Initial training code
Browse files- src/loss.py +2 -13
- src/metrics.py +12 -0
- src/trainer.py +91 -0
src/loss.py
CHANGED
@@ -3,17 +3,6 @@ from torch import nn
|
|
3 |
import torch.nn.functional as F
|
4 |
|
5 |
|
6 |
-
def metrics(similarity: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
|
7 |
-
y = torch.arange(len(similarity)).to(similarity.device)
|
8 |
-
img2cap_match_idx = similarity.argmax(dim=1)
|
9 |
-
cap2img_match_idx = similarity.argmax(dim=0)
|
10 |
-
|
11 |
-
img_acc = (img2cap_match_idx == y).float().mean()
|
12 |
-
cap_acc = (cap2img_match_idx == y).float().mean()
|
13 |
-
|
14 |
-
return img_acc, cap_acc
|
15 |
-
|
16 |
-
|
17 |
def get_similarity_matrix(
|
18 |
image_features: torch.Tensor, text_features: torch.Tensor
|
19 |
) -> torch.Tensor:
|
@@ -34,7 +23,7 @@ class CLIPLoss(nn.Module):
|
|
34 |
super().__init__()
|
35 |
self.logit_temperature = nn.Parameter(torch.tensor(logit_temperature))
|
36 |
|
37 |
-
def forward(self, similarity_matrix: torch.Tensor):
|
38 |
temperature = self.logit_temperature.sigmoid()
|
39 |
|
40 |
caption_loss = contrastive_loss(similarity_matrix / temperature, dim=0)
|
@@ -77,7 +66,7 @@ class SigLIPLoss(nn.Module):
|
|
77 |
super().__init__()
|
78 |
self.logit_temperature = nn.Parameter(torch.tensor(logit_temperature))
|
79 |
|
80 |
-
def forward(self, similarity_matrix: torch.Tensor):
|
81 |
temperature = self.logit_temperature.sigmoid()
|
82 |
return contrastive_sigmoid_loss(similarity_matrix / temperature)
|
83 |
|
|
|
3 |
import torch.nn.functional as F
|
4 |
|
5 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
6 |
def get_similarity_matrix(
|
7 |
image_features: torch.Tensor, text_features: torch.Tensor
|
8 |
) -> torch.Tensor:
|
|
|
23 |
super().__init__()
|
24 |
self.logit_temperature = nn.Parameter(torch.tensor(logit_temperature))
|
25 |
|
26 |
+
def forward(self, similarity_matrix: torch.Tensor, *args):
|
27 |
temperature = self.logit_temperature.sigmoid()
|
28 |
|
29 |
caption_loss = contrastive_loss(similarity_matrix / temperature, dim=0)
|
|
|
66 |
super().__init__()
|
67 |
self.logit_temperature = nn.Parameter(torch.tensor(logit_temperature))
|
68 |
|
69 |
+
def forward(self, similarity_matrix: torch.Tensor, *args):
|
70 |
temperature = self.logit_temperature.sigmoid()
|
71 |
return contrastive_sigmoid_loss(similarity_matrix / temperature)
|
72 |
|
src/metrics.py
ADDED
@@ -0,0 +1,12 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
|
3 |
+
|
4 |
+
def metrics(similarity_matrix: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
|
5 |
+
y = torch.arange(len(similarity_matrix)).to(similarity_matrix.device)
|
6 |
+
img2cap_match_idx = similarity_matrix.argmax(dim=1)
|
7 |
+
cap2img_match_idx = similarity_matrix.argmax(dim=0)
|
8 |
+
|
9 |
+
img_acc = (img2cap_match_idx == y).float().mean()
|
10 |
+
cap_acc = (cap2img_match_idx == y).float().mean()
|
11 |
+
|
12 |
+
return img_acc, cap_acc
|
src/trainer.py
ADDED
@@ -0,0 +1,91 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import pytorch_lightning as pl
|
2 |
+
import torch
|
3 |
+
import torch.nn as nn
|
4 |
+
|
5 |
+
from src import config
|
6 |
+
from src import loss as loss_utils
|
7 |
+
from src import metrics
|
8 |
+
from src import models
|
9 |
+
|
10 |
+
|
11 |
+
class LightningModule(pl.LightningModule):
|
12 |
+
def __init__(
|
13 |
+
self,
|
14 |
+
vision_encoder: models.TinyCLIPVisionEncoder,
|
15 |
+
text_encoder: models.TinyCLIPTextEncoder,
|
16 |
+
loss_fn: nn.Module,
|
17 |
+
hyper_parameters: config.TrainerConfig,
|
18 |
+
len_train_dl: int,
|
19 |
+
) -> None:
|
20 |
+
super().__init__()
|
21 |
+
self.vision_encoder = vision_encoder
|
22 |
+
self.text_encoder = text_encoder
|
23 |
+
self.loss_fn = loss_fn
|
24 |
+
self.hyper_parameters = hyper_parameters
|
25 |
+
self.len_train_dl = len_train_dl
|
26 |
+
|
27 |
+
def common_step(self, batch: tuple[torch.Tensor, list[str]], step_kind: str) -> torch.Tensor:
|
28 |
+
text, images = batch
|
29 |
+
image_features = self.vision_encoder(images)
|
30 |
+
text_features = self.text_encoder(text)
|
31 |
+
similarity_matrix = loss_utils.get_similarity_matrix(image_features, text_features)
|
32 |
+
|
33 |
+
loss = self.loss_fn(similarity_matrix, image_features, text_features)
|
34 |
+
|
35 |
+
img_acc, cap_acc = metrics.metrics(similarity_matrix)
|
36 |
+
|
37 |
+
self.log(f"{step_kind}_loss", loss, on_step=False, on_epoch=True)
|
38 |
+
self.log(f"{step_kind}_img_acc", img_acc, on_step=False, on_epoch=True, prog_bar=True)
|
39 |
+
self.log(f"{step_kind}_cap_acc", cap_acc, on_step=False, on_epoch=True, prog_bar=True)
|
40 |
+
return loss
|
41 |
+
|
42 |
+
def training_step(self, batch: tuple[torch.Tensor, list[str]], *args: list) -> torch.Tensor:
|
43 |
+
loss = self.common_step(batch, step_kind="training")
|
44 |
+
return loss
|
45 |
+
|
46 |
+
def validation_step(self, batch: tuple[torch.Tensor, list[str]], *args: list):
|
47 |
+
_ = self.common_step(batch, step_kind="training")
|
48 |
+
|
49 |
+
def configure_optimizers(self):
|
50 |
+
# TODO: Add loss parameters here
|
51 |
+
vision_params = [
|
52 |
+
{
|
53 |
+
"params": self.vision_encoder.projection.parameters(),
|
54 |
+
"lr": self.hyper_parameters.learning_rate,
|
55 |
+
},
|
56 |
+
{
|
57 |
+
"params": self.vision_encoder.base.parameters(),
|
58 |
+
"lr": self.hyper_parameters.learning_rate / 2,
|
59 |
+
},
|
60 |
+
]
|
61 |
+
caption_params = [
|
62 |
+
{
|
63 |
+
"params": self.text_encoder.projection.parameters(),
|
64 |
+
"lr": self.hyper_parameters.learning_rate,
|
65 |
+
},
|
66 |
+
]
|
67 |
+
if not self.hyper_parameters.freeze_text_base:
|
68 |
+
caption_params += [
|
69 |
+
{
|
70 |
+
"params": self.text_encoder.base.encoder.parameters(),
|
71 |
+
"lr": self.hyper_parameters.learning_rate / 2,
|
72 |
+
},
|
73 |
+
]
|
74 |
+
|
75 |
+
optimizer = torch.optim.Adam(vision_params + caption_params)
|
76 |
+
|
77 |
+
if self.hyper_parameters.lr_scheduler:
|
78 |
+
scheduler = torch.optim.lr_scheduler.OneCycleLR(
|
79 |
+
optimizer,
|
80 |
+
max_lr=self.hyper_parameters.learning_rate,
|
81 |
+
total_steps=self.trainer.estimated_stepping_batches,
|
82 |
+
)
|
83 |
+
return [optimizer], [scheduler]
|
84 |
+
else:
|
85 |
+
return optimizer
|
86 |
+
|
87 |
+
def on_epoch_end(self):
|
88 |
+
if self.current_epoch == 0:
|
89 |
+
for p in self.vision_encoder.base.parameters():
|
90 |
+
p.requires_grad = True
|
91 |
+
self.vision_encoder.base.train()
|