How to evaluate a deep belief network of a stack of BernoulliRBM's performance?

Question

I have this dataset of 1s and 0s (7248x2048) and I want to do a feature reduction from 2048 to 256 . I have already tried an autoencoder which performs well and now I thought that maybe a deep belief network( stack of BernoulliRBMs from scikit-learn) could also reduce the features and maybe faster. I followed this previous implementation of dbn.

1. How can I assess the performance of the dbn? I tried building a pipeline with layers 1024-> 512 -> 256 -> 512 -> 1024 -> 2048 and then calculate the "reconstruction error" of it. Does this make sense?

2. The decreasing pseudo-likelihood in the encoding part is promising? If you know other similar DBN implementations in tensorflow or pytorch, I would appreciate it.

3. The .score_samples function calculates the pseudo likelihood and I am not sure how to interpret it.

    import numpy as np 
   

      import pandas as pd 
      from sklearn.model_selection import train_test_split 
      from sklearn.neural_network import BernoulliRBM
      df = pd.DataFrame(np.random.randint(0,2,size=(7248, 2048)))
      X_train, X_test = train_test_split(df, test_size=0.2, random_state=0)
      X_train, X_val  = train_test_split(X_train, test_size=0.15, random_state=0)
   

    learning_rate = 0.1 
   

      total_units   =  2048 
      total_epochs  =  20 
      batch_size    =  16 
      rbm1 = BernoulliRBM(n_components=total_units // 2, learning_rate=learning_rate, 
      batch_size=batch_size, n_iter=total_epochs, verbose=1)
      rbm2 = BernoulliRBM(n_components=total_units // 4 , learning_rate=learning_rate,  
      batch_size=batch_size, n_iter=total_epochs, verbose=1)
      rbm3 = BernoulliRBM(n_components=total_units // 8, learning_rate=learning_rate, 
      batch_size=batch_size, n_iter=total_epochs, verbose=1)
      rbm4 = BernoulliRBM(n_components=total_units // 4 , learning_rate=learning_rate, 
      batch_size=batch_size, n_iter=total_epochs, verbose=1)
      rbm5 = BernoulliRBM(n_components=total_units // 2, learning_rate=learning_rate, 
      batch_size=batch_size, n_iter=total_epochs, verbose=1)
      rbmout = BernoulliRBM(n_components=total_units , learning_rate=learning_rate, 
      batch_size=batch_size, n_iter=total_epochs, verbose=1)
   

    model = Pipeline(steps=[('rbm1', rbm1),('rbm2', rbm2),('rbm3',rbm3),('rbm4',rbm4),('rbm5', rbm5),('rbmout', rbmout)]) 
   

      model.fit(X_train)
   

    actual = pd.DataFrame(X_val)
   

      preds = pd.DataFrame(model.fit_transform(X_val))
      dif = preds.subtract(actual)
      dif2 = np.square(dif)
      dif2['loss'] = dif2.sum(axis=1)
      dif2['loss'].mean()