dcslin commented on pull request #792: URL: https://github.com/apache/singa/pull/792#issuecomment-688418118
current result: - [x] training with fp16 ok, with graph, comparable accuracy, faster computation, less memory - [x] tensor cuda backend generic support on fp16 with broadcast - [ ] review operations resuing float32 ``` root@1c6aaef3db53:~/singa-hp2# PYTHONPATH=build/python/ python3 examples/cnn/train_cnn.py mlp mnist -m5 Starting Epoch 0: Training loss = 446.399231, training accuracy = 0.870331 Evaluation accuracy = 0.922676, Elapsed Time = 4.054065s Starting Epoch 1: Training loss = 246.745819, training accuracy = 0.926194 Evaluation accuracy = 0.938301, Elapsed Time = 3.921566s Starting Epoch 2: Training loss = 201.893021, training accuracy = 0.939384 Evaluation accuracy = 0.944611, Elapsed Time = 3.735095s Starting Epoch 3: Training loss = 171.419769, training accuracy = 0.948289 Evaluation accuracy = 0.952524, Elapsed Time = 3.625971s Starting Epoch 4: Training loss = 149.009338, training accuracy = 0.955326 Evaluation accuracy = 0.956530, Elapsed Time = 3.582685s root@1c6aaef3db53:~/singa-hp2# PYTHONPATH=build/python/ python3 examples/cnn/train_cnn.py mlp mnist -m5 -pfloat16 Starting Epoch 0: Training loss = 447.799744, training accuracy = 0.869547 Evaluation accuracy = 0.922075, Elapsed Time = 3.899604s Starting Epoch 1: Training loss = 249.704956, training accuracy = 0.925110 Evaluation accuracy = 0.937300, Elapsed Time = 2.524199s Starting Epoch 2: Training loss = 206.520721, training accuracy = 0.938334 Evaluation accuracy = 0.942809, Elapsed Time = 2.410751s Starting Epoch 3: Training loss = 177.916901, training accuracy = 0.946538 Evaluation accuracy = 0.950120, Elapsed Time = 2.390487s Starting Epoch 4: Training loss = 157.046936, training accuracy = 0.952958 Evaluation accuracy = 0.954828, Elapsed Time = 2.396067s root@1c6aaef3db53:~/singa-hp2# ``` ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected]
