tweak film conditioning, and then make sure it is optional

Author: lucidrains

meshgpt_pytorch/meshgpt_pytorch.py

@@ -67,6 +67,9 @@ def default(v, d):
 def first(it):
     return it[0]
 
+def identity(t, *args, **kwargs):
+    return t
+
 def divisible_by(num, den):
     return (num % den) == 0
 
@@ -264,8 +267,8 @@ def forward(self, x, cond):
 
         # for initializing to identity
 
-        gamma = (1 + self.gamma_mult * gamma)
-        beta = beta * self.beta_mult
+        gamma = (1 + self.gamma_mult * gamma.tanh())
+        beta = beta.tanh() * self.beta_mult
 
         # classic film
 
@@ -1067,6 +1070,7 @@ def __init__(
         pad_id = -1,
         num_sos_tokens = None,
         condition_on_text = False,
+        text_cond_with_film = False,
         text_condition_model_types = ('t5',),
         text_condition_cond_drop_prob = 0.25,
         quads = False,
@@ -1124,8 +1128,8 @@ def __init__(
             dim_text = self.conditioner.dim_latent
             cross_attn_dim_context = dim_text
 
-            self.text_coarse_film_cond = FiLM(dim_text, dim)
-            self.text_fine_film_cond = FiLM(dim_text, dim_fine)
+            self.text_coarse_film_cond = FiLM(dim_text, dim) if text_cond_with_film else identity
+            self.text_fine_film_cond = FiLM(dim_text, dim_fine) if text_cond_with_film else identity
 
         # for summarizing the vertices of each face
 

meshgpt_pytorch/version.py

@@ -1 +1 @@
-__version__ = '1.2.14'
+__version__ = '1.2.15'