svc instead of linearsvc -- may be much better

app.py

@@ -2,7 +2,7 @@ DEVICE = 'cuda'
 
 import gradio as gr
 import numpy as np
-from sklearn.svm import LinearSVC
+from sklearn.svm import SVC
 from sklearn import preprocessing
 import pandas as pd
 
@@ -55,6 +55,7 @@ pipe.to(device=DEVICE)
 
 @spaces.GPU
 def compile_em():
+    return None# TODO add back
     pipe.unet = torch.compile(pipe.unet, mode='reduce-overhead')
     pipe.vae = torch.compile(pipe.vae, mode='reduce-overhead')
     autoencoder.model.forward = torch.compile(autoencoder.model.forward, backend='inductor', dynamic=True)
@@ -195,11 +196,11 @@ def get_coeff(embs_local, ys):
     feature_embs = scaler.transform(feature_embs)
     print(len(feature_embs), len(ys))
     
-    lin_class = LinearSVC(max_iter=50000, dual='auto', class_weight='balanced').fit(feature_embs, np.array([ys[i] for i in indices]))
-    lin_class.coef_ = torch.tensor(lin_class.coef_, dtype=torch.double)
-    lin_class.coef_ = (lin_class.coef_.flatten() / (lin_class.coef_.flatten().norm())).unsqueeze(0)
+    lin_class = SVC(max_iter=50000, kernel='linear',  class_weight='balanced', C=.1).fit(feature_embs, np.array([ys[i] for i in indices]))
+    coef_ = torch.tensor(lin_class.coef_, dtype=torch.double)
+    coef_ = (coef_.flatten() / (coef_.flatten().norm())).unsqueeze(0)
     
-    return lin_class.coef_
+    return coef_
 
 # TODO add to state instead of shared across all
 glob_idx = 0