1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
|
"""Lightning model for transformer networks."""
from typing import Callable, Optional, Tuple, Type
from text_recognizer.network.mammut import MaMMUT
import torch
from einops import rearrange
from omegaconf import DictConfig
from torch import einsum, nn, Tensor
from torchmetrics import CharErrorRate, WordErrorRate
import torch.nn.functional as F
from text_recognizer.decoder.greedy_decoder import GreedyDecoder
from text_recognizer.data.tokenizer import Tokenizer
from .base import LitBase
class LitMaMMUT(LitBase):
def __init__(
self,
network: MaMMUT,
loss_fn: Type[nn.Module],
optimizer_config: DictConfig,
tokenizer: Tokenizer,
decoder: Callable = GreedyDecoder,
lr_scheduler_config: Optional[DictConfig] = None,
max_output_len: int = 682,
caption_loss_weight=1.0,
contrastive_loss_weight=1.0,
) -> None:
super().__init__(
network,
loss_fn,
optimizer_config,
lr_scheduler_config,
tokenizer,
)
self.max_output_len = max_output_len
self.val_cer = CharErrorRate()
self.test_cer = CharErrorRate()
self.val_wer = WordErrorRate()
self.test_wer = WordErrorRate()
self.decoder = decoder
self.contrastive_loss = F.cross_entropy
self.caption_loss_weight = caption_loss_weight
self.contrastive_loss_weight = contrastive_loss_weight
self.temperature = nn.Parameter(Tensor([1.0]))
def forward(self, data: Tensor) -> Tensor:
"""Autoregressive forward pass."""
return self.predict(data)
def to_caption_loss(self, logits: Tensor, text: Tensor) -> Tensor:
caption_loss = self.loss_fn(logits, text[:, 1:])
return self.caption_loss_weight * caption_loss
def to_contrastive_loss(
self, image_embeddings: Tensor, text_embeddings: Tensor
) -> Tensor:
b, device = image_embeddings.shape[0], image_embeddings.device
image_latents, text_latents = self.network.to_latents(
image_embeddings, text_embeddings
)
sim = einsum("i d, j d -> i j", text_latents, image_latents)
sim = sim * self.temperature.exp()
contrastive_labels = torch.arange(b, device=device)
contrastive_loss = (
F.cross_entropy(sim, contrastive_labels)
+ F.cross_entropy(sim.t(), contrastive_labels)
) / 2
return self.contrastive_loss_weight * contrastive_loss
def teacher_forward(self, images: Tensor, text: Tensor) -> Tuple[Tensor, Tensor]:
"""Non-autoregressive forward pass."""
text_embeddings = self.network.to_text_cls_features(text[:, :-1])
image_embeddings, image_features = self.network.to_image_features(images)
logits = self.network.decode(text[:, :-1], image_features)
logits = rearrange(logits, "b n c -> b c n")
caption_loss = self.to_caption_loss(logits, text)
contrastive_loss = self.to_contrastive_loss(image_embeddings, text_embeddings)
self.log("train/caption_loss", caption_loss)
self.log("train/contrastive_loss", contrastive_loss)
return logits, caption_loss + contrastive_loss
def training_step(self, batch: Tuple[Tensor, Tensor], batch_idx: int) -> dict:
"""Training step."""
data, targets = batch
logits, loss = self.teacher_forward(data, targets)
self.log("train/loss", loss, prog_bar=True)
outputs = {"loss": loss}
if self.is_logged_batch():
preds, gts = self.tokenizer.decode_logits(
logits
), self.tokenizer.batch_decode(targets)
outputs.update({"predictions": preds, "ground_truths": gts})
return outputs
def validation_step(self, batch: Tuple[Tensor, Tensor], batch_idx: int) -> dict:
"""Validation step."""
data, targets = batch
preds = self(data)
preds, gts = self.tokenizer.batch_decode(preds), self.tokenizer.batch_decode(
targets
)
self.val_cer(preds, gts)
self.val_wer(preds, gts)
self.log("val/cer", self.val_cer, on_step=False, on_epoch=True, prog_bar=True)
self.log("val/wer", self.val_wer, on_step=False, on_epoch=True, prog_bar=True)
outputs = {}
self.add_on_first_batch(
{"predictions": preds, "ground_truths": gts}, outputs, batch_idx
)
return outputs
def test_step(self, batch: Tuple[Tensor, Tensor], batch_idx: int) -> dict:
"""Test step."""
data, targets = batch
preds = self(data)
preds, gts = self.tokenizer.batch_decode(preds), self.tokenizer.batch_decode(
targets
)
self.test_cer(preds, gts)
self.test_wer(preds, gts)
self.log("test/cer", self.test_cer, on_step=False, on_epoch=True, prog_bar=True)
self.log("test/wer", self.test_wer, on_step=False, on_epoch=True, prog_bar=True)
outputs = {}
self.add_on_first_batch(
{"predictions": preds, "ground_truths": gts}, outputs, batch_idx
)
return outputs
@torch.no_grad()
def predict(self, x: Tensor) -> Tensor:
return self.decoder(x)
|