Releases: google-ai-edge/model-explorer
model-explorer-v0.1.16
What's Changed
- We've made the PyTorch dependency optional.
By default,pip install ai-edge-model-explorer
won't install PyTorch, and we will show messages to instruct users to install PyTorch if any PyTorch related feature is used. - Support specifying model paths as separate arguments by @kartynnik in #239
- Allow to reuse server for pytorch models by @jinjingforever in #245
- Add support for showing inputs/outputs for selected layer node by @copybara-service in #244
Enable this feature in settings. After enabling it, the inputs and outputs of a selected layer node will be highlighted, and they will also be listed in the side panel (screenshot) - Add a setting that hides node data entries with empty values. by @copybara-service in #247
disabled vs enabled - Bug fixes related pinned-to-group-top node: by @copybara-service in #254
- Takes the width of the pinned-to-group-top node into account when calculating the width of the group node.
- Properly render the node when the pinned-to-group-top node is the only child node of a group node (when "Keep layers with a single op node child" setting is enabled)
- Fix seenGroupNodeIds mismatch with modelGraph nodes data. by @copybara-service in #246
New Contributors
- @kartynnik made their first contribution in #232
(We skipped v0.1.15 version number due to a bad push)
Full Changelog: model-explorer-v0.1.14...model-explorer-v0.1.16
model-explorer-v0.1.14
What's Changed
-
Add support for custom edge overlays.
Edge overlays provide a way to visualize additional connections between op nodes in a graph, separate from the original model structure. This is helpful for illustrating flows of information or processes that differ from the underlying graph's connections. See the guide for more details.
- Allow users to disable the fallback behavior when finding mapped node ids in sync navigation. See the updated guide.
- Show a message when no mapped node is found in sync navigation.
Full Changelog: model-explorer-v0.1.13...model-explorer-v0.1.14
model-explorer-v0.1.13
What's Changed
- Add support for synchronizing split pane navigation. See the guide here.
Full Changelog: model-explorer-v0.1.12...model-explorer-v0.1.13
adapter-v0.1.6
Add support for synchronizing navigation across split panes. - Allow users to sync navigation by node id, or upload a data json file to specify node id mapping. When user selects one node in one side of the pane, the node with mapped id will be automatically selected in another pane. - Allow visualizer component user to pass mapping data through visualizer config. PiperOrigin-RevId: 682822429
model-explorer-v0.1.12
What's Changed
- Use edit distance to find the best node namespace for TFLite model by @copybara-service in #144
- Always show full namespace instead of optimized namespace of the selected node in info panel. by @copybara-service in #146
- Add namespace as a search target for layer nodes. by @copybara-service in #158
- Allow users to pass node data json string when calling
add_node_data
. by @jinjingforever in #169 - Update API to make it more user-friendly by allowing users to pass a single node data item instead of a list by @jinjingforever in #170
- Don't layout the graph twice when restoring pane state. by @copybara-service in #172
- Remove obsolete op registrations from c_api_no_xla. by @copybara-service in #174
- Use model_builder from compiler/mlir. by @copybara-service in #175
- Retrieve release note and asset download url when checking new version. by @jinjingforever in #178
- Update package version and minor fix for printing release note only when
print_msg
is true by @jinjingforever in #185
Full Changelog: model-explorer-v0.1.11...model-explorer-v0.1.12
model-explorer-v0.1.11
What's Changed
-
Add support for multi-line node labels. For op node, use
\n
in its label field. For layer, use\n
in the corresponding op node's namespace field.
-
Add support to re-use an existing server instead of starting a new server.
The
visualize
API call now has a set of new parameters to support reusing an existing server. For example, by settingreuse_server
parameter toTrue
, the program will try to find a running Model Explorer server and refresh the tab with the new model without starting a new server. See docs here.import model_explorer model_explorer.visualize('/path/to/model/file', reuse_server=True)
Similar flags are also added to the command line:
$ model-explorer /path/to/model/file --reuse_server
Full Changelog: model-explorer-v0.1.10...model-explorer-v0.1.11
model-explorer-v0.1.10
What's Changed
-
Add support for HLO dialects like
mhlo
and otherhlo
family of dialects. -
Add support for adding custom attributes to group/layer nodes.
Use the newgroupNodeAttributes
field ingraph_builder.Graph
. Example:graph = gb.Graph(id='graph', nodes=[top_node, node0, node1, node2], groupNodeAttributes={ # Use empty name for graph-level attributes. '': {'model attr': 'value'}, 'layer1': {'layer attr': 'value'}})
-
Add a setting to show out-of-layer edges for an op node without needing to select it.
Note that this will make the visualization more noisy but you might find it useful in some cases. -
Allow regex filtering node data table rows by node.
Full Changelog: model-explorer-v0.1.9...model-explorer-v0.1.10
adapter-v0.1.5
Use StableHLO adapter implementation as default in MLIR This should support dialects like `mhlo` and other `hlo` family of dialects. PiperOrigin-RevId: 662596997
model-explorer-v0.1.9
What's Changed
-
Now the UI code is fully open-sourced. Thanks for your patience!
-
Add support to pin a single op node to the top of a layer:
SetGraphNode.config.pinToGroupTop
totrue
in your adapter to pin an op node to the top of the corresponding group.
-
Fix ExportedProgram display when tensor is None by @justinchuby in #118
-
Minor update to the README file by @copybara-service in #122
Full Changelog: adapter-v0.1.4...model-explorer-v0.1.9
adapter-v0.1.4
Rollback the autoformatting PiperOrigin-RevId: 658075315