Skip to content
Branch: master
Commits on May 13, 2019
Commits on Apr 22, 2019
  1. Update and release versions of all models (except speech-commands) (#194

    dsmilkov committed Apr 22, 2019
    Update and release versions of all models (except speech-commands)
  2. Upgrade speech-commands to tfjs 1.1 (#193)

    dsmilkov committed Apr 22, 2019
    Upgrade all models to depend on tfjs 1.1 and typescript 3.3.3333
Commits on Apr 19, 2019
  1. Add Cloud Build integration (migrate away from Travis) (#192)

    dsmilkov committed Apr 19, 2019
    Add Cloud Build integration (migrate away from Travis)
    Also standardize some of the yarn scripts around models:
    - move `rollup -c` out of `yarn build` into `yarn publish-local` and `yarn publish-npm`
    - make sure each repo has both `publish-local` and `publish-npm`
Commits on Apr 15, 2019
  1. Allow users of knn to use arbitrary class Id (#188)

    dsmilkov committed Apr 15, 2019
    Fixes tensorflow/tfjs#1478
    Users can now pass arbitrary classId/label when calling `addExample()`. The classId/label can be an arbitrary string or number.
  2. Update version of posenet (#189)

    dsmilkov committed Apr 15, 2019
Commits on Mar 18, 2019
  1. Update (#173)

    dsmilkov committed Mar 18, 2019
    V2 support was added in #161. Update readme to reflect that we support MobileNet V2 as well.
Commits on Mar 5, 2019
  1. Update mobilenet model to use TF-HUB (#161)

    dsmilkov committed Mar 5, 2019
    Update MobileNet model to use TF-HUB and improve/fix several issues.
    - Allow users to specify any batch size > 1.
    - Enable v2 of mobilenet from TF Hub.
    - Change `model.infer(input, embedding: boolean)`. If `embedding` is `true`, it returns the internal features recommended for transfer learning.
    - Improve error messages.
    - Add a demo
Commits on Feb 26, 2019
  1. Fix security vulnerability of our dev deps (#157)

    dsmilkov committed Feb 26, 2019
    Fixed by running `yarn audit` followed by `yarn upgrade PCK_NAME` where PCK_NAME had `lodash` (the vulnerable package) as a dependency.
  2. Update models to use latest 0.x (#156)

    dsmilkov committed Feb 26, 2019
    Update the models to use the latest `0.x`.
    Fixes tensorflow/tfjs#1282
    Fixes tensorflow/tfjs#1283
    - Update each model to depend on tfjs `0.15.3`
    - Remove all deprecated API (fromPixels, tensor.get(), loadModel, loadFrozenModel, data.forEachAsync)
    - Remove private API imports such as from 'tfjs_core/dist/....'
    - Migrate all models hosted on GCP to use the new model.json format
      - `ssd_mobilenet_v1`, `ssd_mobilenet_v2/`, `ssdlite_mobilenet_v2/`, `mobilenet_v1_1.0_224`, `mobilenet_v2_1.0_224`, `posenet_mobilenet_{025,050,075,100}_partmap`, `universal_sentence_encoder`, `toxicity`
    - Align the lint and yarn commands across the models
    - Tested that each model's demo is working by yalc linking the model's npm package `@master`.
Commits on Feb 11, 2019
  1. Update bodypix to depend on tfjs 0.15.1 (#147)

    dsmilkov committed Feb 11, 2019
    - Update bodypix to depend on tfjs 0.15.1
    - Fix deprecated API usage
    - Update body pix to 0.1.1 and release 0.1.1
    - Update demo and upload on GCS:
Commits on Feb 6, 2019
  1. Prepare universal sentence encoder for npm (#144)

    dsmilkov committed Feb 6, 2019
    - Rename the global variable to `use`
    - Add `.npmignore` file to reduce npm pkg size
    - Update the docs to reflect both npm and script tag
Commits on Nov 13, 2018
  1. Improvements to the speech recognition model (#101)

    dsmilkov committed Nov 13, 2018
    Several improvements/fixes to the speech recognition model:
    - Stop the audio extractor immediately after calling `recognizeOnline()`. Otherwise the extractor is in an infinite loop of processing data.
    - When the extractor is stopped, close the audio stream by calling `[0].stop();`. Otherwise the browser shows that the microphone is always on (red dot on the browser tab).
    - Return the original spectrogram data to the user (unnormalized). This makes a large difference in data quality for applications that decide to use only the last K frames out of the 43 frames. Those K frames are not impacted by the mean and stdev of all of the 43 frames. Before, my audio app will react with a delay of ~1sec after I make a sound even when using only the last 3 frames (70ms of data). The delay was because of the silence the preceded my sound which shifted the mean and the stdev.
    - Speedup data collection by 25% by using a queue instead of a circular buffer, which allows us to use the fast built-in `TypedArray.set()` when creating the extractor. Now calling listen with overlapFactor of 0.999 results in ~40 callbacks/sec - before it was ~32.
    - Update the patch version. Now version is `0.2.1`
Commits on Nov 8, 2018
  1. Fix overlapFactor bug and cleanup a bit (#97)

    dsmilkov committed Nov 8, 2018
    Cleanup the speech commands model:
    - Fix an issue with `overlapFactor`, where if the user passed an overlap close to 1, the callback `f` in `startStream(f)` would still be called only once every 1s. Now that callback can be called as frequently as the frame duration (tested in the work-in-progress code lab). This makes the demo run in real-time, which is really nice.
    - Remove `columnBufferLength` and `columnHopLength` as extractor config params since they were used to indirectly talk about `overlapFactor` between the recognizer and extractor. Thus, add `overlapFactor` as config instead.
    - Change `DEFAULT_SUPPRESSION_TIME_MILLIS` from `1000ms` to `0ms`, so the user can get instant predictions.
    - Remove unused params `filterSize` from `RecognizerParams`, and `minSamples` from `StreamingRecognitionConfig`
    - Move top-level `utils/` folder to `training/utils` since only training uses it
    - Remove `utils/util.ts` (all the util functions there were unused), other than `powerOfTwo` which got moved to `audio_util.ts`
    - Make tests (`*_test.ts`) visible to the typescript compiler and linter.
    - Improve `tslint.json`
    - s/SpeechCommands/speechCommands/ in Rollup to align with the other models.
    - minimize use of `any`
  2. person-segmentation: upgrade parcel, and update demo to use released …

    dsmilkov committed Nov 8, 2018
    …version (#98)
    person-segmentation: upgrade parcel, and update demo to use released version
Commits on Jul 19, 2018
  1. Remove postinstall script from models that we publish on npm (#47)

    dsmilkov committed Jul 19, 2018
    - The postinstall script `yarn upgrade` causes problems, especially for users that don't have yarn installed globally. Remove it
    - Update the knn-classifier demo to use `onclick` instead of `mousedown/up` so that it works for both desktop and phones.
    - Add `demo/` to `.npmignore` in knn-classifier
    - Build the demo in knn-classifier as part of the travis presubmit
    - Update dependencies to latest tfjs
Commits on Jun 12, 2018
  1. Update build setup of models to use ES bundle (#31)

    dsmilkov committed Jun 12, 2018
    Update build setup of models to use ES bundle. Also rename `src/posenet.ts` to `src/posenet_model.ts` to avoid conflict with `posenet.js` (the bundle) since all gets stored in `dist/`
Commits on May 31, 2018
  1. Update posenet model and demo (#26)

    dsmilkov committed May 31, 2018
    - Update posenet and mobilenet model to depend on latest tfjs
    - Update posenet demo to depend on the public posenet npm package
    - Release new posenet (0.1.1)
    Fixes tensorflow/tfjs#323
Commits on Apr 30, 2018
  1. Add license and contributing

    dsmilkov committed Apr 30, 2018
Commits on Apr 4, 2018
  1. first commit

    dsmilkov committed Apr 4, 2018
You can’t perform that action at this time.