ResNet18
An 18-layer convolutional neural network model from the deep residual network family, commonly used for image classification and medical imaging analysis tasks.
Timeline (last 90 days)
MedScribe AI uses a ResNet18 model trained on the Kaggle pediatric chest X-ray dataset at its foundation, and with the help of the TorchXRayVision open-source medical imaging foundation model, achieves 90.5% accuracy, 98.4% precision, and 93.2% F1 score according to data published by the project team.
Under identical seed, data order, model initialization, and machine conditions, training ResNet18 twice for 200 steps produces loss curves that completely overlap at step 0 but begin to diverge at step 1.
After 200 steps, the maximum discrepancy between the corresponding weights of the two runs reaches 0.237.
After setting torch.use_deterministic_algorithms(True) and cudnn.benchmark=False, the blogger's two 200-step ResNet18 runs achieved bit-for-bit consistency with a maximum weight difference of 0.
On a ResNet18 training step, deterministic mode incurs an overhead of +16.5% compared to the default under fp32 (17.94 ms vs. 20.90 ms), and +15.6% under bf16 autocast (12.42 ms vs. 14.37 ms).
Disabling cudnn.benchmark accounts for approximately 3% of the performance overhead on this model.
All Facts (6)
MedScribe AI uses a ResNet18 model trained on the Kaggle pediatric chest X-ray dataset at its foundation, and with the help of the TorchXRayVision open-source medical imaging foundation model, achieves 90.5% accuracy, 98.4% precision, and 93.2% F1 score according to data published by the project team.
50%UnverifiedUnder identical seed, data order, model initialization, and machine conditions, training ResNet18 twice for 200 steps produces loss curves that completely overlap at step 0 but begin to diverge at step 1.
50%UnverifiedAfter 200 steps, the maximum discrepancy between the corresponding weights of the two runs reaches 0.237.
50%UnverifiedAfter setting torch.use_deterministic_algorithms(True) and cudnn.benchmark=False, the blogger's two 200-step ResNet18 runs achieved bit-for-bit consistency with a maximum weight difference of 0.
50%UnverifiedOn a ResNet18 training step, deterministic mode incurs an overhead of +16.5% compared to the default under fp32 (17.94 ms vs. 20.90 ms), and +15.6% under bf16 autocast (12.42 ms vs. 14.37 ms).
50%UnverifiedDisabling cudnn.benchmark accounts for approximately 3% of the performance overhead on this model.
50%