titans-pytorch 0.2.14__py3-none-any.whl → 0.2.16__py3-none-any.whl

titans_pytorch/mac_transformer.py

@@ -491,6 +491,7 @@ class MemoryAsContextTransformer(Module):
         neural_memory_layers: tuple[int, ...] | None = None,
         use_flex_attn = False,
         sliding_window_attn = False,
+        neural_mem_weight_residual = False
     ):
         super().__init__()
 
@@ -524,6 +525,8 @@ class MemoryAsContextTransformer(Module):
 
         neural_memory_layers = default(neural_memory_layers, layers)
 
+        self.neural_mem_weight_residual = neural_mem_weight_residual
+
         # mem, attn, and feedforward layers
 
         for layer in layers:
@@ -739,6 +742,10 @@ class MemoryAsContextTransformer(Module):
 
         value_residual = None
 
+        # neural mem weight residual
+
+        mem_weight_residual = None
+
         # when inferencing, only do one token at a time
 
         if is_inferencing:
@@ -761,16 +768,14 @@ class MemoryAsContextTransformer(Module):
 
                 mem_input, add_residual = mem_hyper_conn(x)
 
-                if not is_inferencing:
-                    retrieved, next_neural_mem_cache = mem(
-                        mem_input
-                    )
+                retrieved, next_neural_mem_cache = mem.forward(
+                    mem_input,
+                    state = next(neural_mem_caches, None),
+                    prev_weights = mem_weight_residual
+                )
 
-                else:
-                    (retrieved, next_neural_mem_cache) = mem.forward_inference(
-                        mem_input,
-                        state = next(neural_mem_caches, None),
-                    )
+                if self.neural_mem_weight_residual:
+                    mem_weight_residual = next_neural_mem_cache.updates
 
                 if self.gate_attn_output:
                     attn_out_gates = retrieved.sigmoid()

titans_pytorch/neural_memory.py

@@ -494,7 +494,8 @@ class NeuralMemory(Module):
         seq,
         weights: dict[str, Tensor] | None = None,
         past_state: tuple[dict[str, Tensor], dict[str, Tensor]] | None = None,
-        seq_index = 0
+        seq_index = 0,
+        prev_weights = None
     ):
         batch, seq_len, heads, chunk_size = *seq.shape[:2], self.heads, self.store_chunk_size
 
@@ -560,6 +561,12 @@ class NeuralMemory(Module):
 
         adaptive_lr = rearrange(adaptive_lr, 'b (n c) -> (b n) c', c = chunk_size)
 
+        # maybe add previous layer weight
+
+        if exists(prev_weights):
+            prev_weights = prev_weights.apply(lambda t: t[:, -1:])
+            weights_for_surprise = weights_for_surprise + prev_weights
+
         # flatten batch and time if surprise depends on previous layer memory model
 
         weights_for_surprise = rearrange_dict_values(weights_for_surprise, 'b n ... -> (b n) ...')
@@ -655,15 +662,19 @@ class NeuralMemory(Module):
         self,
         seq,
         past_weights: dict[str, Tensor],
+        chunk_size = None,
+        need_pad = True
     ):
-        chunk_size = self.retrieve_chunk_size
+        chunk_size = default(chunk_size, self.retrieve_chunk_size)
         batch, seq_len = seq.shape[:2]
 
         seq = self.retrieve_norm(seq)
 
-        needs_pad = chunk_size > 1
+        need_pad = need_pad or chunk_size > 1
+
+        if need_pad:
+            seq = pad_at_dim(seq, (1, 0), dim = 1)
 
-        seq = pad_at_dim(seq, (1, 0), dim = 1)
         seq_len_plus_one = seq.shape[-2]
 
         next_seq_len = round_up_multiple(seq_len_plus_one, chunk_size)
@@ -718,94 +729,37 @@ class NeuralMemory(Module):
 
         # restore, pad with empty memory embed
 
-        values = values[:, 1:(seq_len + 1)]
-
-        return values
-
-    @torch.no_grad()
-    def forward_inference(
-        self,
-        token: Tensor,
-        state: NeuralMemCache | None = None,
-    ):
-        # unpack previous state
-
-        if not exists(state):
-            state = (0, None, None, None, None)
-
-        seq_index, weights, cache_store_seq, past_states, updates = state
-
-        curr_seq_len = seq_index + 1
-        batch = token.shape[0]
-
-        if token.ndim == 2:
-            token = rearrange(token, 'b d -> b 1 d')
-
-        assert token.shape[1] == 1
-
-        # increment the sequence cache which is at most the chunk size
-
-        cache_store_seq = safe_cat((cache_store_seq, token), dim = -2)
-
-        # early return empty memory, when no memories are stored for steps < first chunk size
-
-        if curr_seq_len < self.chunk_size:
-            retrieve = self.retrieve_memories(token, weights, chunk_size = 1)
-
-            output = retrieve, NeuralMemCache(curr_seq_len, weights, cache_store_seq, past_states, updates)
-
-            return output
-
-        # store if storage sequence cache hits the chunk size
+        if need_pad:
+            values = values[:, 1:]
 
-        next_states = past_states
-        store_seq_cache_len = cache_store_seq.shape[-2]
-
-        if not exists(updates):
-            updates = weights.clone().zero_()
-            updates = repeat_dict_values(updates, '... -> b 1 ...', b = batch)
-        else:
-            updates = updates.apply(lambda t: t[:, -1:])
-
-        if store_seq_cache_len == self.chunk_size:
-
-            next_updates, store_state = self.store_memories(
-                cache_store_seq,
-                weights,
-                past_state = past_states,
-            )
-
-            updates = next_updates
-            cache_store_seq = None
-            next_states = store_state.states
-
-        # retrieve
-
-        retrieved = self.retrieve_memories(token, updates, chunk_size = 1)
-
-        # next state tuple
-
-        next_store_state = NeuralMemCache(curr_seq_len, weights, cache_store_seq, next_states, updates)
-
-        return retrieved, next_store_state
+        return values[:, :seq_len]
 
     def forward(
         self,
         seq,
         store_seq = None,
         state: NeuralMemCache | None = None,
+        prev_weights = None
     ):
+        if seq.ndim == 2:
+            seq = rearrange(seq, 'b d -> b 1 d')
+
+        is_single_token = seq.shape[1] == 1
+
         if not exists(state):
             state = (0, None, None, None, None)
 
         seq_index, weights, cache_store_seq, past_state, updates = state
 
-        assert not exists(cache_store_seq) or is_empty_tensor(cache_store_seq)
-
         # store
 
         store_seq = default(store_seq, seq)
 
+        # take care of cache
+
+        if exists(cache_store_seq):
+            store_seq = safe_cat((cache_store_seq, store_seq))
+
         # functions
 
         # compute split sizes of sequence
@@ -861,6 +815,7 @@ class NeuralMemory(Module):
                 weights,
                 seq_index = seq_index,
                 past_state = past_state,
+                prev_weights = prev_weights
             )
 
             seq_index = next_neural_mem_state.seq_index
@@ -883,9 +838,21 @@ class NeuralMemory(Module):
 
         # retrieve
 
+        need_pad = True
+        retrieve_chunk_size = None
+
+        if is_single_token:
+            retrieve_chunk_size = 1
+            need_pad = False
+
+            last_update, _ = past_state
+            updates = rearrange_dict_values(last_update, 'b ... -> b 1 ...')
+
         retrieved = self.retrieve_memories(
             seq,
-            updates
+            updates,
+            chunk_size = retrieve_chunk_size,
+            need_pad = need_pad,
         )
 
         return retrieved, next_neural_mem_state

{titans_pytorch-0.2.14.dist-info → titans_pytorch-0.2.16.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: titans-pytorch
-Version: 0.2.14
+Version: 0.2.16
 Summary: Titans
 Project-URL: Homepage, https://pypi.org/project/titans-pytorch/
 Project-URL: Repository, https://github.com/lucidrains/titans-pytorch

titans_pytorch-0.2.16.dist-info/RECORD

@@ -0,0 +1,9 @@
+titans_pytorch/__init__.py,sha256=Y3m_ZlpEqYwp-Md1ARhNGJxq8bQp8ty1o039nZOOJo0,276
+titans_pytorch/associative_scan.py,sha256=Y-iYqmFuG-NoCKu6kgql1mhowXTeJfyawi3eUIXamp0,2650
+titans_pytorch/mac_transformer.py,sha256=QhdSPgWntfWILMJ1t0xLKgvZfPZWu9vhzZWaesftaPg,24724
+titans_pytorch/memory_models.py,sha256=Q9SAIyAbStF5Tz0EhvRbn3yAdE3nk3xKc1ndieIe714,4671
+titans_pytorch/neural_memory.py,sha256=pOoDhYrUgQ5EaUyiEvwDu7mfcaH6Jqqod5NwIFLbD9U,25798
+titans_pytorch-0.2.16.dist-info/METADATA,sha256=TxyjTuJmP0o2NhrHmlzJCU3JivOA1rTY-xQp3Ir_igY,6812
+titans_pytorch-0.2.16.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+titans_pytorch-0.2.16.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
+titans_pytorch-0.2.16.dist-info/RECORD,,

titans_pytorch-0.2.14.dist-info/RECORD

@@ -1,9 +0,0 @@
-titans_pytorch/__init__.py,sha256=Y3m_ZlpEqYwp-Md1ARhNGJxq8bQp8ty1o039nZOOJo0,276
-titans_pytorch/associative_scan.py,sha256=Y-iYqmFuG-NoCKu6kgql1mhowXTeJfyawi3eUIXamp0,2650
-titans_pytorch/mac_transformer.py,sha256=RfJ1SvQH5_4PmlB7g-13wPAqYtCCUJxfmtaL0oBrRCU,24563
-titans_pytorch/memory_models.py,sha256=Q9SAIyAbStF5Tz0EhvRbn3yAdE3nk3xKc1ndieIe714,4671
-titans_pytorch/neural_memory.py,sha256=BGAzdxxZoKo8SmuPpSx1h1Yexf0oRje5qsCjl5r4FSA,26833
-titans_pytorch-0.2.14.dist-info/METADATA,sha256=SE27_ln7ludwaO2J1uX7TJ99MbnCl4GLorYpbBIK920,6812
-titans_pytorch-0.2.14.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-titans_pytorch-0.2.14.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
-titans_pytorch-0.2.14.dist-info/RECORD,,