Upload 14 files

LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

README.md

@@ -0,0 +1,88 @@
+# cmmd-pytorch
+
+(Unofficial) PyTorch implementation of CLIP Maximum Mean Discrepancy (CMMD) for evaluating image generation models, proposed in [Rethinking FID: Towards a Better Evaluation Metric for Image Generation](https://arxiv.org/abs/2401.09603). CMMD stands out to be a better metric than FID and tries to mitigate the longstanding issues of FID.
+
+This implementation is a super simple PyTorch port of the [original codebase](https://github.com/google-research/google-research/tree/master/cmmd). I have only focused on the JAX and TensorFlow specific bits and replaced them PyTorch. Some differences:
+
+* The original codebase relies on [`scenic`](https://github.com/google-research/scenic) for computing CLIP embeddings. This repository uses [`transformers`](https://github.com/huggingface/transformers).
+* For the data loading, the original codebase uses TensorFlow, this one uses PyTorch `Dataset` and `DataLoader`.
+
+## Setup
+
+First, install PyTorch following instructions from the [official website](https://pytorch.org/).
+
+Then install the depdencies:
+
+```bash
+pip install -r requirements.txt
+```
+
+## Running
+
+```bash
+python main.py /path/to/reference/images /path/to/eval/images --batch_size=32 --max_count=30000
+```
+
+A working example command:
+
+```bash
+python main.py reference_images generated_images --batch_size=1
+```
+
+It should output:
+
+```bash
+The CMMD value is:  7.696
+```
+
+This is the same as the original codebase, so, that confirms the implementation correctness 🤗
+
+> [!TIP]
+> GPU execution is supported when a GPU is available.
+
+## Results
+
+Below, we report the CMMD metric for some popular pipelines on the COCO-30k dataset, as commonly used by the community. CMMD, like FID, is better when it's lower.
+
+| **Pipeline** | **Inference Steps** | **Resolution** | **CMMD** |
+|:------------:|:-------------------:|:--------------:|:--------:|
+|   [`stabilityai/stable-diffusion-xl-base-1.0`](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)   |     30     |   1024x1024  | 0.696 |
+|   [`segmind/SSD-1B`](https://huggingface.co/segmind/SSD-1B)   |     30     |   1024x1024  | 0.669 |
+|   [`stabilityai/sdxl-turbo`](https://huggingface.co/stabilityai/sdxl-turbo)   |     1     |   512x512  | 0.548 |
+|   [`runwayml/stable-diffusion-v1-5`](https://huggingface.co/runwayml/stable-diffusion-v1-5)   |     50     |   512x512  | 0.582 |
+|   [`PixArt-alpha/PixArt-XL-2-1024-MS`](https://huggingface.co/PixArt-alpha/PixArt-XL-2-1024-MS)   |     20     |   1024x1024  | 1.140 |
+|   [`SPRIGHT-T2I/spright-t2i-sd2`](https://huggingface.co/SPRIGHT-T2I/spright-t2i-sd2)   |     50     |   768x768  | 0.512 |
+
+**Notes**:
+
+* For SDXL Turbo, `guidance_scale` is set to 0 following the [official guide](https://huggingface.co/docs/diffusers/main/en/using-diffusers/sdxl_turbo) in `diffusers`. 
+* For all other pipelines, default `guidace_scale` was used. Refer to the official pipeline documentation pages [here](https://huggingface.co/docs/diffusers/main/en/index) for more details.
+
+> [!CAUTION]
+> As per the CMMD authors, with models producing high-quality/high-resolution images, COCO images don't seem to be a good reference set (they are of pretty small resolution). This might help explain why SD v1.5 has a better CMMD than SDXL.
+
+## Obtaining CMMD for your pipelines
+
+One can refer to the `generate_images.py` script that generates images from the [COCO-30k randomly sampled captions](https://huggingface.co/datasets/sayakpaul/sample-datasets/raw/main/coco_30k_randomly_sampled_2014_val.csv) using `diffusers`. 
+
+Once the images are generated, run:
+
+```bash
+python main.py /path/to/reference/images /path/to/generated/images --batch_size=32 --max_count=30000
+```
+
+Reference images are COCO-30k images and can be downloaded from [here](https://huggingface.co/datasets/sayakpaul/coco-30-val-2014).
+
+Pre-computed embeddings for the COCO-30k images can be found [here](https://huggingface.co/datasets/sayakpaul/coco-30-val-2014/blob/main/ref_embs_coco_30k.npy).
+
+To use the pre-computed reference embeddings, run:
+
+```bash
+python main.py None /path/to/generated/images ref_embed_file=ref_embs.npy --batch_size=32 --max_count=30000
+```
+
+## Acknowledgements
+
+Thanks to Sadeep Jayasumana (first author of CMMD) for all the helpful discussions.
+
+

distance.py

@@ -0,0 +1,64 @@
+# coding=utf-8
+# Copyright 2024 The Google Research Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Memory-efficient MMD implementation in JAX."""
+
+import torch
+
+# The bandwidth parameter for the Gaussian RBF kernel. See the paper for more
+# details.
+_SIGMA = 10
+# The following is used to make the metric more human readable. See the paper
+# for more details.
+_SCALE = 1000
+
+
+def mmd(x, y):
+    """Memory-efficient MMD implementation in JAX.
+
+    This implements the minimum-variance/biased version of the estimator described
+    in Eq.(5) of
+    https://jmlr.csail.mit.edu/papers/volume13/gretton12a/gretton12a.pdf.
+    As described in Lemma 6's proof in that paper, the unbiased estimate and the
+    minimum-variance estimate for MMD are almost identical.
+
+    Note that the first invocation of this function will be considerably slow due
+    to JAX JIT compilation.
+
+    Args:
+      x: The first set of embeddings of shape (n, embedding_dim).
+      y: The second set of embeddings of shape (n, embedding_dim).
+
+    Returns:
+      The MMD distance between x and y embedding sets.
+    """
+    x = torch.from_numpy(x)
+    y = torch.from_numpy(y)
+
+    x_sqnorms = torch.diag(torch.matmul(x, x.T))
+    y_sqnorms = torch.diag(torch.matmul(y, y.T))
+
+    gamma = 1 / (2 * _SIGMA**2)
+    k_xx = torch.mean(
+        torch.exp(-gamma * (-2 * torch.matmul(x, x.T) + torch.unsqueeze(x_sqnorms, 1) + torch.unsqueeze(x_sqnorms, 0)))
+    )
+    k_xy = torch.mean(
+        torch.exp(-gamma * (-2 * torch.matmul(x, y.T) + torch.unsqueeze(x_sqnorms, 1) + torch.unsqueeze(y_sqnorms, 0)))
+    )
+    k_yy = torch.mean(
+        torch.exp(-gamma * (-2 * torch.matmul(y, y.T) + torch.unsqueeze(y_sqnorms, 1) + torch.unsqueeze(y_sqnorms, 0)))
+    )
+
+    return _SCALE * (k_xx + k_yy - 2 * k_xy)

embedding.py

@@ -0,0 +1,71 @@
+# coding=utf-8
+# Copyright 2024 The Google Research Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Embedding models used in the CMMD calculation."""
+
+from transformers import CLIPImageProcessor, CLIPVisionModelWithProjection
+import torch
+import numpy as np
+
+_CLIP_MODEL_NAME = "openai/clip-vit-large-patch14-336"
+_CUDA_AVAILABLE = torch.cuda.is_available()
+
+
+def _resize_bicubic(images, size):
+    images = torch.from_numpy(images.transpose(0, 3, 1, 2))
+    images = torch.nn.functional.interpolate(images, size=(size, size), mode="bicubic")
+    images = images.permute(0, 2, 3, 1).numpy()
+    return images
+
+
+class ClipEmbeddingModel:
+    """CLIP image embedding calculator."""
+
+    def __init__(self):
+        self.image_processor = CLIPImageProcessor.from_pretrained(_CLIP_MODEL_NAME)
+
+        self._model = CLIPVisionModelWithProjection.from_pretrained(_CLIP_MODEL_NAME).eval()
+        if _CUDA_AVAILABLE:
+            self._model = self._model.cuda()
+
+        self.input_image_size = self.image_processor.crop_size["height"]
+
+    @torch.no_grad()
+    def embed(self, images):
+        """Computes CLIP embeddings for the given images.
+
+        Args:
+          images: An image array of shape (batch_size, height, width, 3). Values are
+            in range [0, 1].
+
+        Returns:
+          Embedding array of shape (batch_size, embedding_width).
+        """
+
+        images = _resize_bicubic(images, self.input_image_size)
+        inputs = self.image_processor(
+            images=images,
+            do_normalize=True,
+            do_center_crop=False,
+            do_resize=False,
+            do_rescale=False,
+            return_tensors="pt",
+        )
+        if _CUDA_AVAILABLE:
+            inputs = {k: v.to("cuda") for k, v in inputs.items()}
+
+        image_embs = self._model(**inputs).image_embeds.cpu()
+        image_embs /= torch.linalg.norm(image_embs, axis=-1, keepdims=True)
+        return image_embs

generate_images.py

@@ -0,0 +1,79 @@
+from diffusers import DiffusionPipeline
+from concurrent.futures import ThreadPoolExecutor
+import pandas as pd
+import argparse
+import torch
+import os
+
+
+ALL_CKPTS = [
+    "runwayml/stable-diffusion-v1-5",
+    "segmind/SSD-1B",
+    "PixArt-alpha/PixArt-XL-2-1024-MS",
+    "stabilityai/stable-diffusion-xl-base-1.0",
+    "stabilityai/sdxl-turbo",
+]
+SEED = 2024
+
+
+def load_dataframe():
+    dataframe = pd.read_csv(
+        "https://huggingface.co/datasets/sayakpaul/sample-datasets/raw/main/coco_30k_randomly_sampled_2014_val.csv"
+    )
+    return dataframe
+
+
+def load_pipeline(args):
+    if "runway" in args.pipeline_id:
+        pipeline = DiffusionPipeline.from_pretrained(
+            args.pipeline_id, torch_dtype=torch.float16, safety_checker=None
+        ).to("cuda")
+    else:
+        pipeline = DiffusionPipeline.from_pretrained(args.pipeline_id, torch_dtype=torch.float16).to("cuda")
+    pipeline.set_progress_bar_config(disable=True)
+    return pipeline
+
+
+def generate_images(args, dataframe, pipeline):
+    all_images = []
+    for i in range(0, len(dataframe), args.chunk_size):
+        if "sdxl-turbo" not in args.pipeline_id:
+            images = pipeline(
+                dataframe.iloc[i : i + args.chunk_size]["caption"].tolist(),
+                num_inference_steps=args.num_inference_steps,
+                generator=torch.manual_seed(SEED),
+            ).images
+        else:
+            images = pipeline(
+                dataframe.iloc[i : i + args.chunk_size]["caption"].tolist(),
+                num_inference_steps=args.num_inference_steps,
+                generator=torch.manual_seed(SEED),
+                guidance_scale=0.0,
+            ).images
+        all_images.extend(images)
+    return all_images
+
+
+def serialize_image(image, path):
+    image.save(path)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--pipeline_id", default="runwayml/stable-diffusion-v1-5", type=str, choices=ALL_CKPTS)
+    parser.add_argument("--num_inference_steps", default=30, type=int)
+    parser.add_argument("--chunk_size", default=2, type=int)
+    parser.add_argument("--root_img_path", default="sdv15", type=str)
+    parser.add_argument("--num_workers", type=int, default=4)
+    args = parser.parse_args()
+
+    dataset = load_dataframe()
+    pipeline = load_pipeline(args)
+    images = generate_images(args, dataset, pipeline)
+    image_paths = [os.path.join(args.root_img_path, f"{i}.jpg") for i in range(len(images))]
+
+    if not os.path.exists(args.root_img_path):
+        os.makedirs(args.root_img_path)
+
+    with ThreadPoolExecutor(max_workers=args.num_workers) as executor:
+        executor.map(serialize_image, images, image_paths)

io_util.py

@@ -0,0 +1,123 @@
+# coding=utf-8
+# Copyright 2024 The Google Research Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""IO utilities."""
+
+import glob
+from torch.utils.data import Dataset, DataLoader
+import numpy as np
+from PIL import Image
+import tqdm
+
+
+class CMMDDataset(Dataset):
+    def __init__(self, path, reshape_to, max_count=-1):
+        self.path = path
+        self.reshape_to = reshape_to
+
+        self.max_count = max_count
+        img_path_list = self._get_image_list()
+        if max_count > 0:
+            img_path_list = img_path_list[:max_count]
+        self.img_path_list = img_path_list
+
+    def __len__(self):
+        return len(self.img_path_list)
+
+    def _get_image_list(self):
+        ext_list = ["png", "jpg", "jpeg"]
+        image_list = []
+        for ext in ext_list:
+            image_list.extend(glob.glob(f"{self.path}/*{ext}"))
+            image_list.extend(glob.glob(f"{self.path}/*.{ext.upper()}"))
+        # Sort the list to ensure a deterministic output.
+        image_list.sort()
+        return image_list
+
+    def _center_crop_and_resize(self, im, size):
+        w, h = im.size
+        l = min(w, h)
+        top = (h - l) // 2
+        left = (w - l) // 2
+        box = (left, top, left + l, top + l)
+        im = im.crop(box)
+        # Note that the following performs anti-aliasing as well.
+        return im.resize((size, size), resample=Image.BICUBIC)  # pytype: disable=module-attr
+
+    def _read_image(self, path, size):
+        im = Image.open(path)
+        if size > 0:
+            im = self._center_crop_and_resize(im, size)
+        return np.asarray(im).astype(np.float32)
+
+    def __getitem__(self, idx):
+        img_path = self.img_path_list[idx]
+
+        x = self._read_image(img_path, self.reshape_to)
+        if x.ndim == 3:
+            return x
+        elif x.ndim == 2:
+            # Convert grayscale to RGB by duplicating the channel dimension.
+            return np.tile(x[Ellipsis, np.newaxis], (1, 1, 3))
+
+
+def compute_embeddings_for_dir(
+    img_dir,
+    embedding_model,
+    batch_size,
+    max_count=-1,
+):
+    """Computes embeddings for the images in the given directory.
+
+    This drops the remainder of the images after batching with the provided
+    batch_size to enable efficient computation on TPUs. This usually does not
+    affect results assuming we have a large number of images in the directory.
+
+    Args:
+      img_dir: Directory containing .jpg or .png image files.
+      embedding_model: The embedding model to use.
+      batch_size: Batch size for the embedding model inference.
+      max_count: Max number of images in the directory to use.
+
+    Returns:
+      Computed embeddings of shape (num_images, embedding_dim).
+    """
+    dataset = CMMDDataset(img_dir, reshape_to=embedding_model.input_image_size, max_count=max_count)
+    count = len(dataset)
+    print(f"Calculating embeddings for {count} images from {img_dir}.")
+
+    dataloader = DataLoader(dataset, batch_size=batch_size)
+
+    all_embs = []
+    for batch in tqdm.tqdm(dataloader, total=count // batch_size):
+        image_batch = batch.numpy()
+
+        # Normalize to the [0, 1] range.
+        image_batch = image_batch / 255.0
+
+        if np.min(image_batch) < 0 or np.max(image_batch) > 1:
+            raise ValueError(
+                "Image values are expected to be in [0, 1]. Found:" f" [{np.min(image_batch)}, {np.max(image_batch)}]."
+            )
+
+        # Compute the embeddings using a pmapped function.
+        embs = np.asarray(
+            embedding_model.embed(image_batch)
+        )  # The output has shape (num_devices, batch_size, embedding_dim).
+        all_embs.append(embs)
+
+    all_embs = np.concatenate(all_embs, axis=0)
+
+    return all_embs

main.py

@@ -0,0 +1,73 @@
+# coding=utf-8
+# Copyright 2024 The Google Research Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""The main entry point for the CMMD calculation."""
+
+from absl import app
+from absl import flags
+import distance
+import embedding
+import io_util
+import numpy as np
+
+
+_BATCH_SIZE = flags.DEFINE_integer("batch_size", 32, "Batch size for embedding generation.")
+_MAX_COUNT = flags.DEFINE_integer("max_count", -1, "Maximum number of images to read from each directory.")
+_REF_EMBED_FILE = flags.DEFINE_string(
+    "ref_embed_file", None, "Path to the pre-computed embedding file for the reference images."
+)
+
+
+def compute_cmmd(ref_dir, eval_dir, ref_embed_file=None, batch_size=32, max_count=-1):
+    """Calculates the CMMD distance between reference and eval image sets.
+
+    Args:
+      ref_dir: Path to the directory containing reference images.
+      eval_dir: Path to the directory containing images to be evaluated.
+      ref_embed_file: Path to the pre-computed embedding file for the reference images.
+      batch_size: Batch size used in the CLIP embedding calculation.
+      max_count: Maximum number of images to use from each directory. A
+        non-positive value reads all images available except for the images
+        dropped due to batching.
+
+    Returns:
+      The CMMD value between the image sets.
+    """
+    if ref_dir and ref_embed_file:
+        raise ValueError("`ref_dir` and `ref_embed_file` both cannot be set at the same time.")
+    embedding_model = embedding.ClipEmbeddingModel()
+    if ref_embed_file is not None:
+        ref_embs = np.load(ref_embed_file).astype("float32")
+    else:
+        ref_embs = io_util.compute_embeddings_for_dir(ref_dir, embedding_model, batch_size, max_count).astype(
+            "float32"
+        )
+    eval_embs = io_util.compute_embeddings_for_dir(eval_dir, embedding_model, batch_size, max_count).astype("float32")
+    val = distance.mmd(ref_embs, eval_embs)
+    return val.numpy()
+
+
+def main(argv):
+    if len(argv) != 3:
+        raise app.UsageError("Too few/too many command-line arguments.")
+    _, dir1, dir2 = argv
+    print(
+        "The CMMD value is: "
+        f" {compute_cmmd(dir1, dir2, _REF_EMBED_FILE.value, _BATCH_SIZE.value, _MAX_COUNT.value):.3f}"
+    )
+
+
+if __name__ == "__main__":
+    app.run(main)

requirements.txt

@@ -0,0 +1,6 @@
+transformers
+accelerate
+Pillow
+tqdm
+numpy
+absl-py