Changes for the Masters Thesis

.devcontainer/Dockerfile

@@ -19,6 +19,7 @@ RUN apt-get -y install --no-install-recommends libgl1-mesa-glx
 
 # Runtime dependencies
 COPY ./requirements/requirements_docker.txt /root/requirements_docker.txt
+RUN pip install Cython
 RUN pip install -r /root/requirements_docker.txt
 
 # Develop dependencies

.devcontainer/devcontainer.json

@@ -6,11 +6,18 @@
  },
  // Set *default* container specific settings.json values on container create.
  "settings": {
- "terminal.integrated.shell.linux": "/usr/bin/zsh",
+ "terminal.integrated.defaultProfile.linux": "zsh",
+ "terminal.integrated.profiles.linux": {
+ "zsh": {
+ "path": "/usr/bin/zsh",
+ "args": []
+ }
+ },
  "editor.rulers": [100],
  // ---- Python specific settings
+ "python.experiments.enabled": false,
  "python.pythonPath": "/usr/local/bin/python",
- "python.formatting.provider": "autopep8",
+ "python.formatting.provider": "black",
  "python.linting.pylintEnabled": false,
  "jupyter.alwaysTrustNotebooks": true,
  "python.languageServer": "Pylance",
@@ -21,7 +28,8 @@
  // ---- Python specific extensions
  "ms-python.python",
  "kevinrose.vsc-python-indent",
- "ms-python.vscode-pylance"
+ "ms-python.vscode-pylance",
+ "njpwerner.autodocstring"
  ],
  // ---- Run with GPU driver
  "runArgs": [

README.md

@@ -3,13 +3,14 @@
 This repository implements two algorithms for image reconstruction using a Gaussian denoiser in TensorFlow 2.
 
 Algorithms:
-* **DMSP:** S. A. Bigdeli, P. Favaro, M. Jin, and M. Zwicker, [“Deep Mean-Shift Priors for Image Restoration,”](http://papers.nips.cc/paper/6678-deep-mean-shift-priors-for-image-restoration.pdf) in Advances in Neural Information Processing Systems 30, 2017, pp. 763–772.
-* **HQS:** K. Zhang, Y. Li, W. Zuo, L. Zhang, L. Van Gool, and R. Timofte, [“Plug-and-Play Image Restoration with Deep Denoiser Prior,”](http://arxiv.org/abs/2008.13751) arXiv:2008.13751 [cs, eess], Aug. 2020
 
+- **DMSP:** S. A. Bigdeli, P. Favaro, M. Jin, and M. Zwicker, [“Deep Mean-Shift Priors for Image Restoration,”](http://papers.nips.cc/paper/6678-deep-mean-shift-priors-for-image-restoration.pdf) in Advances in Neural Information Processing Systems 30, 2017, pp. 763–772.
+- **HQS:** K. Zhang, Y. Li, W. Zuo, L. Zhang, L. Van Gool, and R. Timofte, [“Plug-and-Play Image Restoration with Deep Denoiser Prior,”](http://arxiv.org/abs/2008.13751) IEEE Trans. Pattern Anal. Mach. Intell., pp. 1–1, 2021
 
 ## Getting Started
 
 Install using pip
+
 ```
 $ pip install git+https://github.com/HedgehogCode/deep-plug-and-play-prior
 ```
@@ -21,10 +22,12 @@ See the example [notebooks](notebooks/) for usage examples.
 ## Related Repositories
 
 DMSP Implementations (significantly slower):
-* Matlab: https://github.com/siavashBigdeli/DMSP (only deblur)
-* TensorFlow 1: https://github.com/siavashBigdeli/DMSP-tensorflow (only deblur)
-* TensorFlow 2: https://github.com/siavashBigdeli/DMSP-TF2 (deblur and sr (but not bicubic))
+
+- Matlab: https://github.com/siavashBigdeli/DMSP (only deblur)
+- TensorFlow 1: https://github.com/siavashBigdeli/DMSP-tensorflow (only deblur)
+- TensorFlow 2: https://github.com/siavashBigdeli/DMSP-TF2 (deblur and sr (but not bicubic))
 
 HQS:
-* Matlab: https://github.com/cszn/IRCNN
-* Pytorch: https://github.com/cszn/DPIR
+
+- Matlab: https://github.com/cszn/IRCNN
+- Pytorch: https://github.com/cszn/DPIR

dppp/__init__.py

@@ -1,5 +1,16 @@
 from .utils import *
 from .callback import *
 from .metrics import *
-from .dmsp import dmsp_deblur_nb, dmsp_super_resolve, dmsp_inpaint
-from .hqs import hqs_deblur_nb, hqs_super_resolve, hqs_inpaint
+from .dmsp import (
+ dmsp_deblur_nb,
+ dmsp_deblur_blind,
+ dmsp_super_resolve,
+ dmsp_inpaint,
+)
+from .hqs import (
+ hqs_deblur_nb,
+ hqs_super_resolve,
+ hqs_inpaint,
+ hqs_video_super_resolve,
+ hqs_lf_super_resolve,
+)

dppp/callback.py

@@ -11,10 +11,8 @@
 # TENSORBOARD
 ###############################################################################
 
-def callback_tb_image(tb_log_dir: str,
- summary_name: str,
- frequency: int
- ) -> Callback:
+
+def callback_tb_image(tb_log_dir: str, summary_name: str, frequency: int) -> Callback:
  tb_writer = tf.summary.create_file_writer(tb_log_dir)
 
  def call(x: TensorLike, step: TensorLike):
@@ -25,10 +23,9 @@ def call(x: TensorLike, step: TensorLike):
  return call
 
 
-def callback_tb_scalar(tb_log_dir: str,
- summary_name: str,
- summary_fn: Callable[[TensorLike], TensorLike]
- ) -> Callback:
+def callback_tb_scalar(
+ tb_log_dir: str, summary_name: str, summary_fn: Callable[[TensorLike], TensorLike]
+) -> Callback:
  tb_writer = tf.summary.create_file_writer(tb_log_dir)
 
  def call(x: TensorLike, step: TensorLike):
@@ -38,67 +35,87 @@ def call(x: TensorLike, step: TensorLike):
  return call
 
 
-def callback_tb_psnr(tb_log_dir: str, summary_name: str, gt_image: TensorLike) -> Callback:
- return callback_tb_scalar(tb_log_dir=tb_log_dir, summary_name=summary_name,
- summary_fn=lambda x: tf.reduce_mean(dppp.psnr(x, gt_image)))
+def callback_tb_psnr(
+ tb_log_dir: str, summary_name: str, gt_image: TensorLike
+) -> Callback:
+ return callback_tb_scalar(
+ tb_log_dir=tb_log_dir,
+ summary_name=summary_name,
+ summary_fn=lambda x: tf.reduce_mean(dppp.psnr(x, gt_image)),
+ )
 
 
-def callback_tb_ssim(tb_log_dir: str, summary_name: str, gt_image: TensorLike) -> Callback:
- return callback_tb_scalar(tb_log_dir=tb_log_dir, summary_name=summary_name,
- summary_fn=lambda x: tf.reduce_mean(dppp.ssim(x, gt_image)))
+def callback_tb_ssim(
+ tb_log_dir: str, summary_name: str, gt_image: TensorLike
+) -> Callback:
+ return callback_tb_scalar(
+ tb_log_dir=tb_log_dir,
+ summary_name=summary_name,
+ summary_fn=lambda x: tf.reduce_mean(dppp.ssim(x, gt_image)),
+ )
 
 
 ###############################################################################
 # TF.PRINT
 ###############################################################################
 
-def callback_print_scalar(name: str,
- frequency: int,
-   summary_fn: Callable[[TensorLike], TensorLike]
- ) -> Callback:
+
+def callback_print_scalar(
+ name: str, frequency: int, summary_fn: Callable[[TensorLike], TensorLike]
+) -> Callback:
  def call(x: TensorLike, step: TensorLike):
  if step % frequency == 0: # type: ignore
- tf.print('Step:', step, ', ', name, ':', summary_fn(x))
+ tf.print("Step:", step, ", ", name, ":", summary_fn(x))
 
  return call
 
 
 def callback_print_psnr(name: str, frequency: int, gt_image: TensorLike) -> Callback:
- return callback_print_scalar(name=name, frequency=frequency,
- summary_fn=lambda x: tf.reduce_mean(dppp.psnr(x, gt_image)))
+ return callback_print_scalar(
+ name=name,
+ frequency=frequency,
+ summary_fn=lambda x: tf.reduce_mean(dppp.psnr(x, gt_image)),
+ )
 
 
 def callback_print_ssim(name: str, frequency: int, gt_image: TensorLike) -> Callback:
- return callback_print_scalar(name=name, frequency=frequency,
- summary_fn=lambda x: tf.reduce_mean(dppp.ssim(x, gt_image)))
+ return callback_print_scalar(
+ name=name,
+ frequency=frequency,
+ summary_fn=lambda x: tf.reduce_mean(dppp.ssim(x, gt_image)),
+ )
 
 
 ###############################################################################
 # CSV
 ###############################################################################
 
-def callback_csv_metric(metrics: Dict[str, Tuple[MetricFnType, int]],
- num_steps: int): # TODO typing -> Tuple[Callable[[tf.Tensor, tf.Tensor], None], tf.Variable]:
+
+def callback_csv_metric(metrics: Dict[str, Tuple[MetricFnType, int]], num_steps: int):
  metric_names = list(metrics.keys())
- csv = tf.Variable('Step,Image_Id,' + ','.join(metric_names) + '\n')
+ csv = tf.Variable("Step,Image_Id," + ",".join(metric_names) + "\n")
 
  def for_image(image_id: str, gt_image: TensorLike):
-
- @tf.function(input_signature=(tf.TensorSpec(shape=[None, None, None, None], dtype=tf.float32),
- tf.TensorSpec(shape=[], dtype=tf.int64)))
+ @tf.function(
+ input_signature=(
+ tf.TensorSpec(shape=[None, None, None, None], dtype=tf.float32),
+ tf.TensorSpec(shape=[], dtype=tf.int64),
+ )
+ )
  def call(x: TensorLike, step: TensorLike):
  def val_for_metric(m):
  metric_fn, per_step = metrics[m]
  if step == num_steps - 1 or (step + 1) % per_step == 0: # type: ignore
  val = metric_fn(x, gt_image)[0] # type: ignore
  else:
- val = tf.constant(float('nan'))
+ val = tf.constant(float("nan"))
  return tf.strings.as_string(val)
 
  values = [val_for_metric(m) for m in metric_names]
- line = tf.strings.join([tf.strings.as_string(
- step), image_id, *values], separator=',')
- csv.assign(tf.strings.join([csv, line, '\n']))
+ line = tf.strings.join(
+ [tf.strings.as_string(step), image_id, *values], separator=","
+ )
+ csv.assign(tf.strings.join([csv, line, "\n"]))
 
  return call