Merge with main.

CHANGELOG.rst

@@ -19,13 +19,18 @@ Added
 - Authors at top of source code files.
 - Add paramiko as dependency for remote capture and display.
 - Mask module, for CodedAperture (FlatCam), PhaseContour (PhlatCam), and FresnelZoneAperture.
+- Script for measuring arbitrary dataset (from Raspberry Pi).
+- Support for preprocessing and postprocessing, such as denoising, in ``TrainableReconstructionAlgorithm``. Both trainable and fix postprocessing can be used.
+- Utilities to load a trained DruNet model for use as postprocessing in ``TrainableReconstructionAlgorithm``.
+- Support for unrolled loading and inference in the script ``admm.py``.
 
 
 Changed
 ~~~~~~~
 
 - Simpler remote capture and display scripts with Hydra.
 - Group source code into four modules: ``hardware``, ``recon``, ``utils``, ``eval``.
+- Split scripts into subfolders.
 
 
 Bugfix

CONTRIBUTING.rst

@@ -153,7 +153,9 @@ Building documentation
 
    # build documentation
    (lensless_docs) python setup.py build_sphinx
-
+   # or
+   (lensless_docs) (cd docs && make html)
+   
 To rebuild the documentation from scratch:
 
 .. code:: bash

configs/analyze_dataset.yaml

@@ -0,0 +1,8 @@
+hydra:
+  job:
+    chdir: True    # change to output folder
+
+dataset_path: null
+desired_range: [190, 254]
+delete_saturated: True
+n_files: null

configs/benchmark.yaml

@@ -5,14 +5,15 @@ hydra:
   job:
     chdir: True
 
+device: "cuda"
 # numbers of iterations to benchmark
 n_iter_range: [5, 10, 30, 60, 100, 200, 300]
 # number of files to benchmark
 n_files: 200
 #How much should the image be downsampled
 downsample: 8
 #algorithm to benchmark
-algorithms: ["ADMM", "ADMM_Monakhova2019", "FISTA", "GradientDescent", "NesterovGradientDescent"]
+algorithms: ["ADMM", "ADMM_Monakhova2019", "FISTA"] #["ADMM", "ADMM_Monakhova2019", "FISTA", "GradientDescent", "NesterovGradientDescent"]
 
 # Hyperparameters
 nesterov:

configs/collect_dataset.yaml

@@ -1,3 +1,5 @@
+# python scripts/collect_dataset_on_device.py -cn collect_dataset
+
 input_dir: data/celeba_mini
 input_file_ext: jpg
 
@@ -16,6 +18,11 @@ start_delay: null   # how much time (in minutes) to wait before starting
 # measurement parameters
 dummy: False    # simply copy file to output folder for testing
 
+# ideal image levels
+max_level: 254
+min_level: 200
+max_tries: 6
+
 # -- display parameters
 display:
   output_fp: "~/LenslessPiCam_display/test.png"
@@ -24,15 +31,15 @@ display:
   pad: 0
   hshift: 0
   vshift: -10
-  brightness: 50
+  brightness: 80   # max brightness
   rot90: 3
 
 capture:
   delay: 2    # to allow picture to display
   config_pause: 2
   iso: 100
   res: null
-  down: 8
-  exposure: 0.02
+  down: 4
+  exposure: 0.02    # min exposure
   awb_gains: [1.9, 1.2]  # red, blue
   # awb_gains: null

configs/defaults_recon.yaml

@@ -61,6 +61,15 @@ admm:
   mu2: 1e-5
   mu3: 4e-5
   tau: 0.0001
+  #Loading unrolled model
+  unrolled: false
+  checkpoint_fp: null
+  pre_process_model: 
+    network : null  # UnetRes or DruNet or null
+    depth : 2 # depth of each up/downsampling layer. Ignore if network is DruNet
+  post_process_model: 
+    network : null  # UnetRes or DruNet or null
+    depth : 2 # depth of each up/downsampling layer. Ignore if network is DruNet
 
 apgd:
   # Stopping criteria

configs/demo.yaml

@@ -3,7 +3,7 @@
 #     chdir: True    # change to output folder
 
 
-fp: data/original/thumbs_up.png
+fp: data/original/tree.png
 output: demo_lensless   # output folder for results
 save: True
 plot: False
@@ -31,7 +31,7 @@ capture:
   gamma: null # for visualization
   exp: 0.02
   delay: 2
-  script: ~/LenslessPiCam/scripts/on_device_capture.py
+  script: ~/LenslessPiCam/scripts/measure/on_device_capture.py
   iso: 100
   config_pause: 2
   sensor_mode: "0"

configs/telegram_demo.yaml

@@ -13,6 +13,7 @@ rpi_lensed_hostname: null
 # psf: null
 psf:
   fp: data/psf/tape_rgb_31032023.png
+  # fp: data/psf/tape_rgb.png  # wrong
   downsample: 4
 
 # which hydra config to use and available algos

configs/unrolled_recon.yaml

@@ -27,7 +27,7 @@ preprocess:
 display:
   # How many iterations to wait for intermediate plot. 
   # Set to negative value for no intermediate plots.
-  disp: 100
+  disp: 400
   # Whether to plot results.
   plot: True
   # Gamma factor for plotting.
@@ -54,7 +54,12 @@ reconstruction:
     mu2: 1e-4
     mu3: 1e-4
     tau: 2e-4
-
+  pre_process: 
+    network : UnetRes  # UnetRes or DruNet or null
+    depth : 2 # depth of each up/downsampling layer. Ignore if network is DruNet
+  post_process: 
+    network : UnetRes  # UnetRes or DruNet or null
+    depth : 2 # depth of each up/downsampling layer. Ignore if network is DruNet
 
 # Train Dataset
 
@@ -90,13 +95,17 @@ simulation:
 #Training
 
 training:
-  batch_size: 16
-  epoch: 10
+  batch_size: 8
+  epoch: 50
+  #In case of instable training
+  skip_NAN: True
+  slow_start: False  #float how much to reduce lr for first epoch
+
 
 optimizer:
   type: Adam
-  lr: 1e-4
+  lr: 1e-6
 
 loss: 'l2'
 # set lpips to false to deactivate. Otherwise, give the weigth for the loss (the main loss l2/l1 always having a weigth of 1)
-lpips: 0.6
+lpips: 1.0

docs/source/measurement.rst

@@ -41,7 +41,7 @@ You can remotely capture data with the following script:
 
 .. code:: bash
 
-   python scripts/remote_capture.py \
+   python scripts/measure/remote_capture.py \
       rpi.username=USERNAME \
       rpi.hostname=HOSTNAME \
       plot=True
@@ -66,12 +66,12 @@ you can run the above script with a different configation:
 
 .. code:: bash
 
-   python scripts/remote_capture.py -cn capture_bayer \
+   python scripts/measure/remote_capture.py -cn capture_bayer \
       rpi.username=USERNAME \
       rpi.hostname=HOSTNAME
 
 
-You can then use ``scripts/analyze_image.py`` to play around with the red and
+You can then use ``scripts/measure/analyze_image.py`` to play around with the red and
 blue gains to find the best white balance for your camera.
 
 
@@ -108,7 +108,7 @@ image on the Raspberry Pi:
 
 .. code-block:: bash
 
-    python scripts/remote_display.py \
+    python scripts/measure/remote_display.py \
         rpi.username=USERNAME \
         rpi.hostname=HOSTNAME \
         fp=FP
@@ -134,7 +134,7 @@ Script that can be launched from the Raspberry Pi:
 
 .. code:: bash
 
-   python scripts/collect_mnist_on_device.py --input_dir MNIST_original \
+   python scripts/measure/collect_mnist_on_device.py --input_dir MNIST_original \
    --output_dir MNIST_meas
 
 If the MNIST dataset is not available at ``MNIST_original`` it will be
@@ -151,7 +151,7 @@ copying files back and forth):
 
 .. code:: bash
 
-   python scripts/collect_mnist.py --hostname <IP_ADDRESS> --output_dir MNIST_meas
+   python scripts/measure/collect_mnist.py --hostname <IP_ADDRESS> --output_dir MNIST_meas
 
 
 Collecting arbitrary dataset
@@ -162,7 +162,7 @@ camera. The script can be launched **from the Raspberry Pi**:
 
 .. code:: 
 
-   python scripts/collect_dataset_on_device.py
+   python scripts/measure/collect_dataset_on_device.py
 
 By default this script will collect a subset (100 files) of the `CelebA <https://mmlab.ie.cuhk.edu.hk/projects/CelebA.html>`__
 dataset.

lensless/eval/benchmark.py

@@ -256,7 +256,12 @@ def benchmark(model, dataset, batchsize=1, metrics=None, **kwargs):
         metrics = {
             "MSE": MSELoss().to(device),
             "MAE": L1Loss().to(device),
-            "LPIPS": lpip.LearnedPerceptualImagePatchSimilarity(net_type="vgg").to(device),
+            "LPIPS_Vgg": lpip.LearnedPerceptualImagePatchSimilarity(
+                net_type="vgg", normalize=True
+            ).to(device),
+            "LPIPS_Alex": lpip.LearnedPerceptualImagePatchSimilarity(
+                net_type="alex", normalize=True
+            ).to(device),
             "PSNR": psnr.PeakSignalNoiseRatio().to(device),
             "SSIM": StructuralSimilarityIndexMeasure().to(device),
             "ReconstructionError": None,
@@ -283,7 +288,7 @@ def benchmark(model, dataset, batchsize=1, metrics=None, **kwargs):
         prediction = prediction.reshape(-1, *prediction.shape[-3:]).movedim(-1, -3)
         lensed = lensed.reshape(-1, *lensed.shape[-3:]).movedim(-1, -3)
         # normalization
-        prediction_max = torch.amax(prediction, dim=(1, 2, 3), keepdim=True)
+        prediction_max = torch.amax(prediction, dim=(-1, -2, -3), keepdim=True)
         if torch.all(prediction_max != 0):
             prediction = prediction / prediction_max
         else:

lensless/hardware/utils.py

@@ -38,7 +38,7 @@ def display(
 
     # assumes that `LenslessPiCam` is in home directory and environment inside `LenslessPiCam`
     rpi_python = "~/LenslessPiCam/lensless_env/bin/python"
-    script = "~/LenslessPiCam/scripts/prep_display_image.py"
+    script = "~/LenslessPiCam/scripts/measure/prep_display_image.py"
     remote_tmp_file = "~/tmp_display.png"
     display_path = "~/LenslessPiCam_display/test.png"