create pl model
Browse files
model.py
ADDED
@@ -0,0 +1,111 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import pytorch_lightning as pl
|
2 |
+
import torch
|
3 |
+
|
4 |
+
from datasets import load_metric
|
5 |
+
from torch import nn
|
6 |
+
from transformers import SegformerForSemanticSegmentation
|
7 |
+
from typing import Dict
|
8 |
+
|
9 |
+
|
10 |
+
class SidewalkSegmentationModel(pl.LightningModule):
|
11 |
+
def __init__(
|
12 |
+
self,
|
13 |
+
num_labels: int,
|
14 |
+
id2label: Dict[int, str],
|
15 |
+
model_flavor: int = 0,
|
16 |
+
learning_rate: float = 6e-5,
|
17 |
+
):
|
18 |
+
super().__init__()
|
19 |
+
self.id2label = id2label
|
20 |
+
self.label2id = {v: k for k, v in id2label.items()}
|
21 |
+
self.learning_rate = learning_rate
|
22 |
+
self.metrics = {
|
23 |
+
"train": load_metric("mean_iou"),
|
24 |
+
"val": load_metric("mean_iou"),
|
25 |
+
}
|
26 |
+
|
27 |
+
self.model = SegformerForSemanticSegmentation.from_pretrained(
|
28 |
+
f"nvidia/mit-b{model_flavor}", num_labels=num_labels, id2label=self.id2label, label2id=self.label2id,
|
29 |
+
)
|
30 |
+
self.save_hyperparameters()
|
31 |
+
|
32 |
+
|
33 |
+
def forward(self, *args, **kwargs):
|
34 |
+
return self.model(*args, **kwargs)
|
35 |
+
|
36 |
+
|
37 |
+
def training_step(self, batch, batch_idx):
|
38 |
+
pixel_values = batch["pixel_values"]
|
39 |
+
labels = batch["labels"]
|
40 |
+
|
41 |
+
outputs = self(pixel_values=pixel_values, labels=labels)
|
42 |
+
loss, logits = outputs.loss, outputs.logits
|
43 |
+
|
44 |
+
self.add_batch_to_metric("train", logits, labels)
|
45 |
+
self.log("train_loss", loss, prog_bar=True, on_step=False, on_epoch=True)
|
46 |
+
return {"loss": loss}
|
47 |
+
|
48 |
+
|
49 |
+
def validation_step(self, batch, batch_idx):
|
50 |
+
pixel_values = batch["pixel_values"]
|
51 |
+
labels = batch["labels"]
|
52 |
+
|
53 |
+
outputs = self(pixel_values=pixel_values, labels=labels)
|
54 |
+
loss, logits = outputs.loss, outputs.logits
|
55 |
+
|
56 |
+
self.add_batch_to_metric("val", logits, labels)
|
57 |
+
self.log("val_loss", loss, prog_bar=True, on_step=False, on_epoch=True)
|
58 |
+
return {"val_loss": loss}
|
59 |
+
|
60 |
+
|
61 |
+
def training_epoch_end(self, training_step_outputs):
|
62 |
+
"""
|
63 |
+
Log the training metrics.
|
64 |
+
"""
|
65 |
+
metrics = self.metrics["train"].compute(num_labels=len(self.id2label), ignore_index=255, reduce_labels=False)
|
66 |
+
self.log("train_mean_iou", metrics["mean_iou"], prog_bar=True, on_step=False, on_epoch=True)
|
67 |
+
self.log("train_mean_acc", metrics["mean_accuracy"], prog_bar=True, on_step=False, on_epoch=True)
|
68 |
+
|
69 |
+
|
70 |
+
def validation_epoch_end(self, validation_step_outputs):
|
71 |
+
"""
|
72 |
+
Log the validation metrics.
|
73 |
+
"""
|
74 |
+
metrics = self.metrics["val"].compute(num_labels=len(self.id2label), ignore_index=255, reduce_labels=False)
|
75 |
+
self.log("val_mean_iou", metrics["mean_iou"], prog_bar=True, on_step=False, on_epoch=True)
|
76 |
+
self.log("val_mean_acc", metrics["mean_accuracy"], prog_bar=True, on_step=False, on_epoch=True)
|
77 |
+
|
78 |
+
|
79 |
+
def add_batch_to_metric(self, stage: str, logits: torch.Tensor, labels: torch.Tensor):
|
80 |
+
"""
|
81 |
+
Add the current batch to the metric.
|
82 |
+
|
83 |
+
Parameters
|
84 |
+
----------
|
85 |
+
stage : str
|
86 |
+
Stage of the training. Either "train" or "val".
|
87 |
+
logits : torch.Tensor
|
88 |
+
Predicted logits.
|
89 |
+
labels : torch.Tensor
|
90 |
+
Ground truth labels.
|
91 |
+
"""
|
92 |
+
with torch.no_grad():
|
93 |
+
upsampled_logits = nn.functional.interpolate(
|
94 |
+
logits, size=labels.shape[-2:], mode="bilinear", align_corners=False
|
95 |
+
)
|
96 |
+
predicted = upsampled_logits.argmax(dim=1)
|
97 |
+
self.metrics[stage].add_batch(
|
98 |
+
predictions=predicted.detach().cpu().numpy(), references=labels.detach().cpu().numpy()
|
99 |
+
)
|
100 |
+
|
101 |
+
|
102 |
+
def configure_optimizers(self) -> torch.optim.AdamW:
|
103 |
+
"""
|
104 |
+
Configure the optimizer.
|
105 |
+
|
106 |
+
Returns
|
107 |
+
-------
|
108 |
+
torch.optim.AdamW
|
109 |
+
Optimizer for the model
|
110 |
+
"""
|
111 |
+
return torch.optim.AdamW(self.parameters(), lr=self.learning_rate)
|