mimic-video 0.0.5__tar.gz

mimic_video-0.0.5/.github/workflows/python-publish.yml

@@ -0,0 +1,36 @@
+# This workflow will upload a Python Package using Twine when a release is created
+# For more information see: https://help.github.com/en/actions/language-and-framework-guides/using-python-with-github-actions#publishing-to-package-registries
+
+# This workflow uses actions that are not certified by GitHub.
+# They are provided by a third-party and are governed by
+# separate terms of service, privacy policy, and support
+# documentation.
+
+name: Upload Python Package
+
+on:
+  release:
+    types: [published]
+
+jobs:
+  deploy:
+
+    runs-on: ubuntu-latest
+
+    steps:
+    - uses: actions/checkout@v2
+    - name: Set up Python
+      uses: actions/setup-python@v2
+      with:
+        python-version: '3.x'
+    - name: Install dependencies
+      run: |
+        python -m pip install --upgrade pip
+        pip install build
+    - name: Build package
+      run: python -m build
+    - name: Publish package
+      uses: pypa/gh-action-pypi-publish@27b31702a0e7fc50959f5ad993c78deac1bdfc29
+      with:
+        user: __token__
+        password: ${{ secrets.PYPI_API_TOKEN }}

mimic_video-0.0.5/.github/workflows/test.yml

@@ -0,0 +1,21 @@
+name: Pytest
+on: [push, pull_request]
+
+jobs:
+  build:
+
+    runs-on: ubuntu-latest
+
+    steps:
+    - uses: actions/checkout@v4
+    - name: Set up Python 3.10
+      uses: actions/setup-python@v5
+      with:
+        python-version: "3.10"
+    - name: Install dependencies
+      run: |
+        python -m pip install --upgrade pip
+        python -m pip install -e .[test]
+    - name: Test with pytest
+      run: |
+        python -m pytest tests/

mimic_video-0.0.5/.gitignore

@@ -0,0 +1,207 @@
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[codz]
+*$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
+
+# UV
+#   Similar to Pipfile.lock, it is generally recommended to include uv.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+#uv.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
+#poetry.toml
+
+# pdm
+#   Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
+#   pdm recommends including project-wide configuration in pdm.toml, but excluding .pdm-python.
+#   https://pdm-project.org/en/latest/usage/project/#working-with-version-control
+#pdm.lock
+#pdm.toml
+.pdm-python
+.pdm-build/
+
+# pixi
+#   Similar to Pipfile.lock, it is generally recommended to include pixi.lock in version control.
+#pixi.lock
+#   Pixi creates a virtual environment in the .pixi directory, just like venv module creates one
+#   in the .venv directory. It is recommended not to include this directory in version control.
+.pixi
+
+# 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
+.envrc
+.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/
+
+# Abstra
+# Abstra is an AI-powered process automation framework.
+# Ignore directories containing user credentials, local state, and settings.
+# Learn more at https://abstra.io/docs
+.abstra/
+
+# Visual Studio Code
+#  Visual Studio Code specific template is maintained in a separate VisualStudioCode.gitignore 
+#  that can be found at https://github.com/github/gitignore/blob/main/Global/VisualStudioCode.gitignore
+#  and can be added to the global gitignore or merged into this file. However, if you prefer, 
+#  you could uncomment the following to ignore the entire vscode folder
+# .vscode/
+
+# Ruff stuff:
+.ruff_cache/
+
+# PyPI configuration file
+.pypirc
+
+# Cursor
+#  Cursor is an AI-powered code editor. `.cursorignore` specifies files/directories to
+#  exclude from AI features like autocomplete and code analysis. Recommended for sensitive data
+#  refer to https://docs.cursor.com/context/ignore-files
+.cursorignore
+.cursorindexingignore
+
+# Marimo
+marimo/_static/
+marimo/_lsp/
+__marimo__/

mimic_video-0.0.5/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2025 Phil Wang
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

mimic_video-0.0.5/PKG-INFO

@@ -0,0 +1,82 @@
+Metadata-Version: 2.4
+Name: mimic-video
+Version: 0.0.5
+Summary: Mimic Video
+Project-URL: Homepage, https://pypi.org/project/mimic-video/
+Project-URL: Repository, https://github.com/lucidrains/mimic-video
+Author-email: Phil Wang <lucidrains@gmail.com>
+License: MIT License
+        
+        Copyright (c) 2025 Phil Wang
+        
+        Permission is hereby granted, free of charge, to any person obtaining a copy
+        of this software and associated documentation files (the "Software"), to deal
+        in the Software without restriction, including without limitation the rights
+        to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+        copies of the Software, and to permit persons to whom the Software is
+        furnished to do so, subject to the following conditions:
+        
+        The above copyright notice and this permission notice shall be included in all
+        copies or substantial portions of the Software.
+        
+        THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+        IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+        FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+        AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+        LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+        OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+        SOFTWARE.
+License-File: LICENSE
+Keywords: artificial intelligence,attention mechanism,deep learning,video language action model
+Classifier: Development Status :: 4 - Beta
+Classifier: Intended Audience :: Developers
+Classifier: License :: OSI Approved :: MIT License
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.10
+Requires-Dist: einops>=0.8.1
+Requires-Dist: einx>=0.3.0
+Requires-Dist: torch-einops-utils>=0.0.8
+Requires-Dist: torch>=2.5
+Requires-Dist: x-mlps-pytorch
+Provides-Extra: examples
+Provides-Extra: test
+Requires-Dist: pytest; extra == 'test'
+Description-Content-Type: text/markdown
+
+<img src="./mimic-video.png" width="450px"></img>
+
+## Mimic Video (wip)
+
+Implementation of [Mimic-Video](https://mimic-video.github.io/), Video-Action Models for Generalizable Robot Control Beyond VLAs
+
+## Appreciation
+
+- [Pranoy](https://github.com/pranoyr) for submitting a pull request for proprioception masking
+
+## Contributing
+
+First make sure `pytest` and test dependencies are installed with
+
+```shell
+$ pip install '.[test]'
+```
+
+Then add your test to `tests/test_mimic_video.py` and run
+
+```shell
+$ pytest tests
+```
+
+That's it
+
+## Citations
+
+```bibtex
+@inproceedings{Pai2025mimicvideoVM,
+    title   = {mimic-video: Video-Action Models for Generalizable Robot Control Beyond VLAs},
+    author  = {Jonas Pai and Liam Achenbach and Victoriano Montesinos and Benedek Forrai and Oier Mees and Elvis Nava},
+    year    = {2025},
+    url     = {https://api.semanticscholar.org/CorpusID:283920528}
+}
+```

mimic_video-0.0.5/README.md

@@ -0,0 +1,36 @@
+<img src="./mimic-video.png" width="450px"></img>
+
+## Mimic Video (wip)
+
+Implementation of [Mimic-Video](https://mimic-video.github.io/), Video-Action Models for Generalizable Robot Control Beyond VLAs
+
+## Appreciation
+
+- [Pranoy](https://github.com/pranoyr) for submitting a pull request for proprioception masking
+
+## Contributing
+
+First make sure `pytest` and test dependencies are installed with
+
+```shell
+$ pip install '.[test]'
+```
+
+Then add your test to `tests/test_mimic_video.py` and run
+
+```shell
+$ pytest tests
+```
+
+That's it
+
+## Citations
+
+```bibtex
+@inproceedings{Pai2025mimicvideoVM,
+    title   = {mimic-video: Video-Action Models for Generalizable Robot Control Beyond VLAs},
+    author  = {Jonas Pai and Liam Achenbach and Victoriano Montesinos and Benedek Forrai and Oier Mees and Elvis Nava},
+    year    = {2025},
+    url     = {https://api.semanticscholar.org/CorpusID:283920528}
+}
+```

mimic_video-0.0.5/mimic_video/__init__.py

@@ -0,0 +1,2 @@
+
+from mimic_video.mimic_video import MimicVideo

mimic_video-0.0.5/mimic_video/mimic_video.py

@@ -0,0 +1,422 @@
+import torch
+from torch import nn, cat, stack, is_tensor, tensor
+from torch.nn import Module, ModuleList, Linear
+
+import torch.nn.functional as F
+
+import einx
+from einops import einsum, rearrange, repeat
+from einops.layers.torch import Rearrange
+
+from x_mlps_pytorch import create_mlp
+
+from torch_einops_utils import (
+    pad_left_ndim,
+    align_dims_left,
+    pad_at_dim,
+    pack_with_inverse,
+)
+
+# ein notation
+
+# b - batch
+# h - heads
+# g - groups
+# n - sequence
+# i, j - sequence (source, target)
+# d - feature dimension
+
+# functions
+
+def exists(v):
+    return v is not None
+
+def default(v, d):
+    return v if exists(v) else d
+
+def divisible_by(num, den):
+    return (num % den) == 0
+
+# tensor function
+
+def cast_tensor(val, device = None):
+    return tensor(val, device = device) if not is_tensor(val) else val
+
+def max_neg_value(t):
+    return -torch.finfo(t.dtype).max
+
+def l2norm(t, eps = 1e-10):
+    return F.normalize(t, dim = -1, eps = eps)
+
+# token shift from Peng et al. of RWKV
+# cheap way to generate relative positions
+
+def shift_feature_dim(t):
+    x, x_shift = t.chunk(2, dim = -1)
+    x_shift = pad_at_dim(x_shift, (1, -1), dim = 1)
+    return cat((x, x_shift), dim = -1)
+
+# time
+
+# they follow p0's research finding with the beta distribution
+# lets stick with 0 noise to 1 data instead of the reverse
+
+def default_sample_time_fn(time, s = 0.999):
+    return torch.sqrt(s - time)
+
+class RandomFourierEmbed(Module):
+    def __init__(self, dim):
+        super().__init__()
+        self.proj = nn.Sequential(
+            Rearrange('... -> ... 1'),
+            nn.Linear(1, dim)
+        )
+
+        self.proj.requires_grad_(False)
+
+    def forward(self, times):
+        rand_proj = self.proj(times)
+        return torch.cos(2 * torch.pi * rand_proj)
+
+# adaptive rmsnorm
+
+class AdaptiveRMSNorm(Module):
+    def __init__(
+        self,
+        dim,
+        dim_time_cond,
+        eps = 1e-6,
+        ada_ln_zero_bias = -5.
+    ):
+        super().__init__()
+        self.scale = dim ** 0.5
+        self.eps = eps
+
+        self.to_modulation = Linear(dim_time_cond, dim * 3, bias = False)
+        self.split_modulation = Rearrange('b (three d) -> three b 1 d', three = 3)
+
+        nn.init.zeros_(self.to_modulation.weight)
+
+        self.ada_ln_zero_bias = ada_ln_zero_bias
+
+    def forward(
+        self,
+        tokens,
+        time_cond
+    ):
+
+        if time_cond.ndim == 1:
+            time_cond = pad_left_ndim(time_cond, 1)
+
+        modulations = self.to_modulation(time_cond)
+
+        scale, shift, gate = self.split_modulation(modulations)
+
+        normed = l2norm(tokens, self.eps) * self.scale
+
+        adaptive_normed = normed * (scale + 1.) + shift
+
+        gate_with_bias = gate + self.ada_ln_zero_bias
+
+        return adaptive_normed, gate_with_bias
+
+# attention
+
+class Attention(Module):
+    def __init__(
+        self,
+        dim,
+        *,
+        dim_context = None,
+        dim_head = 64,
+        heads = 8,
+        kv_heads = 2
+    ):
+        super().__init__()
+        dim_q_inner = dim_head * heads
+        dim_kv_inner = dim_head * kv_heads
+        dim_context = default(dim_context, dim)
+
+        self.scale = dim_head ** -0.5
+
+        self.to_queries = Linear(dim, dim_q_inner, bias = False)
+        self.to_keys_values = Linear(dim_context, dim_kv_inner * 2, bias = False)
+        self.to_out = Linear(dim_q_inner, dim, bias = False)
+
+        assert divisible_by(heads, kv_heads)
+        groups = heads // kv_heads
+
+        self.split_q_heads = Rearrange('b n (g h d) -> b g h n d', g = groups, d = dim_head)
+        self.split_kv_heads = Rearrange('b n (h d) -> b h n d', d = dim_head)
+        self.merge_heads = Rearrange('b g h n d -> b n (g h d)')
+
+    def forward(
+        self,
+        tokens,
+        context = None,
+        context_mask = None
+    ):
+        context = default(context, tokens)
+
+        queries = self.to_queries(tokens)
+        keys, values = self.to_keys_values(context).chunk(2, dim = -1)
+
+        queries = self.split_q_heads(queries)
+        keys, values = tuple(self.split_kv_heads(t) for t in (keys, values))
+
+        queries = queries * self.scale
+
+        sim = einsum(queries, keys, 'b g h i d, b h j d -> b g h i j')
+
+        if exists(context_mask):
+            mask_value = max_neg_value(sim)
+            sim = einx.where('b j, b g h i j,', context_mask, sim, mask_value)
+
+        attn = sim.softmax(dim = -1)
+
+        out = einsum(attn, values, 'b g h i j, b h j d -> b g h i d')
+
+        out = self.merge_heads(out)
+
+        return self.to_out(out)
+
+# feedforward
+
+class SwiGLUFeedForward(Module):
+    def __init__(
+        self,
+        dim,
+        *,
+        expansion_factor = 4.,
+    ):
+        super().__init__()
+        dim_inner = int(dim * expansion_factor * 2 / 3)
+
+        self.proj_in = nn.Linear(dim, dim_inner * 2)
+        self.proj_out = nn.Linear(dim_inner, dim)
+
+    def forward(
+        self,
+        tokens
+    ):
+        hidden, gates = self.proj_in(tokens).chunk(2, dim = -1)
+
+        out = hidden * F.gelu(gates)
+
+        return self.proj_out(out)
+
+# classes
+
+class MimicVideo(Module):
+    def __init__(
+        self,
+        dim,
+        *,
+        dim_video_hidden,
+        dim_action = 20,
+        dim_joint_state = 32,
+        proprio_mask_prob = 0.1,
+        depth = 8,
+        dim_head = 64,
+        heads = 8,
+        expansion_factor = 4.,
+        ada_ln_zero_bias = -5.,
+        dim_time_cond = None,
+        sample_time_fn = None
+    ):
+        super().__init__()
+
+        # flow related
+
+        self.sample_time_fn = default(sample_time_fn, default_sample_time_fn)
+
+        # embed
+
+        self.to_action_tokens = Linear(dim_action, dim)
+
+        dim_time_cond = default(dim_time_cond, dim * 2)
+
+        self.to_fourier_embed = RandomFourierEmbed(dim) # used by deepmind, its fine
+        self.to_time_cond = create_mlp(dim_in = dim * 2, dim = dim_time_cond, depth = 2, activation = nn.SiLU())
+
+        # joint token related
+
+        self.to_joint_state_token = Linear(dim_joint_state, dim)
+
+        self.proprio_mask_prob = proprio_mask_prob
+        self.has_proprio_masking = proprio_mask_prob > 0.
+
+        self.proprio_mask_token = nn.Parameter(torch.randn(dim))
+
+        # video norm
+
+        self.video_hidden_norm = nn.RMSNorm(dim_video_hidden)
+
+        # transformer
+
+        layers = []
+
+        for _ in range(depth):
+            attn_adanorm = AdaptiveRMSNorm(dim = dim, dim_time_cond = dim_time_cond)
+
+            attn = Attention(dim = dim, dim_head = dim_head, heads = heads)
+
+            cross_attn_adanorm = AdaptiveRMSNorm(dim = dim, dim_time_cond = dim_time_cond)
+
+            cross_attn = Attention(dim = dim, dim_head = dim_head, dim_context = dim_video_hidden, heads = heads)
+
+            ff_adanorm = AdaptiveRMSNorm(dim = dim, dim_time_cond = dim_time_cond, ada_ln_zero_bias = ada_ln_zero_bias)
+
+            ff = SwiGLUFeedForward(dim = dim, expansion_factor = expansion_factor)
+
+            layers.append(ModuleList([
+                attn_adanorm,
+                attn,
+                cross_attn_adanorm,
+                cross_attn,
+                ff_adanorm,
+                ff
+            ]))
+
+        self.layers = ModuleList(layers)
+
+        # predictions
+
+        self.to_pred_action_flow = nn.Sequential(
+            nn.RMSNorm(dim),
+            Linear(dim, dim_action)
+        )
+
+    def forward(
+        self,
+        actions,
+        video_hiddens, # they use layer 19 of cosmos predict, at first denoising step. that's all
+        *,
+        joint_state,
+        time = None,
+        time_video_denoise = 0., # 0 is noise in the scheme i prefer - default to their optimal choice, but can be changed
+        context_mask = None,
+    ):
+        batch, device = actions.shape[0], actions.device
+
+        is_training = not exists(time)
+
+        # handle flow time conditioning
+
+        if is_training:
+            time = torch.rand((batch,), device = device)
+            time = self.sample_time_fn(time)
+
+            noise = torch.randn_like(actions)
+            flow = actions - noise
+
+            actions, left_aligned_time = align_dims_left((actions, time))
+
+            noised = noise.lerp(actions, left_aligned_time)
+        else:
+            noised = actions
+
+        if time.ndim == 0:
+            time = rearrange(time, '-> b', b = batch)
+
+        # handle the video denoising times
+
+        time_video_denoise = cast_tensor(time_video_denoise)
+
+        if time_video_denoise.ndim == 0:
+            time_video_denoise = rearrange(time_video_denoise, '-> 1')
+
+        if time_video_denoise.shape[0] != batch:
+            time_video_denoise = repeat(time_video_denoise, '1 -> b', b = batch)
+
+        times = stack((time, time_video_denoise), dim = -1)
+
+        # fourier embed and mlp to time condition
+
+        fourier_embed = self.to_fourier_embed(times)
+
+        fourier_embed = rearrange(fourier_embed, '... times d -> ... (times d)')
+
+        time_cond = self.to_time_cond(fourier_embed)
+
+        # handle video hiddens
+
+        video_hiddens = self.video_hidden_norm(video_hiddens)
+
+        # embed
+
+        tokens = self.to_action_tokens(noised)
+
+        #  mask joint state token for proprioception masking training
+
+        joint_state_token = self.to_joint_state_token(joint_state)
+
+        if self.training and self.has_proprio_masking:
+            mask = torch.rand((batch,), device = device) < self.proprio_mask_prob
+
+            joint_state_token = einx.where('b, d, b d', mask, self.proprio_mask_token, joint_state_token)
+
+        # pack joint with action tokens
+
+        tokens, inverse_pack = pack_with_inverse((joint_state_token, tokens), 'b * d')
+
+        # transformer layers
+
+        for (
+            attn_norm,
+            attn,
+            cross_attn_norm,
+            cross_attn,
+            ff_norm,
+            ff
+        ) in self.layers:
+
+            # cross attention
+
+            residual = tokens
+
+            tokens, gate = cross_attn_norm(tokens, time_cond)
+
+            tokens = residual + cross_attn(tokens, context = video_hiddens, context_mask = context_mask) * gate
+
+            # self attention
+
+            residual = tokens
+
+            tokens, gate = attn_norm(tokens, time_cond)
+
+            tokens = residual + attn(tokens) * gate.sigmoid()
+
+            # prepare feedforward
+
+            residual = tokens
+
+            tokens, gate = ff_norm(tokens, time_cond)
+
+            # shift along time for action tokens for cheap relative positioning, which is better than messing with rope with such short action chunks
+
+            joint_state_token, tokens = inverse_pack(tokens)
+
+            tokens = shift_feature_dim(tokens)
+
+            tokens, _ = pack_with_inverse((joint_state_token, tokens), 'b * d')
+
+            # feedforward
+
+            tokens = residual + ff(tokens) * gate.sigmoid()
+
+        # remove joint token
+
+        _, tokens = inverse_pack(tokens)
+
+        # prediction
+
+        pred_flow = self.to_pred_action_flow(tokens)
+
+        if not is_training:
+            return pred_flow
+
+        # mse flow loss
+
+        flow_loss = F.mse_loss(pred_flow, flow)
+        return flow_loss

mimic_video-0.0.5/pyproject.toml

@@ -0,0 +1,61 @@
+[project]
+name = "mimic-video"
+version = "0.0.5"
+description = "Mimic Video"
+authors = [
+    { name = "Phil Wang", email = "lucidrains@gmail.com" }
+]
+readme = "README.md"
+requires-python = ">= 3.10"
+license = { file = "LICENSE" }
+keywords = [
+    'artificial intelligence',
+    'deep learning',
+    'attention mechanism',
+    'video language action model'
+]
+
+classifiers=[
+    'Development Status :: 4 - Beta',
+    'Intended Audience :: Developers',
+    'Topic :: Scientific/Engineering :: Artificial Intelligence',
+    'License :: OSI Approved :: MIT License',
+    'Programming Language :: Python :: 3.10',
+]
+
+dependencies = [
+    "einx>=0.3.0",
+    "einops>=0.8.1",
+    "torch>=2.5",
+    "torch-einops-utils>=0.0.8",
+    "x-mlps-pytorch"
+]
+
+[project.urls]
+Homepage = "https://pypi.org/project/mimic-video/"
+Repository = "https://github.com/lucidrains/mimic-video"
+
+[project.optional-dependencies]
+examples = []
+test = [
+    "pytest"
+]
+
+[tool.pytest.ini_options]
+pythonpath = [
+  "."
+]
+
+[build-system]
+requires = ["hatchling"]
+build-backend = "hatchling.build"
+
+[tool.rye]
+managed = true
+dev-dependencies = []
+
+[tool.hatch.metadata]
+allow-direct-references = true
+
+[tool.hatch.build.targets.wheel]
+packages = ["mimic_video"]

mimic_video-0.0.5/tests/test_mimic_video.py

@@ -0,0 +1,24 @@
+import pytest
+import torch
+
+def test_mimic_video():
+    from mimic_video.mimic_video import MimicVideo
+
+    video_hiddens = torch.randn(2, 64, 77)
+    video_mask = torch.randint(0, 2, (2, 64)).bool()
+
+    mimic_video = MimicVideo(512, dim_video_hidden = 77)
+
+    actions = torch.randn(2, 32, 20)
+
+    joint_state = torch.randn(2, 32)
+
+    forward_kwargs = dict(video_hiddens = video_hiddens, context_mask = video_mask, joint_state = joint_state)
+
+    loss = mimic_video(actions, **forward_kwargs)
+
+    assert loss.numel() == 1
+
+    flow = mimic_video(actions, **forward_kwargs, time = torch.tensor([0.5, 0.5]))
+
+    assert flow.shape == actions.shape