gradexp 0.1.28

Gradient Explorer Python Client Library

gradexp

Like wandb, but for tensors.

gradexp logs tensors from Python and lets you inspect them in Gradient Explorer, either in the hosted app or a local viewer.

Quickstart

pip install gradexp
gradexp login
# For local mode:
gradexp local open

Example:

import gradexp
import numpy as np

gradexp.init(project_name="examples", run_name="hello-gradexp")

for step in range(10):
    tensor = np.random.randn(224, 224, 3).astype(np.float32)
    gradexp.log(tensor, name="tensor")

When gradexp.init() succeeds, it prints a run URL. If a local instance is running, data is routed there automatically.

Note: tensor names are immutable once a tensor is created — they're keys into per-tensor stats and shape metadata. Choose names carefully on the first log call. Project and run names can be renamed in the explorer UI; tensor names cannot.

Semantic Model Tensor Views

gradexp.log_model_tensors(model) and gradexp.log_gradient_stats(model) log model.named_parameters() by default. Models that store several logical layers in one banked parameter can provide a semantic view by defining:

def gradexp_unbank_state_dict(self, state_dict):
    ...

The method receives a mapping of raw parameter names to tensors and returns a mapping of semantic names to tensors. Gradexp calls it automatically when present, for both weights and gradients:

class GPT(nn.Module):
    def gradexp_unbank_state_dict(self, state_dict):
        return _unbank_state_dict(state_dict, self.num_layers)

Use semantic=False or semantic="raw" to force raw PyTorch parameter names.

Common CLI Commands

gradexp login
gradexp local open
gradexp local start
gradexp local status
gradexp local stop
gradexp local sync