ttach
v0.0.3
图像测试时间增加与Pytorch!
类似于增强对培训集的数据,测试时间增加的目的是对测试图像进行随机修改。因此,我们将多次向其展示增强图像,而不是向训练有素的模型显示常规的“干净”图像,而是只向训练有素的模型显示一次。然后,我们将平均每个相应图像的预测,并将其作为我们的最终猜测[1]。
Input
| # input batch of images
/ / /| # apply augmentations (flips, rotation, scale, etc.)
| | | | | | | # pass augmented batches through model
| | | | | | | # reverse transformations for each batch of masks/labels
/ / / # merge predictions (mean, max, gmean, etc.)
| # output batch of masks/labels
Output
import ttach as tta
tta_model = tta . SegmentationTTAWrapper ( model , tta . aliases . d4_transform (), merge_mode = 'mean' ) tta_model = tta . ClassificationTTAWrapper ( model , tta . aliases . five_crop_transform ()) tta_model = tta . KeypointsTTAWrapper ( model , tta . aliases . flip_transform (), scaled = True )注意:该模型必须返回格式torch([x1, y1, ..., xn, yn])
# defined 2 * 2 * 3 * 3 = 36 augmentations !
transforms = tta . Compose (
[
tta . HorizontalFlip (),
tta . Rotate90 ( angles = [ 0 , 180 ]),
tta . Scale ( scales = [ 1 , 2 , 4 ]),
tta . Multiply ( factors = [ 0.9 , 1 , 1.1 ]),
]
)
tta_model = tta . SegmentationTTAWrapper ( model , transforms ) # Example how to process ONE batch on images with TTA
# Here `image`/`mask` are 4D tensors (B, C, H, W), `label` is 2D tensor (B, N)
for transformer in transforms : # custom transforms or e.g. tta.aliases.d4_transform()
# augment image
augmented_image = transformer . augment_image ( image )
# pass to model
model_output = model ( augmented_image , another_input_data )
# reverse augmentation for mask and label
deaug_mask = transformer . deaugment_mask ( model_output [ 'mask' ])
deaug_label = transformer . deaugment_label ( model_output [ 'label' ])
# save results
labels . append ( deaug_mask )
masks . append ( deaug_label )
# reduce results as you want, e.g mean/max/min
label = mean ( labels )
mask = mean ( masks )| 转换 | 参数 | 值 |
|---|---|---|
| 水平流体 | - | - |
| 垂直流动 | - | - |
| 旋转90 | 角度 | 列表[0,90,180,270] |
| 规模 | 秤 插值 | 列表[float] “最近”/“线性” |
| 调整大小 | 尺寸 原始_size 插值 | 列表[元组[INT,INT]] 元组[int,int] “最近”/“线性” |
| 添加 | 值 | 列表[float] |
| 乘 | 因素 | 列表[float] |
| Fivecrops | crop_height crop_width | int int |
PYPI:
$ pip install ttach来源:
$ pip install git+https://github.com/qubvel/ttachdocker build -f Dockerfile.dev -t ttach:dev . && docker run --rm ttach:dev pytest -p no:cacheprovider