tiny-recursive-model 0.0.1__tar.gz

tiny_recursive_model-0.0.1/.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 }}

tiny_recursive_model-0.0.1/.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/

tiny_recursive_model-0.0.1/.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__/

tiny_recursive_model-0.0.1/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.

tiny_recursive_model-0.0.1/PKG-INFO

@@ -0,0 +1,70 @@
+Metadata-Version: 2.4
+Name: tiny-recursive-model
+Version: 0.0.1
+Summary: Tiny Recursive Model
+Project-URL: Homepage, https://pypi.org/project/tiny-recursive-model/
+Project-URL: Repository, https://github.com/lucidrains/tiny-recursive-model
+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,deep learning,reasoning
+Classifier: Development Status :: 4 - Beta
+Classifier: Intended Audience :: Developers
+Classifier: License :: OSI Approved :: MIT License
+Classifier: Programming Language :: Python :: 3.9
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.9
+Requires-Dist: accelerate
+Requires-Dist: einops>=0.8.1
+Requires-Dist: ema-pytorch
+Requires-Dist: torch>=2.4
+Requires-Dist: x-transformers
+Provides-Extra: examples
+Provides-Extra: test
+Requires-Dist: pytest; extra == 'test'
+Description-Content-Type: text/markdown
+
+
+<img width="300" alt="trm-fig1" src="https://github.com/user-attachments/assets/950db79e-5f9c-4fec-a4e4-7b9355b39ce8" />
+
+## Tiny Recursive Model (TRM) wip
+
+Implementation of [Tiny Recursive Model](https://arxiv.org/abs/2510.04871) (TRM), improvement to [HRM](https://github.com/lucidrains/hrm) from Sapient AI, by [Alexia Jolicoeur-Martineau](https://ajolicoeur.wordpress.com/about/)
+
+Official repository is [here](https://github.com/SamsungSAILMontreal/TinyRecursiveModels)
+
+<img width="300" alt="trm-fig3" src="https://github.com/user-attachments/assets/bfe3dd2a-e859-492a-84d5-faf37339f534" />
+
+## Citations
+
+```bibtex
+@misc{jolicoeurmartineau2025morerecursivereasoningtiny,
+    title   = {Less is More: Recursive Reasoning with Tiny Networks}, 
+    author  = {Alexia Jolicoeur-Martineau},
+    year    = {2025},
+    eprint  = {2510.04871},
+    archivePrefix = {arXiv},
+    primaryClass = {cs.LG},
+    url     = {https://arxiv.org/abs/2510.04871}, 
+}
+```

tiny_recursive_model-0.0.1/README.md

@@ -0,0 +1,24 @@
+
+<img width="300" alt="trm-fig1" src="https://github.com/user-attachments/assets/950db79e-5f9c-4fec-a4e4-7b9355b39ce8" />
+
+## Tiny Recursive Model (TRM) wip
+
+Implementation of [Tiny Recursive Model](https://arxiv.org/abs/2510.04871) (TRM), improvement to [HRM](https://github.com/lucidrains/hrm) from Sapient AI, by [Alexia Jolicoeur-Martineau](https://ajolicoeur.wordpress.com/about/)
+
+Official repository is [here](https://github.com/SamsungSAILMontreal/TinyRecursiveModels)
+
+<img width="300" alt="trm-fig3" src="https://github.com/user-attachments/assets/bfe3dd2a-e859-492a-84d5-faf37339f534" />
+
+## Citations
+
+```bibtex
+@misc{jolicoeurmartineau2025morerecursivereasoningtiny,
+    title   = {Less is More: Recursive Reasoning with Tiny Networks}, 
+    author  = {Alexia Jolicoeur-Martineau},
+    year    = {2025},
+    eprint  = {2510.04871},
+    archivePrefix = {arXiv},
+    primaryClass = {cs.LG},
+    url     = {https://arxiv.org/abs/2510.04871}, 
+}
+```

tiny_recursive_model-0.0.1/pyproject.toml

@@ -0,0 +1,60 @@
+[project]
+name = "tiny-recursive-model"
+version = "0.0.1"
+description = "Tiny Recursive Model"
+authors = [
+    { name = "Phil Wang", email = "lucidrains@gmail.com" }
+]
+readme = "README.md"
+requires-python = ">= 3.9"
+license = { file = "LICENSE" }
+keywords = [
+    'artificial intelligence',
+    'deep learning',
+    'reasoning',
+]
+
+classifiers=[
+    'Development Status :: 4 - Beta',
+    'Intended Audience :: Developers',
+    'Topic :: Scientific/Engineering :: Artificial Intelligence',
+    'License :: OSI Approved :: MIT License',
+    'Programming Language :: Python :: 3.9',
+]
+
+dependencies = [
+    "accelerate",
+    "einops>=0.8.1",
+    "ema-pytorch",
+    "torch>=2.4",
+    "x-transformers",
+]
+
+[project.urls]
+Homepage = "https://pypi.org/project/tiny-recursive-model/"
+Repository = "https://github.com/lucidrains/tiny-recursive-model"
+
+[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 = ["tiny_recursive_model"]

tiny_recursive_model-0.0.1/tests/test_trm.py

@@ -0,0 +1,70 @@
+import pytest
+param = pytest.mark.parametrize
+
+import torch
+from tiny_recursive_model.trm import TinyRecursiveModel, Trainer
+
+@param('use_self_attn', (False, True))
+def test_trm(
+    use_self_attn
+):
+    from torch.optim import AdamW
+
+    if use_self_attn:
+        from x_transformers import Encoder
+        network = Encoder(dim = 512, depth = 2)
+    else:
+        from tiny_recursive_model.mlp_mixer_1d import MLPMixer1D
+        network = MLPMixer1D(dim = 512, depth = 2, seq_len = 1024)
+
+    trm = TinyRecursiveModel(
+        dim = 512,
+        num_tokens = 256,
+        network = network
+    )
+
+    optim = AdamW(trm.parameters(), lr = 1e-4)
+
+    seq = torch.randint(0, 256, (2, 1024))
+    answer = torch.randint(0, 256, (2, 1024))
+
+    outputs, latents = trm.get_initial()
+
+    for _ in range(3):
+        loss, losses, outputs, latents, pred, halt = trm(seq, outputs, latents, labels = answer)
+
+        loss.backward()
+        optim.step()
+        optim.zero_grad()
+
+def test_trainer():
+    from torch.utils.data import Dataset
+    from tiny_recursive_model.mlp_mixer_1d import MLPMixer1D
+
+    trm = TinyRecursiveModel(
+        dim = 16,
+        num_tokens = 256,
+        network = MLPMixer1D(
+            dim = 16,
+            depth = 2,
+            seq_len = 256
+        ),
+    )
+
+    class MockDataset(Dataset):
+        def __len__(self):
+            return 16
+
+        def __getitem__(self, idx):
+            inp = torch.randint(0, 256, (256,))
+            out = torch.randint(0, 256, (256,))
+            return inp, out
+
+    trainer = Trainer(
+        trm,
+        MockDataset(),
+        epochs = 1,
+        batch_size = 16
+    )
+
+    trainer()

tiny_recursive_model-0.0.1/tiny_recursive_model/__init__.py

@@ -0,0 +1,4 @@
+from tiny_recursive_model.trm import (
+    TinyRecursiveModel,
+    Trainer
+)

tiny_recursive_model-0.0.1/tiny_recursive_model/mlp_mixer_1d.py

@@ -0,0 +1,46 @@
+from functools import partial
+
+from torch import nn
+from torch.nn import Module, LayerNorm
+from einops.layers.torch import Rearrange, Reduce
+
+pair = lambda x: x if isinstance(x, tuple) else (x, x)
+
+class PreNormResidual(Module):
+    def __init__(self, dim, fn):
+        super().__init__()
+        self.fn = fn
+        self.norm = LayerNorm(dim, bias = False)
+
+    def forward(self, x):
+        return self.fn(self.norm(x)) + x
+
+def FeedForward(dim, dim_hidden, dropout = 0., dense = nn.Linear):
+    return nn.Sequential(
+        dense(dim, dim_hidden),
+        nn.GELU(),
+        nn.Dropout(dropout),
+        dense(dim_hidden, dim),
+        nn.Dropout(dropout)
+    )
+
+def MLPMixer1D(*, dim, depth, seq_len, expansion_factor = 4, expansion_factor_token = 0.5, dropout = 0.):
+    chan_first, chan_last = partial(nn.Conv1d, kernel_size = 1), nn.Linear
+
+    return nn.Sequential(
+        *[nn.Sequential(
+            PreNormResidual(dim, FeedForward(seq_len, int(expansion_factor * dim), dropout, chan_first)),
+            PreNormResidual(dim, FeedForward(dim, int(expansion_factor_token * dim), dropout, chan_last))
+        ) for _ in range(depth)],
+        LayerNorm(dim, bias = False)
+    )
+
+# quick test
+
+if __name__ == '__main__':
+
+    import torch
+    tokens = torch.randn(1, 1024, 512)
+    mixer = MLPMixer1D(dim = 512, depth = 4, seq_len = 1024)
+
+    assert mixer(tokens).shape == tokens.shape

tiny_recursive_model-0.0.1/tiny_recursive_model/trm.py

@@ -0,0 +1,232 @@
+from __future__ import annotations
+from contextlib import nullcontext
+
+import torch
+from torch import nn
+import torch.nn.functional as F
+from torch.nn import Module, ModuleList
+from torch.optim import AdamW
+from torch.utils.data import Dataset, DataLoader
+
+from einops import rearrange, repeat
+from einops.layers.torch import Reduce, Rearrange
+
+# network related
+
+from x_transformers import Encoder
+from tiny_recursive_model.mlp_mixer_1d import MLPMixer1D
+
+# ema - apparently greatly helped with results
+
+from ema_pytorch import EMA
+
+# helpers
+
+def exists(v):
+    return v is not None
+
+def default(v, d):
+    return v if exists(v) else d
+
+def range_from_one(n):
+    return range(1, n + 1)
+
+def is_empty(t):
+    return t.numel() == 0
+
+# classes
+
+class TinyRecursiveModel(Module):
+    def __init__(
+        self,
+        *,
+        dim,
+        num_tokens,
+        network: Module,
+        num_refinement_blocks = 3,   # T in paper
+        num_latent_refinements = 6,  # n in paper - 1 output refinement per N latent refinements
+        halt_loss_weight = 1.
+    ):
+        super().__init__()
+        assert num_refinement_blocks > 1
+
+        self.input_embed = nn.Embedding(num_tokens, dim)
+        self.output_init_embed = nn.Parameter(torch.randn(dim) * 1e-2)
+        self.latent_init_embed = nn.Parameter(torch.randn(dim) * 1e-2)
+
+        self.network = network
+
+        self.num_latent_refinements = num_latent_refinements
+        self.num_refinement_blocks = num_refinement_blocks
+
+        # prediction heads
+
+        self.to_pred = nn.Linear(dim, num_tokens, bias = False)
+
+        self.to_halt_pred = nn.Sequential(
+            Reduce('b n d -> b d', 'mean'),
+            nn.Linear(dim, 1, bias = False),
+            nn.Sigmoid(),
+            Rearrange('... 1 -> ...')
+        )
+
+        self.halt_loss_weight = halt_loss_weight
+
+    def refine_latent_then_output_once(
+        self,
+        inputs,     # (b n d)
+        outputs,    # (b n d)
+        latents,    # (b n d)
+    ):
+
+        # so it seems for this work, they use only one network
+        # the network learns to refine the latents if input is passed in, otherwise it refines the output
+
+        for _ in range(self.num_latent_refinements):
+
+            latents = self.network(outputs + latents + inputs)
+
+        outputs = self.network(outputs + latents)
+
+        return outputs, latents
+
+    def get_initial(self):
+        outputs = self.output_init_embed
+        latents = self.latent_init_embed
+
+        return outputs, latents
+
+    def deep_refinement(
+        self,
+        inputs,    # (b n d)
+        outputs,   # (b n d)
+        latents,   # (b n d)
+    ):
+
+        for i in range(self.num_refinement_blocks):
+
+            # only last round of refinement receives gradients
+
+            is_last = i == (self.num_refinement_blocks - 1)
+            context = torch.no_grad if not is_last else nullcontext
+
+            with context():
+                outputs, latents = self.refine_latent_then_output_once(inputs, outputs, latents)
+
+        return outputs, latents
+
+    def forward(
+        self,
+        seq,
+        outputs,
+        latents,
+        labels = None
+    ):
+        inputs = self.input_embed(seq)
+
+        outputs, latents = self.deep_refinement(inputs, outputs, latents)
+
+        pred = self.to_pred(outputs)
+
+        should_halt = self.to_halt_pred(outputs)
+
+        outputs, latents = outputs.detach(), latents.detach()
+
+        return_package = (outputs, latents, pred, should_halt)
+
+        if not exists(labels):
+            return return_package
+
+        # calculate loss if labels passed in
+
+        loss = F.cross_entropy(rearrange(pred, 'b n l -> b l n'), labels)
+
+        is_all_correct = (pred.argmax(dim = -1) == labels).all(dim = -1)
+
+        halt_loss = F.binary_cross_entropy(should_halt, is_all_correct.float())
+
+        # total loss and loss breakdown
+
+        total_loss = loss + halt_loss * self.halt_loss_weight
+        losses = (loss, halt_loss)
+
+        return (total_loss, losses, *return_package)
+
+# trainer
+
+class Trainer(Module):
+    def __init__(
+        self,
+        model: TinyRecursiveModel | Module,
+        dataset: Dataset,
+        optim_klass = AdamW,
+        learning_rate = 1e-4,
+        weight_decay = 1.,
+        batch_size = 16,
+        epochs = 2,
+        halt_prob_thres = 0.5,
+        max_recurrent_steps = 12,
+        ema_decay_rate = 0.999,
+        ema_update_model_with_ema_every = 10000
+    ):
+        super().__init__()
+
+        self.batch_size = batch_size
+        self.epochs = epochs
+
+        self.dataset = dataset
+        self.dataloader = dataloader = DataLoader(self.dataset, batch_size = self.batch_size, shuffle = True)
+
+        self.optim = optim_klass(
+            model.parameters(),
+            lr = learning_rate,
+            weight_decay = weight_decay
+        )
+
+        self.model = model
+
+        self.ema_model = EMA(
+            model,
+            beta = ema_decay_rate,
+            update_model_with_ema_every = ema_update_model_with_ema_every
+        )
+
+        self.halt_prob_thres = halt_prob_thres
+
+        self.max_recurrent_steps = max_recurrent_steps
+
+    def forward(self):
+
+        for epoch in range_from_one(self.epochs):
+
+            for dataset_input, dataset_output in self.dataloader:
+
+                outputs, latents = self.model.get_initial()
+
+                for recurrent_step in range_from_one(self.max_recurrent_steps):
+
+                    loss, (main_loss, halt_loss), outputs, latents, pred, halt = self.model(dataset_input, outputs, latents, labels = dataset_output)
+
+                    print(f'[{epoch} ({recurrent_step} / {self.max_recurrent_steps})] loss: {main_loss.item():.3f} | halt loss: {halt_loss.item():.3f}')
+
+                    loss.backward()
+
+                    self.optim.step()
+                    self.optim.zero_grad()
+
+                    self.ema_model.update()
+
+                    # handle halting
+
+                    halt_mask = halt >= self.halt_prob_thres
+
+                    if not halt_mask.any():
+                        continue
+
+                    outputs = outputs[~halt_mask]
+                    latents = latents[~halt_mask]
+                    dataset_input = dataset_input[~halt_mask]
+                    dataset_output = dataset_output[~halt_mask]
+
+                    if is_empty(outputs):
+                        break