Coming from lovely-tensors¶

This page is for people who use lovely-tensors to print a one-line summary of a tensor and want to know what TorchLens's stats line is and how the two differ. Install both: lovely-tensors for any tensor at a REPL, TorchLens for the numbers of a whole model's run. TorchLens's stats line complements lovely-tensors and does not replace it. The examples were written against the lovely-tensors range TorchLens's test environment installs (lovely-tensors~=0.1.18, in the viz extra); the lovely-tensors example skips where it is not installed. Checked 2026-11-03.

The same task in both¶

Print the summary line of torch.arange(8) as a float tensor and read its standard deviation.

In lovely-tensors¶

# migration-test: tool=lovely_tensors expected=True
import lovely_tensors as lt
import torch

values = torch.arange(8, dtype=torch.float32)
line = str(lt.lovely(values))
RESULT = "2.449" in line

In TorchLens¶

TorchLens renders the same kind of line with a function, never a monkey-patch: torchlens.stats.glance(t) for any dense tensor, and repr(op) for a captured op.

# migration-test: tool=torchlens expected=True
import torch
from torchlens.stats import glance

values = torch.arange(8, dtype=torch.float32)
line = glance(values)
RESULT = "sd=2.291" in line

Translation table¶

TorchLens value displays converged on one grammar (see docs/reference/lovely_reprs.md).

What TorchLens adds¶

• The line for every op of a model's run, with the op's place in the graph: repr(op) is one line (envelope plus core stats), str(op) a bounded card whose first line is the repr, and print(trace) a bounded overview.
• Exactness stated per number: extremes, NaN and Inf counts, and zero and true counts are exact at every size; sampled quantities carry a ~ mark (see "Exactness model" below).
• Honest markers for values that are not there: (not saved), (freed), (edited), (disk-backed: metadata only until read).
• A live per-firing stream of the line during capture, torchlens.stats.narrate(site, sink=print) (docs/reference/lovely_reprs.md).

What lovely-tensors does that TorchLens does not¶

• It summarizes any tensor at a REPL with one import and, with monkey_patch(), changes how every tensor prints. TorchLens's glance is a function you call.
• It draws images and channel grids from tensors (lt.rgb, lt.chans). TorchLens's stats line carries a sparkline, not an image.
• It is the right tool when you have a tensor and no model run to describe. See where to go instead.

Using both¶

lovely-tensors feeds TorchLens¶

There is no bridge. The two do not conflict: a TorchLens trace holds ordinary tensors (trace[label].out), which lovely-tensors prints like any other tensor.

TorchLens feeds lovely-tensors¶

There is no bridge. Pass any saved activation to lt.lovely(...) to see its lovely line beside TorchLens's. The sample standard deviation lovely-tensors prints will differ from TorchLens's population one, as below.

Semantic differences¶

Behavior changes¶

What changed for existing TorchLens users:

• repr(op) / repr(layer) are now ONE line (envelope + core stats); str(...) is a bounded card whose first line is the repr. The historical 21-line Op dump (8x8 preview, full relation lists, eleven lookup keys) moved behind the card's More: exits (.out, .parents, .children, .func_config, .lookup_keys).
• print(trace) is bounded: cached health in the identity line and at most twelve rows of stored metadata. The module hierarchy lives in .summary().
• Collection reprs (trace.layers, trace.params, trace.modules, trace.buffers) are one-line count and composition cards. Their str() shows at most five members, an exact remainder and .head() / .search() / .to_pandas() exits.
• tensor_stats_summary (the legacy helper line) DROPPED the neg= field from its default output: on real float activations it read ~50% nearly always, and the exact count cost a full extra traversal for a fact the sparkline shows. The record and table column keep the exact zero/sign story.
• Resolved selections iterate in GRAPH (execution) order, not lexicographic site-key order.

Population sd (divergence from lovely-tensors, on purpose)¶

TorchLens prints the POPULATION standard deviation (unbiased=False): a line describing this tensor describes a population, not an estimate of another one. lovely-tensors prints the SAMPLE sd. On torch.arange(8) that is 2.291 (TorchLens) vs 2.449 (lovely) -- neither tool is broken; the denominators differ (N vs N-1).

Exactness model¶

Extremes, NaN/Inf counts, zero/true counts: exact at every size. Mean: exact on every supported default path. Only sd and histogram may sample (above 2^26 / 2^22 elements respectively), from a seeded gathered sample, marked ~ with disclosed precision. Gate constants were derived on CPU.