Refactoring to distangle modules
Browse files- src/config.py +2 -14
- src/models.py +6 -5
- src/tokenizer.py +5 -3
- src/trainer.py +19 -1
- src/vision_model.py +14 -5
src/config.py
CHANGED
@@ -65,7 +65,7 @@ class TinyCLIPConfig(PretrainedConfig):
|
|
65 |
max_len: int = 128,
|
66 |
cls_type: bool = True,
|
67 |
freeze_vision_base: bool = False,
|
68 |
-
freeze_text_base: bool =
|
69 |
loss_type: str = "cyclip",
|
70 |
**kwargs,
|
71 |
):
|
@@ -85,18 +85,6 @@ class TinyCLIPConfig(PretrainedConfig):
|
|
85 |
super().__init__(**kwargs)
|
86 |
|
87 |
|
88 |
-
class ModelConfig(pydantic.BaseModel):
|
89 |
-
text_model: str = "microsoft/xtremedistil-l6-h256-uncased" # 51 mb
|
90 |
-
vision_model: str = "edgenext_small" # 20 mb
|
91 |
-
projection_layers: int = 3
|
92 |
-
embed_dim: int = 256
|
93 |
-
transformer_embed_dim: int = 768
|
94 |
-
max_len: int = 128 # 77
|
95 |
-
cls_type: bool = True
|
96 |
-
freeze_vision_base: bool = False
|
97 |
-
freeze_text_base: bool = False
|
98 |
-
|
99 |
-
|
100 |
class TrainerConfig(pydantic.BaseModel):
|
101 |
epochs: int = 20
|
102 |
batch_size: int = 64
|
@@ -112,5 +100,5 @@ class TrainerConfig(pydantic.BaseModel):
|
|
112 |
|
113 |
run_openai_clip: bool = False
|
114 |
|
115 |
-
_model_config:
|
116 |
_data_config: DataConfig = DataConfig()
|
|
|
65 |
max_len: int = 128,
|
66 |
cls_type: bool = True,
|
67 |
freeze_vision_base: bool = False,
|
68 |
+
freeze_text_base: bool = True,
|
69 |
loss_type: str = "cyclip",
|
70 |
**kwargs,
|
71 |
):
|
|
|
85 |
super().__init__(**kwargs)
|
86 |
|
87 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
88 |
class TrainerConfig(pydantic.BaseModel):
|
89 |
epochs: int = 20
|
90 |
batch_size: int = 64
|
|
|
100 |
|
101 |
run_openai_clip: bool = False
|
102 |
|
103 |
+
_model_config: TinyCLIPConfig = TinyCLIPConfig()
|
104 |
_data_config: DataConfig = DataConfig()
|
src/models.py
CHANGED
@@ -1,14 +1,14 @@
|
|
1 |
from PIL import Image
|
2 |
-
import
|
3 |
-
from timm import data
|
4 |
import torch
|
5 |
import torch.nn as nn
|
6 |
import torch.nn.functional as F
|
7 |
-
|
8 |
from transformers import PreTrainedModel
|
9 |
|
10 |
from src.config import TinyCLIPConfig, TinyCLIPTextConfig, TinyCLIPVisionConfig
|
11 |
from src import loss
|
|
|
12 |
|
13 |
|
14 |
class Projection(nn.Module):
|
@@ -70,9 +70,10 @@ class TinyCLIPVisionEncoder(PreTrainedModel):
|
|
70 |
|
71 |
def __init__(self, config: TinyCLIPVisionConfig):
|
72 |
super().__init__(config)
|
73 |
-
|
|
|
74 |
self.projection = projection_layers(
|
75 |
-
|
76 |
)
|
77 |
|
78 |
def forward(self, images: list[Image.Image]):
|
|
|
1 |
from PIL import Image
|
2 |
+
import transformers
|
|
|
3 |
import torch
|
4 |
import torch.nn as nn
|
5 |
import torch.nn.functional as F
|
6 |
+
|
7 |
from transformers import PreTrainedModel
|
8 |
|
9 |
from src.config import TinyCLIPConfig, TinyCLIPTextConfig, TinyCLIPVisionConfig
|
10 |
from src import loss
|
11 |
+
from src import vision_model
|
12 |
|
13 |
|
14 |
class Projection(nn.Module):
|
|
|
70 |
|
71 |
def __init__(self, config: TinyCLIPVisionConfig):
|
72 |
super().__init__(config)
|
73 |
+
base, num_features = vision_model.get_vision_base(config)
|
74 |
+
self.base = base
|
75 |
self.projection = projection_layers(
|
76 |
+
num_features, config.embed_dims, config.projection_layers
|
77 |
)
|
78 |
|
79 |
def forward(self, images: list[Image.Image]):
|
src/tokenizer.py
CHANGED
@@ -3,11 +3,13 @@ from typing import Union
|
|
3 |
import torch
|
4 |
from transformers import AutoTokenizer
|
5 |
|
|
|
|
|
6 |
|
7 |
class Tokenizer:
|
8 |
-
def __init__(self,
|
9 |
-
self.tokenizer = AutoTokenizer.from_pretrained(
|
10 |
-
self.max_len = max_len
|
11 |
|
12 |
def __call__(self, x: Union[str, list[str]]) -> dict[str, torch.LongTensor]:
|
13 |
return self.tokenizer(
|
|
|
3 |
import torch
|
4 |
from transformers import AutoTokenizer
|
5 |
|
6 |
+
from src.config import TinyCLIPTextConfig
|
7 |
+
|
8 |
|
9 |
class Tokenizer:
|
10 |
+
def __init__(self, text_config: TinyCLIPTextConfig) -> None:
|
11 |
+
self.tokenizer = AutoTokenizer.from_pretrained(text_config.text_model)
|
12 |
+
self.max_len = text_config.max_len
|
13 |
|
14 |
def __call__(self, x: Union[str, list[str]]) -> dict[str, torch.LongTensor]:
|
15 |
return self.tokenizer(
|
src/trainer.py
CHANGED
@@ -1,7 +1,25 @@
|
|
1 |
from src import data
|
2 |
from src import config
|
3 |
from src import vision_model
|
|
|
|
|
|
|
|
|
4 |
|
5 |
|
6 |
def train(config: config.TrainerConfig):
|
7 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
from src import data
|
2 |
from src import config
|
3 |
from src import vision_model
|
4 |
+
from src import tokenizer as tk
|
5 |
+
from src.lightning_module import LightningModule
|
6 |
+
from src import loss
|
7 |
+
from src import models
|
8 |
|
9 |
|
10 |
def train(config: config.TrainerConfig):
|
11 |
+
transform = vision_model.get_vision_transform(config._model_config.vision_config)
|
12 |
+
tokenizer = tk.Tokenizer(config._model_config.text_config)
|
13 |
+
train_dl, valid_dl = data.get_dataset(
|
14 |
+
transform=transform, tokenizer=tokenizer, hyper_parameters=config # type: ignore
|
15 |
+
)
|
16 |
+
vision_encoder = models.TinyCLIPVisionEncoder(config=config._model_config.vision_config)
|
17 |
+
text_encoder = models.TinyCLIPTextEncoder(config=config._model_config.text_config)
|
18 |
+
|
19 |
+
lightning_module = LightningModule(
|
20 |
+
vision_encoder=vision_encoder,
|
21 |
+
text_encoder=text_encoder,
|
22 |
+
loss_fn=loss.get_loss(config._model_config.loss_type),
|
23 |
+
hyper_parameters=config,
|
24 |
+
len_train_dl=len(train_dl),
|
25 |
+
)
|
src/vision_model.py
CHANGED
@@ -1,11 +1,20 @@
|
|
1 |
import timm
|
2 |
from timm import data
|
|
|
|
|
3 |
|
4 |
-
from src import
|
5 |
|
6 |
|
7 |
-
def
|
8 |
-
|
9 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
10 |
transform = data.transforms_factory.create_transform(**timm_config)
|
11 |
-
return
|
|
|
1 |
import timm
|
2 |
from timm import data
|
3 |
+
import torch.nn as nn
|
4 |
+
from torchvision import transforms
|
5 |
|
6 |
+
from src.config import TinyCLIPVisionConfig
|
7 |
|
8 |
|
9 |
+
def get_vision_base(
|
10 |
+
config: TinyCLIPVisionConfig,
|
11 |
+
) -> tuple[nn.Module, int]:
|
12 |
+
base = timm.create_model(config.vision_model, num_classes=0, pretrained=True)
|
13 |
+
num_features = base.num_features
|
14 |
+
return base, num_features
|
15 |
+
|
16 |
+
|
17 |
+
def get_vision_transform(config: TinyCLIPVisionConfig) -> transforms.Compose:
|
18 |
+
timm_config = data.resolve_data_config({}, model=config.vision_model)
|
19 |
transform = data.transforms_factory.create_transform(**timm_config)
|
20 |
+
return transform # type: ignore
|