app

app.py

@@ -6,13 +6,16 @@ import seaborn as sns
 import pandas as pd
 
 uploaded_file = st.file_uploader("Choose original dataset")
-bw = st.number_input('Scale',value=3.05)
+col1,col2,col3 = st.columns(3)
+bw = col1.number_input('Scale',value=3.05)
+wd = col2.number_input('Weight Decay',value=0.0002)
+iters = col3.number_input('Iterations',value=400)
 
 
 
-def compute():
-    api = nflow(dim=8,latent=16,dataset=uploaded_file)
-    api.compile(optim=torch.optim.ASGD,bw=bw,lr=0.0001,wd=None)
+def compute(dim):
+    api = nflow(dim=dim,latent=16,dataset=uploaded_file)
+    api.compile(optim=torch.optim.ASGD,bw=bw,lr=0.0001,wd=wd)
     
     my_bar = st.progress(0)
     
@@ -39,5 +42,6 @@ def compute():
    
     
 if uploaded_file is not None:
-    samples=compute()
+    dim = pd.read_csv(uploaded_file).shape[-1]
+    samples=compute(dim)
     st.download_button('Download generated CSV', pd.DataFrame(samples).to_csv(), 'text/csv')

normflows.py

@@ -330,7 +330,7 @@ class nflow():
             z_k, sum_log_det = self.model(samples)
             log_p_x = self.density.log_prob(z_k)
             # Reverse KL since we can evaluate target density but can't sample
-            loss = (-sum_log_det - (log_p_x)).mean() 
+            loss = (-sum_log_det - (log_p_x)).mean()/self.density.n
 
             self.opt.zero_grad()
             loss.backward()