SURE-tools 2.2.2__py3-none-any.whl → 2.4.3__py3-none-any.whl

SURE/{PerturbFlow.py → PerturbE.py}

@@ -10,7 +10,7 @@ from torch.distributions.utils import logits_to_probs, probs_to_logits, clamp_pr
 from torch.distributions import constraints
 from torch.distributions.transforms import SoftmaxTransform
 
-from .utils.custom_mlp import MLP, Exp, ZeroBiasMLP
+from .utils.custom_mlp import MLP, Exp, ZeroBiasMLP2
 from .utils.utils import CustomDataset, CustomDataset2, CustomDataset4, tensor_to_numpy, convert_to_tensor
 
 
@@ -54,7 +54,7 @@ def set_random_seed(seed):
     # Set seed for Pyro
     pyro.set_rng_seed(seed)
 
-class PerturbFlow(nn.Module):
+class PerturbE(nn.Module):
     def __init__(self,
                  input_size: int,
                  codebook_size: int = 200,
@@ -62,10 +62,10 @@ class PerturbFlow(nn.Module):
                  supervised_mode: bool = False,
                  z_dim: int = 10,
                  z_dist: Literal['normal','studentt','laplacian','cauchy','gumbel'] = 'gumbel',
-                 loss_func: Literal['negbinomial','poisson','multinomial','bernoulli'] = 'poisson',
+                 loss_func: Literal['negbinomial','poisson','multinomial','bernoulli'] = 'multinomial',
                  inverse_dispersion: float = 10.0,
                  use_zeroinflate: bool = False,
-                 hidden_layers: list = [300],
+                 hidden_layers: list = [500],
                  hidden_layer_activation: Literal['relu','softplus','leakyrelu','linear'] = 'relu',
                  nn_dropout: float = 0.1,
                  post_layer_fct: list = ['layernorm'],
@@ -73,7 +73,6 @@ class PerturbFlow(nn.Module):
                  config_enum: str = 'parallel',
                  use_cuda: bool = True,
                  seed: int = 42,
-                 zero_bias: bool|list = True,
                  dtype = torch.float32, # type: ignore
                  ):
         super().__init__()
@@ -97,11 +96,6 @@ class PerturbFlow(nn.Module):
         self.post_layer_fct = post_layer_fct
         self.post_act_fct = post_act_fct
         self.hidden_layer_activation = hidden_layer_activation
-        if type(zero_bias) == list:
-            self.use_bias = [not x for x in zero_bias]
-        else:
-            self.use_bias = [not zero_bias] * self.cell_factor_size
-        #self.use_bias = not zero_bias
         
         self.codebook_weights = None
 
@@ -200,29 +194,14 @@ class PerturbFlow(nn.Module):
         )
 
         if self.cell_factor_size>0:
-            self.cell_factor_effect = nn.ModuleList()
-            for i in np.arange(self.cell_factor_size):
-                if self.use_bias[i]:
-                    self.cell_factor_effect.append(MLP(
-                        [self.latent_dim+1] + self.decoder_hidden_layers + [self.latent_dim],
-                        activation=activate_fct,
-                        output_activation=None,
-                        post_layer_fct=post_layer_fct,
-                        post_act_fct=post_act_fct,
-                        allow_broadcast=self.allow_broadcast,
-                        use_cuda=self.use_cuda,
-                        )
-                    )
-                else:
-                    self.cell_factor_effect.append(ZeroBiasMLP(
-                        [self.latent_dim+1] + self.decoder_hidden_layers + [self.latent_dim],
+            self.cell_factor_effect = ZeroBiasMLP2(
+                        [self.cell_factor_size] + self.decoder_hidden_layers + [self.latent_dim],
                         activation=activate_fct,
                         output_activation=None,
                         post_layer_fct=post_layer_fct,
                         post_act_fct=post_act_fct,
                         allow_broadcast=self.allow_broadcast,
                         use_cuda=self.use_cuda,
-                        )
                     )
             
         self.decoder_concentrate = MLP(
@@ -308,7 +287,7 @@ class PerturbFlow(nn.Module):
         return xs
 
     def model1(self, xs):
-        pyro.module('PerturbFlow', self)
+        pyro.module('PerturbE', self)
 
         eps = torch.finfo(xs.dtype).eps
         batch_size = xs.size(0)
@@ -370,7 +349,8 @@ class PerturbFlow(nn.Module):
                 else:
                     pyro.sample('x', dist.Poisson(rate=rate).to_event(1), obs=xs.round())
             elif self.loss_func == 'multinomial':
-                pyro.sample('x', dist.Multinomial(total_count=int(1e8), probs=theta), obs=xs)
+                #pyro.sample('x', dist.Multinomial(total_count=int(1e8), probs=theta), obs=xs)
+                pyro.sample('x', dist.Multinomial(total_count=int(1e8), logits=concentrate), obs=xs)
             elif self.loss_func == 'bernoulli':
                 if self.use_zeroinflate:
                     pyro.sample('x', dist.ZeroInflatedDistribution(dist.Bernoulli(logits=log_theta),gate_logits=gate_logits).to_event(1), obs=xs)
@@ -387,7 +367,7 @@ class PerturbFlow(nn.Module):
             ns = pyro.sample('n', dist.OneHotCategorical(logits=alpha))
 
     def model2(self, xs, us=None):
-        pyro.module('PerturbFlow', self)
+        pyro.module('PerturbE', self)
 
         eps = torch.finfo(xs.dtype).eps
         batch_size = xs.size(0)
@@ -429,7 +409,7 @@ class PerturbFlow(nn.Module):
                 zns = pyro.sample('zn', dist.Gumbel(zn_loc, zn_scale).to_event(1))
 
             if self.cell_factor_size>0:
-                zus = self._total_effects(zns, us)
+                zus = self._perturb_effects(us)
                 zs = zns+zus
             else:
                 zs = zns
@@ -454,7 +434,8 @@ class PerturbFlow(nn.Module):
                 else:
                     pyro.sample('x', dist.Poisson(rate=rate).to_event(1), obs=xs.round())
             elif self.loss_func == 'multinomial':
-                pyro.sample('x', dist.Multinomial(total_count=int(1e8), probs=theta), obs=xs)
+                #pyro.sample('x', dist.Multinomial(total_count=int(1e8), probs=theta), obs=xs)
+                pyro.sample('x', dist.Multinomial(total_count=int(1e8), logits=concentrate), obs=xs)
             elif self.loss_func == 'bernoulli':
                 if self.use_zeroinflate:
                     pyro.sample('x', dist.ZeroInflatedDistribution(dist.Bernoulli(logits=log_theta),gate_logits=gate_logits).to_event(1), obs=xs)
@@ -471,7 +452,7 @@ class PerturbFlow(nn.Module):
             ns = pyro.sample('n', dist.OneHotCategorical(logits=alpha))
 
     def model3(self, xs, ys, embeds=None):
-        pyro.module('PerturbFlow', self)
+        pyro.module('PerturbE', self)
 
         eps = torch.finfo(xs.dtype).eps
         batch_size = xs.size(0)
@@ -550,7 +531,8 @@ class PerturbFlow(nn.Module):
                 else:
                     pyro.sample('x', dist.Poisson(rate=rate).to_event(1), obs=xs.round())
             elif self.loss_func == 'multinomial':
-                pyro.sample('x', dist.Multinomial(total_count=int(1e8), probs=theta), obs=xs)
+                #pyro.sample('x', dist.Multinomial(total_count=int(1e8), probs=theta), obs=xs)
+                pyro.sample('x', dist.Multinomial(total_count=int(1e8), logits=concentrate), obs=xs)
             elif self.loss_func == 'bernoulli':
                 if self.use_zeroinflate:
                     pyro.sample('x', dist.ZeroInflatedDistribution(dist.Bernoulli(logits=log_theta),gate_logits=gate_logits).to_event(1), obs=xs)
@@ -567,7 +549,7 @@ class PerturbFlow(nn.Module):
                 zns = embeds
 
     def model4(self, xs, us, ys, embeds=None):
-        pyro.module('PerturbFlow', self)
+        pyro.module('PerturbE', self)
 
         eps = torch.finfo(xs.dtype).eps
         batch_size = xs.size(0)
@@ -631,7 +613,7 @@ class PerturbFlow(nn.Module):
                 #        zus = self.cell_factor_effect[i]([zns,us[:,i].reshape(-1,1)])
                 #    else:
                 #        zus = zus + self.cell_factor_effect[i]([zns,us[:,i].reshape(-1,1)])
-                zus = self._total_effects(zns, us)
+                zus = self._perturb_effects(us)
                 zs = zns+zus
             else:
                 zs = zns
@@ -656,7 +638,8 @@ class PerturbFlow(nn.Module):
                 else:
                     pyro.sample('x', dist.Poisson(rate=rate).to_event(1), obs=xs.round())
             elif self.loss_func == 'multinomial':
-                pyro.sample('x', dist.Multinomial(total_count=int(1e8), probs=theta), obs=xs)
+                #pyro.sample('x', dist.Multinomial(total_count=int(1e8), probs=theta), obs=xs)
+                pyro.sample('x', dist.Multinomial(total_count=int(1e8), logits=concentrate), obs=xs)
             elif self.loss_func == 'bernoulli':
                 if self.use_zeroinflate:
                     pyro.sample('x', dist.ZeroInflatedDistribution(dist.Bernoulli(logits=log_theta),gate_logits=gate_logits).to_event(1), obs=xs)
@@ -672,13 +655,8 @@ class PerturbFlow(nn.Module):
             else:
                 zns = embeds
 
-    def _total_effects(self, zns, us):
-        zus = None
-        for i in np.arange(self.cell_factor_size):
-            if i==0:
-                zus = self.cell_factor_effect[i]([zns,us[:,i].reshape(-1,1)])
-            else:
-                zus = zus + self.cell_factor_effect[i]([zns,us[:,i].reshape(-1,1)])
+    def _perturb_effects(self, us):
+        zus = self._cell_response(us)
         return zus
                         
     def _get_codebook_identity(self):
@@ -696,7 +674,7 @@ class PerturbFlow(nn.Module):
         """
         Return the mean part of metacell codebook
         """
-        cb = self._get_metacell_coordinates()
+        cb = self._get_codebook()
         cb = tensor_to_numpy(cb)
         return cb
 
@@ -810,23 +788,13 @@ class PerturbFlow(nn.Module):
         A = np.concatenate(A)
         return A
     
-    def predict(self, xs, us, perturbs_predict:list, perturbs_reference:list, library_sizes=None):
+    def predict(self, xs, perturbs_us, library_sizes=None):
         perturbs_reference = np.array(perturbs_reference)
         
         # basal embedding
         zs = self.get_basal_embedding(xs)
-        for pert in perturbs_predict:
-            pert_idx = int(np.where(perturbs_reference==pert)[0])
-            us_i = us[:,pert_idx].reshape(-1,1)
-            
-            # factor effect of xs 
-            dzs0 = self.get_cell_response(xs, factor_idx=pert_idx, perturb=us_i)
-            
-            # perturbation effect 
-            ps = np.ones_like(us_i)
-            dzs = self.get_cell_response(xs, factor_idx=pert_idx, perturb=ps)
-            
-            zs = zs + dzs0 + dzs
+        dzs = self.get_cell_response(perturbs_us)          
+        zs = zs + dzs
         
         if library_sizes is None:
             library_sizes = np.sum(xs, axis=1, keepdims=True)
@@ -840,47 +808,32 @@ class PerturbFlow(nn.Module):
         
         return counts, zs
     
-    def _cell_response(self, xs, factor_idx, perturb):
-        #zns,_ = self.encoder_zn(xs)    
-        zns,_ = self._get_basal_embedding(xs)   
-        if perturb.ndim==2:
-            ms = self.cell_factor_effect[factor_idx]([zns, perturb])
-        else:
-            ms = self.cell_factor_effect[factor_idx]([zns, perturb.reshape(-1,1)])
-            
+    def _cell_response(self, perturb):
+        ms = self.cell_factor_effect(perturb)
         return ms 
 
     def get_cell_response(self, 
-                             xs, 
-                             factor_idx,
-                             perturb,
+                             perturb_us,
                              batch_size: int = 1024):
         """
         Return cells' changes in the latent space induced by specific perturbation of a factor
 
         """
-        xs = self.preprocess(xs)
-        xs = convert_to_tensor(xs, device=self.get_device())
-        ps = convert_to_tensor(perturb, device=self.get_device())
-        dataset = CustomDataset2(xs,ps)
+        #xs = self.preprocess(xs)
+        ps = convert_to_tensor(perturb_us, device=self.get_device())
+        dataset = CustomDataset(ps)
         dataloader = DataLoader(dataset, batch_size=batch_size, shuffle=False)
 
         Z = []
         with tqdm(total=len(dataloader), desc='', unit='batch') as pbar:
-            for X_batch, P_batch, _ in dataloader:
-                zns = self._cell_response(X_batch, factor_idx, P_batch)
+            for P_batch, _ in dataloader:
+                zns = self._cell_response(P_batch)
                 Z.append(tensor_to_numpy(zns))
                 pbar.update(1)
 
         Z = np.concatenate(Z)
         return Z
     
-    def get_metacell_response(self, factor_idx, perturb):
-        zs = self._get_codebook()
-        ps = convert_to_tensor(perturb, device=self.get_device())         
-        ms = self.cell_factor_effect[factor_idx]([zs,ps])
-        return tensor_to_numpy(ms)
-    
     def _get_expression_response(self, delta_zs):
         return self.decoder_concentrate(delta_zs)
     
@@ -905,36 +858,28 @@ class PerturbFlow(nn.Module):
         R = np.concatenate(R)
         return R
     
-    def _count(self,concentrate, library_size=None):
+    def _count(self, concentrate, library_size=None):
         if self.loss_func == 'bernoulli':
             #counts = self.sigmoid(concentrate)
             counts = dist.Bernoulli(logits=concentrate).to_event(1).mean
+        elif self.loss_func == 'multinomial':
+            theta = dist.Multinomial(total_count=int(1e8), logits=concentrate).mean
+            counts = theta * library_size
         else:
             rate = concentrate.exp()
             theta = dist.DirichletMultinomial(total_count=1, concentration=rate).mean
             counts = theta * library_size
-            #counts = dist.Poisson(rate=rate).to_event(1).mean
-        return counts
-    
-    def _count_sample(self,concentrate):
-        if self.loss_func == 'bernoulli':
-            logits = concentrate
-            counts = dist.Bernoulli(logits=logits).to_event(1).sample()
-        else:
-            counts = self._count(concentrate=concentrate)
-            counts = dist.Poisson(rate=counts).to_event(1).sample()
         return counts
     
     def get_counts(self, zs, library_sizes,
-                        batch_size: int = 1024, 
-                        use_sampler: bool = False):
+                        batch_size: int = 1024):
 
         zs = convert_to_tensor(zs, device=self.get_device())
         
         if type(library_sizes) == list:
-            library_sizes = np.array(library_sizes).view(-1,1)
+            library_sizes = np.array(library_sizes).reshape(-1,1)
         elif len(library_sizes.shape)==1:
-            library_sizes = library_sizes.view(-1,1)
+            library_sizes = library_sizes.reshape(-1,1)
         ls = convert_to_tensor(library_sizes, device=self.get_device())
             
         dataset = CustomDataset2(zs,ls)
@@ -944,10 +889,7 @@ class PerturbFlow(nn.Module):
         with tqdm(total=len(dataloader), desc='', unit='batch') as pbar:
             for Z_batch, L_batch, _ in dataloader:
                 concentrate = self._get_expression_response(Z_batch)
-                if use_sampler:
-                    counts = self._count_sample(concentrate)
-                else:
-                    counts = self._count(concentrate, L_batch)
+                counts = self._count(concentrate, L_batch)
                 E.append(tensor_to_numpy(counts))
                 pbar.update(1)
         
@@ -970,7 +912,7 @@ class PerturbFlow(nn.Module):
             us = None, 
             ys = None,
             zs = None,
-            num_epochs: int = 200, 
+            num_epochs: int = 500, 
             learning_rate: float = 0.0001, 
             batch_size: int = 256, 
             algo: Literal['adam','rmsprop','adamw'] = 'adam', 
@@ -981,7 +923,7 @@ class PerturbFlow(nn.Module):
             threshold: int = 0,
             use_jax: bool = True):
         """
-        Train the PerturbFlow model.
+        Train the PerturbE model.
 
         Parameters
         ----------
@@ -1007,7 +949,7 @@ class PerturbFlow(nn.Module):
             Parameter for optimization.
         use_jax
             If toggled on, Jax will be used for speeding up. CAUTION: This will raise errors because of unknown reasons when it is called in
-            the Python script or Jupyter notebook. It is OK if it is used when runing PerturbFlow in the shell command.
+            the Python script or Jupyter notebook. It is OK if it is used when runing PerturbE in the shell command.
         """
         xs = self.preprocess(xs, threshold=threshold)
         xs = convert_to_tensor(xs, dtype=self.dtype, device=self.get_device())
@@ -1125,12 +1067,12 @@ class PerturbFlow(nn.Module):
 
         
 EXAMPLE_RUN = (
-    "example run: PerturbFlow --help"
+    "example run: PerturbE --help"
 )
 
 def parse_args():
     parser = argparse.ArgumentParser(
-        description="PerturbFlow\n{}".format(EXAMPLE_RUN))
+        description="PerturbE\n{}".format(EXAMPLE_RUN))
 
     parser.add_argument(
         "--cuda", action="store_true", help="use GPU(s) to speed up training"
@@ -1317,7 +1259,7 @@ def main():
     cell_factor_size = 0 if us is None else us.shape[1]
 
     ###########################################
-    perturbflow = PerturbFlow(
+    perturbe = PerturbE(
         input_size=input_size,
         cell_factor_size=cell_factor_size,
         inverse_dispersion=args.inverse_dispersion,
@@ -1336,7 +1278,7 @@ def main():
         dtype=dtype,
     )
 
-    perturbflow.fit(xs, us=us, 
+    perturbe.fit(xs, us=us, 
              num_epochs=args.num_epochs,
              learning_rate=args.learning_rate,
              batch_size=args.batch_size,
@@ -1348,12 +1290,11 @@ def main():
 
     if args.save_model is not None:
         if args.save_model.endswith('gz'):
-            PerturbFlow.save_model(perturbflow, args.save_model, compression=True)
+            PerturbE.save_model(perturbe, args.save_model, compression=True)
         else:
-            PerturbFlow.save_model(perturbflow, args.save_model)
+            PerturbE.save_model(perturbe, args.save_model)
     
 
 
 if __name__ == "__main__":
-
     main()