-
Notifications
You must be signed in to change notification settings - Fork 4.3k
How do I in BrainScript
Nikos Karampatziakis edited this page Dec 21, 2016
·
22 revisions
Express things
- Express the error rate of my binary classifier?
- Express a softmax with a temperature parameter?
- Express a gating mechanism?
- Express a softmax over a dynamic axis?
- Zoneout (http://arxiv.org/abs/1606.01305)?
- Build a constant 3D tensor?
- Combine or concatenate vectors in BrainScript?
- Interpret the training loss?
Train models
- Perform layer-wise training?
- Train with a multitask objective?
- Train a regression model on images?
- Train a multilabel classifier?
- Get started in sequence to sequence modelling?
- Train a DSSM (or a convolutional-DSSM) model?
Evaluate models
- Interpret the training loss?
- Do early stopping?
- Monitor the error on a held out set during training?
- Set the dropout rate to 0 during evaluation/testing?
- Evaluate my newly trained model but output the activations at an intermediate layer?
- Associate an id with a prediction?
- Deploy model evaluation on Windows?
- Evaluate a model in an Azure WebApi
Adapt models
- Use a trained model as a feature extractor?
- Use an already trained network multiple times inside a larger network?
- Adapt a model I trained on one task to another
- Save and reload weights from one model to another
- [Continue training from a previously saved (snapshot from auto adjust) model](Continue-training-from-a-previously-saved model)?
Read things
- Specify multiple label streams with the HTKMLFReader?
- Use the built-in readers to train a network model using multiple input files?
Deal with errors
- Deal with the error "No node named 'x'; skipping"?
- Avoid the "AddSequence: Sequences must be a least one frame long." exception in sequence to sequence?
- Deal with the "No Output nodes found" error?
- Deal with the error "Reached the maximum number of allowed errors"?
- Deal with "InputValue operation had its row dimension x changed by the reader to y"?
- Avoid the error CURAND failure 201?