Tensor Spy
inspect NumPy and PyTorch tensors in the browser, no upload
Last verified:
What is Tensor Spy?
Tensor Spy is a browser‑based tool that lets you inspect .npy, .npz, .pt, and .pth tensor files directly in your web browser without uploading them to any server. It shows tensor shapes, dtypes, storage order, and per‑slice or per‑tensor statistics such as min, max, mean ± std, and counts of zeros, NaNs, and ±Inf values. You can visually slice and tile higher‑dimensional tensors, flip channels‑first/last layouts, and zoom in on tiles to inspect fine structure.
The tool is designed for data scientists, ML engineers, and researchers who need to quickly verify tensor contents, debug data loading pipelines, or compare intermediate activations from PyTorch or NumPy workflows. It supports multi‑tensor archives so you can open an .npz or a .pt/.pth bundle, select specific tensors, and then inspect or export them individually. Because everything runs locally in the browser, your data stays private and is never sent to external servers.
Tensor Spy also helps during model development by enabling visual comparison of related tensors, such as different checkpoints or feature‑map activations, using Page Up/Down to switch tensors and Shift+Page Up/Down to jump to compatible view sizes. You can reshape tensors before export, apply channel‑order fixes where auto‑detection fails, and export selected views as individual .npy files without ever loading the full tensor into memory. This makes it useful both for interactive debugging and for lightweight tensor surgery workflows around existing PyTorch or NumPy artifacts.
Tensor Spy pricing
Pricing model: Freemium
Tensor Spy is currently offered as a free web application with no paid tiers listed on the site. The authors state that they intend to keep it free and do not mention any subscription or usage‑based pricing, metered credits, or feature‑gated plans. There is no signup or account requirement to upload or inspect tensor files, and the only cost is the traffic of the web page itself; user data remains local and is not processed on the provider’s servers.
Tensor Spy pros
- Browser‑based so no installation is required
- Runs entirely client‑side; no upload of tensor data
- Supports .npy, .npz, .pt, and .pth file formats
- Handles multi‑tensor archives (.npz bundles, checkpoint .pt/.pth files)
- Displays tensor shapes, dtypes, and storage order at a glance
- Shows per‑slice and full‑tensor statistics: min/max, mean ± std, zero/NaN/Inf counts
- Interactive hover overlay for exact values and indices on canvas
- Allows slicing and tiling of higher‑dimensional tensors (e.g., time, depth, channels)
- Lets you flip channels‑first vs channels‑last layouts when auto‑detection is wrong
- Supports reshaping tensors into new shapes without full materialization in memory
- Enables visual comparison of related tensors using Page Up/Down and Shift+Page Up/Down
- Highlights NaN and Inf values with high‑contrast colors for easy anomaly spotting
- Pan and zoom in tile space to inspect fine‑scale structure
- Keyboard shortcuts for slice navigation (arrow keys) and tensor switching (Page keys)
- Allows exporting selected tensors or views as individual .npy files
- Fast local caching of inspected tensors for quick revisits
- Lowers setup cost versus writing custom Python scripts for each inspection
Tensor Spy cons
- Limited to NumPy and PyTorch tensor formats (.npy, .npz, .pt, .pth)
- No native support for TensorFlow or MXNet tensor formats
- No built‑in model execution or training integration
- No collaborative or multi‑user capabilities (pure local workflow)
- No export formats beyond .npy for tensor dumps
- No version control or diffing features for tensor files themselves
- No batch processing or scriptable automation interface
- No guarantees for extremely large tensor files that may strain browser memory
- Relies on browser‑local caching which may be cleared by the user or by browser policy
- No advanced layout or visualization options beyond the built‑in slicing/tile view
Frequently asked questions about Tensor Spy
What file types does Tensor Spy support?
Tensor Spy reads .npy, .npz, .pt, and .pth tensor files, including multi‑tensor archives such as compressed .npz bundles and PyTorch checkpoint files. It can also handle non‑image numeric data as well as tensors that represent common image formats when they are saved in these formats.
Is my tensor data uploaded to a server?
No; Tensor Spy processes tensor data entirely in your browser and caches tensors locally on your machine. The tensors are never uploaded to any external server, so your data remains private and under your control.
What kind of tensor information can I see?
For each tensor you load, Tensor Spy shows shape, dtype, storage order, and basic metadata at a glance. It also reports per‑slice and full‑tensor statistics such as minimum and maximum values, mean ± standard deviation, and counts of zeros, NaNs, and ±Inf values.
Can I inspect values inside a tensor interactively?
Yes; when you hover over the visualization canvas, an overlay shows the exact value at that position along with its indices. You can also inspect how each slice across higher‑dimensional axes behaves in terms of value ranges and statistics.
How does Tensor Spy handle slicing and higher‑dimensional tensors?
You can choose which dimensions are used as the primary x/y axes and then slice or tile the remaining dimensions. The interface lets you step through slice indices and tile extra dimensions so you can inspect multi‑dimensional data such as time series, feature maps, or channels in a compact tiled view.
Can I compare multiple tensors visually?
Yes; Tensor Spy lets you navigate between tensors using Page Up/Down, and Shift+Page Up/Down jumps to the next compatible view size. This allows you to keep the same layout and compare tiles, slices, or channel orders side by side for a quick visual diff.
Is it possible to export selected tensors or views?
Yes; when you open an .npz or a .pt/.pth bundle, you can select specific tensors, optionally reshape them, and then export them as individual .npy files. The export path is designed so that you do not need to materialize the full tensor in memory before saving it.
What if the channel order is wrong?
Tensor Spy can auto‑detect channels‑first or channels‑last layouts, but if the detection is incorrect you can manually flip the channel order. This lets you correct visualization issues such as grayscale slices being interpreted as multiple channels or RGB channels being squashed together.
Does Tensor Spy support very large tensors?
Tensor Spy is browser‑based and may be constrained by your device’s memory and browser limits. Extremely large tensors that cannot be efficiently mapped or cached in the browser may load slowly or fail to render, even though the tool is designed to avoid full materialization of tensors whenever possible.
Is there any pricing or account requirement?
Tensor Spy is currently free to use with no paid plans or account requirement. The site does not mention any subscription tiers, usage limits, or enforced quotas, and the only restriction is that everything runs client‑side in your browser using local storage.