torchexplorer 1.0.0

Interactively inspect pytorch modules during training.

Made by Samuel Pfrommer as part of Somayeh Sojoudi's group at Berkeley.

Try it yourself.

Curious about what's happening in your network? TorchExplorer is a simple tool that allows you to interactively inspect the inputs, outputs, parameters, and gradients for each nn.Module in your network during training. It integrates with weights and biases and can also operate locally as a standalone solution. If your use case fits (see limitations below), it's very simple to try:

torchexplorer.setup() # Call once before wandb.init(), not needed for standalone
wandb.init()

model = ...

torchexplorer.watch(model, backend='wandb') # Or 'standalone'

# Training loop...

TorchExplorer's interactive view of the model structure is also useful in its own right (without the histogram bells-and-whistles). Here's a self-contained example for how to get an interactive view of a ResNet18. Mousing over a particular node reveals input/output tensor shapes and Module parameters. Get a feel for what this looks like with an interactive demo.

import torch
import torchvision
from torchexplorer import watch

model = torchvision.models.resnet18(pretrained=False)
dummy_X = torch.randn(5, 3, 32, 32)

# Only log input/output and parameter histograms, if you don't want even these set log=[].
watch(model, log_freq=1, log=['io', 'params'], backend='standalone')

# Do one forwards and backwards pass
model(dummy_X).sum().backward()

# Your model will be available at http://localhost:5000

For more usage examples, see /tests and /examples.

Install

Installing requires one external graphviz dependency, which should be available on most package managers.

sudo apt-get install libgraphviz-dev graphviz
pip install torchexplorer

For Mac, brew install graphviz should suffice.

User interface

Explorer

The left-hand panel contains a module-level graph of your network architecture, automatically extracted from the autograd graph. Clicking on a module will open its "internal" submodules. To return to a parent module, click on the appropriate element in the top-left expanding list.

Nodes. A node in the explorer graph is either a) an input/output placeholder for the visualized module, or b) a specific invocation of a submodule of the visualized module. If the visualized module has multiple inputs to its forward function, these will appear as multiple nodes ("Input 0", "Input 1", ...). A similar logic applies to outputs. All other nodes represent a distinct submodule invocation. This means that if a particular submodule is called twice in one forwards pass, these two invocations show up separately in the explorer graph. Their histograms and "internal" submodules will also be distinct.

Edges. An edge between two nodes means that there exists a autograd trace from some output of the parent node to some input of the child mode. The number of incoming / outgoing edges to a node is unrelated to how many inputs/outputs the forward function takes. To illustrate this, let's consider a Linear node with two incoming edges from two distinct parent nodes. This can arise if, say, the outputs of the parent modules are added together and then passed to the single forward function input. Conversely, consider a TransformerEncoderLayer node, which accepts multiple inputs. There may still only be one incoming edge from a parent module, if all the inputs to the TransformerEncoderLayer are computed from this source.

Tooltips. Mousing over explorer graph nodes displays a helpful tooltip. The first few lines summarize the shapes of the input / output tensors, recorded once from the first forwards pass through the network. The subsequent lines parse key information from module.extra_repr(). This string parsing is designed around common PyTorch extra_repr() implementations (e.g., nn.Conv2d). The string is first split on commas, with each resulting string becoming one row in the tooltip. If a resulting substring is of the form "{key}={value}", these become the key and value pairs for the tooltip. Otherwise the entire string is treated as a value with an empty key, visualized using a dash. This occurs for the in_channels and out_channels attributes for Conv2d.

Panels

To inspect a module in more detail, just drag and drop it into one of the columns on the right. The histogram colors don't represent anything intrinsically—they're just to help identify in the explorer which modules are being visualized.

Histograms

Each vertical "slice" of a histogram encodes the distribution of values at the corresponding x-axis time. The y-axis displays the minimum / maximum bounds of the histogram. Completely white squares mean that no data fell in that bin. A bin with one entry will be shaded light gray, with the color intensifying as more values fall in that bin (this encodes the "height" of the histogram). The dashed horizontal line is the $y=0$ line.

[!NOTE] The tensors populating histograms are processed in two ways. First, for performance reasons they are randomly subsampled according to the sample_n parameter. This is 100 by default, and passing None will disable sub-sampling. Note that this sampling means that histograms that should be the same may look slightly different (e.g., output of parent node and input of child node). Second, a fraction of the most extreme values from the median are rejected to prevent outliers from compressing the histogram. This fraction is 0.1 by default, and can be disabled by passing 0.0 to the reject_outlier_propertion parameter.

For the following explanations, I'll be referencing this module:

class TestModule(nn.Module):
    def __init__(self):
        super().__init__()
        self.fc = nn.Linear(20, 20)
        self.activation = nn.ReLU()

    def forward(self, x):
        x1 = self.fc(x)
        x2 = self.activation(x1)
        return x2  

Input/output histograms. These histograms represent the values passed into and out of the module's forward method, captured using hooks. For instance, if we are visualizing the fc layer in the above TestModule, the input 0 histogram will be the histogram of x, and the output 0 histogram will be the histogram of x1. If fc accepted two inputs self.fc(x, y), then the histogram would show input 0 and input 1. Note that the input 0 histogram on the activation module will look very close to the output 0 histogram on the fc module, with some small differences due to random sampling.

Input/output gradient norm histograms. These histograms capture tensor gradients from backward passes through the module. Unlike parameter gradients, we record here the $\ell_2$-norm of the gradients, averaged over the batch dimension. This means that if the gradient of the loss with respect to the module input is of dimension $b \times d_1 \times d_2$, we first flatten to a $b \times (d_1 \cdot d_2)$ vector and take the row-wise norm to get a length $b$ vector. These values then populate the histogram. For the fc layer in the above example, input 0 (grad norm) would apply this procedure to the gradient of the loss with respect to x, while output 0 (grad norm) would apply this procedure to the gradient of the loss with respect to y.

Parameter histograms. After the input/output histograms are extracted, all submodules will have their immediate parameters (module._parameters) logged as histograms. Note that this is not the same as module.parameters(), which would also recurse to include all child parameters. Some modules (particularly activations) have no parameters and nothing will show up in the interface. For instance, TestModule above has no trainable immediate parameters; fc will have weight and bias parameters; and activation will again have nothing.

Parameter gradient histograms. After the backward call is completed, each parameter will have a .grad attribute storing the gradient of the loss with respect to that parameter. This tensor is directly passed to the histogram. Unlike the input/output gradients, no norms are computed.

API

[!NOTE] For wandb training, make sure to call torchexplorer.setup() before wandb.init(). This will configure subprocess open file limits to work around some wandb limitations.

Then the main api surface is just one function call, inspired by wandb's watch.

def watch(
    module: nn.Module,
    log: list[str] = ['io', 'io_grad', 'params', 'params_grad'],
    log_freq: int = 500,
    ignore_io_grad_classes: list[type] = [],
    disable_inplace: bool = False,
    bins: int = 20,
    sample_n: Optional[int] = 100,
    reject_outlier_proportion: float = 0.1,
    time_log: tuple[str, Callable] = ('step', lambda module, step: step),
    backend: Literal['wandb', 'standalone', 'none'] = 'wandb',
    standalone_dir: str = './torchexplorer_standalone',
    standalone_port: int = 5000,
    verbose: bool = True,
) -> StructureWrapper:
"""Watch a module and log its structure and histograms to a backend.

Args:
    module (nn.Module): The module to watch.
    log (list[str]): What to log. Can be a subset of
        ['io', 'io_grad', 'params', 'params_grad'].
    log_freq (int): How many backwards passes to wait between logging.
    ignore_io_grad_classes (list[type]): A list of classes to ignore when logging
        io_grad. This is useful for ignoring classes which do inplace operations,
        which will throw an error.
    disable_inplace (bool): disables the 'inplace' attribute for all activations in
        the module.
    bins (int): The number of bins to use for histograms.
    sample_n (Optional[int]): The number of tensor elements to randomly sample for
        histograms. Passing "None" will sample all elements.
    reject_outlier_proportion (float): The proportion of outliners to reject when
        computing histograms, based on distance to the median. 0.0 means reject
        nothing, 1.0 rejects everything. Helps chart stay in a reasonable range.
    time_log: ([tuple[str, Callable]): A tuple of (time_unit, Callable) to use for
        logging. The allable should take in the module and step and return a value
        to log. The time_unit string is just the axis label on the histogram graph.
        If "module" is a pytorch lightning modules, torchexplorer.LIGHTNING_EPOCHS
        should work to change the time axis to epochs.
    backend (Literal['wandb', 'standalone', 'none']): The backend to log to. If
        'wandb', there must be an active wandb run. Otherwise, a standalone web app
        will be created in the standalone_dir.
    standalone_dir (str): The directory to create the standalone web app in. Only
        matters if the 'standalone' backend is selected.
    standalone_port (int): The port to run the standalone server on. Only matters if
        the 'standalone' backend is selected.
    verbose (bool): Whether to print out standalone server start message.
"""

Features and limitations

Notes on some corner cases. If something isn't covered here, feel free to open a GitHub issue.

Supported

1. Performing multiple invocations of the same module is supported. Inputs/outputs will be displayed separately for each invocation, but the parameters and parameter gradients will of course be shared. So something like this should work:

class TestModule(nn.Module):
    def __init__(self):
        super().__init__()
        self.fc = nn.Linear(20, 20)
        self.activation = nn.ReLU()

    def forward(self, x):
        x = self.activation(x)
        x = self.fc(x)
        x = self.activation(x)
        return x  

2. Nondifferentiable operations which break the autograd graph are permissible and should not cause a crash. However, the resulting module-level graph will be correspondingly disconnected.
3. Multiple inputs and outputs will display correctly (i.e., "Input 0", "Input 1", ...)

Unsupported

1. Having multiple .backward() calls in one training step is not supported.
2. Recursive operations are not supported, and anything which dynamically changes the module-level control flow over training is not supported. For instance, something like this isn't permissible:

if x > 0:
    return self.module1(x)
else:
    return self.module2(x)

3. Inplace operations are not supported and should be corrected or filtered (see "Common errors" below).
4. Keyword tensor arguments to the forward method are not supported. Only positional arguments will be tracked. Behavior for keyword tensor arguments is untested as of now.
5. The tool isn't usable in the "Workspace" section of wandb when multiple runs are selected. To inspect a model, you must first open the specific run that you want in wandb.
6. Stopping and resumption of training runs is untested.

Other notes

1. When invoking a module, don't use the module.forward(x) method. Always call the forward method as module(x). The former does not call the hooks that torchexplorer uses.
2. Histograms will only be updated during training, not validation. This is directly checked using module.training. This means that if your validation dataset has a different distribution than your training dataset, what you see in the tool might not tell you what's going on during validation.

Common errors

This section includes some errors that I've run into. For something not covered here, feel free to open a GitHub issue.

1. Inplace operations in the computational graph

RuntimeError: Output 0 of BackwardHookFunctionBackward is a view and is being modified inplace...

This indicates that an inplace operation is occurring somewhere in the computational graph, which messes with the input/output gradient capturing (io_grad) feature. This commonly comes from inplace activations (e.g. nn.ReLU(inplace=True)), or residual inplace additions (e.g. out += identity). If you don't care about gradients you can just omit 'io_grad' in log argument to the watch function. Otherwise, there are two additional tools available. You can use the disable_inplace argument to automatically turn off the inplace flag on all activations. If this still doesn't cut it, you must figure out what submodules are doing inplace operations and either manually fix them or pass those classes to the ignore_io_grad_classes argument. For example, the BasicBlock in the torchvision resnet implementation has an inplace residual connection. The ReLU activations in the resnet also have inplace=True. So we would do the following:

model = torchvision.models.resnet18(pretrained=False)
watch(
    model,
    disable_inplace=True,
    ignore_io_grad_classes=[torchvision.models.resnet.BasicBlock]
)

2. Weights and biases chart glitches

"No data available." in the Custom Chart.

This occasionally shows up for me in the weights and biases interface and seems to be a difficult-to-reproduce bug in their custom charts support. Sometimes waiting fixes it. If possible, just restarting training when you notice this. Also, make sure that your browser is up-to-date. Updating Chrome to the latest version fixed this for me completely.

"Something went wrong..." and Google Chrome crashes.

It happens occasionally that the wandb website crashes with torchexplorer active. Reloading the page seems to always work. Again, this was completely fixed for me when I updated Chrome to the latest version.

3. Graphviz overflow errors

"Trapezoid overflow" error in the graphviz call.

This is a known bug in Graphviz 2.42.2, an ancient version which is still the default on most package managers. If you're getting this error, you can fix it by installing a newer release.

4. Too many open files

OSError: [Errno 24] Too many open files: b'/proc'
...
wandb: WARNING Failed to cache... too many open files

This is a known bug in wandb when uploading many tables. The only workaround I have for now is to modify the ulimit from the default of 1024 to 50000 by calling torchexplorer.setup() before wandb.init(). You can also try increasing log_freq so that fewer tables are logged. If you're still getting issues, you might have to edit /etc/security/limits.conf as described here.