tiny-recursive-model 0.0.1__py3-none-any.whl

tiny_recursive_model/__init__.py

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

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/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

tiny_recursive_model-0.0.1.dist-info/METADATA

@@ -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.dist-info/RECORD

@@ -0,0 +1,7 @@
+tiny_recursive_model/__init__.py,sha256=UufV6--ilPn4quRWyhvaFRMKRfHvfLsAmF9RU-L31rM,77
+tiny_recursive_model/mlp_mixer_1d.py,sha256=6ivDK9dgHdVl1axg2ayifJ7H5QI3hXptHnb6lfNrno0,1398
+tiny_recursive_model/trm.py,sha256=YwzTod4CeeXlbAiM-TBB7rEEHWsxnPxavaGiVCTPMEM,6350
+tiny_recursive_model-0.0.1.dist-info/METADATA,sha256=G-cM7okuLAiOxhofXoRh2Ih-bwYifcA3AAhmYmKo-v4,3107
+tiny_recursive_model-0.0.1.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+tiny_recursive_model-0.0.1.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
+tiny_recursive_model-0.0.1.dist-info/RECORD,,

tiny_recursive_model-0.0.1.dist-info/WHEEL

@@ -0,0 +1,4 @@
+Wheel-Version: 1.0
+Generator: hatchling 1.27.0
+Root-Is-Purelib: true
+Tag: py3-none-any

tiny_recursive_model-0.0.1.dist-info/licenses/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.