python train.py val --exp-config ../config_base.yaml --resume path_to_checkpoint
python train.py val --exp-config ../config_base.yaml --resume path_to_checkpoint--save-on
```
You can find the checkpoints in `exp/base/weights/`, the log files in `exp/base/logs`, and the output change maps in `exp/outs`.
You can find the checkpoints in `exp/base/weights/`, the log files in `exp/base/logs`, and the output change maps in `exp/base/outs`.
# Train on Air Change Dataset and OSCD Dataset
To carry out a full training on these two datasets and with all three architectures, run the `train9.sh` script under the root folder of this repo.
```bash
. ./train9.sh
```
And check the results in different subdirectories of `./exp/`.
# Create your own configuration file
During scientific research, it is common case that we have to do a lot of experiments with different settings, and that's why we need the configuration files to better manage those settings. In this repo, you can create a `yaml` file under the naming convention below:
`config_TAG{_SUFFIX}.yaml`
Those in the curly braces can be omitted. `TAG` usually stands for an experiment group. For example, a set of experiments for an architecture, a dataset, etc. It will be the name of the subdirectory that holds all the checkpoints, log files, and output images. `SUFFIX` can be used to distinguish different experiments in an experiment group. If it is specified, the generated files of this experiment will be tagged with `SUFFIX` in their file names. In plain English, `TAG1` and `TAG2` have major differences, while `SUFFIX1` and `SUFFIX2` of the same `TAG` share most of the configurations. By combining `TAG` and `SUFFIX`, it is convenient for both coarse-grained and find-grained control of experimental configurations.
Here is an example to help you understand. Suppose I'm going to finish my experiments on two datasets, OSCD and Lebedev, and I'm not sure which batch size achieves best performance. So I create these 5 config files.
```
config_OSCD_bs4.yaml
config_OSCD_bs8.yaml
config_OSCD_bs16.yaml
config_Lebedev_bs16.yaml
config_Lebedev_bs32.yaml
```
After training, I get my `exp/` folder like this:
```
-exp/
--OSCD/
---weights/
----model_best_bs4.pth
----model_best_bs8.pth
----model_best_bs16.pth
---outs/
---logs/
---config_OSCD_bs4.yaml
---config_OSCD_bs8.yaml
---config_OSCD_bs16.yaml
--Lebedev/
---weights/
----model_best_bs16.pth
----model_best_bs32.pth
---outs/
---logs/
---config_Lebedev_bs16.yaml
---config_Lebedev_bs32.yaml
```
Now the experiment results are organized in a more structured way, and I think it would be a little bit easier to collect the statistics. Also, since the historical experiments are arranged in neat order, you will soon remember what you'd done when you come back to these results, even after a long time.
---
# Changed
2020.3.14 Add the configuration files of my experiments.
\ No newline at end of file
2020.3.14 Add the configuration files of my experiments.
