Instructions to use Devil-Assassin/iVision with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Keras
How to use Devil-Assassin/iVision with Keras:
# !pip install -U keras tensorflow huggingface_hub # Keras needs TensorFlow installed to read "hf://" paths, so the tensorflow backend is selected here; # "jax" and "torch" also work for computation once TensorFlow is installed. import os os.environ["KERAS_BACKEND"] = "tensorflow" import keras model = keras.saving.load_model("hf://Devil-Assassin/iVision") - Notebooks
- Google Colab
- Kaggle
Commit History
-
Object Detection: Uploaded OD v2 models.
Verified signature0be0818 -
Delete object_detection\od_v2 with huggingface_hub
Verified signaturebb48161 -
Object Detection: Uploaded OD v2 models.
Verified signature2eb8e66
-
Depth Estimation: Moved files to dedicated directory.
Verified signature023885d -
Scene Classification: Moved files to dedicated directory.
Verified signature671065c -
Scene Classification: Moved files to dedicated directory.
Verified signatureebefd70 -
COmplexity Estimation: Cleaned up redundant models.
Verified signatured47acde -
Complexity Estimation: Uploaded all models.
Verified signatureb255445 -
Cleaned up redundant object detection models.
Verified signature98d4b64 -
Object Detection: Uploaded re-exported YOLO models (COCO)
Verified signatureRe-exported the original YOLO models using the latest Ultralytics export, which first modifies the model itself for full GPU delegation, then uses litert-torch as the exporter. I modified the export process so that I can export the models with NHWC input instead of NCHW. - COCO: Original YOLO models pretrained on COCO, exported to LiteRT. - Precision: float32 and w8a32 - Resolution: 320x320 (letterboxed with value=(114, 114, 114)) - Input: 1x320x320x3 (NHWC) - Output (det): 1x84x2100 (pre-NMS) - Output (seg): 1x116x2100 (pre-NMS) and 1x32x80x803a22027 -
Object Detection: Uploaded Legacy Files
Verified signatureThe YOLO models that we previously used but now we don't, either due to issues with the export process used or due to them becoming obsolete in our app. Some of the issues were: 1. Old YOLO export was incompatible with the new LiteRT API, so Ultralytics updated their export to use litert-torch instead, and they fixed inherent incompatibilities with the YOLO model itself, to be able to export it with full GPU delegation. 2. The new export they introduced caused the models to be exported with NCHW inputs followed by an unnecessary transpose operation, causing NCHW inputs to turn back to NHWC anyways. At the time, I couldn't find a solution around it, but later on I did and now we have NHWC models with no useless transpose operation. 3. The new export process removed the nms argument, so the default for LiteRT exports was that the output had to be postprocessed with NMS later on. At the time, I forced nms=False anyways but this stopped full GPU d…6bfba57 -
Delete object_detection with huggingface_hub
Verified signature4735c5d -
Object Detection: Uploaded Legacy Files
Verified signatureThe YOLO models that we previously used but now we don't, either due to issues with the export process used or due to them becoming obsolete in our app. Some of the issues were: - Old YOLO export was incompatible with the new LiteRT API, so Ultralytics updated their export to use litert-torch instead, and they fixed inherent incompatibilities with the YOLO model itself, to be able to export it with full GPU delegation. - The new export they introduced caused the models to be exported with NCHW inputs followed by an unnecessary transpose operation, causing NCHW inputs to turn back to NHWC anyways. At the time, I couldn't find a solution around it, but later on I did and now we have NHWC models with no useless transpose operation. - The new export process removed the nms argument, so the default for LiteRT exports was that the output had to be postprocessed with NMS later on. At the time, I forced nms=False anyways but this stopped full GPU delegation …b1b5411 -
Delete object_detection with huggingface_hub
Verified signaturee31168e -
Delete _legacy with huggingface_hub
Verified signaturea424075 -
Copy files from models/Devil-Assassin/wandering-mode
Verified signature782b1e0 -
Object Detection: Uploaded Legacy Files
Verified signatureThe YOLO models that we previously used but now we don't, either due to issues with the export process used or due to them becoming obsolete in our app. Some of the issues were: - Old YOLO export was incompatible with the new LiteRT API, so Ultralytics updated their export to use litert-torch instead, and they fixed inherent incompatibilities with the YOLO model itself, to be able to export it with full GPU delegation. - The new export they introduced caused the models to be exported with NCHW inputs followed by an unnecessary transpose operation, causing NCHW inputs to turn back to NHWC anyways. At the time, I couldn't find a solution around it, but later on I did and now we have NHWC models with no useless transpose operation. - The new export process removed the nms argument, so the default for LiteRT exports was that the output had to be postprocessed with NMS later on. At the time, I forced nms=False anyways but this stopped full GPU delegation and I wasn't …b74a6a2 -
Delete object_detection/yolo26n-seg_float32.tflite with huggingface_hub
Verified signature9cbae9d -
Copy files from models/Devil-Assassin/wandering-mode
Verified signature66b58fb
-
Uploaded new YOLO LiteRT models (OD v1), exports were in NCHW
Verified signature94eef9f
-
Uploaded YOLO TFLite models.
Verified signature02154ad
-
Uploaded TFLite NHWC versions of MobileVIT models (Scene Understanding).
Verified signature7704d23
-
Uploaded tflite depth estimation models.
Verified signature7dc524d
-
Uploaded correct complexity estimation models.
Verified signature1a36416 -
Delete complexity_estimation_v1.keras
Verified signature38d25c4 -
Delete complexity_estimation_v1.tflite
Verified signature39b1621 -
Upload complexity_estimation_v1.tflite
Verified signature9e93890
-
Upload complexity_estimation_v1.keras
Verified signature70ed478
-
Upload best_mobilevit_outdoor.pth
Verified signatureecce08a -
Upload best_mobilevit_merged.pth
Verified signaturee16757b -
Upload best_mobilevit_indoor.pth
Verified signaturee0d0abc -
initial commit
Verified signature5d8268d