Clean up DiffuserCam dataset upload script.

configs/upload_diffusercam_huggingface.yaml

@@ -6,8 +6,8 @@ hydra:
 repo_id: "bezzam/DiffuserCam-Lensless-Mirflickr-Dataset"
 dir_diffuser: "/scratch/bezzam/DiffuserCam_mirflickr/dataset/diffuser_images"
 dir_lensed: "/scratch/bezzam/DiffuserCam_mirflickr/dataset/ground_truth_lensed"
-psf_fp: "data/psf/diffusercam_psf.tiff"
+psf_fp: "/home/bezzam/LenslessPiCam/data/psf/diffusercam_psf.tiff"
 hf_token: null
 file_ext: ".npy"
 n_files: null
-n_jobs: 4     # for parallelizing conversion to PNG
+n_jobs: 8     # for parallelizing conversion to PNG

docs/requirements.txt

@@ -1,4 +1,4 @@
-sphinx==4.0.1
+sphinx==5.0.2
 sphinx_rtd_theme==0.4.3
 docutils==0.16     # >0.17 doesn't render bullets
 numpy>=1.22      # so that default dtype are correctly rendered

docs/source/data.rst

@@ -53,21 +53,12 @@ They both correspond to the PSF which can be found `here <https://drive.switch.c
 distance of 2 mm).
 
 
-DiffuserCam Lensless Mirflickr Dataset (DLMD)
----------------------------------------------
+DiffuserCam Lensless Mirflickr Dataset
+--------------------------------------
 
-You can download a subset for the `DiffuserCam Lensless Mirflickr
-Dataset <https://waller-lab.github.io/LenslessLearning/dataset.html>`__
-that we've prepared
-`here <https://drive.switch.ch/index.php/s/vmAZzryGI8U8rcE>`__ with
-``scripts/prepare_mirflickr_subset.py``. The original dataset is quite 
-large (25000 files, 100 GB). So we've prepared a more manageable
-dataset (200 files, 725 MB). It was prepared with the following script:
-
-.. code:: bash
-
-    python scripts/prepare_mirflickr_subset.py \
-    --data ~/Documents/DiffuserCam/DiffuserCam_Mirflickr_Dataset
+The original dataset is available `here <https://waller-lab.github.io/LenslessLearning/dataset.html>`__.
+However, it is quite large (100 GB). We've prepared a more manageable (6GB)
+and viewable version on `Hugging Face <https://huggingface.co/datasets/bezzam/DiffuserCam-Lensless-Mirflickr-Dataset>`__.
 
 
 3D data

lensless/utils/io.py

@@ -536,17 +536,19 @@ def load_data(
     return psf, data
 
 
-def save_image(img, fp, max_val=255):
+def save_image(img, fp, max_val=255, normalize=True):
     """Save as uint8 image."""
 
     img_tmp = img.copy()
 
-    if img_tmp.dtype == np.uint16:
+    if img_tmp.dtype == np.uint16 or img_tmp.dtype == np.uint8:
         img_tmp = img_tmp.astype(np.float32)
 
-    if img_tmp.dtype == np.float64 or img_tmp.dtype == np.float32:
+    if normalize:
         img_tmp -= img_tmp.min()
         img_tmp /= img_tmp.max()
+
+    if img_tmp.dtype == np.float64 or img_tmp.dtype == np.float32:
         img_tmp *= max_val
         img_tmp = img_tmp.astype(np.uint8)
 

scripts/data/upload_diffusercam_huggingface.py

@@ -22,7 +22,6 @@
 import glob
 from lensless.utils.io import save_image
 import cv2
-import PIL
 from datasets import Dataset, DatasetDict, Image
 from huggingface_hub import upload_file
 from lensless.utils.dataset import natural_sort
@@ -44,17 +43,12 @@ def upload_dataset(config):
     file_ext = config.file_ext
     n_files = config.n_files
     n_jobs = config.n_jobs
+    normalize = False
 
     assert hf_token is not None, "Please provide a HuggingFace token."
 
     start_time = time.time()
 
-    # load PSF, convert to RGB, save as PNG
-    psf_img = np.array(PIL.Image.open(psf_fp))
-    psf_img = cv2.cvtColor(psf_img, cv2.COLOR_BGR2RGB)  # convert to RGB
-    psf_fp_png = psf_fp.replace(".tiff", ".png")
-    save_image(psf_img, psf_fp_png)
-
     # get all lensless-lensed pairs
     files_diffuser = glob.glob(os.path.join(dir_diffuser, "*" + file_ext))
     files_lensed = glob.glob(os.path.join(dir_lensed, "*" + file_ext))
@@ -69,50 +63,78 @@ def upload_dataset(config):
         print(f"Only keeping {n_files} files...")
         common_files = common_files[:n_files]
 
+    # load PSF, convert to RGB, save as PNG
+    # psf_img = np.array(PIL.Image.open(psf_fp))
+    psf_img = cv2.imread(psf_fp, cv2.IMREAD_UNCHANGED)
+    psf_img = cv2.cvtColor(psf_img, cv2.COLOR_BGR2RGB)  # convert to RGB
+    psf_fp_png = psf_fp.replace(".tiff", ".png")
+    save_image(psf_img, psf_fp_png, normalize=True)  # need normalize=True
+
     # save as PNG
     dir_diffuser_png = dir_diffuser.replace("diffuser_images", "diffuser_png")
     os.makedirs(dir_diffuser_png, exist_ok=True)
     dir_lensed_png = dir_lensed.replace("ground_truth_lensed", "lensed_png")
     os.makedirs(dir_lensed_png, exist_ok=True)
-    diffuser_png_files = []
-    lensed_png_files = []
 
     # -- parallelize with joblib
     def save_png(f, dir_diffuser, dir_diffuser_png, dir_lensed, dir_lensed_png):
+
         diffuser_img = np.load(os.path.join(dir_diffuser, f))
         diffuser_img = cv2.cvtColor(diffuser_img, cv2.COLOR_BGR2RGB)  # convert to RGB
         diffuser_fn = os.path.join(dir_diffuser_png, f.replace(file_ext, ".png"))
-        diffuser_png_files.append(diffuser_fn)
-        save_image(diffuser_img, diffuser_fn)
+        save_image(diffuser_img, diffuser_fn, normalize=normalize)
 
         lensed_img = np.load(os.path.join(dir_lensed, f))
         lensed_img = cv2.cvtColor(lensed_img, cv2.COLOR_BGR2RGB)  # convert to RGB
         lensed_fn = os.path.join(dir_lensed_png, f.replace(file_ext, ".png"))
-        lensed_png_files.append(lensed_fn)
-        save_image(lensed_img, lensed_fn)
+        save_image(lensed_img, lensed_fn, normalize=normalize)
 
     Parallel(n_jobs=n_jobs)(
         delayed(save_png)(f, dir_diffuser, dir_diffuser_png, dir_lensed, dir_lensed_png)
         for f in tqdm(common_files)
     )
 
+    # get file paths
+    diffuser_files = [
+        os.path.join(dir_diffuser_png, f.replace(file_ext, ".png")) for f in common_files
+    ]
+    lensed_files = [os.path.join(dir_lensed_png, f.replace(file_ext, ".png")) for f in common_files]
+    diffuser_files = natural_sort(diffuser_files)
+    lensed_files = natural_sort(lensed_files)
+
     # step 1: create Dataset objects
-    dataset = Dataset.from_dict(
-        {
-            "lensless": diffuser_png_files,
-            "lensed": lensed_png_files,
-        }
-    )
-    dataset = dataset.cast_column("lensless", Image())
-    dataset = dataset.cast_column("lensed", Image())
+    def create_dataset(diffuser_files, lensed_files):
+        dataset = Dataset.from_dict(
+            {
+                "lensless": diffuser_files,
+                "lensed": lensed_files,
+            }
+        )
+        dataset = dataset.cast_column("lensless", Image())
+        dataset = dataset.cast_column("lensed", Image())
+        return dataset
+
+    # according to original split test files are up to idx=1000, for some reason im1 is missing?
+    test_dataset = create_dataset(diffuser_files[:999], lensed_files[:999])
+    train_dataset = create_dataset(diffuser_files[999:], lensed_files[999:])
 
     # step 2: create DatasetDict
     dataset_dict = DatasetDict(
         {
-            "all": dataset,
+            "train": train_dataset,
+            "test": test_dataset,
         }
     )
+
     # step 3: push to hub
+    upload_file(
+        path_or_fileobj=psf_fp,
+        path_in_repo="psf.tiff",
+        repo_id=repo_id,
+        repo_type="dataset",
+        token=hf_token,
+    )
+
     # -- dataset
     dataset_dict.push_to_hub(
         repo_id,
@@ -126,14 +148,14 @@ def save_png(f, dir_diffuser, dir_diffuser_png, dir_lensed, dir_lensed_png):
         token=hf_token,
     )
     upload_file(
-        path_or_fileobj=diffuser_png_files[0],
+        path_or_fileobj=diffuser_files[0],
         path_in_repo="lensless_example.png",
         repo_id=repo_id,
         repo_type="dataset",
         token=hf_token,
     )
     upload_file(
-        path_or_fileobj=lensed_png_files[0],
+        path_or_fileobj=lensed_files[0],
         path_in_repo="lensed_example.png",
         repo_id=repo_id,
         repo_type="dataset",