Spaces:
Runtime error
Runtime error
Upload lightning_utils.py
Browse files- lightning_utils.py +171 -0
lightning_utils.py
ADDED
|
@@ -0,0 +1,171 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
|
| 2 |
+
from torch.utils.data import Dataset, DataLoader
|
| 3 |
+
from loss import YoloLoss
|
| 4 |
+
import config
|
| 5 |
+
import torch
|
| 6 |
+
from dataset import YOLODataset
|
| 7 |
+
from torch.optim.lr_scheduler import OneCycleLR
|
| 8 |
+
import random
|
| 9 |
+
from model import YOLOv3
|
| 10 |
+
import lightning.pytorch as pl
|
| 11 |
+
|
| 12 |
+
def criterion(out, y, anchors):
|
| 13 |
+
loss_fn = YoloLoss()
|
| 14 |
+
loss = (
|
| 15 |
+
loss_fn(out[0], y[0], anchors[0])
|
| 16 |
+
+ loss_fn(out[1], y[1], anchors[1])
|
| 17 |
+
+ loss_fn(out[2], y[2], anchors[2]))
|
| 18 |
+
return loss
|
| 19 |
+
|
| 20 |
+
|
| 21 |
+
def get_loader(train_dataset, test_dataset):
|
| 22 |
+
train_loader = DataLoader(
|
| 23 |
+
dataset=train_dataset,
|
| 24 |
+
batch_size=config.BATCH_SIZE,
|
| 25 |
+
num_workers=config.NUM_WORKERS,
|
| 26 |
+
pin_memory=config.PIN_MEMORY,
|
| 27 |
+
shuffle=True,
|
| 28 |
+
drop_last=False,
|
| 29 |
+
)
|
| 30 |
+
|
| 31 |
+
test_loader = DataLoader(
|
| 32 |
+
dataset=test_dataset,
|
| 33 |
+
batch_size=config.BATCH_SIZE,
|
| 34 |
+
num_workers=config.NUM_WORKERS,
|
| 35 |
+
pin_memory=config.PIN_MEMORY,
|
| 36 |
+
shuffle=False,
|
| 37 |
+
drop_last=False,
|
| 38 |
+
|
| 39 |
+
)
|
| 40 |
+
|
| 41 |
+
return(train_loader, test_loader)
|
| 42 |
+
|
| 43 |
+
|
| 44 |
+
def accuracy_fn(y, out, threshold,
|
| 45 |
+
correct_class, correct_obj,
|
| 46 |
+
correct_noobj, tot_class_preds,
|
| 47 |
+
tot_obj, tot_noobj):
|
| 48 |
+
|
| 49 |
+
for i in range(3):
|
| 50 |
+
|
| 51 |
+
obj = y[i][..., 0] == 1 # in paper this is Iobj_i
|
| 52 |
+
noobj = y[i][..., 0] == 0 # in paper this is Iobj_i
|
| 53 |
+
|
| 54 |
+
correct_class += torch.sum(
|
| 55 |
+
torch.argmax(out[i][..., 5:][obj], dim=-1) == y[i][..., 5][obj]
|
| 56 |
+
)
|
| 57 |
+
tot_class_preds += torch.sum(obj)
|
| 58 |
+
|
| 59 |
+
obj_preds = torch.sigmoid(out[i][..., 0]) > threshold
|
| 60 |
+
correct_obj += torch.sum(obj_preds[obj] == y[i][..., 0][obj])
|
| 61 |
+
tot_obj += torch.sum(obj)
|
| 62 |
+
correct_noobj += torch.sum(obj_preds[noobj] == y[i][..., 0][noobj])
|
| 63 |
+
tot_noobj += torch.sum(noobj)
|
| 64 |
+
|
| 65 |
+
return((correct_class/(tot_class_preds+1e-16))*100,
|
| 66 |
+
(correct_noobj/(tot_noobj+1e-16))*100,
|
| 67 |
+
(correct_obj/(tot_obj+1e-16))*100)
|
| 68 |
+
|
| 69 |
+
|
| 70 |
+
def get_datasets(train_loc="/train.csv", test_loc="/test.csv"):
|
| 71 |
+
|
| 72 |
+
train_dataset = YOLODataset(
|
| 73 |
+
config.DATASET + train_loc,
|
| 74 |
+
transform=config.train_transform,
|
| 75 |
+
img_dir=config.IMG_DIR,
|
| 76 |
+
label_dir=config.LABEL_DIR,
|
| 77 |
+
anchors=config.ANCHORS,
|
| 78 |
+
)
|
| 79 |
+
|
| 80 |
+
test_dataset = YOLODataset(
|
| 81 |
+
config.DATASET + test_loc,
|
| 82 |
+
transform=config.test_transform,
|
| 83 |
+
img_dir=config.IMG_DIR,
|
| 84 |
+
label_dir=config.LABEL_DIR,
|
| 85 |
+
anchors=config.ANCHORS,
|
| 86 |
+
train=False
|
| 87 |
+
)
|
| 88 |
+
|
| 89 |
+
return(train_dataset, test_dataset)
|
| 90 |
+
|
| 91 |
+
|
| 92 |
+
|
| 93 |
+
class YOLOv3Lightning(pl.LightningModule):
|
| 94 |
+
def __init__(self, dataset=None, lr=config.LEARNING_RATE):
|
| 95 |
+
super().__init__()
|
| 96 |
+
|
| 97 |
+
self.save_hyperparameters()
|
| 98 |
+
|
| 99 |
+
self.model = YOLOv3(num_classes=config.NUM_CLASSES)
|
| 100 |
+
self.lr = lr
|
| 101 |
+
self.criterion = criterion
|
| 102 |
+
self.losses = []
|
| 103 |
+
self.threshold = config.CONF_THRESHOLD
|
| 104 |
+
self.iou_threshold = config.NMS_IOU_THRESH
|
| 105 |
+
self.train_idx = 0
|
| 106 |
+
self.box_format="midpoint"
|
| 107 |
+
self.dataset = dataset
|
| 108 |
+
self.criterion = criterion
|
| 109 |
+
self.accuracy_fn = accuracy_fn
|
| 110 |
+
self.tot_class_preds, self.correct_class = 0, 0
|
| 111 |
+
self.tot_noobj, self.correct_noobj = 0, 0
|
| 112 |
+
self.tot_obj, self.correct_obj = 0, 0
|
| 113 |
+
self.scaled_anchors = 0
|
| 114 |
+
|
| 115 |
+
def forward(self, x):
|
| 116 |
+
return self.model(x)
|
| 117 |
+
|
| 118 |
+
def set_scaled_anchor(self, scaled_anchors):
|
| 119 |
+
self.scaled_anchors = scaled_anchors
|
| 120 |
+
|
| 121 |
+
def on_train_epoch_start(self):
|
| 122 |
+
# Set a new image size for the dataset at the beginning of each epoch
|
| 123 |
+
size_idx = random.choice(range(len(config.IMAGE_SIZES)))
|
| 124 |
+
self.dataset.set_image_size(size_idx)
|
| 125 |
+
self.set_scaled_anchor((
|
| 126 |
+
torch.tensor(config.ANCHORS)
|
| 127 |
+
* torch.tensor(config.S[size_idx]).unsqueeze(1).unsqueeze(1).repeat(1, 3, 2)
|
| 128 |
+
))
|
| 129 |
+
|
| 130 |
+
def on_validation_epoch_start(self):
|
| 131 |
+
self.set_scaled_anchor((
|
| 132 |
+
torch.tensor(config.ANCHORS)
|
| 133 |
+
* torch.tensor(config.S[1]).unsqueeze(1).unsqueeze(1).repeat(1, 3, 2)
|
| 134 |
+
))
|
| 135 |
+
|
| 136 |
+
|
| 137 |
+
def training_step(self, batch, batch_idx):
|
| 138 |
+
x, y = batch
|
| 139 |
+
out = self(x)
|
| 140 |
+
loss = self.criterion(out, y, self.scaled_anchors)
|
| 141 |
+
|
| 142 |
+
self.log('train_loss', loss, prog_bar=True, on_epoch=True, on_step=True, logger=True)
|
| 143 |
+
|
| 144 |
+
return loss
|
| 145 |
+
|
| 146 |
+
def validation_step(self, val_batch, batch_idx):
|
| 147 |
+
x, labels = val_batch
|
| 148 |
+
out = self(x)
|
| 149 |
+
|
| 150 |
+
loss = self.criterion(out, labels, self.scaled_anchors)
|
| 151 |
+
self.log('val_loss', loss, prog_bar=True, on_epoch=True)
|
| 152 |
+
|
| 153 |
+
self.evaluate(x, labels, out, 'val')
|
| 154 |
+
|
| 155 |
+
|
| 156 |
+
def evaluate(self, x, y, out, stage=None):
|
| 157 |
+
|
| 158 |
+
# Class Accuracy
|
| 159 |
+
class_accuracy, no_obj_accuracy, obj_accuracy = self.accuracy_fn(y,
|
| 160 |
+
out,
|
| 161 |
+
self.threshold,
|
| 162 |
+
self.correct_class,
|
| 163 |
+
self.correct_obj,
|
| 164 |
+
self.correct_noobj,
|
| 165 |
+
self.tot_class_preds,
|
| 166 |
+
self.tot_obj,
|
| 167 |
+
self.tot_noobj, )
|
| 168 |
+
if stage:
|
| 169 |
+
self.log(f'{stage}_class_accuracy', class_accuracy, prog_bar=True, on_epoch=True, on_step=True, logger=True)
|
| 170 |
+
self.log(f'{stage}_no_obj_accuracy', no_obj_accuracy, prog_bar=True, on_epoch=True, on_step=True, logger=True)
|
| 171 |
+
self.log(f'{stage}_obj_accuracy', obj_accuracy, prog_bar=True, on_epoch=True, on_step=True, logger=True)
|