Train neural network code: common questions answered

The training cycle has four parts: a forward pass turns inputs into predictions, a loss function compares them with the correct answers, backpropagation computes the gradients, and an optimiser changes the weights by a small step set by the learning rate. Your Python code does this work, not Sigmora. Sigmora reads, edits and runs that code, and you judge the loss on validation data.

Backpropagation is the step that computes how much each weight contributed to the loss, which gives the gradients. The optimiser then uses those gradients and the learning rate to change the weights. It runs once per batch, after the forward pass. Sigmora does not do this itself; it runs your training script and shows the printed loss in a card, and you read the numbers.

A convolutional neural network is a network for images in which small filters scan the picture and produce feature maps that the next layers combine. Most first bugs are shape mismatches between the image data and the first layer. If your neural network from scratch is not working, check those shapes first. Sigmora can read the layer definitions and run a small batch through them.

An LSTM network is a recurrent network for sequences, such as text or measurements over time, with gates that decide what to keep from earlier steps. Its input has a time dimension, so shape errors often come from the sequence length. People ask how you debug a neural network like this: print the shape at each layer and try to overfit a tiny batch first.

A neural network learns by changing its weights so that the loss becomes smaller on the training data. One epoch is one full pass over that data, split into batches. Learning only counts if the loss on separate validation data improves too, so keep the test set untouched until the end. Sigmora can log the loss of every epoch to a CSV file for you to read.

Yes, neuralnetworkz.com is the official Sigmora website. The download button gives the setup program, about 34 MB, and records the time, IP address, browser user agent and referral source before it redirects you to the installer. Compare the SHA-256 checksum with the published manifest, and use the Send feedback button in the app to report a problem or an idea. Check it first.

Yes, Sigmora is built for Windows 10 and 11 in 64-bit (x64) form, and no account is needed to run it. Your training command can run on the PC, in WSL or over SSH on a server you own, as long as Python is installed there. The interface comes in 13 languages, and the current version is 2.9.26. Choose the environment in the app before you start.

No. Sigmora runs commands and edits files, but it does not train models and it cannot promise that code or results are correct. A zero exit code only means the script finished, and a model can look good on training data and still fail on new data. Read every diff, run your tests and judge results on data the model has not seen.

The agent works with your files and your terminal. read_file and grep read scripts and logs, edit_file and write_file change code and show a diff, and Shell runs your training command in the environment you chose and returns the real output and exit code. Every action appears as a card in the chat, and Ask first pauses before actions that change something. The expanded card shows the full output.

Plan only mode is the approval mode in which Sigmora writes a plan and runs nothing: no commands and no file changes. Use it when you want an explanation of a change, such as how to fix a shape error, before anything happens. Plans survive restarts, and you switch to Ask first or Auto-run to carry the plan out. Auto-run, by contrast, acts at once.

Yes, if you choose a cloud model: your prompts, including the text of files the agent reads, go to that provider. With a local model nothing leaves the PC. Chats, working folders and API keys stay on the PC, and secrets in tool output are redacted before they reach the model. Sigmora has no telemetry, so pick the endpoint that fits how sensitive your code and data are.

The setup program is about 34 MB, and that is what the download button on this site gives you. Before you run it, check it with certutil -hashfile on the downloaded file and compare the SHA-256 value with the published manifest. The installer is not code-signed yet, so Windows SmartScreen may warn; choose More info and then Run anyway only after the checksum matches.

The current version is 2.9.26, released on 2 Nov 2026. New versions are published automatically, and the app shows an Update button when one is ready, so you press it to install the new version. The installer build enables these automatic updates, which is why the download button on this site gives you the setup program. Check the version number on this page against the one in the app.

The Send feedback button inside the app is the way to report a problem or an idea, and no account is needed. Describe what you did, which approval mode you used and what the card showed. Do not paste secrets or private data into the message. This website has no other support form, so the button is the place to start.