Assignment 10  classification with a dense, non-linear model

Original price was: $35.00.Current price is: $30.00.

Rate this product

CSCI 362: Machine Learning Assignment 10
classification with a dense, non-linear model
Post both your code and a
screenshot of its output along with the answer to the bolded question below (about the number of weights).
We now wish to build and train a dense — that is, fully-connected — feed-forward neural net to solve the
general digit classification problem. Let us agree to go ahead and use a non-linear model. This means
that we need to split out and evaluate on testing data to gauge true performance on data unseen during
Please refer to Exercise 7 (on this DL@DU project page) where we define a so-called LogSoftmax model.
As discussed in class, piping the output of our hidden layers through softmax allows us to naturally interpret of the output as a discrete probability distribution. Since in our current project involves 10 categories,
softmax looks like this:
softmax(x1, x2,…, x10) =

i=1 e
i=1 e
i=1 e

However, for optimal training in practice we output through log_softmax and use NLLLoss negative loglikelihood loss as outlined here.
You may use the model on the DL@DU page (which has a single hidden layer of width 200) or any nonlinear dense model you wish. As part of your solution, compute (or approximate) and post the number
of weights (i.e., trainable parameters) in your model.
• Since we are using a non-linear model, we need to split out test data and use it to validate our trained
Suppose, as usual, that the features of our data are housed in a tensor xss in such a way that the
examples are indexed by the first axis, and that the corresponding targets live in yss.
In our current digit classification problem, xss is a torch.Tensor of size 5000×400, while yss is a
torch.LongTensor of size 5000.
A sensible and readable way to perform a 80/20 train/test split is:
indices = torch.randperm(len(xss))
xss = xss[indices]; yss = yss[indices] # coherently randomize the data
xss_train = xss[:4000]; yss_train = yss[:4000]
xss_test = xss[4000:]; yss_test = yss[4000:]
• In a classification problem, you do not need to center or normalize the outputs yss.
• You should likely center the features xss but there is no real need to normalize them.
• You can use DUlib’s train function if you wish:
import du.lib as dulib

model = dulib.train(
model = model,
crit = nn.NLLLoss(),
train_data = (xss_train, yss_train),
valid_data = (xss_test, yss_test),
# your other training parameters

To graph, you can add the parameter graph=1.
• Feel free to use Dulib’s class_accuracy function:
# gauge performance on training data
pct_training = dulib.class_accuracy(model, (xss_train, yss_train), show_cm=False)
print(f”Percentage correct on training data: {100*pct_training:.2f}”)
When validating on test data, you may wish to print the confusion matrix by setting show_cm=True.


There are no reviews yet.

Be the first to review “Assignment 10  classification with a dense, non-linear model”

Your email address will not be published. Required fields are marked *

Scroll to Top