Don't guess why your neural network fails to learn. Instead, have a look inside.
demo.webm #
Video 1. The main nansense UI. Clicking the layers in the architecture shows activation/gradient maps. The size of the receptive field can be measured by perturbing the input image and measuring the diff. Watched layers collect histograms and min/max activating pixel statistics for interpretability. You can even run deep dream at any point during the training run to visualize what exactly each neuron is looking for.
Nansense is a PyTorch debugger that visualizes activations, gradients, weights, optimizer state and various statistics. You can , step batch-by-batch, and time-travel to a different epoch while training, and see exactly what every layer is doing.
Here's how nansense can help:
See what is actually going on.Visualize activations and gradients,find image patches with minimal or maximal activation for a given channelandsimulate what each neuron is searching for (deep dream)Spot optimization bottlenecks.Discover insufficient receptive fields,measure neuron death,discover padding artifactsandspot gradient underflow
You can easily try out the examples yourself. Or wire it into your own training loop. Adding nansense support is just a few lines of code. Here's an example for integrating with raw PyTorch and with Lightning.
Loggers like Weights & Biases and TensorBoard record scalar curves of loss and accuracy that you scroll through after the run. Nansense works inside the live training loop instead: it s so you can step batch-by-batch and time-travel while inspecting the activations, gradients, weights and optimizer state of every layer. You can even run experiments like deep dream or Grad-CAM on the d model to probe what a given neuron has learned.
Persisting all this data on disk is infeasible, as a single batch of activations and gradients can easily be several gigabytes. Nansense sidesteps that by pausing and inspecting the tensors on demand, instead of writing everything to disk.
A layer's activations (top row) and gradients (bottom row) for a single input. Here, an image of a paraglider passes through an intermediate batch normalization layer. Each column is a channel, drawn on a diverging red/blue scale. Step through training to watch what each channel responds to and how strong the backward signal reaching it is.
Figure 1. Intermediate layer's activations and gradient from an image of a golf ball. Each column is a separate channel. Due to the next layer being a ReLU, the gradient exists only where the activation is positive.
Figure 2. Activations of a CIFAR10 trained network layer, with the input shown for comparison as the rightmost image. The augmentation used here zero-pads on the left and bottom of the image, which lights up as strong edge activations on every channel. Maybe use reflection padding next time?
For any channel, nansense collects the input patches that drove it to its strongest (and weakest) responses over an epoch. Reading off the gallery is the quickest way to tell what a specific neuron has learned to detect.
Figure 3. For each of the 6 first channels/neurons in a specific layer, the 4 strongest activating patches from the training set have been collected. The heatmap coloring shows the activation strength. As an example, both CHANNEL 1 and CHANNEL 4 both seem to be optimized for detecting french horns, however CHANNEL 1 is more centered on the instrument itself, while CHANNEL 4 seems to also be activated by human faces. See also Figure 5.
Deep dream optimizes the input itself to maximally excite a chosen neuron, synthesizing the pattern it is looking for.
Figure 4. Deep dream on exactly the same channels/neurons that were used to select maximally activating patches for Figure 4. CHANNEL 0 creates a lot of vertical red structures, loosely resembling the typical gas station presented in figure 4. In CHANNEL 1 we can yellowish curved structures, picked up from french horns. 3 and 5 have circular structures with dots inside, analogous to golf balls.
Any layer can be visualized this way, but here we use the network's final output layer, where the result is easiest to interpret. On MNIST, it produces ghostly digits between 0 and 9.
Figure 5. Deep dream on the final layer of a lenet network on the mnist dataset.
Those numbers look strange because deep dream does not necessarily make the features realistic; it maximizes them. A good example is the number 4: there are many different ways you could combine these strokes into a 4, which is why it excites the neuron even more than a typical 4 would.
Here's a visualization of other layers:
demo_deep_dream.webm #
To measure the receptive field of a neuron, nansense has support for perturbing a single pixel, and watching the diff between the original propagate through the neural network.
Figure 6. Here we perturb a single pixel of an image, and visualize how the perturbation transmits through the network. As we go deeper down the layers, the diff spreads throughout most of the image, which indicates a reasonably healthy receptive field (at least some part of the network can see the whole image).
Nansense can measure each channel's activation and gradient distribution over a full epoch. This makes it easy to discover optimization problems, such as some neurons being driven to zero.
Figure 7. The activation histogram of a dead channel in a layer. Apparently all activations are negative, which causes the next ReLU layer to clamp everything to zero. Because this eliminates any gradients, the channel will likely never recover from this state.
In low-precision training (fp16) a layer's gradients can collapse into the subnormal range (below the dtype's smallest normal value) where precision drains toward zero and the layer's learning quality quietly drops. nansense checks activations and gradients for NaNs, infinities and this subnormal/overflow band every few batches, and s with a warning banner once a meaningful share of a layer's gradient magnitude lands there.
The examples run with uv, a fast Python package manager. uv
does not pollute your other Python environments, and automatically installs the necessary packages when running a script.
curl -LsSf https://astral.sh/uv/install.sh | sh
Pick the dependency group that matches your hardware and pass it as --group
:
| Group | Hardware |
|---|---|
cpu |
|
| No GPU, CPU-only, any platform | |
cuda-legacy |
|
| Older NVIDIA GPUs: Maxwell, Pascal, Volta (CUDA 12.6) | |
cuda |
|
| Current NVIDIA GPUs: Turing through Blackwell (CUDA 13.0) | |
rocm |
|
| AMD GPUs (ROCm 7.2) |
Then launch any example; the requirements, datasets and any pretrained networks are downloaded automatically, and the UI serves on --nansense-port
.
uv run --group [group] examples/standard/main.py --nansense-port 8080
uv run --group [group] examples/game_of_life/main.py --nansense-port 8080
uv run --group [group] examples/audio_keywords/main.py --nansense-port 8080
uv run --group [group] examples/depth_make3d/main.py --nansense-port 8080
uv run --group [group] examples/multimodal/main.py --nansense-port 8080
A focused browser tab opens automatically at the boxed URL it prints (open it yourself if your environment has no browser); training s on the first batch. Drive it from the top bar. See the UI tutorial for more info.
If you hit out-of-memory errors, lower --batch-size
. If training is slow and you have GPU VRAM left, increase --batch-size
. Both memory and training speed can be improved with --dtype bf16
(older GPUs don't support it).
Figure 8. Main view of the UI, with stepping controls, architecture, individual activations/gradients, inputs and input controls. Each layer in the architecture can be clicked to open the respective layer card.
When a session starts, nansense serves a web page and s on the first batch. You drive the run from the top bar: Step Batch advances one batch, Run runs to the end and then s, and Stop s a free-running session. The dropdown next to Step Batch steps a whole epoch or up to a custom point.
Time Travel jumps back to the start of any cached epoch. It is enabled once the training loop is wrapped in a restorer, which checkpoints each epoch start to disk.
The left pane shows the model as a clickable architecture graph. Click a node to watch that layer: its activations and gradients appear as a card, and from that point on every batch feeds them into running statistics. Watched views refresh on every and, while training runs, on the cadence set under Update frequency in the settings.
Watching slows down the training and consumes memory, so it's generally better to watch only a number of layers at a time. Open a watched layer's stats view for a closer look: a histogram of its activation and gradient values over the epoch (down to a single channel), and a gallery of the input patches that drove each channel to its most extreme responses. Its Current batch phase shows the last captured batch's distribution for any layer, watched or not, and the top bar's stats button s or resumes collection without hiding the cards.
Each layer card has an Experiment button. On the experiment page, pick a method (deep dream, or a Captum attribution: Grad-CAM, Neuron Gradient, Neuron Integrated Gradients, Occlusion), set its parameters, and run it on the layer. Experiments run between batches, so training must be d; results show one card per input sample.
The right sidebar controls which input the layer views are computed from. A
model with several inputs gets an Input dropdown to choose which one the
pane shows and perturbs; a non-RGB image needs an input_transform
to display
(see the Python API), and a flat (N, C)
input shows as a clickable per-feature strip. Select sample in batch picks which sample of the current batch to show. The views follow the live training batch by default; Pin freezes the current batch as a fixed input that nansense re-runs at every update, so you can watch one input's activations evolve as training proceeds and across time travel, and Forward mode (Unchanged / Eval / Train) sets how BatchNorm and dropout behave on those re-runs.
Perturb lets you click pixels to edit the input; nansense re-runs the model and the layer cards switch to the diff, so you can trace a single changed pixel through the network.
The settings dialog records any view to an MP4, one frame per visualization
update, written under nansense_recordings/
. Start a recording with a layer watched or an experiment open, then save or discard it from the same dialog.
pip install nansense
Note:Install your PyTorch build first (see[pytorch.org]) so your CUDA / ROCm / CPU choice is preserved: nansense bundlescaptum
for the experiment page's attribution methods, and captum needs torch ≥ 2.3, so a pre-existing torch keepspip
from pulling a default CPU build.pip install lightning
additionally enablesnansense.lightning
. Runs on Python 3.10–3.14.
import torch
import nansense
model = ...
optimizer = ...
criterion = ...
train_dl, val_dl = ...
session = nansense.start(model, optimizer=optimizer, port=8080, enabled=True)
for epoch in session.epochs(50, cache_dir=".nansense_cache"):
with session.restore_point():
for inputs, targets in session.batches(train_dl, phase="train"):
optimizer.zero_grad() # keep zero_grad at the beginning of the batch
loss = criterion(model(inputs), targets) # as nansense reads .grad when
loss.backward() # the batch exits, so zeroing after step() would
optimizer.step() # leave the weight-gradient views empty.
session.close()
See the Python API for more information.
import lightning as L
from nansense.lightning import NansenseCallback, fit_with_time_travel
module = ...
datamodule = ...
callback = NansenseCallback(port=8080, model="net", enabled=True)
trainer_factory = lambda: L.Trainer(max_epochs=50)
fit_with_time_travel(trainer_factory, module, datamodule=datamodule, callback=callback)
See the Python API for more information.
nansense.start(model, ...)
creates the Session
and, when port=
is given, serves the UI. The arguments worth knowing:
optimizer
(optional): adds per-parameter optimizer state and live hyperparameters to the weights page.scheduler
(optional): lets time-travel checkpoints restore the LR schedule.enabled
:False
makes the session a near-zero-overhead no-op, so you can leave the wiring in place and switch the UI off with one flag.port
/host
/open_browser
: serve the UI immediately (the banner and auto-opened tab are skipped if a concurrent session already holds the port); omitport
and callnansense.serve(session, port=...)
separately for finer control.input_mean
/input_std
: the input normalization, so images display in their original colors.input_transform
: a callable mapping a non-RGB image input(N, C, H, W)
to a displayable(N, 1|3, H, W)
image in[0, 1]
(keepingH × W
); without it, an input whose channel count isn't 1 or 3 shows a hint to add one. A flat(N, C)
input needs none; it renders as a colormapped strip. For a multi-input model,input_mean
/input_std
/input_transform
each take either one value for all inputs or adict
keyed by input name, and the input pane gains a dropdown to pick which input to view and perturb.
Iterate each phase with session.batches(, phase=...)
, and call
session.close()
when training finishes (the served page stays up for
post-mortem browsing). For time travel, drive the epoch loop with
for epoch in session.epochs(N, cache_dir=...)
(default .nansense_cache
) and
wrap each iteration's body in with session.restore_point():
as shown above.
The schedule is discovered as you go: phase names and per-phase batch counts are learned while you iterate session.batches
, so the UI's per-phase progress and boundary stops become exact after the first epoch. Pass phases={"train": a, "val": b}
to start()
if you want that precision from the very first epoch, an optional up-front declaration (it's what the PyTorch Lightning integration uses).
For PyTorch Lightning, attach a NansenseCallback(model="<attr path to the network>", ...)
to your trainer and run the fit through fit_with_time_travel
,
which owns the jump-and-replay loop. Both accept the same port
/ host
/
open_browser
/ enabled
/ input_mean
/ input_std
/ input_transform
arguments as start
.
Distributed (DDP) needs no special wiring: call nansense.start()
on every
rank (the DDP-wrapped model is unwrapped automatically). Rank 0 serves the UI and
drives pausing and stepping; the other ranks follow its pace and fold their data
shard into the watch-page statistics. See examples/standard/main.py --distributed
. Keep in mind that DDP support is currently experimental.
See INTERNALS.md for how it works under the hood (it's long).