Copyright (c) Microsoft Corporation. All rights reserved.
Licensed under the MIT License.
Transforms¶
In this tutorial, we demonstrate how to use TorchGeo’s data augmentation transforms and provide examples of how to utilize them in your experiments with multispectral imagery.
It’s recommended to run this notebook on Google Colab if you don’t have your own GPU. Click the “Open in Colab” button above to get started.
Setup¶
Install TorchGeo
[ ]:
%pip install torchgeo
Imports¶
[16]:
from typing import Dict, List
import kornia.augmentation as K
import torch
import torch.nn as nn
import torchvision.transforms as T
from PIL import Image
from torch.utils.data import DataLoader
from torchgeo.datasets import EuroSAT
from torchgeo.transforms import AugmentationSequential, indices
Custom Transforms¶
Here we create an transform to show an example of how you can chain custom operations along with TorchGeo and Kornia transforms/augmentations. Note how our transform takes as input a Dict of Tensors. We specify our data by the keys [“image”, “mask”, “label”, etc.] and follow this standard across TorchGeo datasets.
[2]:
class MinMaxNormalize(nn.Module):
"""Normalize channels to the range [0, 1] using min/max values."""
def __init__(self, min: List[float], max: List[float]) -> None:
super().__init__()
self.min = torch.tensor(min)[:, None, None]
self.max = torch.tensor(max)[:, None, None]
self.denominator = self.max - self.min
def forward(self, inputs: Dict[str, torch.Tensor]) -> Dict[str, torch.Tensor]:
x = inputs["image"]
# Batch
if x.ndim == 4:
x = (x - self.min[None, ...]) / self.denominator[None, ...]
# Sample
else:
x = (x - self.min) / self.denominator
inputs["image"] = x.clamp(0, 1)
return inputs
Dataset Bands and Statistics¶
Below we have min/max values calculated across the dataset per band. The values were clipped to the interval [0, 98] to stretch the band values and avoid outliers influencing the band histograms.
[3]:
mins = [
1013.0,
676.0,
448.0,
247.0,
269.0,
253.0,
243.0,
189.0,
61.0,
4.0,
33.0,
11.0,
186.0,
]
maxs = [
2309.0,
4543.05,
4720.2,
5293.05,
3902.05,
4473.0,
5447.0,
5948.05,
1829.0,
23.0,
4894.05,
4076.05,
5846.0,
]
bands = {
"B1": "Coastal Aerosol",
"B2": "Blue",
"B3": "Green",
"B4": "Red",
"B5": "Vegetation Red Edge 1",
"B6": "Vegetation Red Edge 2",
"B7": "Vegetation Red Edge 3",
"B8": "NIR 1",
"B8A": "NIR 2",
"B9": "Water Vapour",
"B10": "SWIR 1",
"B11": "SWIR 2",
"B12": "SWIR 3",
}
Load the EuroSat MS dataset and dataloader¶
Here we load the EuroSat Multispectral (MS) dataset. The dataset contains 27,000 64x64 Sentinel-2 multispectral patches with 10 land cover classes.
[4]:
dataset = EuroSAT(download=True)
dataloader = DataLoader(dataset, batch_size=4, shuffle=True, num_workers=2)
dataloader = iter(dataloader)
print(f"Number of images in dataset: {len(dataset)}")
print(f"Dataset Classes: {dataset.classes}")
Number of images in dataset: 27000
Dataset Classes: ['AnnualCrop', 'Forest', 'HerbaceousVegetation', 'Highway', 'Industrial', 'Pasture', 'PermanentCrop', 'Residential', 'River', 'SeaLake']
Load a sample and batch of images and labels¶
Here we test our dataset by loading a single image and label. Note how the image is of shape (13, 64, 64) containing a 64x64 shape with 13 multispectral bands.
[16]:
sample = dataset[0]
x, y = sample["image"], sample["label"]
print(x.shape, x.dtype, x.min(), x.max())
print(y, dataset.classes[y])
torch.Size([13, 64, 64]) torch.int32 tensor(9, dtype=torch.int32) tensor(3490, dtype=torch.int32)
tensor(0) AnnualCrop
Here we test our dataloader by loading a single batch of images and labels. Note how the image is of shape (4, 13, 64, 64) containing 4 samples due to our batch_size.
[17]:
batch = next(dataloader)
x, y = batch["image"], batch["label"]
print(x.shape, x.dtype, x.min(), x.max())
print(y, [dataset.classes[i] for i in y])
torch.Size([4, 13, 64, 64]) torch.int32 tensor(6, dtype=torch.int32) tensor(4696, dtype=torch.int32)
tensor([6, 1, 8, 4]) ['PermanentCrop', 'Forest', 'River', 'Industrial']
Transforms Usage¶
Transforms are able to operate across batches of samples and singular samples. This allows them to be used inside the dataset itself or externally, chained together with other transform operations using nn.Sequential
.
[19]:
transforms = MinMaxNormalize(mins, maxs)
print(batch["image"].shape)
batch = transforms(batch)
print(batch["image"].dtype, batch["image"].min(), batch["image"].max())
torch.Size([4, 13, 64, 64])
torch.float32 tensor(0.0079) tensor(0.9734)
Indices can also be computed on batches of images and appended as an additional band to the specified channel dimension. Notice how the number of channels increases from 13 -> 14.
[20]:
transform = indices.AppendNDVI(index_red=3, index_nir=7)
batch = next(dataloader)
print(batch["image"].shape)
batch = transform(batch)
print(batch["image"].shape)
torch.Size([4, 13, 64, 64])
torch.Size([4, 14, 64, 64])
This makes it incredibly easy to add indices as additional features during training by chaining multiple Appends together.
[21]:
transforms = nn.Sequential(
MinMaxNormalize(mins, maxs),
indices.AppendNDBI(index_swir=11, index_nir=7),
indices.AppendNDSI(index_green=3, index_swir=11),
indices.AppendNDVI(index_red=3, index_nir=7),
indices.AppendNDWI(index_green=2, index_nir=7),
)
batch = next(dataloader)
print(batch["image"].shape)
batch = transforms(batch)
print(batch["image"].shape)
torch.Size([4, 13, 64, 64])
torch.Size([4, 17, 64, 64])
It’s even possible to chain indices along with augmentations from kornia for a single callable during training.
[34]:
augmentations = AugmentationSequential(
K.RandomHorizontalFlip(p=0.5), K.RandomVerticalFlip(p=0.5), data_keys=["image"]
)
transforms = nn.Sequential(
MinMaxNormalize(mins, maxs),
indices.AppendNDBI(index_swir=11, index_nir=7),
indices.AppendNDSI(index_green=3, index_swir=11),
indices.AppendNDVI(index_red=3, index_nir=7),
indices.AppendNDWI(index_green=2, index_nir=7),
augmentations,
)
batch = next(dataloader)
print(batch["image"].shape)
batch = transforms(batch)
print(batch["image"].shape)
torch.Size([4, 13, 64, 64])
torch.Size([4, 17, 64, 64])
All of our transforms are nn.Modules
. This allows us to push them and the data to the GPU to see significant gains for large scale operations.
[8]:
%nvidia-smi
Tue Sep 28 20:52:49 2021
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 470.63.01 Driver Version: 460.32.03 CUDA Version: 11.2 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|===============================+======================+======================|
| 0 Tesla K80 Off | 00000000:00:04.0 Off | 0 |
| N/A 33C P8 27W / 149W | 3MiB / 11441MiB | 0% Default |
| | | N/A |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=============================================================================|
| No running processes found |
+-----------------------------------------------------------------------------+
[10]:
augmentations = AugmentationSequential(
K.RandomHorizontalFlip(p=0.5),
K.RandomVerticalFlip(p=0.5),
K.RandomAffine(degrees=(0, 90), p=0.25),
K.RandomGaussianBlur(kernel_size=(3, 3), sigma=(0.1, 2.0), p=0.25),
K.RandomResizedCrop(size=(512, 512), scale=(0.8, 1.0), p=0.25),
data_keys=["image"],
)
transforms = nn.Sequential(
MinMaxNormalize(mins, maxs),
indices.AppendNDBI(index_swir=11, index_nir=7),
indices.AppendNDSI(index_green=3, index_swir=11),
indices.AppendNDVI(index_red=3, index_nir=7),
indices.AppendNDWI(index_green=2, index_nir=7),
augmentations,
)
device = "cpu" if torch.cuda.is_available() else "cuda"
batch_cpu = dict(image=torch.randn(64, 13, 512, 512).to("cpu"))
batch_gpu = dict(image=torch.randn(64, 13, 512, 512).to(device))
transforms_gpu = transforms.to(device)
[11]:
%%timeit -n 1 -r 1
_ = transforms(batch_cpu)
1 loop, best of 1: 12.3 s per loop
[12]:
%%timeit -n 1 -r 1
_ = transforms_gpu(batch_gpu)
1 loop, best of 1: 8.58 s per loop
Visualize Images and Labels¶
This is a Google Colab browser for the EuroSAT dataset. Adjust the slider to visualize images in the dataset.
[20]:
dataset = EuroSAT(transforms=MinMaxNormalize(mins, maxs))
# @title EuroSat Multispectral (MS) Browser { run: "auto", vertical-output: true }
idx = 16199 # @param {type:"slider", min:0, max:16199, step:1}
sample = dataset[idx]
rgb = sample["image"][1:4]
rgb = rgb[[2, 1, 0], ...]
image = T.ToPILImage()(rgb)
print(f"Class Label: {dataset.classes[sample['label']]}")
image.resize((256, 256), resample=Image.BILINEAR)
Class Label: PermanentCrop
[20]: