x-transformers 2.8.0__py3-none-any.whl → 2.8.2__py3-none-any.whl

x_transformers/gpt_vae.py

@@ -46,32 +46,39 @@ class GPTVAE(Module):
         vae_kl_loss_weight = 1.,
         latents_dropout_prob = 0.5, # what percentage of the time to dropout the latents completely
         pad_id = -1,
+        encoder: Module | None = None,
         **kwargs
     ):
         super().__init__()
         dim_latent = default(dim_latent, dim)
 
-        self.encoder = TransformerWrapper(
-            num_tokens = num_tokens,
-            max_seq_len = max_seq_len + 1,
-            return_only_embed = True,
-            average_pool_embed = True,
-            attn_layers = Encoder(
-                dim = dim,
-                depth = enc_depth,
-                attn_dim_head = attn_dim_head,
-                heads = heads,
-                **kwargs,
-                **enc_kwargs
-            ),
-        )
+        if not exists(encoder):
+            encoder = TransformerWrapper(
+                num_tokens = num_tokens,
+                max_seq_len = max_seq_len + 1,
+                return_only_embed = True,
+                average_pool_embed = True,
+                attn_layers = Encoder(
+                    dim = dim,
+                    depth = enc_depth,
+                    attn_dim_head = attn_dim_head,
+                    heads = heads,
+                    **kwargs,
+                    **enc_kwargs
+                ),
+            )
+
+        self.encoder = encoder
 
         self.to_latent_mean_log_variance = nn.Sequential(
             nn.Linear(dim, dim_latent * 2),
-            Rearrange('b (two d) -> two b 1 d', two = 2)
+            Rearrange('b (two d) -> two b d', two = 2)
         )
 
-        self.from_latent_to_prepend_token = nn.Linear(dim_latent, dim)
+        self.from_latent_to_prepend_token = nn.Sequential(
+            nn.Linear(dim_latent, dim),
+            Rearrange('b d -> b 1 d')
+        )
 
         self.decoder = TransformerWrapper(
             num_tokens = num_tokens,
@@ -126,11 +133,19 @@ class GPTVAE(Module):
         prompts,
         seq_len,
         latents = None,
+        seq_for_latents = None,
         **generate_kwargs
     ):
         assert prompts.ndim in {1, 2}
         batch = prompts.shape[0] if prompts.ndim == 2 else 1
 
+        # if seq_for_latents passed in, derive latents from it
+
+        if exists(seq_for_latents):
+            assert not exists(latents), 'latents should not be passed in if given the seq from which to derive them'
+
+            latents = self.encode_to_latents(seq_for_latents)
+
         # prepend embeds
 
         prepend_embeds = None
@@ -143,9 +158,6 @@ class GPTVAE(Module):
 
             prepend_embeds = self.from_latent_to_prepend_token(latents)
 
-        if exists(prepend_embeds):
-            prepend_embeds = rearrange(prepend_embeds, 'b d -> b 1 d')
-
         # generated
 
         generated = self.ar_wrapped_decoder.generate(

{x_transformers-2.8.0.dist-info → x_transformers-2.8.2.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: x-transformers
-Version: 2.8.0
+Version: 2.8.2
 Summary: X-Transformers
 Project-URL: Homepage, https://pypi.org/project/x-transformers/
 Project-URL: Repository, https://github.com/lucidrains/x-transformers

{x_transformers-2.8.0.dist-info → x_transformers-2.8.2.dist-info}/RECORD

@@ -5,7 +5,7 @@ x_transformers/belief_state_wrapper.py,sha256=YLUMk6t2MhFBEw5lHDDHJHcoCxTIkHvxTN
 x_transformers/continuous.py,sha256=WwpQCjyVY4PtuEAOFY68zqgklbF9I7AL5w6874YlDe8,13249
 x_transformers/dpo.py,sha256=xt4OuOWhU8pN3OKN2LZAaC2NC8iiEnchqqcrPWVqf0o,3521
 x_transformers/entropy_based_tokenizer.py,sha256=F2lO8-v3aLIcVDVNhu7RR-UtRdlmaaYJzBK9m7OnLE8,5018
-x_transformers/gpt_vae.py,sha256=yqL1K2yJ6RSP_MC6XSHI3hjiUnaptddg6CUnbEX4Bsk,5281
+x_transformers/gpt_vae.py,sha256=myYSgcx66V0M4zeEGKyhY1P2HlPDHcezhaZEoo_uMdo,5715
 x_transformers/multi_input.py,sha256=tCh-fTJDj2ib4SMGtsa-AM8MxKzJAQSwqAXOu3HU2mg,9252
 x_transformers/neo_mlp.py,sha256=XCNnnop9WLarcxap1kGuYc1x8GHvwkZiDRnXOxSl3Po,3452
 x_transformers/nonautoregressive_wrapper.py,sha256=hMQqNimGtchNIe13cR5LZule1V7I1qM5LmY8VQfVdnA,11698
@@ -13,7 +13,7 @@ x_transformers/up_wrapper.py,sha256=YC2LN14_7Xx9Wtiek2rtEJ_qHqdfSmKlh3d7Cgxwd80,
 x_transformers/x_transformers.py,sha256=odnCZAKZKrQLXmpaWhiPVB5elGjt8kerDbO3-yeC-60,124764
 x_transformers/xl_autoregressive_wrapper.py,sha256=CvZMJ6A6PA-Y_bQAhnORwjJBSl6Vjq2IdW5KTdk8NI8,4195
 x_transformers/xval.py,sha256=AwwYUm8yDAtKQyKJDIhYMsiLTJ_skh3scUFMjp5sda8,8597
-x_transformers-2.8.0.dist-info/METADATA,sha256=jPo0ZPhD1d_aocaDqFYWXA7EXPAcxWeUYNDzKpY1yi8,94136
-x_transformers-2.8.0.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-x_transformers-2.8.0.dist-info/licenses/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
-x_transformers-2.8.0.dist-info/RECORD,,
+x_transformers-2.8.2.dist-info/METADATA,sha256=CxS0kKTbwKj2sX0HexDhQksRAYGc-Gj3fLLqpnbu5zM,94136
+x_transformers-2.8.2.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+x_transformers-2.8.2.dist-info/licenses/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
+x_transformers-2.8.2.dist-info/RECORD,,