Add wandb support.

CHANGELOG.rst

@@ -21,12 +21,13 @@ Added
 - ``lensless.utils.dataset.simulate_dataset`` for simulating a dataset given a mask/PSF.
 - Support for training/testing with multiple mask patterns in the dataset.
 - Multi-GPU support for training.
-- DigiCam dataset which interfaces with Hugging Face.
+- Dataset which interfaces with Hugging Face (``lensless.utils.dataset.HFDataset``).
 - Scripts for authentication.
 - DigiCam support for Telegram demo.
 - DiffuserCamMirflickr Hugging Face API.
 - Fallback for normalization if data not in 8bit range (``lensless.utils.io.save_image``).
 - Add utilities for fabricating masks with 3D printing (``lensless.hardware.fabrication``).
+- WandB support.
 
 Changed
 ~~~~~~~
@@ -151,7 +152,7 @@ Added
 -  Option to warm-start reconstruction algorithm with ``initial_est``.
 -  TrainableReconstructionAlgorithm class inherited from ReconstructionAlgorithm and torch.module for use with pytorch autograd and optimizers.
 -  Unrolled version of FISTA and ADMM as TrainableReconstructionAlgorithm with learnable parameters.
-- ``train_unrolled.py`` script for training unrolled algorithms.
+- ``train_learning_based.py`` script for training unrolled algorithms.
 - ``benchmark_recon.py`` script for benchmarking and comparing reconstruction algorithms.
 - Added ``reconstruction_error`` to ``ReconstructionAlgorithm`` .
 - Added support for npy/npz image in load_image.

README.rst

@@ -45,7 +45,7 @@ The toolkit includes:
 * Measurement scripts (`link <https://lensless.readthedocs.io/en/latest/measurement.html>`__).
 * Dataset preparation and loading tools, with `Hugging Face <https://huggingface.co/bezzam>`__ integration (`slides <https://docs.google.com/presentation/d/18h7jTcp20jeoiF8dJIEcc7wHgjpgFgVxZ_bJ04W55lg/edit?usp=sharing>`__ on uploading a dataset to Hugging Face with `this script <https://github.com/LCAV/LenslessPiCam/blob/main/scripts/data/upload_dataset_huggingface.py>`__).
 * `Reconstruction algorithms <https://lensless.readthedocs.io/en/latest/reconstruction.html>`__ (e.g. FISTA, ADMM, unrolled algorithms, trainable inversion, pre- and post-processors).
-* `Training script <https://github.com/LCAV/LenslessPiCam/blob/main/scripts/recon/train_unrolled.py>`__ for learning-based reconstruction.
+* `Training script <https://github.com/LCAV/LenslessPiCam/blob/main/scripts/recon/train_learning_based.py>`__ for learning-based reconstruction.
 * `Pre-trained models <https://github.com/LCAV/LenslessPiCam/blob/main/lensless/recon/model_dict.py>`__ that can be loaded from `Hugging Face <https://huggingface.co/bezzam>`__, for example in `this script <https://github.com/LCAV/LenslessPiCam/blob/main/scripts/recon/diffusercam_mirflickr.py>`__.
 * Mask `design <https://lensless.readthedocs.io/en/latest/mask.html>`__ and `fabrication <https://lensless.readthedocs.io/en/latest/fabrication.html>`__ tools.
 * `Simulation tools <https://lensless.readthedocs.io/en/latest/simulation.html>`__.

configs/fine-tune_PSF.yaml

@@ -1,4 +1,4 @@
-# python scripts/recon/train_unrolled.py -cn fine-tune_PSF
+# python scripts/recon/train_learning_based.py -cn fine-tune_PSF
 defaults:
   - train_unrolledADMM
   - _self_
@@ -12,25 +12,7 @@ trainable_mask:
 
 #Training
 training:
-  save_every: 10
-  epoch: 50
-  crop_preloss: False
+  save_every: 1   # to see how PSF evolves
 
 display:
   gamma: 2.2
-
-reconstruction:
-  method: unrolled_admm
-
-  pre_process:
-    network: UnetRes
-    depth: 2
-  post_process:
-    network: DruNet
-    depth: 4
-
-optimizer:
-  slow_start: 0.01
-
-loss: l2
-lpips: 1.0

configs/train_celeba_digicam_hitl.yaml

@@ -1,7 +1,7 @@
 # Learn mask with HITL training by setting measure configuration (set to null for learning in simulation)
 # 
 # EXAMPLE COMMAND: 
-# python scripts/recon/train_unrolled.py -cn train_celeba_digicam_hitl measure.rpi_username=USERNAME measure.rpi_hostname=HOSTNAME files.vertical_shift=SHIFT
+# python scripts/recon/train_learning_based.py -cn train_celeba_digicam_hitl measure.rpi_username=USERNAME measure.rpi_hostname=HOSTNAME files.vertical_shift=SHIFT
 
 defaults:
   - train_celeba_digicam

configs/train_celeba_digicam_mask.yaml

@@ -1,5 +1,5 @@
 # fine-tune mask for PSF, but don't re-simulate
-# python scripts/recon/train_unrolled.py -cn train_celeba_digicam_mask
+# python scripts/recon/train_learning_based.py -cn train_celeba_digicam_mask
 defaults:
   - train_celeba_digicam
   - _self_

configs/train_coded_aperture.yaml

@@ -1,4 +1,4 @@
-# python scripts/recon/train_unrolled.py -cn train_coded_aperture
+# python scripts/recon/train_learning_based.py -cn train_coded_aperture
 defaults:
   - train_unrolledADMM
   - _self_

configs/train_digicam_celeba.yaml

@@ -1,4 +1,4 @@
-# python scripts/recon/train_unrolled.py -cn train_digicam_singlemask
+# python scripts/recon/train_learning_based.py -cn train_digicam_celeba
 defaults:
   - train_unrolledADMM
   - _self_
@@ -13,6 +13,7 @@ files:
   huggingface_psf: "psf_simulated.png"
   huggingface_dataset: True
   split_seed: 0
+  test_size: 0.15
   downsample: 2
   rotate: True   # if measurement is upside-down
   save_psf: False
@@ -34,14 +35,14 @@ alignment:
     random_vflip: False
     random_hflip: False
     quantize: False
-    # shifting when there is no files.downsample
+    # shifting when there is no files to downsample
     vertical_shift: -117
     horizontal_shift: -25
 
 training:
   batch_size: 4
   epoch: 25
-  eval_batch_size: 4
+  eval_batch_size: 16
   crop_preloss: True
 
 reconstruction:

configs/train_digicam_multimask.yaml

@@ -1,17 +1,18 @@
-# python scripts/recon/train_unrolled.py -cn train_digicam_multimask
+# python scripts/recon/train_learning_based.py -cn train_digicam_multimask
 defaults:
   - train_unrolledADMM
   - _self_
 
-
 torch_device: 'cuda:0'
 device_ids: [0, 1, 2, 3]
 eval_disp_idx: [1, 2, 4, 5, 9]
 
+
 # Dataset
 files:
   dataset: bezzam/DigiCam-Mirflickr-MultiMask-25K
   huggingface_dataset: True
+  huggingface_psf: null
   downsample: 1
   # TODO: these parameters should be in the dataset?
   image_res: [900, 1200]  # used during measurement
@@ -55,4 +56,5 @@ reconstruction:
   post_process: 
     network : UnetRes  # UnetRes or DruNet or null
     depth : 4 # depth of each up/downsampling layer. Ignore if network is DruNet
-    nc: [32,64,116,128]
+    nc: [32,64,116,128]
+

configs/train_digicam_singlemask.yaml

@@ -1,4 +1,4 @@
-# python scripts/recon/train_unrolled.py -cn train_digicam_singlemask
+# python scripts/recon/train_learning_based.py -cn train_digicam_singlemask
 defaults:
   - train_unrolledADMM
   - _self_
@@ -11,12 +11,13 @@ eval_disp_idx: [1, 2, 4, 5, 9]
 files:
   dataset: bezzam/DigiCam-Mirflickr-SingleMask-25K
   huggingface_dataset: True
+  huggingface_psf: null
   downsample: 1
+  # TODO: these parameters should be in the dataset?
   image_res: [900, 1200]  # used during measurement
   rotate: True   # if measurement is upside-down
   save_psf: False
 
-  # extra_eval: null
   extra_eval: 
     multimask: 
       huggingface_repo: bezzam/DigiCam-Mirflickr-MultiMask-25K
@@ -26,6 +27,7 @@ files:
         topright: [80, 100]  # height, width
         height: 200
 
+# TODO: these parameters should be in the dataset?
 alignment:
   # when there is no downsampling
   topright: [80, 100]  # height, width

configs/train_psf_from_scratch.yaml

@@ -1,11 +1,15 @@
-# python scripts/recon/train_unrolled.py -cn train_psf_from_scratch
+# python scripts/recon/train_learning_based.py -cn train_psf_from_scratch
 defaults:
   - train_unrolledADMM
   - _self_
 
 # Train Dataset
 files:
   dataset: mnist   # Simulated : "mnist", "fashion_mnist", "cifar10", "CelebA". Measure :"DiffuserCam"
+  huggingface_dataset: False
+  n_files: 1000
+  test_size: 0.15
+
   celeba_root: /scratch/bezzam
   downsample: 8
 
@@ -24,8 +28,6 @@ simulation:
   object_height: 0.30
 
 training:
-  crop_preloss: False # crop region for computing loss
-  batch_size: 8
-  epoch: 25
+  batch_size: 2
   eval_batch_size: 16
   save_every: 5

configs/train_unrolledADMM.yaml

@@ -1,39 +1,52 @@
-# python scripts/recon/train_unrolled.py
+# python scripts/recon/train_learning_based.py
 hydra:
   job:
     chdir: True    # change to output folder
 
 
+wandb_project: lensless
 seed: 0
 start_delay: null
 
 # Dataset
 files:
-  dataset: /scratch/bezzam/DiffuserCam_mirflickr/dataset  # Simulated : "mnist", "fashion_mnist", "cifar10", "CelebA". Measure :"DiffuserCam"
-  celeba_root: null   # path to parent directory of CelebA: https://mmlab.ie.cuhk.edu.hk/projects/CelebA.html
-  psf: data/psf/diffusercam_psf.tiff
-  diffusercam_psf: True
-
-  huggingface_dataset: null
-  huggingface_psf: null
+  # -- using local dataset
+  # dataset: /scratch/bezzam/DiffuserCam_mirflickr/dataset  # Simulated : "mnist", "fashion_mnist", "cifar10", "CelebA". Measure :"DiffuserCam"
+  # celeba_root: null   # path to parent directory of CelebA: https://mmlab.ie.cuhk.edu.hk/projects/CelebA.html
+  # psf: data/psf/diffusercam_psf.tiff
+  # diffusercam_psf: True
+
+  # -- using huggingface dataset
+  dataset: bezzam/DiffuserCam-Lensless-Mirflickr-Dataset-NORM
+  huggingface_dataset: True
+  huggingface_psf: psf.tiff
+
+  # -- train/test split
   split_seed: null   # if null use train/test split from dataset
-
   n_files: null    # null to use all for both train/test
-  downsample: 2    # factor by which to downsample the PSF, note that for DiffuserCam the PSF has 4x the resolution
-  test_size: 0.15
+  test_size: null
 
+  # -- processing parameters
+  downsample: 2    # factor by which to downsample the PSF, note that for DiffuserCam the PSF has 4x the resolution
+  downsample_lensed: 2
   input_snr: null    # adding shot noise at input (for measured dataset) at this SNR in dB
   vertical_shift: null
   horizontal_shift: null
+  rotate: False
+  save_psf: False
   crop: null
     # vertical: null
     # horizontal: null
   image_res: null   # for measured data, what resolution used at screen
-
   extra_eval: null  # dict of extra datasets to evaluate on
 
+alignment: null
+#   topright: null    # height, width
+  # height: null
+
 torch: True
 torch_device: 'cuda'
+device_ids: null    # for multi-gpu set list, e.g. [0, 1, 2, 3]
 measure: null       # if measuring data on-the-fly
 
 # test set example to visualize at the end of every epoch
@@ -130,14 +143,13 @@ simulation:
 
 training:
   batch_size: 8
-  epoch: 50
+  epoch: 25
   eval_batch_size: 10
   metric_for_best_model: null   # e.g. LPIPS_Vgg, null does test loss
   save_every: null
   #In case of instable training
   skip_NAN: True
   clip_grad: 1.0
-
   crop_preloss: False  # crop region for computing loss, files.crop should be set
 
 optimizer:

configs/train_unrolled_pre_post.yaml

@@ -0,0 +1,14 @@
+# python scripts/recon/train_learning_based.py -cn train_unrolled_pre_post
+defaults:
+  - train_unrolledADMM
+  - _self_
+
+reconstruction:
+  method: unrolled_admm
+
+  pre_process:
+    network: UnetRes
+    depth: 2
+  post_process:
+    network: UnetRes
+    depth: 2

docs/source/dataset.rst

@@ -19,6 +19,26 @@ or measured).
     :special-members: __init__, __len__
 
 
+Measured dataset objects
+------------------------
+
+.. autoclass:: lensless.utils.dataset.HFDataset
+    :members:
+    :special-members: __init__
+
+.. autoclass:: lensless.utils.dataset.MeasuredDataset
+    :members:
+    :special-members: __init__
+
+.. autoclass:: lensless.utils.dataset.MeasuredDatasetSimulatedOriginal
+    :members:
+    :special-members: __init__
+
+.. autoclass:: lensless.utils.dataset.DiffuserCamTestDataset
+    :members:
+    :special-members: __init__
+
+
 Simulated dataset objects
 -------------------------
 
@@ -43,19 +63,3 @@ mask / PSF.
 .. autoclass:: lensless.utils.dataset.SimulatedDatasetTrainableMask
     :members:
     :special-members: __init__
-
-
-Measured dataset objects
-------------------------
-
-.. autoclass:: lensless.utils.dataset.MeasuredDataset
-    :members:
-    :special-members: __init__
-
-.. autoclass:: lensless.utils.dataset.MeasuredDatasetSimulatedOriginal
-    :members:
-    :special-members: __init__
-
-.. autoclass:: lensless.utils.dataset.DiffuserCamTestDataset
-    :members:
-    :special-members: __init__