imacdonald/mobilenet_v2_1.0_224-tflite

MobileNetV2 1.0 224 (TFLite, fp32), tflite-server model directory

ImageNet classifier for tflite-server's POST /classify/image.

• model.tflite: mobilenet_v2_1.0_224.tflite from the TensorFlow hosted-models archive https://storage.googleapis.com/download.tensorflow.org/models/tflite_11_05_08/mobilenet_v2_1.0_224.tgz (13978596 bytes, sha256 9f3bc29e38e90842a852bfed957dbf5e36f2d97a91dd17736b1e5c0aca8d3303). Input float32 [1,224,224,3] NHWC RGB on the -1..1 scale; output float32 [1,1001], already softmax. It has no SignatureDefs; LiteRT serves it through its default signature.
• labels.txt: https://storage.googleapis.com/download.tensorflow.org/data/ImageNetLabels.txt (sha256 536feacc519de3d418de26b2effb4d75694a8c4c0063e36499a46fa8061e2da9), 1001 lines. Index 0 is background. Some names repeat (crane, maillot), so clients should key on the index.
• manifest.json: stretch resize to 224x224 (Pillow BILINEAR semantics), then (x - 127.5) / 127.5.

License: the weights are from tensorflow/models research/slim, Apache-2.0 (see LICENSE). The ImageNet dataset they were trained on has its own, non-commercial research terms.

Expected top-5 for TensorFlow's grace_hopper.jpg (see validation.json when present): 653 military uniform 0.805430, 440 bearskin 0.036332, 668 mortarboard 0.016773 (tflite-server v0.2.0, which decodes JPEG with stb_image). A Pillow decode gives the same top-5 with scores up to 2e-3 apart; validation.json carries both.

Served by tflite-server v0.2.0 (POST /classify/image), and through lemonade's tflite recipe as POST /v1/images/classify on the fork branch release/prpl-demo. This is an unofficial repackaging of Google's published file; the weights are unchanged.