delve 0.1.40

Delve lets you monitor PyTorch model layer saturation during training

Delve: Deep Live Visualization and Evaluation

Delve is a Python package for visualizing deep learning model training.

Use Delve if you need a lightweight PyTorch or Keras extension that:

• Plots live statistics of network layer inputs to TensorBoard or terminal
• Performs spectral analysis to identify layer saturation for network pruning
• Is easily extendible and configurable

---

Motivation

Designing a deep neural network involves optimizing over a wide range of parameters and hyperparameters. Delve allows you to visualize your layer saturation during training so you can grow and shrink layers as needed.

Demo

Getting Started

pip install delve

Layer Saturation

PyTorch

delve.CheckLayerSat can be configured as follows:

savefile (str)  : destination for summaries, depending on the saving strategy, this may be a directory or file
save_to (str)   : specifies the saving strategy
    supported writers are:
        console     : print the stats to console everytime save() is called
        tensorboard : logs everything in tensorboard format, in this case the savefile must be a directory
        csv         : creates a csv file with each column corresponding to a logged variable. Everytime save() is called a new 
                      line in the file is created

        
        
layerwise_sat (bool)    : toggles if layerwise sautration should be saved by the writer
average_sat (bool)      : toggles if average saturation should be saved by the writer
ignore_layer_names (list) : a list of layer names, as specified in the modules. The layers specified will be excluded in the 
                          computation. Usefull for excluding layers which are force into a speciic saturation like softmaxes                                 or other output layers.
include_conv (bool)     : toggle if convolutional layers should be included
conv_method (str)       : the method used to pool the latent space of convolutional layers. Default is 'median", valid inputs
                          'median', 'mean' and 'max'
sat_threshold (float)   : the saturation theshold for computing the dimensionality of the latent representations. Default is
                          .99. This value may be any floating point in 0 and 1.
modules (torch modules or list of modules) : layer-containing object (may contain submodules)
log_interval (int) : steps between writing summaries
stats (list of str): list of stats to collect 

    supported stats are:
        lsat       : layer saturation

conv_method         : Method for calculating saturation. Use `cumvar99``, or `all`.
                        See https://github.com/justinshenk/playground for a comparison of how they work.
include_conv       : bool, setting to False includes only linear layers
verbose (bool)     : print saturation for every layer during training

Pass either a PyTorch model or torch.nn.Linear layers to CheckLayerSat:

from delve import CheckLayerSat

model = TwoLayerNet() # PyTorch network
stats = CheckLayerSat('runs', model) #logging directory and input

... # setup data loader

for i, data in enumerate(train_loader):    
    stats.saturation() # output saturation

Only fully-connected and convolutional layers are currently supported.

To log the saturation to console, call stats.saturation(). For example:

Regression - SixLayerNet - Hidden layer size 10                        │
loss=0.231825:  68%|████████████████████▎         | 1350/2000 [00:04<00:02, 289.30it/s]│
linear1:  90%|█████████████████████████████████▎   | 90.0/100 [00:00<00:00, 453.47it/s]│
linear2:  18%|██████▊                               | 18.0/100 [00:00<00:00, 90.68it/s]│
linear3:  32%|███████████▊                         | 32.0/100 [00:00<00:00, 161.22it/s]│
linear4:  32%|███████████▊                         | 32.0/100 [00:00<00:00, 161.24it/s]│
linear5:  28%|██████████▎                          | 28.0/100 [00:00<00:00, 141.11it/s]│
linear6:  90%|██████████████████████████████████▏   | 90.0/100 [00:01<00:00, 56.04it/s]

Keras

Two classes are provided in delve.kerascallback: CustomTensorBoard,SaturationLogger.

CustomTensorBoard takes two parameters:

Argument	Description
log_dir	location for writing summaries
user_defined_freq	frequency for writing summaries
kwargs	passed to tf.keras.callbacks.TensorBoard

SaturationLogger contains two parameters:

Argument	Description
model	Keras model
input_data	data for passing through the model
print_freq	frequency for printing

Example usage:

from delve.kerascallback import CustomTensorBoard, SaturationLogger

...

# Tensorboard logging
tbCallBack = CustomTensorBoard(log_dir='./runs', user_defined_freq=1)

# Console logging
saturation_logger = SaturationLogger(model, input_data=input_x_train[:2], print_freq=1)

...

# Add callback to Keras `fit` method
model.fit(x_train, y_train,
          epochs=100,
          batch_size=128,
          callbacks=[saturation_logger]) # can also pass tbCallBack

Output:

Epoch 29/100
 128/1000 [==>...........................] - ETA: 0s - loss: 2.2783 - acc: 0.1406
dense_1  : %0.83 | dense_2  : %0.79 | dense_3  : %0.67 |

Optimize neural network topology

Ever wonder how big your fully-connected layers should be? Delve helps you visualize the effect of modifying the layer size on your layer saturation.

For example, see how modifying the hidden layer size of this network affects the second layer saturation but not the first. Multiple runs show that the fully-connected "linear2" layer (light blue is 256-wide and orange is 8-wide) saturation is sensitive to layer size:

Log spectral analysis

Writes the top 5 eigenvalues of each layer to TensorBoard summaries:

# PyTorch-only
stats = CheckLayerSat('runs', layers, 'spectrum')

Other options

Intrinsic dimensionality

View the intrinsic dimensionality of models in realtime:

This comparison suggests that the 8-unit layer (light blue) is too saturated and that a larger layer is needed.

Why this name, Delve?

delve (verb):

• reach inside a receptacle and search for something
• to carry on intensive and thorough research for data, information, or the like