End of training

lr_scheduler/lr_scheduler.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6a7d0d9cabdbf4fb4d054a96278cf07983c703d71347b7308aa9dcf192d70138
+oid sha256:cc9a5d647349c541df37221f857f1528d69a64ee8921808d7169158f833bb344
 size 1076

optimizer/optimizer.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:731cb919eff7ff204a18d7a99a1b62717839c367823d744953745f379fbb3af6
-size 1149290750
+oid sha256:31f24ae8c9be356b80823017a6443f945de9d160bf797c00ddb5c588e5ddf595
+size 1149290878

scheduler/scheduler_config.json

@@ -1,6 +1,6 @@
 {
   "_class_name": "TorchSDE_PARAM",
-  "_config_class_name": "SDEParameterizedBaseLineConfig",
+  "_config_class_name": "SDEPolynomialConfig",
   "_diffusers_version": "0.25.0.dev0",
   "data_dimension": 3072
 }

scheduler/scheduler_config.py

@@ -8,37 +8,48 @@ from sympy import Matrix, Symbol
 import math
 from sde_redefined_param import SDEDimension
 @dataclass
-class SDEParameterizedBaseLineConfig:
+class SDEPolynomialConfig:
     name = "Custom"
-    variable = Symbol('t', nonnegative=True, real=True)
 
+    initial_variable_value = 0
+    max_variable_value = 1# math.inf
+    min_sample_value = 1e-6
+
+    variable = Symbol('t', nonnegative=True, real=True, domain=sympy.Interval(initial_variable_value, max_variable_value, left_open=False, right_open=False))
     drift_dimension = SDEDimension.SCALAR 
     diffusion_dimension = SDEDimension.SCALAR
     diffusion_matrix_dimension = SDEDimension.SCALAR 
 
-    # TODO (KLAUS): HANDLE THE PARAMETERS BEING Ø
-    drift_parameters = Matrix([sympy.symbols("f1", real=True)])
-    diffusion_parameters = Matrix([sympy.symbols("sigma_min sigma_max", real=True)])
-    
-    drift = 0
+    drift_degree = 20
+    diffusion_degree = 20
+
+    drift_parameters = Matrix([sympy.symbols(f"f:{drift_degree}", real=True, nonzero=True)])
+
+    diffusion_parameters = Matrix([sympy.symbols("l0", real=True, nonzero=True)])
 
-    sigma_min = sympy.Abs(diffusion_parameters[0]) #0.002 
-    sigma_max = sympy.Abs(diffusion_parameters[1]) #80
-    diffusion = sigma_min * (sigma_max/sigma_min)**variable * sympy.sqrt(2 * sympy.Abs(sympy.log(sigma_max/sigma_min))) 
+
+    @property
+    def drift(self): 
+        transformed_variable = self.variable
+        return -sympy.Abs(sum(sympy.HadamardProduct(Matrix([[transformed_variable**i for i in range(1,self.drift_degree+1)]]), self.drift_parameters).doit()))
+        
+
+    @property
+    def diffusion(self):
+        
+        return self.variable**(self.diffusion_parameters[0]**2)
 
     # TODO (KLAUS) : in the SDE SAMPLING CHANGING Q impacts how we sample z ~ N(0, Q*(delta t))
     diffusion_matrix = 1 
 
-    initial_variable_value = 0
-    max_variable_value = 1 # math.inf
-    min_sample_value = 0
 
     module = 'jax'
 
-    drift_integral_form=False
-    diffusion_integral_form=False
+    drift_integral_form=True
+    diffusion_integral_form=True
     diffusion_integral_decomposition = 'cholesky' # ldl
 
-    non_symbolic_parameters = {'diffusion': torch.tensor([0.002, 80.])}
+
 
     target = "epsilon" # x0
+    non_symbolic_parameters = {'drift': torch.ones(drift_degree)}

scheduler/sdeparameters.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5574e08de37650fa21ba29800959ee5ddb031e0e18d8b6a200240649e4afcc5c
-size 140
+oid sha256:8b59e815a0d5d903799db1da0cfaa61298044b04691741e1b76dd1404f6e6d38
+size 220

text_encoder/config.json

@@ -1,5 +1,5 @@
 {
-  "_name_or_path": "AltLuv/pokemon-test",
+  "_name_or_path": "AltLuv/pokemon-test-tti",
   "architectures": [
     "CLIPTextModel"
   ],

unet/diffusion_pytorch_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b8a074b69da1eb235189af21eaa1c0a22d5dc4a300b197e868a70f74537e80a8
+oid sha256:77aac723d4efa78d5f3f90d242c24cc384f2826c2459f7e2ee1b3e3dc2f8f022
 size 574476604