A pytorch model profiler with information about flops, energy, and e.t.c
Please see the files at /examples like test_linear.py and test_transformer.py for more information.
Basically, we use the profiler function in torch analyze.
simply
pip3 install torchanalyse
| Op Type | Dimension | Bound | C/M ratio | Op Intensity | Latency (msec) | Cycles | C Effcy | Flops (MFLOP) | Input_a (MB) | Input_w (MB) | Output (MB) | Total Data (MB) | Throughput (Tflops) | Roofline Throughput offchip (Tflops) | Roofline Throughput onchip (Tflops) | Compute Cycles | Memory Cycles | Sparsity | Total energy (mJ) | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 0 | aten::linear | "([1, 16], [32, 16], [1, 32])" | M | 0.006689895470383274 | 0.9142857142857143 | 1.2444444444444445e-06 | 1.1697777777777778 | 1.0 | 0.001024 | 1.6e-05 | 0.000512 | 3.2e-05 | 0.00056 | 0.8228571428571428 | 0.8228571428571428 | 0.8228571428571428 | 0.00782569105691057 | 1.1697777777777778 | 0.0 | 154980.04707236143 |
For now the profile function will provide a datafram with several information for each aten operators. You could see the flops of each at the line of Flops.
I may try to refine the datafram structure in the future.