paleo
paleo copied to clipboard
An analytical performance modeling tool for deep neural networks.
Dear developers, I installed paleo on a machine that did not have cuDNN (since the install instructions say it is optional). Then I ran #! /bin/sh NET_FILE=nets/inception_v3.json OUT_FILE=results/case_inception PPP_COMP=0.62 PPP_COMM=0.72...
It'd be nice to have the latency for different networks to see how much that influences the results. Currently, the network object only looks at bandwidth: https://github.com/TalwalkarLab/paleo/blob/984bbf5d2942f28b8599db4374c0ad788efe0f6e/paleo/device.py#L38-L54
I could not run 3 of the 4 Paleo commands with the provided test inputs as reported in #3 #4 #5. Could you add integration tests to your `.travis.yml` so...
``` $ ./paleo.sh profile nets/alex_v2_1gpu.json Network: nets/alex_v2_1gpu.json Direction: forward Traceback (most recent call last): File "paleo/profiler.py", line 427, in cli() File "/usr/local/lib/python2.7/dist-packages/click/core.py", line 716, in __call__ return self.main(*args, **kwargs) File...
I tried to run ` ./scripts/exp_case_alexnet_scalability.sh ` and raised a butch of Warning and got no profiling results: > WARNING build/bdist.linux-x86_64/egg/paleo/profilers/flops_profiler.py:192] Unsupported algorithm: CUDNN_CONVOLUTION_FWD_ALGO_WINOGRAD > WARNING build/bdist.linux-x86_64/egg/paleo/profilers/flops_profiler.py:192] Unsupported algorithm: CUDNN_CONVOLUTION_FWD_ALGO_WINOGRAD...