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

Matryoshka Representation Learning for Open CLIP #299

Draft
wants to merge 5 commits into
base: main
Choose a base branch
from
Draft

Matryoshka Representation Learning for Open CLIP #299

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
cf4fd04
Hardcoding # of MRL dims -- 4
adityakusupati Dec 15, 2022
ae1907b
Modifying training script to hard code 4 granular MRL
adityakusupati Dec 15, 2022
6581243
Fixing integer division for dimensions
adityakusupati Dec 15, 2022
bc72c7e
Fixing unbalanced parenthesis.
adityakusupati Dec 15, 2022
039eb1d
Fixed logit_scale reference issues
adityakusupati Dec 15, 2022
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
5 changes: 3 additions & 2 deletions src/open_clip/model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -172,8 +172,9 @@ def __init__(
self.ln_final = text.ln_final
self.text_projection = text.text_projection
self.register_buffer('attn_mask', text.attn_mask, persistent=False)

self.logit_scale = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))

# Num Nested Dims in MRL -- hard coding to 4
self.logit_scale = nn.Parameter(torch.ones([4]) * np.log(1 / 0.07))

def lock_image_tower(self, unlocked_groups=0, freeze_bn_stats=False):
# lock image tower as per LiT - https://arxiv.org/abs/2111.07991
Expand Down
14 changes: 10 additions & 4 deletions src/training/train.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -95,7 +95,11 @@ def train_one_epoch(model, data, epoch, optimizer, scaler, scheduler, args, tb_w
if args.accum_freq == 1:
with autocast():
image_features, text_features, logit_scale = model(images, texts)
total_loss = loss(image_features, text_features, logit_scale)
# Hard coding 4 MRL dims -- [dim, dim/2, dim/4, dim/8]
rep_size = image_features.shape[1]
total_loss = loss(image_features, text_features, logit_scale[0])
for mrl_i in range(1, 4):
total_loss += loss(image_features[:, :(rep_size//(2**mrl_i))], text_features[:, :(rep_size//(2**mrl_i))], logit_scale[mrl_i])

backward(total_loss, scaler)
else:
Expand Down Expand Up @@ -127,7 +131,7 @@ def train_one_epoch(model, data, epoch, optimizer, scaler, scheduler, args, tb_w
accum_image_features[:j] + [chunk_image_features] + accum_image_features[j + 1:])
text_features = torch.cat(
accum_text_features[:j] + [chunk_text_features] + accum_text_features[j + 1:])
total_loss = loss(image_features, text_features, logit_scale)
total_loss = loss(image_features, text_features, logit_scale[0])
backward(total_loss, scaler)

if scaler is not None:
Expand Down Expand Up @@ -155,7 +159,9 @@ def train_one_epoch(model, data, epoch, optimizer, scaler, scheduler, args, tb_w

# Note: we clamp to 4.6052 = ln(100), as in the original paper.
with torch.no_grad():
unwrap_model(model).logit_scale.clamp_(0, math.log(100))
for mrl_i in range(0, 4):
unwrap_model(model).logit_scale[mrl_i].clamp_(0, math.log(100))


batch_time_m.update(time.time() - end)
end = time.time()
Expand All @@ -168,7 +174,7 @@ def train_one_epoch(model, data, epoch, optimizer, scaler, scheduler, args, tb_w

# NOTE loss is coarsely sampled, just master node and per log update
loss_m.update(total_loss.item(), batch_size)
logit_scale_scalar = logit_scale.item()
logit_scale_scalar = logit_scale[0].item()
logging.info(
f"Train Epoch: {epoch} [{num_samples:>{sample_digits}}/{samples_per_epoch} ({percent_complete:.0f}%)] "
f"Loss: {loss_m.val:#.5g} ({loss_m.avg:#.4g}) "
Expand Down