logits)
probs = F.softmax(filtered_logits / temperature, dim=-1)

out = torch.cat((out, sample), dim=-1)
if (cur_len + 1 == seq_len):
sample[~mask, :] = torch.ones((sum(~mask), 1), device=device, dtype=torch.long) * eos_token_id
else:
sample[~mask, :] = torch.multinomial(probs, 1)

cur_len += 1
out = torch.cat((out, sample), dim=-1)

cur_len += 1

if stopping_criteria(out, None):
break

if stopping_criteria(out, None):
break
outputs += out.cpu()

if num_dims == 1:
out = out.squeeze(0)
self.train(was_training)

outputs = nn.utils.rnn.pad_sequence(outputs, batch_first=True, padding_value=eos_token_id)

self.train(was_training)
return out
if num_dims == 1:
return outputs.squeeze(0)

return outputs

def _generate_beamsearch(
self,
Expand Down