From e38a4f62d7e513eac7be85aa680749678997df61 Mon Sep 17 00:00:00 2001 From: Remi Cresson Date: Wed, 24 Mar 2021 18:17:09 +0100 Subject: [PATCH 1/2] DOC: pretrained models documentation --- doc/PRETRAINED_MODELS.md | 11 +++++++++++ 1 file changed, 11 insertions(+) create mode 100644 doc/PRETRAINED_MODELS.md diff --git a/doc/PRETRAINED_MODELS.md b/doc/PRETRAINED_MODELS.md new file mode 100644 index 0000000..2e2ce37 --- /dev/null +++ b/doc/PRETRAINED_MODELS.md @@ -0,0 +1,11 @@ +# SR4RS Pre-models + +This section in currently in contruction! + + + +| Model name | Nb. patches used | Area | Source Sensor | Target Sensor | L1 loss wgt. | L2 loss wgt. | VGG loss wgt. | Loss type | +| ------------- | ---------------- | --------------------- | ------------------ | -------------- | ------------ | ------------ | ------------- | --------- | +| mini-mtp-2.5 | 3984 | Montpellier area (Fr) | Sentinel-2 (B4328) | Spot-7 (B1234) | 0.0 | 100.0 | 0.0003 | WGAN-GP | + + From e116ecef4abfb2b97fb8fce5e4b8190a9f7a1739 Mon Sep 17 00:00:00 2001 From: Remi Cresson Date: Wed, 24 Mar 2021 18:51:04 +0100 Subject: [PATCH 2/2] DOC: add pretrained models links --- doc/PRETRAINED_MODELS.md | 13 +++++++++---- 1 file changed, 9 insertions(+), 4 deletions(-) diff --git a/doc/PRETRAINED_MODELS.md b/doc/PRETRAINED_MODELS.md index 2e2ce37..b436b64 100644 --- a/doc/PRETRAINED_MODELS.md +++ b/doc/PRETRAINED_MODELS.md @@ -1,11 +1,16 @@ # SR4RS Pre-models -This section in currently in contruction! +This section in currently in contruction. +## Notes +- The checkpoints can be used to start the training in `train.py` (use the `--load_ckpt` parameter). You can retrieve the model parameters values from the checkpoint file name. +- The SavedModels can be used directly on remote sensing images from `sr.py` to generate high-resolution images. +- Before applying a model, check that your input image has the same spectral content/bands + in the same order as indicated in the table below. -| Model name | Nb. patches used | Area | Source Sensor | Target Sensor | L1 loss wgt. | L2 loss wgt. | VGG loss wgt. | Loss type | -| ------------- | ---------------- | --------------------- | ------------------ | -------------- | ------------ | ------------ | ------------- | --------- | -| mini-mtp-2.5 | 3984 | Montpellier area (Fr) | Sentinel-2 (B4328) | Spot-7 (B1234) | 0.0 | 100.0 | 0.0003 | WGAN-GP | +## Models +| Model name | Nb. patches used | Area | Source Sensor | Target Sensor | L1 loss weight | L2 loss weight | VGG loss weight (feats) | Loss type | Depth | Nb. res. blocks | Comment | Links | +| ---------- | ---------------- | -----| -------------- | ------------ | -------------- | -------------- | ----------------------- | --------- | ----- | --------------- | ------- | ----- | +| mini-mtp-2.5 | 3984 | Montpellier area, Fr. | Sentinel-2 (B4328, TOC reflectance) from THEIA Land data center | Spot-7 (B1234, corrected) | 0.0 | 1000.0 | 0.00001 ("1234") | WGAN-GP | 64 | 16 | Quickly trained model, not very nice. For testing purposes. You can apply it on Sentinel-2 images from ESA hub, though it was trained on a TOC reflectance product. | [checkpoint](https://nextcloud.inrae.fr/s/MWaqnKCsRmkQmtm) / [SavedModel](https://nextcloud.inrae.fr/s/JLsak68H2KYzPyG) |