From ade238e5c2bb6f1069a686263d6c41832e09e066 Mon Sep 17 00:00:00 2001
From: Cresson Remi <remi.cresson@irstea.fr>
Date: Mon, 14 Aug 2023 09:51:42 +0200
Subject: [PATCH 1/4] DOC: update api_tutorial
---
doc/api_tutorial.md | 54 ++++++++++++++++++++++++++-------------------
1 file changed, 31 insertions(+), 23 deletions(-)
diff --git a/doc/api_tutorial.md b/doc/api_tutorial.md
index bcd6ea45..78dcb674 100644
--- a/doc/api_tutorial.md
+++ b/doc/api_tutorial.md
@@ -388,6 +388,7 @@ size of our model:
| Conv. name | Conv. type | Kernel | Stride | Out. size | Valid out. size |
|------------|-------------------|--------|--------|-----------|-----------------|
+| *input* | / | / | / | 128 | 128 |
| *conv1* | Conv2D | 3 | 2 | 64 | 63 |
| *conv2* | Conv2D | 3 | 2 | 32 | 30 |
| *conv3* | Conv2D | 3 | 2 | 16 | 14 |
@@ -395,12 +396,18 @@ size of our model:
| *tconv1* | Transposed Conv2D | 3 | 2 | 16 | 10 |
| *tconv2* | Transposed Conv2D | 3 | 2 | 32 | 18 |
| *tconv3* | Transposed Conv2D | 3 | 2 | 64 | 34 |
+| *tconv4* | Transposed Conv2D | 3 | 2 | 128 | 66 |
This shows that our model can be applied in a fully convolutional fashion
without generating blocking artifacts, using the central part of the output of
-size 34. This is equivalent to remove \((128 - 24)/2 = 47\) pixels from
-the borders of the output. We can hence use the output cropped with **64**
-pixels, named ***predictions_crop64***.
+size 66. This is equivalent to remove \((128 - 66)/2 = 31\) pixels from
+the borders of the output. We keep the upper nearest power of 2 to keep the
+convolutions consistent between two adjacent image chunks, hence we can remove 32
+pixels from the borders. We can hence use the output cropped with **32** pixels,
+named ***predictions_crop32*** in the model outputs.
+By default, cropped outputs in `otbtf.ModelBase` are generated for the following
+values: `[16, 32, 64, 96, 128]` but that can be changed setting `inference_cropping`
+in the model `__init__()` (see the reference API documentation for details).
!!! Info
@@ -427,10 +434,11 @@ In the following subsections, we run `TensorflowModelServe` over the input
image, with the following parameters:
- the input name is ***input_xs***
-- the output name is ***predictions_crop64*** (cropping margin of 64 pixels)
-- we choose a receptive field of ***256*** and an expression field of
-***128*** so that they match the cropping margin of 64 pixels.
-
+- the output name is ***predictions_crop32*** (cropping margin of 32 pixels)
+- we choose a receptive field of ***128*** and an expression field of
+***64*** so that they match the cropping margin of 32 pixels (since we remove
+32 pixels from each side in x and y dimensions, we remove a total of 64 pixels
+from each borders in x/y dimensions).
### Command Line Interface
@@ -439,14 +447,14 @@ Open a terminal and run the following command:
```commandline
otbcli_TensorflowModelServe \
-source1.il $DATADIR/fake_spot6.jp2 \
--source1.rfieldx 256 \
--source1.rfieldy 256 \
+-source1.rfieldx 128 \
+-source1.rfieldy 128 \
-source1.placeholder "input_xs" \
-model.dir /tmp/my_1st_savedmodel \
-model.fullyconv on \
--output.names "predictions_crop64" \
--output.efieldx 128 \
--output.efieldy 128 \
+-output.names "predictions_crop32" \
+-output.efieldx 64 \
+-output.efieldy 64 \
-out softmax.tif
```
@@ -459,14 +467,14 @@ python wrapper:
import otbApplication
app = otbApplication.Registry.CreateApplication("TensorflowModelServe")
app.SetParameterStringList("source1.il", ["fake_spot6.jp2"])
-app.SetParameterInt("source1.rfieldx", 256)
-app.SetParameterInt("source1.rfieldy", 256)
+app.SetParameterInt("source1.rfieldx", 128)
+app.SetParameterInt("source1.rfieldy", 128)
app.SetParameterString("source1.placeholder", "input_xs")
app.SetParameterString("model.dir", "/tmp/my_1st_savedmodel")
app.EnableParameter("fullyconv")
-app.SetParameterStringList("output.names", ["predictions_crop64"])
-app.SetParameterInt("output.efieldx", 128)
-app.SetParameterInt("output.efieldy", 128)
+app.SetParameterStringList("output.names", ["predictions_crop32"])
+app.SetParameterInt("output.efieldx", 64)
+app.SetParameterInt("output.efieldy", 64)
app.SetParameterString("out", "softmax.tif")
app.ExecuteAndWriteOutput()
```
@@ -479,14 +487,14 @@ Using PyOTB is nicer:
import pyotb
pyotb.TensorflowModelServe({
"source1.il": "fake_spot6.jp2",
- "source1.rfieldx": 256,
- "source1.rfieldy": 256,
+ "source1.rfieldx": 128,
+ "source1.rfieldy": 128,
"source1.placeholder": "input_xs",
"model.dir": "/tmp/my_1st_savedmodel",
"fullyconv": True,
- "output.names": ["predictions_crop64"],
- "output.efieldx": 128,
- "output.efieldy": 128,
+ "output.names": ["predictions_crop32"],
+ "output.efieldx": 64,
+ "output.efieldy": 64,
"out": "softmax.tif",
})
```
@@ -499,4 +507,4 @@ pyotb.TensorflowModelServe({
control the output image chunk size and tiling/stripping layout. Combined
with the `optim` parameters, you will likely always find the best settings
suited for the hardware. Also, the receptive and expression fields sizes
- have a major contribution.
\ No newline at end of file
+ have a major contribution.
--
GitLab
From c6fca592f93b2ff0c10289d82bbbcd70b9ca4dba Mon Sep 17 00:00:00 2001
From: Cresson Remi <remi.cresson@irstea.fr>
Date: Mon, 14 Aug 2023 09:54:36 +0200
Subject: [PATCH 2/4] DOC: fix table in api_tutorial
---
doc/api_tutorial.md | 22 +++++++++++-----------
1 file changed, 11 insertions(+), 11 deletions(-)
diff --git a/doc/api_tutorial.md b/doc/api_tutorial.md
index 78dcb674..8c95190c 100644
--- a/doc/api_tutorial.md
+++ b/doc/api_tutorial.md
@@ -386,17 +386,17 @@ $$
Let's consider a chunk of input image of size 128, and check the valid output
size of our model:
-| Conv. name | Conv. type | Kernel | Stride | Out. size | Valid out. size |
-|------------|-------------------|--------|--------|-----------|-----------------|
-| *input* | / | / | / | 128 | 128 |
-| *conv1* | Conv2D | 3 | 2 | 64 | 63 |
-| *conv2* | Conv2D | 3 | 2 | 32 | 30 |
-| *conv3* | Conv2D | 3 | 2 | 16 | 14 |
-| *conv4* | Conv2D | 3 | 2 | 8 | 6 |
-| *tconv1* | Transposed Conv2D | 3 | 2 | 16 | 10 |
-| *tconv2* | Transposed Conv2D | 3 | 2 | 32 | 18 |
-| *tconv3* | Transposed Conv2D | 3 | 2 | 64 | 34 |
-| *tconv4* | Transposed Conv2D | 3 | 2 | 128 | 66 |
+| Conv. name | Conv. type | Kernel | Stride | Out. size | Valid out. size |
+|----------------|-------------------|--------|--------|-----------|-----------------|
+| *input* | / | / | / | 128 | 128 |
+| *conv1* | Conv2D | 3 | 2 | 64 | 63 |
+| *conv2* | Conv2D | 3 | 2 | 32 | 30 |
+| *conv3* | Conv2D | 3 | 2 | 16 | 14 |
+| *conv4* | Conv2D | 3 | 2 | 8 | 6 |
+| *tconv1* | Transposed Conv2D | 3 | 2 | 16 | 10 |
+| *tconv2* | Transposed Conv2D | 3 | 2 | 32 | 18 |
+| *tconv3* | Transposed Conv2D | 3 | 2 | 64 | 34 |
+| *classifier* | Transposed Conv2D | 3 | 2 | 128 | 66 |
This shows that our model can be applied in a fully convolutional fashion
without generating blocking artifacts, using the central part of the output of
--
GitLab
From 39de16c6c589d4678a7e5e84fb7994d08dbb3d30 Mon Sep 17 00:00:00 2001
From: Cresson Remi <remi.cresson@irstea.fr>
Date: Mon, 14 Aug 2023 12:10:38 +0200
Subject: [PATCH 3/4] CI: use godzilla for testing
---
.gitlab-ci.yml | 1 +
1 file changed, 1 insertion(+)
diff --git a/.gitlab-ci.yml b/.gitlab-ci.yml
index b1dff867..88a7f5a0 100644
--- a/.gitlab-ci.yml
+++ b/.gitlab-ci.yml
@@ -146,6 +146,7 @@ pages:
- public
.tests_base:
+ tags: [godzilla]
artifacts:
paths:
- $ARTIFACT_TEST_DIR/*.*
--
GitLab
From 9a736f1e5b18d5f00a5c1a0fbdc31962d5cffcdc Mon Sep 17 00:00:00 2001
From: Cresson Remi <remi.cresson@irstea.fr>
Date: Mon, 14 Aug 2023 12:46:59 +0200
Subject: [PATCH 4/4] Apply 1 suggestion(s) to 1 file(s)
---
.gitlab-ci.yml | 1 -
1 file changed, 1 deletion(-)
diff --git a/.gitlab-ci.yml b/.gitlab-ci.yml
index 88a7f5a0..b1dff867 100644
--- a/.gitlab-ci.yml
+++ b/.gitlab-ci.yml
@@ -146,7 +146,6 @@ pages:
- public
.tests_base:
- tags: [godzilla]
artifacts:
paths:
- $ARTIFACT_TEST_DIR/*.*
--
GitLab