Replies: 1 comment 2 replies
-
It's not using the full range of bit widths, because the compression ratio is too aggressive - 1.9. I recommend using a pretrained model in order to approximate Hessian traces more correctly and get a more reliable metric of sensitivity based on that. "compression": {
"algorithm": "quantization",
"initializer": {
"batchnorm_adaptation": {
"num_bn_adaptation_samples": 3500
},
"range": {
"type": "mean_min_max",
"num_init_samples": 1000
},
"precision": {
"type": "hawq",
"bits": [4, 8, 12, 16, 20],
"iter_number": 300,
"compression_ratio": 0.75,
"eval_subset_ratio": 0.20,
"dump_init_precision_data": true
}
}, |
Beta Was this translation helpful? Give feedback.
-
Config file:
The bitwidths distribution in the
output.log
file looks like this after 4 epochs:Clearly, it uses only 4-, 8-, and 16-bit values. Why isn't it using the full range of bit widths I am providing in the config file?
Beta Was this translation helpful? Give feedback.
All reactions