Skip to content

Commit

Permalink
Merge pull request #7 from allenai/refactor-upgrade
Browse files Browse the repository at this point in the history
Refactor upgrade
  • Loading branch information
pbeukema authored Jan 4, 2024
2 parents 1d287fe + 55fd134 commit 190095c
Show file tree
Hide file tree
Showing 3 changed files with 19 additions and 161 deletions.
159 changes: 11 additions & 148 deletions README.md
Original file line number Diff line number Diff line change
Expand Up @@ -24,19 +24,7 @@ This is the first planned release of these models. We anticipate further updates

## **Dataset**

The Sentinel-1 models in this repository have been trained on a dataset which we are also releasing, and will therefore describe in more detail here.

**Note**: If you are only interested in performing Sentinel-1 inference, you can safely skip this section.

The _metadata_ for this dataset dataset consists of:

1. Approximately 60,000 point labels associated with Sentinel-1 imagery, corresponding to the centers of vessels in the imagery. These labels were produced via hand annotations by subject matter experts.
2. Attribute labels for a subset of approximately 20,000 of the point labels. The attribute labels describe additional information about the vessels in the imagery, such as their length, width, heading, speed, and vessel type. These attributes were produced by cross-referencing detection labels with AIS records.

All of this metadata is stored in a sqlite database included with this repository: [`./data/metadata.sqlite3`](./data/metadata.sqlite3).

A brief report of some summary statistics of the dataset is also included: [`./data/reports/dataset_report.ipynb`](./data/reports/dataset_report.ipynb).
<br>
The annotations are included in data/metadata.sqlite3 with the following schema

### Metadata Schema

Expand Down Expand Up @@ -297,23 +285,12 @@ or use an x86 VM to build.
To use the docker image:
1. Acquire the docker image.
Pull the latest container from Github container registry:
```bash
export REPO_OWNER=your-repo-owner-name
export REPO_NAME=sentinel-vessel-detection
export IMAGE_TAG=latest
docker pull ghcr.io/$REPO_OWNER/$REPO_NAME/vessel-detection-download:$IMAGE_TAG
```
or build it from source by running the following from the top level of this repo:
1. Build the docker image.
```bash
export GIT_HASH=$(git rev-parse --short HEAD)
export IMAGE_TAG=$GIT_HASH
docker build . -f docker/downloader.dockerfile -t vessel-detection-download:$IMAGE_TAG
docker build . -f dockerfile -t vessel-detection-sentinels:$IMAGE_TAG
```
2. Create a [Copernicus Hub](https://scihub.copernicus.eu/dhus/#/home) account, and export your user credentials:
Expand Down Expand Up @@ -375,7 +352,7 @@ run specified in the default config [./data/cfg/](./data/config) on a particular
We describe below how to set up an environment on your machine for training (either manually, or via docker).
**Note:** Our docker images implicitly assume an x86 architechture. While you may have luck building these images on other architechtures ( e.g. by using docker's `--platform` arg),
**Note:** Our docker images implicitly assume an x86 architecture. While you may have luck building these images on other architechtures ( e.g. by using docker's `--platform` arg),
we have not tested this functionality. For example, we have not tested these builds on ARM machines.
If you have trouble building the image from source on your machine, you can pull it from ghcr, use the scripts without docker,
or use an x86 VM to build.
Expand Down Expand Up @@ -482,22 +459,11 @@ saved weights, and training and model config. Here `{model_name}` will indicate

You can also use our docker container to avoid setting up the necessary environment manually.

1. Acquire the docker image.

Pull the latest container from Github container registry:

```bash
export REPO_OWNER=your-repo-owner-name
export REPO_NAME=sentinel-vessel-detection
export IMAGE_TAG=latest
docker pull ghcr.io/$REPO_OWNER/$REPO_NAME/vessel-detection-train:$IMAGE_TAG
```

or build it from source by running the following from the top level of this repo:
1. Build the docker image.

```bash
export IMAGE_TAG=$(git rev-parse --short HEAD)
docker build . -f docker/train.dockerfile -t vessel-detection-train:$IMAGE_TAG
docker build . -f Dockerfile -t vessel-detection-sentinels:$IMAGE_TAG
```

2. Place the entire `preprocess` folder generated by the data download script, as well as the [`./data/metadata.sqlite3`](./data/metadata.sqlite3) file, somewhere accessible on your machine.
Expand Down Expand Up @@ -589,12 +555,7 @@ saved weights, and training and model config. Here `{model_name}` will indicate

You can perform inference right off the bat with our provided trained weights, or with weights you produce from running our training scripts.

The script `detect.py` allows one to perform inference on a raw (but de-compressed) Sentinel-1 and Sentinel-2 image products. It has
a few command line options, which you can read about in the source file, or using

```bash
python3 detect.py --help
```
The script `src/main.py` allows one to perform inference on a raw (but de-compressed) Sentinel-1 and Sentinel-2 image products.

### System Requirements and Performance

Expand Down Expand Up @@ -678,62 +639,13 @@ To get started in your own environment:
5. Install the python packages required for training via pip:

```bash
pip install -r inference_requirements.txt
pip install -r requirements.txt
```

6. Run the inference script for the detection model.

To run inference on a single image, without access to historical overlaps:

```bash
python detect.py \
--raw_path=data/ \
--scratch_path=data/scratch/ \
--output=data/output/ \
--scene_id=S1B_IW_GRDH_1SDV_20211130T025211_20211130T025236_029811_038EEF_D350.SAFE \
--conf=.9 \
--nms_thresh=10 \
--save_crops=True \
--detector_model_dir=data/models/frcnn_cmp2/3dff445 \
--postprocess_model_dir=data/models/attr/c34aa37 \
--catalog=sentinel1
```

To run inference on a single image, using additional context from a single historical overlap of the image:

```bash
python detect.py \
--raw_path=data/ \
--scratch_path=data/scratch/ \
--output=data/output/ \
--scene_id=S1B_IW_GRDH_1SDV_20211130T025211_20211130T025236_029811_038EEF_D350.SAFE \
--historical1=S1B_IW_GRDH_1SDV_20211118T025212_20211118T025237_029636_03896D_FCBE.SAFE \
--conf=.9 \
--nms_thresh=10 \
--save_crops=True \
--detector_model_dir=data/models/frcnn_cmp2/3dff445 \
--postprocess_model_dir=data/models/attr/c34aa37 \
--catalog=sentinel1
```

To run inference on a single image, using additional context from two historical overlaps of the image:

```bash
python detect.py \
--raw_path=data/ \
--scratch_path=data/scratch/ \
--output=data/output/ \
--scene_id=S1B_IW_GRDH_1SDV_20211130T025211_20211130T025236_029811_038EEF_D350.SAFE \
--historical1=S1B_IW_GRDH_1SDV_20211118T025212_20211118T025237_029636_03896D_FCBE.SAFE \
--historical2=S1B_IW_GRDH_1SDV_20211106T025212_20211106T025237_029461_03840B_6E73.SAFE \
--conf=.9 \
--nms_thresh=10 \
--save_crops=True \
--detector_model_dir=data/models/frcnn_cmp2/3dff445 \
--postprocess_model_dir=data/models/attr/c34aa37 \
--catalog=sentinel1
```

### Sentinel-2

1. Download a Sentinel-2 scene (and optionally two historical overlaps of that scene) from the Copernicus Hub UI or API, and place it in a folder of your choosing.
Expand Down Expand Up @@ -788,62 +700,13 @@ To get started in your own environment:
5. Install the python packages required for training via pip:

```bash
pip install -r inference_requirements.txt
pip install -r requirements.txt
```

6. Run the inference script for the detection model.

To run inference on a single image, without access to historical overlaps:

```bash
python detect.py \
--raw_path=data/ \
--scratch_path=data/scratch/ \
--output=data/output/ \
--scene_id=S2A_MSIL1C_20230108T060231_N0509_R091_T42RUN_20230108T062956.SAFE \
--conf=.9 \
--nms_thresh=10 \
--save_crops=True \
--detector_model_dir=data/models/frcnn_cmp2/15cddd5-sentinel2-swinv2-small-fpn \
--postprocess_model_dir=data/models/attr/e609150-sentinel2-attr-resnet \
--catalog=sentinel2
```

To run inference on a single image, using additional context from a single historical overlap of the image:

```bash
python detect.py \
--raw_path=data/ \
--scratch_path=data/scratch/ \
--output=data/output/ \
--scene_id=S2A_MSIL1C_20230108T060231_N0509_R091_T42RUN_20230108T062956.SAFE \
--historical1=S2A_MSIL1C_20230111T061221_N0509_R134_T42RUN_20230111T064344.SAFE \
--conf=.9 \
--nms_thresh=10 \
--save_crops=True \
--detector_model_dir=data/models/frcnn_cmp2/15cddd5-sentinel2-swinv2-small-fpn \
--postprocess_model_dir=data/models/attr/e609150-sentinel2-attr-resnet \
--catalog=sentinel2
```

To run inference on a single image, using additional context from two historical overlaps of the image:

```bash
python detect.py \
--raw_path=data/ \
--scratch_path=data/scratch/ \
--output=data/output/ \
--scene_id=S2A_MSIL1C_20230108T060231_N0509_R091_T42RUN_20230108T062956.SAFE \
--historical1=S2A_MSIL1C_20230111T061221_N0509_R134_T42RUN_20230111T064344.SAFE \
--historical2=S2B_MSIL1C_20230106T061239_N0509_R134_T42RUN_20230106T063926.SAFE \
--conf=.9 \
--nms_thresh=10 \
--save_crops=True \
--detector_model_dir=data/models/frcnn_cmp2/15cddd5-sentinel2-swinv2-small-fpn \
--postprocess_model_dir=data/models/attr/e609150-sentinel2-attr-resnet \
--catalog=sentinel2
```

### Notes (common to both imagery sources):

Here `--raw_path` must point to a directory containing the directory specified by `--scene_id` (and `--historical1` and `--historical2`, if they are provided).
Expand Down Expand Up @@ -884,7 +747,7 @@ or use an x86 VM to build.

```bash
export IMAGE_TAG=$(git rev-parse --short HEAD)
docker build . -f docker/inference.dockerfile -t vessel-detection:$IMAGE_TAG
docker build . -f Dockerfile -t vessel-detection:$IMAGE_TAG
```

2. Prepare a machine with at least 16GB RAM, and a GPU w/ >= 8GB memory.
Expand Down Expand Up @@ -1133,4 +996,4 @@ Running the inference script will produce at most three types of artifact in the
1. Skylight-ML (especially Mike Gartner who wrote most of this codebase
2. PRIOR at AI2 (especially Favyen Bastani and Piper Wolters) contributed considerable expertise to the foundational architectures of both models
3. European Space Agency for making Sentinel-1 and Sentinel-2 data available to the public.
4. The [Defense Innovation Unit](https://www.diu.mil/) (DIU) who built the foundation for this work through [xView3](https://iuu.xview.us/).
4. The [Defense Innovation Unit](https://www.diu.mil/) (DIU) who built the foundation for this work through [xView3](https://iuu.xview.us/).
13 changes: 0 additions & 13 deletions example/s1_request.py
Original file line number Diff line number Diff line change
@@ -1,18 +1,5 @@
""" Use this script to inference the API with locally stored data
This is the docker command that was used before the API was created:
docker run --shm-size 16G --gpus='"device=0"' \
-v /path/to/your/data:/home/vessel_detection/data vessel-detection:$IMAGE_TAG \
--raw_path=/home/vessel_detection/data/ \
--scratch_path=/home/vessel_detection/data/scratch/ \
--output=/home/vessel_detection/data/output/ \
--detector_model_dir=/home/vessel_detection/data/models/frcnn_cmp2/3dff445 \
--postprocess_model_dir=/home/vessel_detection/data/models/attr/c34aa37 \
--historical1=S1B_IW_GRDH_1SDV_20211118T025212_20211118T025237_029636_03896D_FCBE.SAFE \
--historical2=S1B_IW_GRDH_1SDV_20211106T025212_20211106T025237_029461_03840B_6E73.SAFE \
--scene_id=S1B_IW_GRDH_1SDV_20211130T025211_20211130T025236_029811_038EEF_D350.SAFE \
--conf=.9 --nms_thresh=10 --save_crops=True --catalog=sentinel1
"""
import json
Expand Down
8 changes: 8 additions & 0 deletions src/config/config.yml
Original file line number Diff line number Diff line change
@@ -0,0 +1,8 @@
main:
sentinel2_detector: "/home/vessel_detection/src/model_artifacts/multihead4/221e8ac9-sentinel2-swinv2-base-nohistory-satlas-weights"
sentinel2_postprocessor: "/home/vessel_detection/src/model_artifacts/attr/2feb47b-sentinel2-attr-resnet-tci-IR"
sentinel1_detector: "/home/vessel_detection/src/model_artifacts/frcnn_cmp2/3dff445"
sentinel1_postprocessor: "/home/vessel_detection/src/model_artifacts/attr/c34aa37"

pipeline:
CLOUD_MASK_THRESHOLD: 0.5

0 comments on commit 190095c

Please sign in to comment.