titans-pytorch 0.0.16__tar.gz → 0.0.17__tar.gz

{titans_pytorch-0.0.16 → titans_pytorch-0.0.17}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: titans-pytorch
-Version: 0.0.16
+Version: 0.0.17
 Summary: Titans
 Project-URL: Homepage, https://pypi.org/project/titans-pytorch/
 Project-URL: Repository, https://github.com/lucidrains/titans-pytorch

{titans_pytorch-0.0.16 → titans_pytorch-0.0.17}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "titans-pytorch"
-version = "0.0.16"
+version = "0.0.17"
 description = "Titans"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }

titans_pytorch-0.0.17/titans_pytorch/titans_attn_memory.py

@@ -0,0 +1,419 @@
+from __future__ import annotations
+import math
+from functools import partial
+
+import torch
+from torch import nn, Tensor
+import torch.nn.functional as F
+from torch.nn import Linear, Module
+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
+)
+
+import einx
+from einops import rearrange, pack, unpack
+from einops.layers.torch import Rearrange, Reduce
+
+"""
+ein notation:
+b - batch
+n - sequence
+d - feature dimension
+c - intra-chunk
+"""
+
+# constants
+
+LinearNoBias = partial(Linear, bias = False)
+
+# functions
+
+def exists(v):
+    return v is not None
+
+def default(v, d):
+    return v if exists(v) else d
+
+def round_down_multiple(seq, mult):
+    return seq // mult * mult
+
+def round_up_multiple(seq, mult):
+    return math.ceil(seq / mult) * mult
+
+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
+
+# classes
+
+# improvised attention as memory module
+# todo - expand if see signal in experiments (update: not seeing it)
+
+class MemoryAttention(Module):
+    def __init__(
+        self,
+        dim
+    ):
+        super().__init__()
+        self.weights = nn.ParameterList([
+            nn.Parameter(torch.randn(dim, dim)), # queries
+            nn.Parameter(torch.randn(dim, dim)), # keys
+            nn.Parameter(torch.randn(dim, dim)), # values weight 1
+            nn.Parameter(torch.randn(dim, dim)), # values weight 2
+        ])
+
+    def forward(self, x):
+
+        assert x.shape[-2] > 1, 'chunk size needs to be greater than 1 for using attention as memory'
+
+        wq, wk, wv1, wv2 = self.weights
+
+        q = x @ wq
+        k = x @ wk
+        v = x @ wv1
+
+        hidden = F.scaled_dot_product_attention(
+            q, k, v,
+            is_causal = True
+        )
+
+        return F.silu(hidden) @ wv2
+
+# main neural memory
+
+def default_loss_fn(pred, target):
+    return (pred - target).pow(2).mean(dim = -1).sum()
+
+class NeuralMemory(Module):
+    def __init__(
+        self,
+        dim,
+        chunk_size = 1,
+        dim_head = None,
+        heads = 1,
+        model: MemoryAttention | None = None,
+        store_memory_loss_fn: Callable = default_loss_fn,
+        pre_rmsnorm = True,
+        post_rmsnorm = True,
+        use_accelerated_scan = False,
+        default_model_kwargs: dict = dict()
+    ):
+        super().__init__()
+
+        # norms
+
+        self.retrieve_norm = nn.RMSNorm(dim) if pre_rmsnorm else nn.Identity()
+        self.store_norm = nn.RMSNorm(dim) if pre_rmsnorm else nn.Identity()
+
+        self.post_rmsnorm = nn.RMSNorm(dim) if post_rmsnorm else nn.Identity()
+
+        # maybe multi-headed
+
+        dim_head = default(dim_head, dim)
+        dim_inner = dim_head * heads
+
+        self.split_heads = Rearrange('b n (h d) -> (b h) n d', h = heads)
+        self.merge_heads = Rearrange('(b h) n d -> b n (h d)', h = heads)
+        self.combine_heads = LinearNoBias(dim_inner, dim) if heads > 1 else nn.Identity()
+
+        # memory mlp
+
+        if not exists(model):
+            model = MemoryAttention(dim_head, **default_model_kwargs)
+
+        assert not exists(next(model.buffers(), None)), 'model cannot have buffers for now'
+
+        # the memory is the weights of the model
+
+        self.memory_model = model
+
+        # the chunk size within the paper where adaptive step, momentum, weight decay are shared
+
+        self.chunk_size = chunk_size
+
+        # prepare function for per sample gradients from model above, using torch.func
+
+        def forward_and_loss(params, inputs, target):
+            pred = functional_call(self.memory_model, params, inputs)
+            loss = self.store_memory_loss_fn(pred, target) # simple mse loss in paper - eq (12) - |M(k) - v|²
+            return loss
+
+        self.per_sample_grad_fn = vmap(grad(forward_and_loss), in_dims = (None, 0, 0))
+
+        # queries for retrieving from the model
+
+        self.to_queries = LinearNoBias(dim, dim_inner)
+
+        # keys and values for storing to the model
+
+        self.to_keys_values = LinearNoBias(dim, dim_inner * 2)
+        self.store_memory_loss_fn = store_memory_loss_fn
+
+        # learned adaptive learning rate and momentum
+        # todo - explore mlp layerwise learned lr / momentum
+
+        self.to_momentum = nn.Sequential(
+            Reduce('b (n c) ... -> b n ...', 'mean', c = chunk_size),
+            LinearNoBias(dim, heads),
+            Rearrange('b n h -> (b h) n 1')
+        )
+
+        self.to_adaptive_step = nn.Sequential(
+            Reduce('b (n c) ... -> b n ...', 'mean', c = chunk_size),
+            LinearNoBias(dim, heads),
+            Rearrange('b n h -> (b h) n')
+        )
+
+        # weight decay factor
+
+        self.to_decay_factor = nn.Sequential(
+            Reduce('b (n c) ... -> b n ...', 'mean', c = chunk_size),
+            LinearNoBias(dim, heads),
+            Rearrange('b n h -> (b h) n 1')
+        )
+
+        # maybe use accelerated scan
+
+        self.use_accelerated_scan = use_accelerated_scan
+
+    def init_weights_and_momentum(self):
+        params = TensorDict(dict(self.memory_model.named_parameters()))
+
+        init_weights = params.clone().zero_()
+        init_momentum = params.clone().zero_()
+
+        return init_weights, init_momentum
+
+    def store_memories(
+        self,
+        seq,
+        past_state: tuple[dict[str, Tensor], dict[str, Tensor]]
+    ):
+
+        seq = self.store_norm(seq)
+
+        # curtail sequence by multiple of the chunk size
+        # only a complete chunk of the sequence provides the memory for the next chunk
+
+        seq_len, chunk_size = seq.shape[-2], self.chunk_size
+        round_down_seq_len = round_down_multiple(seq_len, self.chunk_size)
+
+        seq = seq[:, :round_down_seq_len]
+
+        # curr weights + past weights, in the case that the initial weights are learned
+
+        curr_weights = TensorDict(dict(self.memory_model.named_parameters()))
+
+        past_state = tuple(TensorDict(d) for d in past_state)
+        past_weights, past_momentum = past_state
+
+        curr_weights = curr_weights + past_weights
+
+        # pack batch and sequence dimension
+
+        adaptive_lr = (self.to_adaptive_step(seq).sigmoid() * -15).exp() # from 1. - 1e-7
+
+        adaptive_momentum = self.to_momentum(seq).sigmoid()
+        decay_factor = self.to_decay_factor(seq).sigmoid()
+
+        # keys and values
+
+        keys, values = self.to_keys_values(seq).chunk(2, dim = -1)
+
+        # maybe multi head
+
+        keys, values = map(self.split_heads, (keys, values))
+
+        batch = keys.shape[0]
+
+        # take care of chunking
+
+        keys, values = tuple(rearrange(t, 'b (n c) d -> (b n) c d', c = self.chunk_size) for t in (keys, values))
+
+        # get grads and extra auxiliary loss (for backwarding through qkv projection in base neural memory module)
+
+        grads = self.per_sample_grad_fn(dict(curr_weights), keys, values)
+
+        grads = TensorDict(grads)
+
+        # restore batch and sequence dimension
+
+        grads = grads.apply(lambda t: rearrange(t, '(b n) ... -> b n ...', b = batch))
+
+        # multiply gradients with learned adaptive step size
+
+        surprises = grads.apply(lambda t: einx.multiply('b n ..., b n -> b n ...', t, -adaptive_lr))
+
+        # determine scan function
+
+        def default_associative_scan(gates, inputs):
+            _, outputs = associative_scan(binary_operator, (gates, inputs))
+            return outputs
+
+        if self.use_accelerated_scan:
+            from accelerated_scan.triton import scan as triton_scan
+            from accelerated_scan.warp import scan as warp_scan
+
+            scan = triton_scan if seq.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, inputs)
+
+                outputs = outputs[..., :seq_len]
+                outputs = rearrange(outputs, 'b d n -> b n d')
+                return outputs
+
+            scan_fn = accelerate_scan_fn
+        else:
+            scan_fn = default_associative_scan
+
+        # momentum + weight decay - momentum is the new contribution, as most linear RNNs have learned forgetting gates
+
+        next_momentum = TensorDict()
+        updates = TensorDict()
+
+        for param_name, surprise in surprises.items():
+
+            surprise, inverse_pack = pack_one_with_inverse(surprise, 'b n *')
+
+            # derive momentum with associative scan - eq (10)
+
+            momentum = scan_fn(adaptive_momentum, surprise) # momentum is S / surprise in the paper
+
+            # use associative scan again for learned forgetting (weight decay) - eq (13)
+
+            update = scan_fn(1. - decay_factor, momentum) # momentum is S / surprise in the paper
+
+            updates[param_name] = inverse_pack(update)
+            next_momentum[param_name] = inverse_pack(momentum)
+
+        # compute the next weight per batch
+
+        last_update = updates.apply(lambda t: t[:, -1])
+
+        next_state = (curr_weights + last_update, next_momentum)
+
+        return updates, next_state
+
+    def retrieve_memories(
+        self,
+        seq,
+        past_weights: dict[str, Tensor] | None = None,
+    ):
+        chunk_size = self.chunk_size
+        seq_len = seq.shape[1]
+
+        seq = self.retrieve_norm(seq)
+
+        assert seq_len > chunk_size
+
+        seq = seq[:, chunk_size:]
+        curtailed_seq_len = seq.shape[-2]
+
+        next_seq_len = round_up_multiple(curtailed_seq_len + 1, chunk_size)
+
+        padding = next_seq_len - curtailed_seq_len
+
+        seq = pad_at_dim(seq, (0, padding), dim = 1)
+
+        # the parameters of the memory model stores the memories of the key / values
+        # when the MLP has only 1 weight matrix, it is equivalent to `kv` fast weight memories from linear attention literature (recall fetching of memories is q @ (kv)) / schmidhuber's paper
+
+        curr_weights = TensorDict(dict(self.memory_model.named_parameters()))
+
+        if exists(past_weights):
+            past_weights = TensorDict(past_weights)
+            assert past_weights.keys() == curr_weights.keys()
+
+            curr_weights = curr_weights + past_weights
+
+        # sequence Float['b n d'] to queries
+
+        queries = self.to_queries(seq)
+
+        # maybe multihead
+
+        queries = self.split_heads(queries)
+
+        batch = queries.shape[0]
+
+        # fetch values from memory model
+
+        curr_weights = curr_weights.apply(lambda t: rearrange(t, 'b n ... -> (b n) ...'))
+        queries = rearrange(queries, 'b (n c) d -> (b n) c d', c = chunk_size)
+
+        # forward functional call
+
+        values = functional_call(self.memory_model, dict(curr_weights), queries)
+
+        # reconstitute batch dimension
+
+        values = rearrange(values, '(b n) c d -> b (n c) d', b = batch)
+
+        # maybe merge heads and combine
+
+        values = self.merge_heads(values)
+
+        values = self.combine_heads(values)
+
+        # post norm, somehow could not stabilize this without it, not in paper
+
+        values = self.post_rmsnorm(values)
+
+        # restore
+
+        values = pad_at_dim(values, (chunk_size, 0), dim = 1, value = 0.) # todo, used a learned null memory embedding instead of 0s for retrieving from empty neural memory
+        values = values[:, :-padding]
+
+        return values
+
+    def forward(
+        self,
+        seq,
+        store_seq = None,
+        past_state: tuple[dict[str, Tensor], dict[str, Tensor]] | None = None,
+        return_next_memories = False
+    ):
+        batch, seq_len = seq.shape[:2]
+
+        if seq_len <= self.chunk_size:
+            return torch.zeros_like(seq)
+
+        if exists(past_state):
+            past_state = tuple(TensorDict(d) for d in past_state)
+
+        if not exists(past_state):
+            past_state = self.init_weights_and_momentum()
+
+        store_seq = default(store_seq, seq)
+
+        updates, next_memories = self.store_memories(store_seq, past_state)
+
+        past_weights, _ = past_state
+
+        retrieved = self.retrieve_memories(seq, past_weights + updates)
+
+        if not return_next_memories:
+            return retrieved
+
+        return retrieved, next_memories