TorchVista

Demos

Basic Examples

Basic examples demonstrating core TorchVista functionality

Simple linear model
Artificially made complex model
CNN with Branching
MLP with Skip Connections
Simple Resnet
Self Attention Classifier
Positional Transformer
Deep CNN

Visual control

  • Initial visual expansion depth of modules to keep the initial graph small and ready for selective expansion
  • Show/hide non-gradient (constant) nodes
MobileVit (All modules fully collapsed)
MobileVit (Collapse modules at depth = 1)
MobileVit (Collapse modules at depth = 2)
MobileVit (Collapse depth = 2) + hide non-gradient (constant) nodes
MobileVit (Collapse modules at depth = 3)

Forced module tracing up to a depth

By default Pytorch standard or inbuilt modules' internals aren't traced. Use this flag to force trace all modules (including standard Pytorch modules) up to a desired depth
Force trace to see the contents of the TransformerEncoder module
Force trace deeper (depth = 2)
Force trace deeper (depth = 2) with increased visual collapse depth
Force trace deeper (depth = 3) with increased visual collapse depth
Force trace (depth = 4) on a deep CNN model

Error Handling

Demonstrations of how TorchVista handles models that encounter runtime errors
Tensor size mismatch error
Index out of bounds error

Complex Input & Output

Models with atypical input/output objects and unusual structures
Complex input and output
Same output tensor repeated and returned
Model nodes that don't connect to the output

Real world Examples

Real-world models
XLNetBaseCased (Large model)
Unet PP
FasterRCNN Mobilenet V3 Large 320 FPN

Computer vision models

Real-world CV models
LeNet5
AlexNet
ConvextTiny
Restnet small
SqueezeNet
Vision Transformer
EfficientNetB0
MobilenetV3Small

Unusual situations

Some unusual situations
Model returns new tensor