Skip to content
GitLab
Commit e75906c7 authored by Cresson Remi's avatar Cresson Remi
Browse files

Initial commit

parents
master
No related merge requests found
Hide whitespace changes
Inline Side-by-side
Showing
with 802 additions and 0 deletions
+802 -0
LICENSE 0 → 100644
+ 202
− 0
View file @ e75906c7
Apache License
Version 2.0, January 2004
http://www.apache.org/licenses/
TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
1. Definitions.
"License" shall mean the terms and conditions for use, reproduction,
and distribution as defined by Sections 1 through 9 of this document.
"Licensor" shall mean the copyright owner or entity authorized by
the copyright owner that is granting the License.
"Legal Entity" shall mean the union of the acting entity and all
other entities that control, are controlled by, or are under common
control with that entity. For the purposes of this definition,
"control" means (i) the power, direct or indirect, to cause the
direction or management of such entity, whether by contract or
otherwise, or (ii) ownership of fifty percent (50%) or more of the
outstanding shares, or (iii) beneficial ownership of such entity.
"You" (or "Your") shall mean an individual or Legal Entity
exercising permissions granted by this License.
"Source" form shall mean the preferred form for making modifications,
including but not limited to software source code, documentation
source, and configuration files.
"Object" form shall mean any form resulting from mechanical
transformation or translation of a Source form, including but
not limited to compiled object code, generated documentation,
and conversions to other media types.
"Work" shall mean the work of authorship, whether in Source or
Object form, made available under the License, as indicated by a
copyright notice that is included in or attached to the work
(an example is provided in the Appendix below).
"Derivative Works" shall mean any work, whether in Source or Object
form, that is based on (or derived from) the Work and for which the
editorial revisions, annotations, elaborations, or other modifications
represent, as a whole, an original work of authorship. For the purposes
of this License, Derivative Works shall not include works that remain
separable from, or merely link (or bind by name) to the interfaces of,
the Work and Derivative Works thereof.
"Contribution" shall mean any work of authorship, including
the original version of the Work and any modifications or additions
to that Work or Derivative Works thereof, that is intentionally
submitted to Licensor for inclusion in the Work by the copyright owner
or by an individual or Legal Entity authorized to submit on behalf of
the copyright owner. For the purposes of this definition, "submitted"
means any form of electronic, verbal, or written communication sent
to the Licensor or its representatives, including but not limited to
communication on electronic mailing lists, source code control systems,
and issue tracking systems that are managed by, or on behalf of, the
Licensor for the purpose of discussing and improving the Work, but
excluding communication that is conspicuously marked or otherwise
designated in writing by the copyright owner as "Not a Contribution."
"Contributor" shall mean Licensor and any individual or Legal Entity
on behalf of whom a Contribution has been received by Licensor and
subsequently incorporated within the Work.
2. Grant of Copyright License. Subject to the terms and conditions of
this License, each Contributor hereby grants to You a perpetual,
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
copyright license to reproduce, prepare Derivative Works of,
publicly display, publicly perform, sublicense, and distribute the
Work and such Derivative Works in Source or Object form.
3. Grant of Patent License. Subject to the terms and conditions of
this License, each Contributor hereby grants to You a perpetual,
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
(except as stated in this section) patent license to make, have made,
use, offer to sell, sell, import, and otherwise transfer the Work,
where such license applies only to those patent claims licensable
by such Contributor that are necessarily infringed by their
Contribution(s) alone or by combination of their Contribution(s)
with the Work to which such Contribution(s) was submitted. If You
institute patent litigation against any entity (including a
cross-claim or counterclaim in a lawsuit) alleging that the Work
or a Contribution incorporated within the Work constitutes direct
or contributory patent infringement, then any patent licenses
granted to You under this License for that Work shall terminate
as of the date such litigation is filed.
4. Redistribution. You may reproduce and distribute copies of the
Work or Derivative Works thereof in any medium, with or without
modifications, and in Source or Object form, provided that You
meet the following conditions:
(a) You must give any other recipients of the Work or
Derivative Works a copy of this License; and
(b) You must cause any modified files to carry prominent notices
stating that You changed the files; and
(c) You must retain, in the Source form of any Derivative Works
that You distribute, all copyright, patent, trademark, and
attribution notices from the Source form of the Work,
excluding those notices that do not pertain to any part of
the Derivative Works; and
(d) If the Work includes a "NOTICE" text file as part of its
distribution, then any Derivative Works that You distribute must
include a readable copy of the attribution notices contained
within such NOTICE file, excluding those notices that do not
pertain to any part of the Derivative Works, in at least one
of the following places: within a NOTICE text file distributed
as part of the Derivative Works; within the Source form or
documentation, if provided along with the Derivative Works; or,
within a display generated by the Derivative Works, if and
wherever such third-party notices normally appear. The contents
of the NOTICE file are for informational purposes only and
do not modify the License. You may add Your own attribution
notices within Derivative Works that You distribute, alongside
or as an addendum to the NOTICE text from the Work, provided
that such additional attribution notices cannot be construed
as modifying the License.
You may add Your own copyright statement to Your modifications and
may provide additional or different license terms and conditions
for use, reproduction, or distribution of Your modifications, or
for any such Derivative Works as a whole, provided Your use,
reproduction, and distribution of the Work otherwise complies with
the conditions stated in this License.
5. Submission of Contributions. Unless You explicitly state otherwise,
any Contribution intentionally submitted for inclusion in the Work
by You to the Licensor shall be under the terms and conditions of
this License, without any additional terms or conditions.
Notwithstanding the above, nothing herein shall supersede or modify
the terms of any separate license agreement you may have executed
with Licensor regarding such Contributions.
6. Trademarks. This License does not grant permission to use the trade
names, trademarks, service marks, or product names of the Licensor,
except as required for reasonable and customary use in describing the
origin of the Work and reproducing the content of the NOTICE file.
7. Disclaimer of Warranty. Unless required by applicable law or
agreed to in writing, Licensor provides the Work (and each
Contributor provides its Contributions) on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
implied, including, without limitation, any warranties or conditions
of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
PARTICULAR PURPOSE. You are solely responsible for determining the
appropriateness of using or redistributing the Work and assume any
risks associated with Your exercise of permissions under this License.
8. Limitation of Liability. In no event and under no legal theory,
whether in tort (including negligence), contract, or otherwise,
unless required by applicable law (such as deliberate and grossly
negligent acts) or agreed to in writing, shall any Contributor be
liable to You for damages, including any direct, indirect, special,
incidental, or consequential damages of any character arising as a
result of this License or out of the use or inability to use the
Work (including but not limited to damages for loss of goodwill,
work stoppage, computer failure or malfunction, or any and all
other commercial damages or losses), even if such Contributor
has been advised of the possibility of such damages.
9. Accepting Warranty or Additional Liability. While redistributing
the Work or Derivative Works thereof, You may choose to offer,
and charge a fee for, acceptance of support, warranty, indemnity,
or other liability obligations and/or rights consistent with this
License. However, in accepting such obligations, You may act only
on Your own behalf and on Your sole responsibility, not on behalf
of any other Contributor, and only if You agree to indemnify,
defend, and hold each Contributor harmless for any liability
incurred by, or claims asserted against, such Contributor by reason
of your accepting any such warranty or additional liability.
END OF TERMS AND CONDITIONS
APPENDIX: How to apply the Apache License to your work.
To apply the Apache License to your work, attach the following
boilerplate notice, with the fields enclosed by brackets "[]"
replaced with your own identifying information. (Don't include
the brackets!) The text should be enclosed in the appropriate
comment syntax for the file format. We also recommend that a
file or class name and description of purpose be included on the
same "printed page" as the copyright notice for easier
identification within third-party archives.
Copyright [2019] [Rémi Cresson (IRSTEA)]
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
README.md 0 → 100644
+ 143
− 0
View file @ e75906c7
Deep-Gapfill
======
Official implementation of **Optical image gap filling using deep convolutional autoencoder from optical and radar images** presented at __IGARSS 2019__.
R. Cresson<sup>a</sup> , D. Ienco<sup>a</sup> , R. Gaetano<sup>b</sup> , K. Ose<sup>a</sup> , D. Ho Tong Minh<sup>a</sup>
a: _IRSTEA, UMR TETIS, University of Montpellier, France_
b: _CIRAD, UMR TETIS, University of Montpellier, France_
<img src ="schema.png" />
# Dependencies
* [The Orfeo ToolBox](https://www.orfeo-toolbox.org/) (OTB)
* [TensorFlow](https://www.tensorflow.org/)
* [OTBTF](https://github.com/remicres/otbtf) remote module of OTB (can be built without TensorFlow support, but you would be able only to train the model, not to generate images with **TensorflowModelServe** or fine-tune an existing SavedModel with **TensorflowModelTrain**.)
# How to install it
The easiest way is to use [the dockerfile of OTBTF](https://github.com/remicres/otbtf/tree/master/tools/dockerfiles) since compiling OTB+TensorFlow takes some time.
[Build the docker image](https://www.google.com/search?q=how+to+build+docker+image+from+dockerfile), then just enter the docker and clone this repository.
# How to use it
Let's explain a bit how you can use Deep-Gapfill
## Step 1. Normalize the input images
Use the OTB application **ComputeImagesStatistics**, or **gdalinfo -stats** (or whatever your favorite stats application is) to compute some stats of your input images.
These statistics will be used to rescale your input images between 0 and 1.
Keep the values somewhere, because later you will need them to denormalize the output images.
The **BandMathX** OTB application is nice to perform the rescaling.
Note that the input images must be resampled at the same resolution.
### Step 2. Patches selection
You can use QGIS for that. The goal is to use the images clouds masks to select where you want to use training patches and validation patches.
In the end you should have two vector data files (one for training __patches_train.shp__, and one for validation __patches_valid.shp__) containing points centered on the patches you want to use.
The **grid**, **centroids** and **Random selection** tools from QGIS are your friends here. Be careful with no-data values! Patches should not contain any no-data pixel.
### Step 3. Patches extraction
Use the OTBTF **PatchesExtraction** application to extract patches in the images.
First, tell **OTBTF** that you need 4 sources (one for SAR: _s1t_, and three for optical images: _s2t-1_,_s2t_,_s2t+1_).
```
export OTB_TF_NSOURCES=3
```
Then run **PatchesExtraction** with $s1t, $s2t, $s2tA and $s2tB corresponding to your normalized images.
The following shows the command line for the training dataset.
```
otbcli_PatchesExtraction \
-source1.il $s1t -source1.patchsizex 64 -source1.patchsizey 64 -source1.out /data/patches/training/s1.tif \
-source2.il $s2t -source2.patchsizex 64 -source2.patchsizey 64 -source2.out /data/patches/training/im1.tif \
-source3.il $s2tA -source3.patchsizex 64 -source3.patchsizey 64 -source3.out /data/patches/training/im2.tif \
-source4.il $s2tB -source4.patchsizex 64 -source4.patchsizey 64 -source4.out /data/patches/training/im3.tif \
-vec /data/vector/patches_train.shp -field "fid" # Or whatever existing field there is... you need to put one existing field. I'll correct this issue soon.
```
Do the same for validation dataset.
```
otbcli_PatchesExtraction \
-source1.il $s1t -source1.patchsizex 64 -source1.patchsizey 64 -source1.out /data/patches/validation/s1.tif \
-source2.il $s2t -source2.patchsizex 64 -source2.patchsizey 64 -source2.out /data/patches/validation/im1.tif \
-source3.il $s2tA -source3.patchsizex 64 -source3.patchsizey 64 -source3.out /data/patches/validation/im2.tif \
-source4.il $s2tB -source4.patchsizex 64 -source4.patchsizey 64 -source4.out /data/patches/validation/im3.tif \
-vec /data/vector/patches_valid.shp -field "fid"
```
### Step 4. Train your model
Now use the script provided to build and train the model.
```
python /path/to/deep-gapfill/model.py \
--s1_t /data/patches/training/s1.tif \
--s2_t_before /data/patches/training/im1.tif \
--s2_t /data/patches/training/im2.tif \
--s2_t_after /data/patches/training/im3.tif \
--valid_s1_t /data/patches/validation/s1.tif \
--valid_s2_t_before /data/patches/validation/im1.tif \
--valid_s2_t /data/patches/validation/im2.tif \
--valid_s2_t_after /data/patches/validation/im3.tif \
--logdir /data/logs/gapfill_log --save_ckpt /data/ckpts/gapfill_model
```
You can use TensorBoard to follow the training and the validation.
In case the _tricks.py_ is missing, just add to the _PYTHONPATH_ the path of **OTBTF** python directory.
```
tensorboard --logdir /data/logs/
```
Then open yout favorite web browser.
### Step 5. Export a SavedModel
Use **OTBTF**'s **ckpt2savedmodel.py** to convert the checkpoint into a TensorFlow SavedModel.
Specify what inputs/outputs of the model you need.
```
python /work/otb/otb/Modules/Remote/otbtf/python/ckpt2savedmodel.py \
--ckpt /data/ckpts/gapfill_model-199 \
--inputs "s1:0" "s2_before:0" "s2_after:0" \
--outputs "gen_fcn:0" \
--model /data/GapFill_SavedModel
```
Note that you can now use **TensorflowModelTrain** OTBTF application to do the training of the exported SavedModel.
### Step 6. Generate the estimated image
Let's use **TensorflowModelServe** to generate the full optical image.
The model has a receptive field of 576x576 pixels and an expression field of 64x64 pixels since we use the exact FCN output ("gen_fcn" tensor).
```
otbcli_TensorflowModelServe \
-source1.il $s2tA -source1.placeholder "s2_before" -source1.rfieldx 576 -source1.rfieldy 576 \
-source2.il $s2tB -source2.placeholder "s2_after" -source2.rfieldx 576 -source2.rfieldy 576 \
-source3.il $s1t -source3.placeholder "s1" -source3.rfieldx 576 -source3.rfieldy 576 \
-model.dir /data/GapFill_SavedModel/ -model.fullyconv on \
-output.names "gen_fcn" -output.efieldx 64 -output.efieldy 64 \
-out /data/s2_estimated.tif
```
You can tweak a bit the application to speed ud the process: force the use of large tiles in adding e.g. ` -optim.tilesizex 2048 -optim.tilesizey 2048` to the former command line.
### Step 7. Denormalize the output image
Use **BandMathX** to denormalize the __s2_estimated.tif__ generated image.
You can also use **BandMathX** to keep non polluted values of the s2t image, and replace the polluted ones with __s2_estimated.tif__.
# Licence
Please see the license for legal issues on the use of the software.
# Contact
You can contact Remi Cresson if you have any issues with this remote module at remi [dot] cresson [at] irstea [dot] fr
model.py 0 → 100644
+ 457
− 0
View file @ e75906c7
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
# Imports
import tensorflow as tf
import numpy as np
import sys
import time
from tricks import read_samples # You must add the path contraining tricks.py to PYTHONPATH
import argparse
tf.logging.set_verbosity(tf.logging.INFO)
parser = argparse.ArgumentParser()
# Learning dataset
parser.add_argument("--s1_t", help="s1 patches (date t)", required=False)
parser.add_argument("--s2_t_before", help="s2 patches (date t-1)", required=True)
parser.add_argument("--s2_t", help="s2 patches (date t)", required=True)
parser.add_argument("--s2_t_after", help="s2 patches (date t+1)", required=True)
# Validation dataset
parser.add_argument("--valid_s1_t", help="s1 patches for validation (date t)", required=False)
parser.add_argument("--valid_s2_t_before", help="s2 patches for validation (date t-1)", required=True)
parser.add_argument("--valid_s2_t", help="s2 patches for validation (date t)", required=True)
parser.add_argument("--valid_s2_t_after", help="s2 patches for validation (date t+1)", required=True)
parser.add_argument("--save_ckpt", help="save the checkpoint to", required=True)
parser.add_argument("--load_ckpt", help="load an existing checkpoint from")
parser.add_argument("--logdir", help="Output logs directory (for tensorboard)")
# Parameters
parser.add_argument("--epochs", type=int, default=200, help="number of epochs")
parser.add_argument("--batchsize", type=int, default=16, help="batch size")
parser.add_argument("--adam_lr", type=float, default=0.0001, help="Adam learning rate")
parser.add_argument("--adam_b1", type=float, default=0.9, help="Adam beta1")
parser.add_argument("--droprate", type=float, default=0.5, help="Dropout rate")
parser.add_argument("--depth", type=int, default=64, help="deep net depth")
parser.add_argument("--combine", type=str, default="concat", help="deep net combination mode (available: sum, concat)")
parser.add_argument("--weight_loss_l1", type=float, default=0.0, help="Weight for L1 loss")
parser.add_argument("--weight_loss_l2", type=float, default=1.0, help="Weight for L2 loss")
params = parser.parse_args()
# Number of channels are global constants
s1_nbc = 2 # Sentinel-1
s2_nbc = 4 # Sentinel-2
###############################################################################
# Batch normalization layer
###############################################################################
def norm_layer(x):
return tf.layers.batch_normalization(x)
###############################################################################
# Convolution layer
###############################################################################
def conv_layer(x, nfilters, strides, kernel_size, activ, norm, padding="same" ):
conv1 = tf.layers.conv2d(
inputs=x,
filters=nfilters,
strides=(strides,strides),
kernel_size=[kernel_size, kernel_size],
padding=padding,
kernel_initializer=tf.random_normal_initializer(0, 0.02))
if (norm):
conv1 = norm_layer(conv1)
return activ(conv1)
###############################################################################
# Dropout layer
###############################################################################
def drop_layer(x, is_training, drop_rate) :
return tf.layers.dropout(x, rate=drop_rate, training=is_training)
###############################################################################
# Transposed convolution layer
###############################################################################
def deconv_layer(x, nfilters, strides, kernel_size, activ, norm):
# Perform the transpose convolution
deconv1 = tf.layers.conv2d_transpose(
inputs=x,
filters=nfilters,
strides=(strides,strides),
kernel_size=[kernel_size, kernel_size],
padding="same",
kernel_initializer=tf.random_normal_initializer(0, 0.02))
if (norm):
deconv1 = norm_layer(deconv1)
return activ(deconv1)
###############################################################################
# Extract the HW center of a 4-D tensor BHWC
###############################################################################
def extractCenter(x, pad, pad2=0):
if (pad2==0):
return x[:,pad:-pad,pad:-pad,:]
else:
return x[:,pad:-pad2,pad:-pad2,:]
###############################################################################
# Feature combination
###############################################################################
def combine (tList, method="sum"):
if (method == "sum"):
return tf.add_n(tList)
if (method == "concat") :
return tf.concat(tList, axis=3)
###############################################################################
# The model
###############################################################################
def model(s1, s2_before, s2_after, is_training, drop_rate, depth = 64, skip_method = "sum"):
with tf.variable_scope("model"):
# Encoder block
###########################################################################
def _encoder_block(x, nfilters, strides=2, kernel_size=4, activ=tf.nn.leaky_relu, norm=True, use_dropout=True):
conv = conv_layer(x, nfilters=nfilters, strides=strides, kernel_size=kernel_size, activ=activ, norm=norm)
if use_dropout: conv = drop_layer(conv, is_training=is_training, drop_rate=drop_rate)
return conv
###########################################################################
# S1 ENCODING
if (s1 is not None):
s1_conv1 = _encoder_block(s1, nfilters=depth, norm=False, use_dropout=False)
s1_conv2 = _encoder_block(s1_conv1, nfilters=depth*2, use_dropout=False)
s1_conv3 = _encoder_block(s1_conv2, nfilters=depth*4, use_dropout=False)
s1_conv4 = _encoder_block(s1_conv3, nfilters=depth*8, use_dropout=False)
s1_conv5 = _encoder_block(s1_conv4, nfilters=depth*8)
s1_conv6 = _encoder_block(s1_conv5, nfilters=depth*8)
# S2B ENCODING
s2b_conv1 = _encoder_block(s2_before, nfilters=depth, norm=False, use_dropout=False)
s2b_conv2 = _encoder_block(s2b_conv1, nfilters=depth*2, use_dropout=False)
s2b_conv3 = _encoder_block(s2b_conv2, nfilters=depth*4, use_dropout=False)
s2b_conv4 = _encoder_block(s2b_conv3, nfilters=depth*8, use_dropout=False)
s2b_conv5 = _encoder_block(s2b_conv4, nfilters=depth*8)
s2b_conv6 = _encoder_block(s2b_conv5, nfilters=depth*8)
# S2A ENCODING
s2a_conv1 = _encoder_block(s2_after, nfilters=depth, norm=False, use_dropout=False)
s2a_conv2 = _encoder_block(s2a_conv1, nfilters=depth*2, use_dropout=False)
s2a_conv3 = _encoder_block(s2a_conv2, nfilters=depth*4, use_dropout=False)
s2a_conv4 = _encoder_block(s2a_conv3, nfilters=depth*8, use_dropout=False)
s2a_conv5 = _encoder_block(s2a_conv4, nfilters=depth*8)
s2a_conv6 = _encoder_block(s2a_conv5, nfilters=depth*8)
# Decoder block
###########################################################################
def _decoder_block(net, s2b_conv, s2a_conv, s1_conv , nfilters, activ=tf.nn.relu, use_dropout=True):
dc_in = []
if net is not None: dc_in.append(net)
dc_in.append(s2b_conv)
dc_in.append(s2a_conv)
if (s1 is not None): dc_in.append(s1_conv)
dc = combine(dc_in, skip_method)
deconv = deconv_layer(dc, nfilters=nfilters, strides=2, kernel_size=4, activ=activ, norm=True)
if use_dropout: deconv = drop_layer(deconv, is_training=is_training, drop_rate=drop_rate)
return deconv
###########################################################################
deconv1 = _decoder_block(None, s2b_conv6, s2a_conv6, s1_conv6, nfilters=depth*8)
deconv2 = _decoder_block(deconv1, s2b_conv5, s2a_conv5, s1_conv5, nfilters=depth*8)
deconv3 = _decoder_block(deconv2, s2b_conv4, s2a_conv4, s1_conv4, nfilters=depth*4)
deconv4 = _decoder_block(deconv3, s2b_conv3, s2a_conv3, s1_conv3, nfilters=depth*2)
deconv5 = _decoder_block(deconv4, s2b_conv2, s2a_conv2, s1_conv2, nfilters=depth)
deconv6 = _decoder_block(deconv5, s2b_conv1, s2a_conv1, s1_conv1, nfilters=s2_nbc, activ=tf.nn.tanh, use_dropout=False)
estimated = tf.identity(deconv6, name="estimated")
return estimated
###############################################################################
# Load the dataset
###############################################################################
def load_dataset(s1, s2b, s2, s2a):
"""
Load a dataset from filenames and perform some checks
"""
# Import patches images into numpy arrays using read_samples() from tricks.py
# which is part of OTBTF (https://github.com/remicres/otbtf)
data_s1_t = None
if (s1 is not None):
data_s1_t = read_samples(s1)
data_s2_t_before = read_samples(s2b)
data_s2_t = read_samples(s2)
data_s2_t_after = read_samples(s2a)
# Check Number of samples
n2 = int(data_s2_t.shape[0])
n2a = int(data_s2_t_after.shape[0])
n2b = int(data_s2_t_before.shape[0])
n1 = n2
if (s1 is not None):
n1 = int(data_s1_t.shape[0])
if ((n1 != n2b) or (n1 != n2) or (n1 != n2a)):
print("Number of samples should be the same !")
sys.exit(1)
# Check patches sizes
for dim in range(1,3):
psz = data_s2_t_before.shape[dim]
if ((psz != data_s2_t.shape[dim]) or (psz != data_s2_t_after.shape[dim])):
print("S2 patches dims are inconsistent")
sys.exit(1)
return n1, data_s1_t, data_s2_t_before, data_s2_t, data_s2_t_after
###############################################################################
# Train the model
###############################################################################
def main(unused_argv):
"""
Main function.
-Parse parameters
-Train the model
"""
##################### import a dataset #####################
print("Loading dataset.")
# Import learning dataset
(n_data_train,
learning_data_s1_t,
learning_data_s2_t_before,
learning_data_s2_t,
learning_data_s2_t_after) = load_dataset(params.s1_t, params.s2_t_before, params.s2_t, params.s2_t_after)
# Import validation dataset
(n_data_valid,
validation_data_s1_t,
validation_data_s2_t_before,
validation_data_s2_t,
validation_data_s2_t_after) = load_dataset(params.valid_s1_t, params.valid_s2_t_before, params.valid_s2_t, params.valid_s2_t_after)
print("Done.")
##################### Build the graph ######################
with tf.Graph().as_default():
# placeholder for images and labels
s1 = None
if (learning_data_s1_t is not None):
s1 = tf.placeholder(tf.float32, shape=(None, None, None, s1_nbc), name="s1")
s2_before = tf.placeholder(tf.float32, shape=(None, None, None, s2_nbc), name="s2_before")
s2 = tf.placeholder(tf.float32, shape=(None, None, None, s2_nbc), name="s2")
s2_after = tf.placeholder(tf.float32, shape=(None, None, None, s2_nbc), name="s2_after")
is_training = tf.placeholder_with_default(tf.constant(False , dtype=tf.bool, shape=[]), shape=[], name="is_training")
dropout = tf.placeholder_with_default(tf.constant(params.droprate, dtype=tf.float32, shape=[]), shape=[], name="drop_rate")
# Generator
out = model(s1, s2_before, s2_after, is_training, dropout, params.depth, params.combine)
# Generator (with pad for FCN)
# Will be used in TensorflowModelServe OTB application to generate a
# seamless output image in a streamable fashion.
gen_fcn = tf.identity(extractCenter(out, pad=256), name="gen_fcn");
# l1 loss
loss_l1_batch = tf.abs(s2 - out)
loss_l1_batch_sum = tf.reduce_sum(loss_l1_batch)
loss_l1 = tf.reduce_mean(loss_l1_batch)
# l2 loss
loss_l2_batch = (tf.square(tf.subtract(s2, out)))
loss_l2_batch_sum = tf.reduce_sum(loss_l2_batch)
loss_l2 = tf.reduce_mean(loss_l2_batch)
# Total loss
w_loss_l1 = params.weight_loss_l1
w_loss_l2 = params.weight_loss_l2
loss_tot_batch_sum = loss_l1_batch_sum * w_loss_l1 + loss_l2_batch_sum * w_loss_l2
loss_tot = loss_l2 * w_loss_l2 +loss_l1 * w_loss_l1
# Losses accumulators
loss_acc = tf.Variable(0, trainable=False, dtype=tf.float32)
loss_l1_acc = tf.Variable(0, trainable=False, dtype=tf.float32)
loss_l2_acc = tf.Variable(0, trainable=False, dtype=tf.float32)
# Accumulators updates
losses_acc = tf.group(tf.assign(loss_acc, loss_acc + loss_tot_batch_sum),
tf.assign(loss_l1_acc, loss_l1_acc + loss_l1_batch_sum),
tf.assign(loss_l2_acc, loss_l2_acc + loss_l2_batch_sum))
# Accumulators reset
raz_losses = tf.group(tf.assign(loss_acc, 0),
tf.assign(loss_l1_acc, 0),
tf.assign(loss_l2_acc, 0))
# Global losses
global_loss = tf.Variable(0, name='global_loss', trainable=False, dtype=tf.float32)
global_loss_l1 = tf.Variable(0, name='global_loss_l1', trainable=False, dtype=tf.float32)
global_loss_l2 = tf.Variable(0, name='global_loss_l2', trainable=False, dtype=tf.float32)
global_loss_valid = tf.Variable(0, name='global_loss_valid', trainable=False, dtype=tf.float32)
global_loss_valid_l1 = tf.Variable(0, name='global_loss_valid_l1', trainable=False, dtype=tf.float32)
global_loss_valid_l2 = tf.Variable(0, name='global_loss_valid_l2', trainable=False, dtype=tf.float32)
# Reduce accumulators to compute global losses
update_global_losses_train = tf.group(tf.assign(global_loss, loss_acc / n_data_train),
tf.assign(global_loss_l1, loss_l1_acc / n_data_train),
tf.assign(global_loss_l2, loss_l2_acc / n_data_train))
update_global_losses_valid = tf.group(tf.assign(global_loss_valid, loss_acc / n_data_valid),
tf.assign(global_loss_valid_l1, loss_l1_acc / n_data_valid),
tf.assign(global_loss_valid_l2, loss_l2_acc / n_data_valid))
# Global losses summaries
tf.summary.scalar('loss', global_loss, collections=['per_epoch'])
tf.summary.scalar('loss_l1', global_loss_l1, collections=['per_epoch'])
tf.summary.scalar('loss_l2', global_loss_l2, collections=['per_epoch'])
tf.summary.scalar('loss_valid', global_loss_valid, collections=['per_epoch'])
tf.summary.scalar('loss_l1_valid', global_loss_valid_l1, collections=['per_epoch'])
tf.summary.scalar('loss_l2_valid', global_loss_valid_l2, collections=['per_epoch'])
# Optimizer
adam_optimizer = tf.train.AdamOptimizer(learning_rate=params.adam_lr, beta1=params.adam_b1)
gen_optim = adam_optimizer.minimize(loss_tot)
# Train op
train = tf.group(gen_optim, losses_acc, name="train_op")
# Merge summaries
merged_pe = tf.summary.merge_all(key='per_epoch')
############### Variable initializer Op ##################
init = tf.global_variables_initializer()
######################### Saver ##########################
saver = tf.train.Saver( max_to_keep=1)
#################### Create a session ####################
sess = tf.Session()
# Writer
train_writer = tf.summary.FileWriter(params.logdir, sess.graph)
sess.run(init)
if (params.load_ckpt != None):
saver.restore(sess, params.load_ckpt)
sequence = np.arange(n_data_train)
for curr_epoch in range(params.epochs):
print("Epoch #" + str(curr_epoch))
############## Here we start the training ################
# Start the training loop.
n_steps = int(n_data_train / params.batchsize)
if (n_data_train % params.batchsize != 0):
n_steps = n_steps + 1
for step in range(n_steps):
start_time = time.time()
# Batch start and end
start_idx = params.batchsize * step
if (n_data_train % params.batchsize != 0 and step == n_steps):
end_idx = n_data_train
else:
end_idx = start_idx + params.batchsize
# Shuffle the training dataset
np.random.shuffle(sequence)
indices = sequence[start_idx:end_idx]
# Feed dictionary
feed_dict = {
s2_before: learning_data_s2_t_before[indices,:],
s2: learning_data_s2_t [indices,:],
s2_after: learning_data_s2_t_after [indices,:],
is_training: True,
}
if (learning_data_s1_t is not None):
feed_dict.update({s1: learning_data_s1_t[indices,:]})
# Run the session for training
_, e_loss_value, e_loss_l1_value, e_loss_l2_value = sess.run([train,
loss_tot,
loss_l1,
loss_l2],
feed_dict=feed_dict)
duration = time.time() - start_time
# Print an overview
if step % 10 == 0:
print("[Training step {}] Losses: Tot={:.4f}, l1={:.4f}, l2={:.4f}, in {:.3f}s".format(step, e_loss_value, e_loss_l1_value, e_loss_l2_value, duration))
if (curr_epoch % 1 == 0):
# Save model variables
saver.save(sess, params.save_ckpt, global_step=curr_epoch)
# Update global losses
sess_targets = [update_global_losses_train, merged_pe]
_, summary = sess.run(sess_targets)
# Reset losses accumulators
_ = sess.run([raz_losses])
train_writer.add_summary(summary, curr_epoch)
# Compute losses on validation dataset
n_steps = int(n_data_valid / params.batchsize)
if (n_data_valid % params.batchsize != 0):
n_steps = n_steps + 1
for step in range(n_steps):
start_idx = params.batchsize * step
if (n_data_valid % params.batchsize != 0 and step == n_steps):
end_idx = n_data_valid
else:
end_idx = start_idx + params.batchsize
# Feed dictionary
feed_dict = {
s2_before: validation_data_s2_t_before[start_idx:end_idx,:],
s2: validation_data_s2_t [start_idx:end_idx,:],
s2_after: validation_data_s2_t_after [start_idx:end_idx,:],
is_training: False,
}
if (validation_data_s1_t is not None):
feed_dict.update({s1: validation_data_s1_t[start_idx:end_idx,:]})
# Run the session to compute losses on validation dataset
targets = [loss_tot, loss_l1, loss_l2, losses_acc]
vloss_tot, vloss_l1, vloss_l2, _ = sess.run(targets, feed_dict=feed_dict)
print("[Validation step {}] Losses: Tot={:.4f}, l1={:.4f}, l2={:.4f}".format(step, vloss_tot, vloss_l1, vloss_l2))
# Update the losses on validation dataset
_, summary = sess.run([update_global_losses_valid, merged_pe])
train_writer.add_summary(summary, curr_epoch)
# Reset losses accumulators
sess.run(raz_losses)
quit()
if __name__ == "__main__":
tf.add_check_numerics_ops()
tf.app.run(main)
schema.png 0 → 100644
+ 0
− 0
View file @ e75906c7
schema.png

149 KB

Supports Markdown
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment