Skip to content
Snippets Groups Projects
Commit 97e37b2c authored by Rony Abecidan's avatar Rony Abecidan
Browse files

[Add] : Add hyperparameters and results

parent 984bd0a4
No related branches found
No related tags found
No related merge requests found
Showing
with 411 additions and 18 deletions
......@@ -4,7 +4,7 @@
target : qf(5)
--- DATA ---
source : 30000 images sampled from the "Splicing" category of the database DEFACTO
source : random half images sampled from the "Splicing" category of the database DEFACTO
repartition : 1-1/3 1/3
A 3-fold cutting is applied to the source to split it into 3 train and test sets
......@@ -18,7 +18,7 @@
patches.
-We kept only 2 patches for each class at maximum by image
target : 30000 images different of the source and potentially presenting a different preprocessing compared to the source
target : other half of the images from the "Splicing" category of the database DEFACTO and potentially presenting a different preprocessing compared to the source
(for instance a change in the quality factor for the compression).
repartition : 1-1/3 1/3
......@@ -30,9 +30,10 @@
--- RESULTS ---
qf(5) : 81.89999999999999% +/- 0.0%
qf(10) : 83.39999999999999% +/- 0.0%
qf(20) : 84.6% +/- 0.0%
qf(5) : 82.1% +/- 0.0%
qf(10) : 83.6% +/- 1.0%
qf(20) : 84.7% +/- 1.0%
qf(50) : 85.39999999999999% +/- 0.0%
qf(100) : 86.1% +/- 0.0%
none : 86.9% +/- 0.0%
qf(100) : 86.2% +/- 0.0%
none : 87.1% +/- 0.0%
{"seed": 2021, "N_fold": 3, "im_size": 128, "max_epochs": 30, "earlystop_patience": 5, "lr": 0.0001, "train_batch_size": 128, "eval_batch_size": 512, "detector_name": "Bayar", "source_path": "source-none.hdf5", "target_path": "target-qf(5).hdf5", "source_name": "none", "target_name": "qf(5)", "setup": "Mix", "domain_paths": ["target-qf(5).hdf5", "target-qf(10).hdf5", "target-qf(20).hdf5", "target-qf(50).hdf5", "target-qf(100).hdf5", "target-none.hdf5"], "domain_names": ["qf(5)", "qf(10)", "qf(20)", "qf(50)", "qf(100)", "none"], "nb_source_max": 100000000, "nb_target_max": 100000000, "save_at_each_epoch": true, "precisions": "s=none_t=qf(5)"}
\ No newline at end of file
{"seed": 2021, "N_fold": 3, "im_size": 128, "max_epochs": 30, "earlystop_patience": 5, "lr": 0.0001, "train_batch_size": 128, "eval_batch_size": 512, "detector_name": "Bayar", "source_path": "source-none.hdf5", "target_path": "target-qf(5).hdf5", "source_name": "none", "target_name": "qf(5)", "setup": "Mix", "domain_paths": ["target-qf(5).hdf5", "target-qf(10).hdf5", "target-qf(20).hdf5", "target-qf(50).hdf5", "target-qf(100).hdf5", "target-none.hdf5"], "domain_names": ["qf(5)", "qf(10)", "qf(20)", "qf(50)", "qf(100)", "none"], "nb_source_max": 100000000, "nb_target_max": 100000000, "save_at_each_epoch": false, "precisions": "s=none_t=qf(5)"}
\ No newline at end of file
=== FORGERY DETECTION TASK ===
source : qf(100)
target : qf(5)
--- DATA ---
source : random half images sampled from the "Splicing" category of the database DEFACTO
repartition : 1-1/3 1/3
A 3-fold cutting is applied to the source to split it into 3 train and test sets
Then, the images in each cuts are transformed into batches of 128x128 patches.
**In each set, there is a perfect balance between forged and non-forged patches : **
-A patch associated to a forged region is kept in the sets only if the forged region occupy a space between 20% and 80%
of the total space (128x128).
-The real patches are chosen randomly so that there is an equal amount of forged and non-forged
patches.
-We kept only 2 patches for each class at maximum by image
target : other half of the images from the "Splicing" category of the database DEFACTO and potentially presenting a different preprocessing compared to the source
(for instance a change in the quality factor for the compression).
repartition : 1-1/3 1/3
The preprocessing of the target images is the same as the one presented above
--- TRAINING ---
trainings_epochs on each fold : 30
hyperparameters_file : hyperparameters-Mix-s=qf(100)_t=qf(5).txt
--- RESULTS ---
qf(5) : 81.69999999999999% +/- 0.0%
qf(10) : 83.6% +/- 0.0%
qf(20) : 84.7% +/- 0.0%
qf(50) : 85.5% +/- 0.0%
qf(100) : 86.4% +/- 0.0%
none : 86.5% +/- 0.0%
......@@ -4,7 +4,7 @@
target : qf(5)
--- DATA ---
source : 30000 images sampled from the "Splicing" category of the database DEFACTO
source : random half images sampled from the "Splicing" category of the database DEFACTO
repartition : 1-1/3 1/3
A 3-fold cutting is applied to the source to split it into 3 train and test sets
......@@ -18,7 +18,7 @@
patches.
-We kept only 2 patches for each class at maximum by image
target : 30000 images different of the source and potentially presenting a different preprocessing compared to the source
target : other half of the images from the "Splicing" category of the database DEFACTO and potentially presenting a different preprocessing compared to the source
(for instance a change in the quality factor for the compression).
repartition : 1-1/3 1/3
......
=== FORGERY DETECTION TASK ===
source : qf(100)
target : qf(5)
--- DATA ---
source : random half images sampled from the "Splicing" category of the database DEFACTO
repartition : 1-1/3 1/3
A 3-fold cutting is applied to the source to split it into 3 train and test sets
Then, the images in each cuts are transformed into batches of 128x128 patches.
**In each set, there is a perfect balance between forged and non-forged patches : **
-A patch associated to a forged region is kept in the sets only if the forged region occupy a space between 20% and 80%
of the total space (128x128).
-The real patches are chosen randomly so that there is an equal amount of forged and non-forged
patches.
-We kept only 2 patches for each class at maximum by image
target : other half of the images from the "Splicing" category of the database DEFACTO and potentially presenting a different preprocessing compared to the source
(for instance a change in the quality factor for the compression).
repartition : 1-1/3 1/3
The preprocessing of the target images is the same as the one presented above
--- TRAINING ---
trainings_epochs on each fold : 30
hyperparameters_file : hyperparameters-SrcOnly-s=qf(100)_t=qf(5).txt
--- RESULTS ---
qf(5) : 66.8% +/- 3.0%
qf(10) : 73.7% +/- 3.0%
qf(20) : 81.2% +/- 1.0%
qf(50) : 86.2% +/- 1.0%
qf(100) : 88.4% +/- 0.0%
none : 88.5% +/- 0.0%
......@@ -4,7 +4,7 @@
target : qf(5)
--- DATA ---
source : 30000 images sampled from the "Splicing" category of the database DEFACTO
source : random half images sampled from the "Splicing" category of the database DEFACTO
repartition : 1-1/3 1/3
A 3-fold cutting is applied to the source to split it into 3 train and test sets
......@@ -18,7 +18,7 @@
patches.
-We kept only 2 patches for each class at maximum by image
target : 30000 images different of the source and potentially presenting a different preprocessing compared to the source
target : other half of the images from the "Splicing" category of the database DEFACTO and potentially presenting a different preprocessing compared to the source
(for instance a change in the quality factor for the compression).
repartition : 1-1/3 1/3
......
......@@ -4,7 +4,7 @@
target : qf(5)
--- DATA ---
source : 30000 images sampled from the "Splicing" category of the database DEFACTO
source : random half images sampled from the "Splicing" category of the database DEFACTO
repartition : 1-1/3 1/3
A 3-fold cutting is applied to the source to split it into 3 train and test sets
......@@ -18,7 +18,7 @@
patches.
-We kept only 2 patches for each class at maximum by image
target : 30000 images different of the source and potentially presenting a different preprocessing compared to the source
target : other half of the images from the "Splicing" category of the database DEFACTO and potentially presenting a different preprocessing compared to the source
(for instance a change in the quality factor for the compression).
repartition : 1-1/3 1/3
......
=== FORGERY DETECTION TASK ===
source : none
target : qf(5)
--- DATA ---
source : random half images sampled from the "Splicing" category of the database DEFACTO
repartition : 1-1/3 1/3
A 3-fold cutting is applied to the source to split it into 3 train and test sets
Then, the images in each cuts are transformed into batches of 128x128 patches.
**In each set, there is a perfect balance between forged and non-forged patches : **
-A patch associated to a forged region is kept in the sets only if the forged region occupy a space between 20% and 80%
of the total space (128x128).
-The real patches are chosen randomly so that there is an equal amount of forged and non-forged
patches.
-We kept only 2 patches for each class at maximum by image
target : other half of the images from the "Splicing" category of the database DEFACTO and potentially presenting a different preprocessing compared to the source
(for instance a change in the quality factor for the compression).
repartition : 1-1/3 1/3
The preprocessing of the target images is the same as the one presented above
--- TRAINING ---
trainings_epochs on each fold : 30
hyperparameters_file : hyperparameters-Update-s=none_t=qf(5).txt
--- RESULTS ---
qf(5) : 76.9% +/- 0.0%
qf(10) : 81.0% +/- 1.0%
qf(20) : 82.69999999999999% +/- 1.0%
qf(50) : 83.2% +/- 1.0%
qf(100) : 83.0% +/- 1.0%
none : 88.7% +/- 0.0%
=== FORGERY DETECTION TASK ===
source : none
target : qf(5)
--- DATA ---
source : random half images sampled from the "Splicing" category of the database DEFACTO
repartition : 1-1/3 1/3
A 3-fold cutting is applied to the source to split it into 3 train and test sets
Then, the images in each cuts are transformed into batches of 128x128 patches.
**In each set, there is a perfect balance between forged and non-forged patches : **
-A patch associated to a forged region is kept in the sets only if the forged region occupy a space between 20% and 80%
of the total space (128x128).
-The real patches are chosen randomly so that there is an equal amount of forged and non-forged
patches.
-We kept only 2 patches for each class at maximum by image
target : other half of the images from the "Splicing" category of the database DEFACTO and potentially presenting a different preprocessing compared to the source
(for instance a change in the quality factor for the compression).
repartition : 1-1/3 1/3
The preprocessing of the target images is the same as the one presented above
--- TRAINING ---
trainings_epochs on each fold : 30
hyperparameters_file : hyperparameters-Update-s=none_t=qf(5)_N_t=10.txt
--- RESULTS ---
qf(5) : 73.6% +/- 3.0%
qf(10) : 79.2% +/- 2.0%
qf(20) : 82.5% +/- 1.0%
qf(50) : 84.0% +/- 0.0%
qf(100) : 84.39999999999999% +/- 0.0%
none : 85.6% +/- 1.0%
=== FORGERY DETECTION TASK ===
source : none
target : qf(5)
--- DATA ---
source : random half images sampled from the "Splicing" category of the database DEFACTO
repartition : 1-1/3 1/3
A 3-fold cutting is applied to the source to split it into 3 train and test sets
Then, the images in each cuts are transformed into batches of 128x128 patches.
**In each set, there is a perfect balance between forged and non-forged patches : **
-A patch associated to a forged region is kept in the sets only if the forged region occupy a space between 20% and 80%
of the total space (128x128).
-The real patches are chosen randomly so that there is an equal amount of forged and non-forged
patches.
-We kept only 2 patches for each class at maximum by image
target : other half of the images from the "Splicing" category of the database DEFACTO and potentially presenting a different preprocessing compared to the source
(for instance a change in the quality factor for the compression).
repartition : 1-1/3 1/3
The preprocessing of the target images is the same as the one presented above
--- TRAINING ---
trainings_epochs on each fold : 30
hyperparameters_file : hyperparameters-Update-s=none_t=qf(5)_N_t=100.txt
--- RESULTS ---
qf(5) : 72.39999999999999% +/- 1.0%
qf(10) : 80.10000000000001% +/- 0.0%
qf(20) : 83.89999999999999% +/- 0.0%
qf(50) : 85.0% +/- 1.0%
qf(100) : 85.39999999999999% +/- 1.0%
none : 89.2% +/- 0.0%
=== FORGERY DETECTION TASK ===
source : none
target : qf(5)
--- DATA ---
source : random half images sampled from the "Splicing" category of the database DEFACTO
repartition : 1-1/3 1/3
A 3-fold cutting is applied to the source to split it into 3 train and test sets
Then, the images in each cuts are transformed into batches of 128x128 patches.
**In each set, there is a perfect balance between forged and non-forged patches : **
-A patch associated to a forged region is kept in the sets only if the forged region occupy a space between 20% and 80%
of the total space (128x128).
-The real patches are chosen randomly so that there is an equal amount of forged and non-forged
patches.
-We kept only 2 patches for each class at maximum by image
target : other half of the images from the "Splicing" category of the database DEFACTO and potentially presenting a different preprocessing compared to the source
(for instance a change in the quality factor for the compression).
repartition : 1-1/3 1/3
The preprocessing of the target images is the same as the one presented above
--- TRAINING ---
trainings_epochs on each fold : 30
hyperparameters_file : hyperparameters-Update-s=none_t=qf(5)_N_t=1000.txt
--- RESULTS ---
qf(5) : 74.8% +/- 1.0%
qf(10) : 80.7% +/- 1.0%
qf(20) : 83.6% +/- 0.0%
qf(50) : 84.89999999999999% +/- 0.0%
qf(100) : 85.2% +/- 1.0%
none : 88.4% +/- 1.0%
=== FORGERY DETECTION TASK ===
source : none
target : qf(5)
--- DATA ---
source : random half images sampled from the "Splicing" category of the database DEFACTO
repartition : 1-1/3 1/3
A 3-fold cutting is applied to the source to split it into 3 train and test sets
Then, the images in each cuts are transformed into batches of 128x128 patches.
**In each set, there is a perfect balance between forged and non-forged patches : **
-A patch associated to a forged region is kept in the sets only if the forged region occupy a space between 20% and 80%
of the total space (128x128).
-The real patches are chosen randomly so that there is an equal amount of forged and non-forged
patches.
-We kept only 2 patches for each class at maximum by image
target : other half of the images from the "Splicing" category of the database DEFACTO and potentially presenting a different preprocessing compared to the source
(for instance a change in the quality factor for the compression).
repartition : 1-1/3 1/3
The preprocessing of the target images is the same as the one presented above
--- TRAINING ---
trainings_epochs on each fold : 30
hyperparameters_file : hyperparameters-Update-s=none_t=qf(5)_sigma=0.01.txt
--- RESULTS ---
qf(5) : 56.8% +/- 1.0%
qf(10) : 56.599999999999994% +/- 1.0%
qf(20) : 56.49999999999999% +/- 0.0%
qf(50) : 56.49999999999999% +/- 1.0%
qf(100) : 56.699999999999996% +/- 2.0%
none : 56.8% +/- 2.0%
{"seed": 2021, "N_fold": 3, "im_size": 128, "max_epochs": 30, "earlystop_patience": 5, "lr": 0.0001, "train_batch_size": 128, "eval_batch_size": 512, "detector_name": "Bayar", "source_path": "source-none.hdf5", "target_path": "target-qf(5).hdf5", "source_name": "none", "target_name": "qf(5)", "setup": "Update", "domain_paths": ["target-qf(5).hdf5", "target-qf(10).hdf5", "target-qf(20).hdf5", "target-qf(50).hdf5", "target-qf(100).hdf5", "target-none.hdf5"], "domain_names": ["qf(5)", "qf(10)", "qf(20)", "qf(50)", "qf(100)", "none"], "nb_source_max": 100000000, "nb_target_max": 100000000, "save_at_each_epoch": false, "precisions": "s=none_t=qf(5)_sigma=0.01", "sigmas": [0.01, 0.01, 0.01]}
\ No newline at end of file
=== FORGERY DETECTION TASK ===
source : none
target : qf(5)
--- DATA ---
source : random half images sampled from the "Splicing" category of the database DEFACTO
repartition : 1-1/3 1/3
A 3-fold cutting is applied to the source to split it into 3 train and test sets
Then, the images in each cuts are transformed into batches of 128x128 patches.
**In each set, there is a perfect balance between forged and non-forged patches : **
-A patch associated to a forged region is kept in the sets only if the forged region occupy a space between 20% and 80%
of the total space (128x128).
-The real patches are chosen randomly so that there is an equal amount of forged and non-forged
patches.
-We kept only 2 patches for each class at maximum by image
target : other half of the images from the "Splicing" category of the database DEFACTO and potentially presenting a different preprocessing compared to the source
(for instance a change in the quality factor for the compression).
repartition : 1-1/3 1/3
The preprocessing of the target images is the same as the one presented above
--- TRAINING ---
trainings_epochs on each fold : 30
hyperparameters_file : hyperparameters-Update-s=none_t=qf(5)_sigma=1000.txt
--- RESULTS ---
qf(5) : 66.2% +/- 1.0%
qf(10) : 74.1% +/- 2.0%
qf(20) : 78.7% +/- 1.0%
qf(50) : 80.2% +/- 1.0%
qf(100) : 79.7% +/- 1.0%
none : 94.0% +/- 0.0%
{"seed": 2021, "N_fold": 3, "im_size": 128, "max_epochs": 30, "earlystop_patience": 5, "lr": 0.0001, "train_batch_size": 128, "eval_batch_size": 512, "detector_name": "Bayar", "source_path": "source-none.hdf5", "target_path": "target-qf(5).hdf5", "source_name": "none", "target_name": "qf(5)", "setup": "Update", "domain_paths": ["target-qf(5).hdf5", "target-qf(10).hdf5", "target-qf(20).hdf5", "target-qf(50).hdf5", "target-qf(100).hdf5", "target-none.hdf5"], "domain_names": ["qf(5)", "qf(10)", "qf(20)", "qf(50)", "qf(100)", "none"], "nb_source_max": 100000000, "nb_target_max": 100000000, "save_at_each_epoch": false, "precisions": "s=none_t=qf(5)_sigma=1000", "sigmas": [1000, 1000, 1000]}
\ No newline at end of file
=== FORGERY DETECTION TASK ===
source : qf(100)
target : qf(5)
--- DATA ---
source : random half images sampled from the "Splicing" category of the database DEFACTO
repartition : 1-1/3 1/3
A 3-fold cutting is applied to the source to split it into 3 train and test sets
Then, the images in each cuts are transformed into batches of 128x128 patches.
**In each set, there is a perfect balance between forged and non-forged patches : **
-A patch associated to a forged region is kept in the sets only if the forged region occupy a space between 20% and 80%
of the total space (128x128).
-The real patches are chosen randomly so that there is an equal amount of forged and non-forged
patches.
-We kept only 2 patches for each class at maximum by image
target : other half of the images from the "Splicing" category of the database DEFACTO and potentially presenting a different preprocessing compared to the source
(for instance a change in the quality factor for the compression).
repartition : 1-1/3 1/3
The preprocessing of the target images is the same as the one presented above
--- TRAINING ---
trainings_epochs on each fold : 30
hyperparameters_file : hyperparameters-Update-s=qf(100)_t=qf(5).txt
--- RESULTS ---
qf(5) : 75.2% +/- 0.0%
qf(10) : 79.9% +/- 0.0%
qf(20) : 82.8% +/- 1.0%
qf(50) : 84.3% +/- 0.0%
qf(100) : 84.89999999999999% +/- 0.0%
none : 85.39999999999999% +/- 1.0%
=== FORGERY DETECTION TASK ===
source : qf(5)
target : qf(100)
--- DATA ---
source : random half images sampled from the "Splicing" category of the database DEFACTO
repartition : 1-1/3 1/3
A 3-fold cutting is applied to the source to split it into 3 train and test sets
Then, the images in each cuts are transformed into batches of 128x128 patches.
**In each set, there is a perfect balance between forged and non-forged patches : **
-A patch associated to a forged region is kept in the sets only if the forged region occupy a space between 20% and 80%
of the total space (128x128).
-The real patches are chosen randomly so that there is an equal amount of forged and non-forged
patches.
-We kept only 2 patches for each class at maximum by image
target : other half of the images from the "Splicing" category of the database DEFACTO and potentially presenting a different preprocessing compared to the source
(for instance a change in the quality factor for the compression).
repartition : 1-1/3 1/3
The preprocessing of the target images is the same as the one presented above
--- TRAINING ---
trainings_epochs on each fold : 30
hyperparameters_file : hyperparameters-Update-s=qf(5)_t=qf(100).txt
--- RESULTS ---
qf(5) : 81.5% +/- 0.0%
qf(10) : 82.19999999999999% +/- 0.0%
qf(20) : 81.6% +/- 0.0%
qf(50) : 81.3% +/- 0.0%
qf(100) : 81.0% +/- 0.0%
none : 81.3% +/- 0.0%
{"0": "SrcOnly-s=none_t=qf(5)", "1": "SrcOnly-s=qf(5)_t=qf(5)", "2": "SrcOnly-s=qf(5)_t=qf(5)_N_s=1000", "3": "SrcOnly-s=qf(100)_t=qf(5)", "4": "Update-s=none_t=qf(5)", "5": "Update-s=none_t=qf(5)_N_t=10", "6": "Update-s=none_t=qf(5)_N_t=100", "7": "Update-s=none_t=qf(5)_N_t=1000", "8": "Update-s=qf(100)_t=qf(5)", "9": "Update-s=qf(5)_t=qf(100)", "10": "Mix-s=none_t=qf(5)", "11": "Mix-s=qf(100)_t=qf(5)"}
\ No newline at end of file
{"0": "SrcOnly-s=none_t=qf(5)", "1": "SrcOnly-s=qf(5)_t=qf(5)", "2": "SrcOnly-s=qf(5)_t=qf(5)_N_s=1000", "3": "SrcOnly-s=qf(100)_t=qf(5)", "4": "Update-s=none_t=qf(5)", "5": "Update-s=none_t=qf(5)_N_t=10", "6": "Update-s=none_t=qf(5)_N_t=100", "7": "Update-s=none_t=qf(5)_N_t=1000","8": "Update-s=none_t=qf(5)_sigma=0.01","9": "Update-s=none_t=qf(5)_sigma=1000", "10": "Update-s=qf(100)_t=qf(5)", "11": "Update-s=qf(5)_t=qf(100)", "12": "Mix-s=none_t=qf(5)", "13": "Mix-s=qf(100)_t=qf(5)"}
\ No newline at end of file
......@@ -38,7 +38,7 @@ You can be at least sure to obtain the same datasets and to start the training w
| Update($`\sigma=8`$) with Source=None and Target=QF(5) using only 100 patches for the target training set | 6 |
| Update($`\sigma=8`$) with Source=None and Target=QF(5) using only 1000 patches for the target training set | 7 |
| Update($`\sigma=0.01`$) with Source=None and Target=QF(5) | 8 |
| Update($`\sigma=100`$) with Source=None and Target=QF(5) | 9 |
| Update($`\sigma=1000`$) with Source=None and Target=QF(5) | 9 |
| Update($`\sigma=8`$) with Source=QF(100) and Target=QF(5) | 10 |
| Update($`\sigma=8`$) with Source=QF(5) and Target=QF(100) | 11 |
| Mix with Source=None and Target=QF(5) | 12 |
......
0% Loading or .
You are about to add 0 people to the discussion. Proceed with caution.
Please register or to comment