This Pull Request (PR) introduces a significant revamp of Picsellia's existing Docker images, delivering the following key enhancements:
Introduction of a Base Image: Our Docker images are now built upon a foundational "base" image. This base image incorporates diverse versions of CUDA and cuDNN, tailored to specific requirements.
Simplified Dockerfiles: The complexity of our Dockerfiles has been considerably reduced. Initiating a training session can now be accomplished with a single, succinct command, thereby streamlining the process.
Improved Logging System: The existing logging mechanism has been superseded with an upgraded system which resides within the base images.
This overhaul aims to optimize the efficiency, usability, and readbility of Picsellia's Docker images.
Here is a list of all the new images that have been tested:
yolov8-classification ✅
yolov8-segmentation ✅
yolov8-detection ✅
yolov7 ❌ (not tested because I couldn't find any yolov7 model in Picsellia's public repository)
yolov5 ❌ (not tested, seems deprecated since I got import errors on startup with the picsellia_yolov5 package)
yolov5-segmentation ⚠️ (the logs are correctly retrieved but I got a RuntimeError: CUDA error: no kernel image is available for execution on the device)
yolov5-detection ⚠️ (the logs are correctly retrieved but I got aRuntimeError: CUDA error: no kernel image is available for execution on the device)
For yolov5-segmentation and yolov5-detection, the logs are correctly retrieved inside the Telemetry, however the error RuntimeError: CUDA error: no kernel image is available for execution on the device was triggered for both of them.
@PN-picsell, the run.sh script now assesses the available Python version within the container, which varies depending on the model used (3.10 for yolov8, 3.8 for yolov5, etc). Right now, the script only evaluates the availability of versions 3.10 and 3.8. However, it can also be updated to automatically select the highest accessible Python version from a broader range (3.10, 3.9, 3.8, 3.7, etc.). What are your thoughts on this potential adjustment?
This Pull Request (PR) introduces a significant revamp of Picsellia's existing Docker images, delivering the following key enhancements:
This overhaul aims to optimize the efficiency, usability, and readbility of Picsellia's Docker images.
Here is a list of all the new images that have been tested:
picsellia_yolov5package)RuntimeError: CUDA error: no kernel image is available for execution on the device)RuntimeError: CUDA error: no kernel image is available for execution on the device)For yolov5-segmentation and yolov5-detection, the logs are correctly retrieved inside the Telemetry, however the error
RuntimeError: CUDA error: no kernel image is available for execution on the devicewas triggered for both of them.@PN-picsell, the run.sh script now assesses the available Python version within the container, which varies depending on the model used (3.10 for yolov8, 3.8 for yolov5, etc). Right now, the script only evaluates the availability of versions 3.10 and 3.8. However, it can also be updated to automatically select the highest accessible Python version from a broader range (3.10, 3.9, 3.8, 3.7, etc.). What are your thoughts on this potential adjustment?