Initial commit

.gitignore

@@ -0,0 +1,167 @@
+# Will go to HF someday
+data/
+data/all_images_8.bin
+data/all_images_64.bin
+
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+cover/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+.pybuilder/
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+#   For a library or package, you might want to ignore these files since the code is
+#   intended to run in multiple environments; otherwise, check them in:
+# .python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# poetry
+#   Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+#   https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
+#poetry.lock
+
+# pdm
+#   Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
+#pdm.lock
+#   pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it
+#   in version control.
+#   https://pdm.fming.dev/latest/usage/project/#working-with-version-control
+.pdm.toml
+.pdm-python
+.pdm-build/
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/
+
+# PyCharm
+#  JetBrains specific template is maintained in a separate JetBrains.gitignore that can
+#  be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
+#  and can be added to the global gitignore or merged into this file.  For a more nuclear
+#  option (not recommended) you can uncomment the following to ignore the entire idea folder.
+#.idea/

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

@@ -1,3 +1,63 @@
----
-license: apache-2.0
----
+---
+license: apache-2.0
+datasets:
+  - huggan/anime-faces
+---
+
+# Mamba face kiss 
+
+## KISS
+
+This repo contains two Keep It Simple Stupid anime face generators that generates 64x64 faces from 8x8 provided images.
+
+Basic idea was to take 64x64 anime faces dataset(https://huggingface.co/datasets/huggan/anime-faces), resize it to 8x8, then teach the model to restore original images, intuition is that after that if new unseen images are provided, it will make some face.
+
+![Validation](./valid.png)
+
+Mamba is being fed a sequence `[A][A]...[A][SEP][B][B][B]...[B]` where there are 64 `[A]` that came from the 8x8 draft. there are 64x64 `[B]`s that are initially are upscaled draft(nearest neighbor) with addition of PAE. Model run several layers of mamba, and spits last 64x64 into RGB image. (`[SEP]` is not used for anything significant other than BERT has it to separate sentences, so I used it too as placeholder for command "Upscale from here")
+
+Two models are used. 
+
+### RNN goess brr (one way)
+
+One(`imgen3test.ipynb` and `imgen3.py`) always feeds images from top-left pixel to bottom-right pixel row by row
+
+![Non-flip image](./krita-nonflip.png)
+
+
+### "Bi-directional" 
+
+Another take(`imgen3test_flip.ipynb` and `imgen3_flip.py`) feed from top-left pixel to bottom-right pixel in every even layer and every odd layer sees upscaled images in reverse order
+
+![Flip image](./krita-flip.png)
+ 
+This flip version also uses way more parameters and different dtype. I didn't notice that much difference.
+
+
+#### Command line tool
+
+Simple script can be used to call the model on a single image
+
+```console
+$ cli_imgen3_flip ./krita/face1.png face1.out.png
+
+python cli_imgen3_flip.py ./krita/face1.png /tmp/face1.png
+Weight path is data/image-flip-weights-1024x4-torch.bfloat16.bin
+Loading the model
+Loading 8x8 input image from ./krita/face1.png
+Writing 64x64 image to /tmp/face1.png
+```
+
+It's not really good way to use, comparing to calling through jupyter it though: mamba2 is implemented using triton and it takes around 30 seconds to initialize the model each time (on Raider GE76).
+
+
+## Recreating
+
+Training is done in `imgen3(_flip)?.py`. Testing is in notebook. `Image_utils` should provide path to anime faces dataset.
+
+## Naming and configuring
+
+Name imgen3 comes from "image generation 3". 
+Two other attemts are not that interesting to even backup them.
+
+I'm too lazy to pass configuration around so parameters are hardcoded in the beginning of the file.

cli_imgen3_flip.py

@@ -0,0 +1,54 @@
+from imgen3flip import weights_path, Model, ImageBatch, OPTS
+import torch
+import torchvision as TV
+import torchvision.transforms.functional as VF
+import sys
+
+
+assert weights_path.exists(), "Model weights do not exist"
+
+assert len(sys.argv) == 3, f"Usage: {
+    sys.argv[0]} <input-filename> <output-filename>"
+
+input_filename = sys.argv[1]
+output_filename = sys.argv[2]
+
+assert input_filename != output_filename, f"Use different file names"
+
+print("Loading the model")
+model = Model()
+model.load_state_dict(torch.load(weights_path))
+
+print(f"Loading 8x8 input image from {input_filename}")
+# read image and ditch alpha-channel if it presents
+image = TV.io.read_image(input_filename)[:3]
+# Convert range from 0..255 to 0.0..1.0
+image = image / 255.0
+assert image.shape[0] == 3, "RGB image expected"
+# Convert C H W -> H W C
+image = image.permute(1, 2, 0)
+# Now add batch dimension(B=1): H W C -> 1 H W C
+# We also specify H, W, C explicitly as model expect them to be 8x8x3
+image = image.view(1, 8, 8, 3)
+
+# Now construct batch that model uses
+# Target and loss are not used in inference, as model code always calculates loss
+dummy_target = torch.zeros(1, 64, 64, 3, **OPTS)
+dummy_loss = torch.tensor(-1, **OPTS)
+inference_batch = ImageBatch(
+    im8=image.to(**OPTS),
+    im64=dummy_target,
+    loss=dummy_loss)
+result = model(inference_batch)
+
+# Now convert image to PIL format so we can save it
+new_image = result.im64.detach().float().cpu()
+# new_image: 1 H W C -> H W C
+new_image = new_image[0]
+# new_image: H W C -> C H W
+new_image = new_image.permute(2, 0, 1)
+assert new_image.shape == (3, 64, 64)
+img = VF.to_pil_image(new_image)
+# Save
+print(f"Writing {img.height}x{img.width} image to {output_filename}")
+img.save(output_filename)

image_utils.py

@@ -0,0 +1,130 @@
+import torch.nn as nn
+from torch import Tensor
+from pathlib import Path
+import torch
+import random
+import torchvision.io as VIO
+import torchvision.transforms.functional as VF
+from dataclasses import dataclass
+from tqdm.auto import tqdm
+
+# https://huggingface.co/datasets/huggan/anime-faces
+RAW_IMAGES_PATH = Path(
+    "~/Downloads/datasets/anime/anime-faces/images").expanduser()
+RESOLUTIONS = [64, 8]
+
+AS_TENSORS_64 = Path(f"data/all_images_64.bin")
+AS_TENSORS_8 = Path(f"data/all_images_8.bin")
+
+
+@dataclass
+class ImageBatch:
+    im8: Tensor
+    im64: Tensor
+    loss: Tensor
+
+    @property
+    def n_batch(self):
+        return self.im8.shape[0]
+
+    def as_1d(self):
+        return ImageBatch(
+            im8=self.im8.view(self.n_batch, 8*8, self.im8.shape[-1]),
+            im64=self.im64.view(self.n_batch, 64*64, self.im64.shape[-1]),
+            loss=self.loss
+        )
+
+    def as_2d(self):
+        return ImageBatch(
+            im8=self.im8.view(self.n_batch, 8, 8, self.im8.shape[-1]),
+            im64=self.im64.view(self.n_batch, 64, 64, self.im64.shape[-1]),
+            loss=self.loss
+        )
+
+
+class ImageDB:
+    def __init__(self, val_ratio=0.05, dtype=None) -> None:
+        if not AS_TENSORS_64.exists():
+            self.make_tensor_version()
+        print("Load tensors file")
+        self.dtype = dtype or torch.bfloat16
+        self.all_images_64 = torch.load(AS_TENSORS_64).to(self.dtype)
+        self.all_images_8 = torch.load(AS_TENSORS_8).to(self.dtype)
+        self.n_val = int(len(self.all_images_64) * val_ratio)
+
+    def split(self, s: str):
+        if s == "train":
+            return {
+                8: self.all_images_8[:-self.n_val],
+                64: self.all_images_64[:-self.n_val]
+            }
+        if s == "valid":
+            return {
+                8: self.all_images_8[-self.n_val:],
+                64: self.all_images_64[-self.n_val:]
+            }
+        raise ValueError(f"Invalid split {s}")
+
+    @property
+    def train_ds(self):
+        return self.split("train")
+
+    @property
+    def valid_ds(self):
+        return self.split("valid")
+
+    @torch.no_grad()
+    def make_tensor_version(self, path=RAW_IMAGES_PATH):
+        items = list(path.glob("*.png"))
+        all_tensors = [load_single_image(item) for item in tqdm(items)]
+        t64 = torch.stack([t[64] for t in all_tensors])
+        t8 = torch.stack([t[8] for t in all_tensors])
+        torch.save(t64, AS_TENSORS_64)
+        torch.save(t8, AS_TENSORS_8)
+        return {8: t8, 64: t64}
+
+    def random_batch(self, bs: int, split: str = "train"):
+        split_dict = self.split(split)
+        im8 = split_dict[8]
+        im64 = split_dict[64]
+        keys = list(range(len(im8)))
+        random.shuffle(keys)
+        keys = keys[: bs]
+        return ImageBatch(
+            im64=im64[keys].cuda(),
+            im8=im8[keys].cuda(),
+            loss=torch.tensor(-1))
+
+
+def load_single_image(path: Path):
+    im = VIO.read_image(str(path))
+    im = im / 255.0
+    # resize to 8x8
+    im8 = VF.resize(im, [8, 8], VF.InterpolationMode.NEAREST_EXACT)
+    # C H W -> H W C
+    im = im.permute(1, 2, 0).contiguous()
+    im8 = im8.permute(1, 2, 0).contiguous()
+
+    return {64: im, 8: im8}
+
+
+class RGBToModel(nn.Module):
+    def __init__(self, d_model, device=None, dtype=None):
+        super().__init__()
+        self.fc = nn.Linear(3, d_model, device=device, dtype=dtype)
+
+    def forward(self, x):
+        return self.fc(x)
+
+
+class ModelToRGB(nn.Module):
+    def __init__(self, d_model, device=None, dtype=None):
+        super().__init__()
+        self.norm = nn.LayerNorm(d_model, device=device, dtype=dtype)
+        self.fc = nn.Linear(d_model, 3, device=device, dtype=dtype)
+
+    def forward(self, x):
+        x = self.norm(x)
+        x = self.fc(x)
+        x = x.sigmoid()
+        return x

imgen3.py

@@ -0,0 +1,100 @@
+from mamba_ssm.modules.mamba_simple import Mamba
+from mamba_ssm.modules.mamba2_simple import Mamba2Simple
+from mamba_ssm.modules.mamba2 import Mamba2
+import torch
+from torch import Tensor
+import random
+import torch.nn as nn
+import torch.nn.functional as F
+from tqdm.auto import tqdm
+from pathlib import Path
+from einops import rearrange, repeat
+from typing import Optional
+
+
+from image_utils import ImageDB, ImageBatch, RGBToModel
+from image_utils import ModelToRGB
+from torch_utils import model_numel
+
+epochs = 10_000
+bs = 16
+d_model = 768
+weights_path = Path(f"data/weights-{d_model}.bin")
+
+OPTS = {
+    'device': "cuda",
+    'dtype': torch.float32
+}
+
+
+class MambaWrap(nn.Module):
+    def __init__(self) -> None:
+        super().__init__()
+        self.mamba = Mamba2Simple(d_model, **OPTS, headdim=64)
+        self.norm = nn.LayerNorm(d_model, **OPTS)
+
+    def forward(self, x):
+        residual = x
+        x = self.norm(x)
+        x = self.mamba(x)
+        x = residual + x
+        return x
+
+
+class Model(nn.Module):
+    def __init__(self) -> None:
+        super().__init__()
+        self.from_rgb = RGBToModel(d_model, **OPTS)
+        self.to_rgb = ModelToRGB(d_model, **OPTS)
+        self.s0 = nn.Parameter(torch.randn(1, 1, d_model, **OPTS))
+        self.suffix = nn.Parameter(torch.randn(64*64, d_model, **OPTS))
+        self.layers = nn.ModuleList([MambaWrap() for _ in range(4)])
+        self.norm0 = nn.LayerNorm(d_model, **OPTS)
+
+    def forward(self, batch: ImageBatch):
+        B = batch.n_batch
+        batch = batch.as_1d()
+        batch.im8 = self.from_rgb(batch.im8)
+
+        s0 = self.s0.repeat(B, 1, 1)
+        s1 = self.zoom(batch.im8)
+
+        x = torch.cat((s0, batch.im8, s1), 1)
+        x = self.norm0(x)
+        x = self.mamba(x)
+        x = x[:, -64*64:]
+        y_hat = self.to_rgb(x)
+        y_true = batch.im64
+        batch.loss = F.mse_loss(y_hat, y_true)
+        batch.im64 = y_hat
+        return batch.as_2d()
+
+    def zoom(self, im8):
+        im8 = im8.view(im8.shape[0], 8, 8, im8.shape[-1])
+        im8 = repeat(
+            im8, "B H W C -> B (H 8) (W 8) C").view(im8.shape[0], 64*64, im8.shape[-1])
+        im8 = im8 + self.suffix
+        return im8
+
+    def mamba(self, x):
+        for layer in self.layers:
+            x = layer(x)
+        return x
+
+if __name__ == "__main_":
+    image_db = ImageDB(dtype=OPTS["dtype"])
+    model = Model()
+    if weights_path.exists():
+        print(f"*** Load {weights_path:s}")
+        model.load_state_dict(torch.load(weights_path))
+    opt = torch.optim.AdamW(model.parameters(), fused=True)
+    
+    for e in (bar := tqdm(range(epochs))):
+        b = model(image_db.random_batch(bs))
+        b.loss.backward()
+        opt.step()
+        opt.zero_grad()
+        bar.set_description(f'L:{b.loss.item():.4f}')
+        if e and e % 100 == 0:
+            torch.save(model.state_dict(), weights_path)
+    torch.save(model.state_dict(), weights_path)

imgen3flip.py

@@ -0,0 +1,132 @@
+from mamba_ssm.modules.mamba2_simple import Mamba2Simple
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from tqdm.auto import tqdm
+from pathlib import Path
+from einops import repeat
+
+
+from image_utils import ImageDB, ImageBatch, RGBToModel
+from image_utils import ModelToRGB
+
+epochs = 10_000
+bs = 16
+# orig;
+# bs = 16
+# d_model = 768
+# headdim = 64
+# n_layer = 4
+
+d_model = 1024
+headdim = 64
+n_layer = 4
+
+OPTS = {
+    'device': "cuda",
+    'dtype': torch.bfloat16
+}
+# Since we have KISS flip/flop think that number of mamba layers are actually 2 times higher
+# This is somewhat relatable to LLM model where 1 block had two mamba layers: one replaced ATTN, one replaced MLP
+
+weights_path = Path(
+    f"data/image-flip-weights-{d_model}x{n_layer}-{str(OPTS['dtype'])}.bin")
+print(f"Weight path is {str(weights_path)}")
+
+
+class MambaWrap(nn.Module):
+    def __init__(self) -> None:
+        super().__init__()
+        self.mamba = Mamba2Simple(d_model, **OPTS, headdim=headdim)
+        self.norm = nn.LayerNorm(d_model, **OPTS)
+
+    def forward(self, x):
+        residual = x
+        x = self.norm(x)
+        x = self.mamba(x)
+        x = residual + x
+        return x
+
+
+class MambaFlipFlop(nn.Module):
+    def __init__(self, n_values) -> None:
+        super().__init__()
+        self.mb_forward = MambaWrap()
+        self.mb_backward = MambaWrap()
+        self.n_values = n_values
+
+    def forward(self, x):
+        x = self.mb_forward(x)
+        x = self.swap_order(x)
+        x = self.mb_backward(x)
+        x = self.swap_order(x)
+        return x
+
+    def swap_order(self, x):
+        T = x.shape[1]
+        head = torch.arange(0, T - self.n_values)
+        tail = torch.arange(T - 1, T - self.n_values - 1, -1)
+        seq = torch.cat((head, tail))
+        x = x[:, seq]
+        return x
+
+
+class Model(nn.Module):
+    def __init__(self) -> None:
+        super().__init__()
+        self.from_rgb = RGBToModel(d_model, **OPTS)
+        self.to_rgb = ModelToRGB(d_model, **OPTS)
+        self.s0 = nn.Parameter(torch.randn(1, 1, d_model, **OPTS))
+        self.suffix = nn.Parameter(torch.randn(64*64, d_model, **OPTS))
+        self.layers = nn.ModuleList([MambaFlipFlop(64*64)
+                                    for _ in range(n_layer)])
+        self.norm0 = nn.LayerNorm(d_model, **OPTS)
+
+    def forward(self, batch: ImageBatch):
+        B = batch.n_batch
+        batch = batch.as_1d()
+        batch.im8 = self.from_rgb(batch.im8)
+
+        s0 = self.s0.repeat(B, 1, 1)
+        s1 = self.zoom(batch.im8)
+
+        x = torch.cat((s0, batch.im8, s1), 1)
+        x = self.norm0(x)
+        x = self.mamba(x)
+        x = x[:, -64*64:]
+        y_hat = self.to_rgb(x)
+        y_true = batch.im64
+        batch.loss = F.mse_loss(y_hat, y_true)
+        batch.im64 = y_hat
+        return batch.as_2d()
+
+    def zoom(self, im8):
+        im8 = im8.view(im8.shape[0], 8, 8, im8.shape[-1])
+        im8 = repeat(im8, "B H W C -> B (H 8) (W 8) C")
+        im8 = im8.view(im8.shape[0], 64*64, im8.shape[-1])
+        im8 = im8 + self.suffix
+        return im8
+
+    def mamba(self, x):
+        for layer in self.layers:
+            x = layer(x)
+        return x
+
+
+if __name__ == "__main__":
+    image_db = ImageDB(dtype=OPTS["dtype"])
+    model = Model()
+    if weights_path.exists():
+        print(f"*** Load {str(weights_path)}")
+        model.load_state_dict(torch.load(weights_path))
+    opt = torch.optim.AdamW(model.parameters(), fused=True)
+
+    for e in (bar := tqdm(range(epochs))):
+        b = model(image_db.random_batch(bs))
+        b.loss.backward()
+        opt.step()
+        opt.zero_grad()
+        bar.set_description(f'L:{b.loss.item():.4f}')
+        if e and e % 100 == 0:
+            torch.save(model.state_dict(), weights_path)
+    torch.save(model.state_dict(), weights_path)

torch_utils.py

@@ -0,0 +1,13 @@
+import torch.nn as nn
+import torch
+
+
+def model_device(m: nn.Module):
+    return next(iter(m.parameters())).device
+
+
+def model_numel(m: nn.Module, requires_grad=False):
+    if requires_grad:
+        return sum(p.numel() for p in m.parameters() if p.requires_grad)
+    else:
+        return sum(p.numel() for p in m.parameters())