titans-pytorch 0.3.3__tar.gz → 0.3.5__tar.gz

{titans_pytorch-0.3.3 → titans_pytorch-0.3.5}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: titans-pytorch
-Version: 0.3.3
+Version: 0.3.5
 Summary: Titans
 Project-URL: Homepage, https://pypi.org/project/titans-pytorch/
 Project-URL: Repository, https://github.com/lucidrains/titans-pytorch

{titans_pytorch-0.3.3 → titans_pytorch-0.3.5}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "titans-pytorch"
-version = "0.3.3"
+version = "0.3.5"
 description = "Titans"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }

{titans_pytorch-0.3.3 → titans_pytorch-0.3.5}/tests/test_titans.py

@@ -31,6 +31,7 @@ def torch_default_dtype(dtype):
 @pytest.mark.parametrize('qk_rmsnorm', (False, True))
 @pytest.mark.parametrize('max_grad_norm', (None, 2.))
 @pytest.mark.parametrize('per_parameter_lr_modulation', (False, True))
+@pytest.mark.parametrize('per_head_learned_parameters', (False, True))
 def test_titans(
     seq_len,
     silu,
@@ -39,7 +40,8 @@ def test_titans(
     momentum,
     qk_rmsnorm,
     max_grad_norm,
-    per_parameter_lr_modulation
+    per_parameter_lr_modulation,
+    per_head_learned_parameters
 ):
     mem = NeuralMemory(
         dim = 16,
@@ -50,6 +52,7 @@ def test_titans(
         momentum = momentum,
         qk_rmsnorm = qk_rmsnorm,
         per_parameter_lr_modulation = per_parameter_lr_modulation,
+        per_head_learned_parameters = per_head_learned_parameters
     )
 
     seq = torch.randn(2, seq_len, 16)

{titans_pytorch-0.3.3 → titans_pytorch-0.3.5}/titans_pytorch/associative_scan.py

@@ -3,18 +3,39 @@ from typing import Callable
 
 import torch
 from torch import Tensor
+from torch.nn import Module
 import torch.nn.functional as F
 
+from einops import rearrange, repeat, reduce, pack, unpack
+
 # taken from S5-pytorch repository
 # https://github.com/i404788/s5-pytorch/blob/74e2fdae00b915a62c914bf3615c0b8a4279eb84/s5/jax_compat.py#L51-L134
 
 # helper functions
 
+def exists(v):
+    return v is not None
+
+def default(*args):
+    for arg in args:
+        if exists(arg):
+            return arg
+    return None
+
 def pad_at_dim(t, pad, dim = -1, value = 0.):
     dims_from_right = (- dim - 1) if dim < 0 else (t.ndim - dim - 1)
     zeros = ((0, 0) * dims_from_right)
     return F.pad(t, (*zeros, *pad), value = value)
 
+def pack_one_with_inverse(t, pattern):
+    packed, packed_shape = pack([t], pattern)
+
+    def inverse(out, inv_pattern = None):
+        inv_pattern = default(inv_pattern, pattern)
+        return unpack(out, packed_shape, inv_pattern)[0]
+
+    return packed, inverse
+
 # the operator that is needed
 
 @torch.jit.script
@@ -88,3 +109,69 @@ def _interleave(a, b):
     interleaved = torch.flatten(stacked, start_dim=1, end_dim=2)
 
     return interleaved[:, :output_axis_len]
+
+# associative scan wrapper around naive and accelerated version
+
+class AssocScan(Module):
+    def __init__(
+        self,
+        use_accelerated = False
+    ):
+        super().__init__()
+        self.use_accelerated = use_accelerated
+
+    def forward(
+        self,
+        gates,
+        inputs,
+        prev = None,
+        remove_prev = None
+    ):
+        remove_prev = default(remove_prev, exists(prev))
+
+        inputs, inverse_pack_weight_shape = pack_one_with_inverse(inputs, 'b n *')
+        gates, _ = pack_one_with_inverse(gates, 'b n *')
+
+        if exists(prev):
+            prev, _ = pack_one_with_inverse(prev, 'b *')
+
+        if exists(prev):
+            inputs, _ = pack([prev, inputs], 'b * d')
+            gates = pad_at_dim(gates, (1, 0), value = 1., dim = -2)
+
+        if not self.use_accelerated:
+            _, out = associative_scan(binary_operator, (gates, inputs))
+
+            if remove_prev:
+                out = out[:, 1:]
+
+            return inverse_pack_weight_shape(out)
+
+        from accelerated_scan.triton import scan as triton_scan
+        from accelerated_scan.warp import scan as warp_scan
+
+        scan = triton_scan if gates.is_cuda else warp_scan
+
+        def accelerate_scan_fn(gates, inputs):
+            gates = gates.expand_as(inputs)
+            gates, inputs = tuple(rearrange(t, 'b n d -> b d n') for t in (gates, inputs))
+
+            seq_len = gates.shape[-1]
+            next_power_two_seq_len = 2 ** max(5, int(math.ceil(math.log2(seq_len))))
+
+            gates = F.pad(gates, (0, next_power_two_seq_len - seq_len))
+            inputs = F.pad(inputs, (0, next_power_two_seq_len - seq_len))
+
+            outputs = scan(gates.contiguous(), inputs.contiguous())
+
+            outputs = outputs[..., :seq_len]
+            outputs = rearrange(outputs, 'b d n -> b n d')
+
+            return outputs
+
+        out = accelerate_scan_fn(gates, inputs)
+
+        if remove_prev:
+            out = out[:, 1:]
+
+        return inverse_pack_weight_shape(out)

{titans_pytorch-0.3.3 → titans_pytorch-0.3.5}/titans_pytorch/memory_models.py

@@ -30,6 +30,25 @@ class LayerNorm(Module):
 
         return self.ln(x) * (gamma + 1.)
 
+# norm + residual wrapper, as used in original TTT paper
+# but could be removed
+
+class ResidualNorm(Module):
+    def __init__(
+        self,
+        dim,
+        model: Module
+    ):
+        super().__init__()
+        self.norm = LayerNorm(dim)
+        self.model = model
+
+    def forward(self, x):
+
+        out = self.model(x)
+
+        return self.norm(out) + x
+
 # memory mlp proposed in TTT
 
 class MemoryMLP(Module):
@@ -45,8 +64,6 @@ class MemoryMLP(Module):
 
         self.weights = ParameterList([Parameter(torch.randn(dim_in, dim_out)) for dim_in, dim_out in zip(dims[:-1], dims[1:])])
 
-        self.ln = LayerNorm(dim)
-
         for weight in self.weights:
             nn.init.xavier_uniform_(weight)
 
@@ -54,8 +71,6 @@ class MemoryMLP(Module):
         self,
         x
     ):
-        residual = x
-
         for ind, weight in enumerate(self.weights):
             is_first = ind == 0
 
@@ -64,7 +79,7 @@ class MemoryMLP(Module):
 
             x = x @ weight
 
-        return self.ln(x) + residual
+        return x
 
 # memory mlp, but with gated residual + final projection
 
@@ -97,7 +112,6 @@ class GatedResidualMemoryMLP(Module):
         self,
         x
     ):
-        residual = x
 
         for weight1, weight2, to_gates in self.weights:
             res = x
@@ -111,9 +125,7 @@ class GatedResidualMemoryMLP(Module):
             gates = cat((branch_out, res), dim = -1) @ to_gates
             x = res.lerp(branch_out, gates.sigmoid())
 
-        out = x @ self.final_proj
-
-        return self.ln(out) + residual
+        return x @ self.final_proj
 
 # memory mlp with factorized weights
 # so can tradeoff capacity for smaller chunk sizes
@@ -143,7 +155,6 @@ class FactorizedMemoryMLP(Module):
         self,
         x
     ):
-        residual = x
 
         for ind, (weight1, weight2) in enumerate(self.weights):
             is_first = ind == 0
@@ -153,7 +164,7 @@ class FactorizedMemoryMLP(Module):
 
             x = x @ weight1 @ weight2
 
-        return self.ln(x) + residual
+        return x
 
 # improvised attention as memory module
 
@@ -182,7 +193,6 @@ class MemoryAttention(Module):
             nn.init.xavier_uniform_(weight)
 
     def forward(self, x):
-        residual = x
 
         wq, wk, wv, ffw1, ffw2 = self.weights
 
@@ -202,4 +212,4 @@ class MemoryAttention(Module):
         h = F.gelu(x @ ffw1)
         ff_out = h @ ffw2
 
-        return self.ln(attn_out + ff_out) + residual
+        return attn_out + ff_out

{titans_pytorch-0.3.3 → titans_pytorch-0.3.5}/titans_pytorch/neural_memory.py

@@ -8,19 +8,16 @@ from collections import namedtuple
 import torch
 from torch import nn, cat, tensor, Tensor
 import torch.nn.functional as F
-from torch.nn import Linear, Module, Parameter, ParameterList
+from torch.nn import Linear, Module, Parameter, ParameterList, ParameterDict
 from torch.func import functional_call, vmap, grad
 
 from tensordict import TensorDict
 
-from titans_pytorch.associative_scan import (
-    associative_scan,
-    binary_operator,
-    pad_at_dim
-)
+from titans_pytorch.associative_scan import AssocScan
 
 from titans_pytorch.memory_models import(
-    MemoryMLP
+    MemoryMLP,
+    ResidualNorm
 )
 
 import einx
@@ -79,8 +76,8 @@ def safe_cat(inputs, dim = -2):
 def is_empty_tensor(t):
     return t.numel() == 0
 
-def dict_get_shape(td):
-    return {k: v.shape for k, v in td.items()}
+def dict_get_value_shapes(td):
+    return [v.shape for k, v in td.items()]
 
 def rearrange_dict_values(td, pattern, **kwargs):
     return td.apply(lambda t: rearrange(t, pattern, **kwargs))
@@ -97,6 +94,11 @@ def round_down_multiple(seq, mult):
 def round_up_multiple(seq, mult):
     return math.ceil(seq / mult) * mult
 
+def pad_at_dim(t, pad, dim = -1, value = 0.):
+    dims_from_right = (- dim - 1) if dim < 0 else (t.ndim - dim - 1)
+    zeros = ((0, 0) * dims_from_right)
+    return F.pad(t, (*zeros, *pad), value = value)
+
 def pack_one_with_inverse(t, pattern):
     packed, packed_shape = pack([t], pattern)
 
@@ -197,72 +199,6 @@ class AttentionPool(Module):
 
         return reduce(x * attn, 'b n c d -> b n d', 'sum')
 
-# associative scan wrapper
-
-class AssocScan(Module):
-    def __init__(
-        self,
-        use_accelerated = False
-    ):
-        super().__init__()
-        self.use_accelerated = use_accelerated
-
-    def forward(
-        self,
-        gates,
-        inputs,
-        prev = None,
-        remove_prev = None
-    ):
-        remove_prev = default(remove_prev, exists(prev))
-
-        inputs, inverse_pack_weight_shape = pack_one_with_inverse(inputs, 'b n *')
-        gates, _ = pack_one_with_inverse(gates, 'b n *')
-
-        if exists(prev):
-            prev, _ = pack_one_with_inverse(prev, 'b *')
-
-        if exists(prev):
-            inputs, _ = pack([prev, inputs], 'b * d')
-            gates = pad_at_dim(gates, (1, 0), value = 1., dim = -2)
-
-        if not self.use_accelerated:
-            _, out = associative_scan(binary_operator, (gates, inputs))
-
-            if remove_prev:
-                out = out[:, 1:]
-
-            return inverse_pack_weight_shape(out)
-
-        from accelerated_scan.triton import scan as triton_scan
-        from accelerated_scan.warp import scan as warp_scan
-
-        scan = triton_scan if gates.is_cuda else warp_scan
-
-        def accelerate_scan_fn(gates, inputs):
-            gates = gates.expand_as(inputs)
-            gates, inputs = tuple(rearrange(t, 'b n d -> b d n') for t in (gates, inputs))
-
-            seq_len = gates.shape[-1]
-            next_power_two_seq_len = 2 ** max(5, int(math.ceil(math.log2(seq_len))))
-
-            gates = F.pad(gates, (0, next_power_two_seq_len - seq_len))
-            inputs = F.pad(inputs, (0, next_power_two_seq_len - seq_len))
-
-            outputs = scan(gates.contiguous(), inputs.contiguous())
-
-            outputs = outputs[..., :seq_len]
-            outputs = rearrange(outputs, 'b d n -> b n d')
-
-            return outputs
-
-        out = accelerate_scan_fn(gates, inputs)
-
-        if remove_prev:
-            out = out[:, 1:]
-
-        return inverse_pack_weight_shape(out)
-
 # main neural memory
 
 def default_adaptive_step_transform(adaptive_step, max_lr = 1e-2):
@@ -285,6 +221,7 @@ class NeuralMemory(Module):
         default_step_transform_max_lr = 1.,
         per_parameter_lr_modulation = False, # allow outer network to control learning rate per weight matrix of memory network
         max_mem_layer_modulation = 1., # max of 10.
+        per_head_learned_parameters = True,
         attn_pool_chunks = False,
         momentum = True,
         pre_rmsnorm = True,
@@ -298,6 +235,7 @@ class NeuralMemory(Module):
         init_decay_bias = None,
         accept_weight_residual = False,
         gated_transition = False,
+        mem_model_norm_add_residual = True, # by default, layernorm output and add residual as proposed in TTT paper, but could be removed
         default_model_kwargs: dict = dict(
             depth = 2,
             expansion_factor = 4.
@@ -368,11 +306,26 @@ class NeuralMemory(Module):
 
         # the memory is the weights of the model
 
+        if mem_model_norm_add_residual:
+            model = ResidualNorm(dim = dim_head, model = model)
+
         self.memory_model = model
 
-        self.num_memory_parameter_tensors = len(set(model.parameters()))
+        mem_model_params = dict(model.named_parameters())
+
+        self.num_memory_parameter_tensors = len(mem_model_params)
 
-        self.init_weight_shape = dict_get_shape(dict(model.named_parameters()))
+        self.memory_model_parameter_names = [*mem_model_params.keys()]
+
+        memory_model_parameters = [*mem_model_params.values()]
+
+        if per_head_learned_parameters:
+            memory_model_parameters = [repeat(p, '... -> h ...', h = heads) for p in memory_model_parameters]
+
+        self.init_weight_shape = [p.shape for p in memory_model_parameters]
+
+        self.memory_model_parameters = ParameterList(memory_model_parameters)
+        self.per_head_learned_parameters = per_head_learned_parameters
 
         # the chunk size within the paper where adaptive step, momentum, weight decay are shared
 
@@ -488,21 +441,32 @@ class NeuralMemory(Module):
 
         self.register_buffer('zero', torch.tensor(0.), persistent = False)
 
+    @property
+    def memory_model_parameter_dict(self):
+        return TensorDict(dict(zip(self.memory_model_parameter_names, self.memory_model_parameters)))
+
     def init_weights(
         self,
         batch,
     ):
-        weights = TensorDict(dict(self.memory_model.named_parameters()))
-        weights = repeat_dict_values(weights, '... -> bh ...', bh = batch * self.heads)
+        if self.per_head_learned_parameters:
+            weights = repeat_dict_values(self.memory_model_parameter_dict, 'h ... -> (b h) ...', b = batch)
+        else:
+            weights = repeat_dict_values(self.memory_model_parameter_dict, '... -> bh ...', bh = batch * self.heads)
+
         return weights
 
     def init_momentum(
         self,
         batch,
     ):
-        weights = TensorDict(dict(self.memory_model.named_parameters()))
-        zeros = weights.clone().zero_()
-        zeros = repeat_dict_values(zeros, '... -> bh ...', bh = batch * self.heads)
+        zeros = self.memory_model_parameter_dict.clone().zero_()
+
+        if self.per_head_learned_parameters:
+            zeros = repeat_dict_values(zeros, 'h ... -> (b h) ...', b = batch)
+        else:
+            zeros = repeat_dict_values(zeros, '... -> bh ...', bh = batch * self.heads)
+
         return zeros
 
     def store_memories(
@@ -694,7 +658,7 @@ class NeuralMemory(Module):
     ):
         chunk_size = self.retrieve_chunk_size
 
-        weights_have_expanded_shape = dict_get_shape(weights) != self.init_weight_shape
+        weights_have_expanded_shape = dict_get_value_shapes(weights) != self.init_weight_shape
 
         batch, seq_len = seq.shape[:2]
 

{titans_pytorch-0.3.3 → titans_pytorch-0.3.5}/train_mac.py

@@ -10,7 +10,11 @@ from torch.utils.data import DataLoader, Dataset
 
 from adam_atan2_pytorch import AdoptAtan2
 
-from titans_pytorch import MemoryAsContextTransformer, MemoryMLP
+from titans_pytorch import (
+    MemoryAsContextTransformer,
+    MemoryMLP,
+    MemoryAttention
+)
 
 # constants
 
@@ -35,6 +39,7 @@ NEURAL_MEM_GATE_ATTN_OUTPUT = False
 NEURAL_MEM_MOMENTUM = True
 NEURAL_MEM_QK_NORM = True
 NEURAL_MEM_MAX_LR = 1e-1
+USE_MEM_ATTENTION_MODEL = False
 WINDOW_SIZE = 32
 NEURAL_MEM_SEGMENT_LEN = 4                      # set smaller for more granularity for learning rate / momentum etc
 NEURAL_MEM_BATCH_SIZE = 128                     # set smaller to update the neural memory weights more often as it traverses the sequence
@@ -75,6 +80,18 @@ def decode_token(token):
 def decode_tokens(tokens):
     return ''.join(list(map(decode_token, tokens)))
 
+# memory model
+
+if USE_MEM_ATTENTION_MODEL:
+    neural_memory_model = MemoryAttention(
+        dim = 64
+    )
+else:
+    neural_memory_model = MemoryMLP(
+        dim = 64,
+        depth = NEURAL_MEMORY_DEPTH
+    )
+
 # instantiate memory-as-context transformer
 
 model = MemoryAsContextTransformer(
@@ -91,10 +108,7 @@ model = MemoryAsContextTransformer(
     neural_mem_weight_residual = NEURAL_MEM_WEIGHT_RESIDUAL,
     use_flex_attn = USE_FLEX_ATTN,
     sliding_window_attn = SLIDING_WINDOWS,
-    neural_memory_model = MemoryMLP(
-        dim = 64,
-        depth = NEURAL_MEMORY_DEPTH
-    ),
+    neural_memory_model = neural_memory_model,
     neural_memory_kwargs = dict(
         dim_head = 64,
         heads = 4,