Deep learning frameworks have often focused on either usability or speed, but not both. PyTorch is a machine learning library that shows that these two goals are in fact compatible: it provides an imperative and Pythonic programming style that supports code as a model, makes debugging easy and is consistent with other popular scientific computing libraries, while remaining efficient and supporting hardware accelerators such as GPUs. In this paper, we detail the principles that drove the implementation of PyTorch and how they are reflected in its architecture. We emphasize that every aspect of PyTorch is a regular Python program under the full control of its user. We also explain how the careful and pragmatic implementation of the key components of its runtime enables them to work together to achieve compelling performance. We demonstrate the efficiency of individual subsystems, as well as the overall speed of PyTorch on several common benchmarks.

During hemolysis, hemoglobin and heme released from red blood cells promote oxidative stress, inflammation and thrombosis. Plasma haptoglobin and hemopexin scavenge free hemoglobin and heme, respectively, but can be depleted in hemolytic states. Haptoglobin and hemopexin supplementation protect tissues, including the vasculature, liver and kidneys. It is widely assumed that these protective effects are due primarily to hemoglobin and heme clearance from the vasculature. However, this simple assumption does not account for the consequent cytoprotective adaptation seen in cells and organs. To further address the mechanism, we used a hyperhemolytic murine model (Townes-SS) of sickle cell disease to examine cellular responses to haptoglobin and hemopexin supplementation. A single infusion of haptoglobin or hemopexin (± equimolar hemoglobin) in SS-mice increased heme oxygenase-1 (HO-1) in the liver, kidney and skin several fold within 1 hour and decreased nuclear NF-ĸB phospho-p65, and vaso-occlusion for 48 hours after infusion. Plasma hemoglobin and heme levels were not significantly changed 1 hour after infusion of haptoglobin or hemopexin. Haptoglobin and hemopexin also inhibited hypoxia/reoxygenation and lipopolysaccharide-induced vaso-occlusion in SS-mice. Inhibition of HO-1 activity with tin protoporphyrin blocked the protections afforded by haptoglobin and hemopexin in SS-mice. The HO-1 reaction product carbon monoxide, fully restored the protection, in part by inhibiting Weibel-Palade body mobilization of P-selectin and von Willebrand factor to endothelial cell surfaces. Thus, the mechanism by which haptoglobin and hemopexin supplementation in hyperhemolytic SS-mice induces cytoprotective cellular responses is linked to increased HO-1 activity.

DMF activates Nrf2, enhances antioxidant defenses, and inhibits inflammation and vaso-occlusion in SCD mice. Antioxid. Redox Signal. 26, 748-762.

Innate immune complement activation may contribute to sickle cell disease (SCD) pathogenesis. Ischemia‐reperfusion physiology is a key component of the inflammatory and vaso‐occlusive milieu in SCD and is associated with complement activation. C5a is an anaphylatoxin, a potent pro‐inflammatory mediator that can activate leukocytes, platelets, and endothelial cells, all of which play a role in vaso‐occlusion. We hypothesize that hypoxia‐reoxygenation (H/R) in SCD mice activates complement, promoting inflammation and vaso‐occlusion. At baseline and after H/R, sickle Townes‐SS mice had increased C3 activation fragments and C5b‐9 deposition in kidneys, livers and lungs and alternative pathway Bb fragments in plasma compared to control AA‐mice. Activated complement promoted vaso‐occlusion (microvascular stasis) in SS‐mice; infusion of zymosan‐activated, but not heat‐inactivated serum, induced substantial vaso‐occlusion in the skin venules of SS‐mice. Infusion of recombinant C5a induced stasis in SS, but not AA‐mice that was blocked by anti‐C5a receptor (C5aR) IgG. C5a‐mediated stasis was accompanied by inflammatory responses in SS‐mice including NF‐κB activation and increased expression of TLR4 and adhesion molecules VCAM‐1, ICAM‐1, and E‐selectin in the liver. Anti‐C5aR IgG blocked these inflammatory responses. Also, C5a rapidly up‐regulated Weibel‐Palade body P‐selectin and von Willebrand factor on the surface of human umbilical vein endothelial cells in vitro and on vascular endothelium in vivo. In SS‐mice, a blocking antibody to P‐selectin inhibited C5a‐induced stasis. Similarly, an antibody to C5 that blocks murine C5 cleavage or an antibody that blocks C5aR inhibited H/R‐induced stasis in SS‐mice. These results suggest that inhibition of C5a may be beneficial in SCD.

Fresh-frozen plasma rather than thawed plasma is optimal for Hp and Hx replacement. Patient data are consistent with Hp and Hx increases via TPE limiting clinical toxicity of worsened hemolysis associated with severe vaso-occlusive complications refractory to RCE in SCD.