Unable to run the resnet model with cifar10 inputs

[aymane-eljerari]

I'm running inside the container. All build steps executed without any errors except for the last one.

root@8eb36c0a69c3:/app/release# /app/release/dataset/resnet20_cifar10.test test_batch.bin 0 9
[ERROR] Fail to initialize cifar-10 reader. please check file test_batch.bin

It seems like test_batch.bin is missing.

[aymane-eljerari]

It looks like the last wget commands inside the dockerfile did not run, because I had no cifar/ directory. I had to download the cifar data by running the command manually. I can now run the example, but the the process is killed before anything happens.

root@8eb36c0a69c3:/app/cifar# /app/release/dataset/resnet20_cifar10.test test_batch.bin 0 1
INFO: infer images [0, 1] from cifar-10 test_batch.bin.
ckks_param: _provider = 0, _poly_degree = 65536, _sec_level = 0, mul_depth = 33, _first_mod_size = 60, _
scaling_mod_size = 56, _num_q_parts = 3, _num_p = 12, _num_rot_idx = 197,_hamming_wieght = 192
Killed

[ywdyang]

I guess it may due to the memory size. How much memory do you have? I have about 128GB memory and can pass it successfully.

[aymane-eljerari]

It was a memory capacity issue. I am now able to run the resnet20 example. The programs outputs execution times of Tensor::conv and add operations.

root@c4a8a81726f1:/app/release# ./dataset/resnet20_cifar10.test ../cifar/test_batch.bin 0
INFO: infer images [0, 0] from cifar-10 ../cifar/test_batch.bin.
ckks_param: _provider = 0, _poly_degree = 65536, _sec_level = 0, mul_depth = 33, _first_mod_size = 60,
 _scaling_mod_size = 56, _num_q_parts = 3, _num_p = 12, _num_rot_idx = 197,_hamming_wieght = 192
[RT_STAT] Tensor::conv takes 18.810 seconds.
[RT_STAT] Tensor::conv takes 25.154 seconds.
[RT_STAT] Tensor::conv takes 8.223 seconds.
[RT_STAT] Tensor::add takes 0.020 seconds.
[RT_STAT] Tensor::conv takes 5.739 seconds.
[RT_STAT] Tensor::conv takes 5.746 seconds.
[RT_STAT] Tensor::add takes 0.007 seconds.
[RT_STAT] Tensor::conv takes 5.750 seconds.
[RT_STAT] Tensor::conv takes 5.743 seconds.
[RT_STAT] Tensor::add takes 0.007 seconds.
[RT_STAT] Tensor::conv takes 14.795 seconds.
[RT_STAT] Tensor::conv takes 10.506 seconds.
[RT_STAT] Tensor::conv takes 11.286 seconds.
[RT_STAT] Tensor::add takes 0.000 seconds.
[RT_STAT] Tensor::conv takes 10.491 seconds.
[RT_STAT] Tensor::conv takes 10.501 seconds.
[RT_STAT] Tensor::add takes 0.007 seconds.
[RT_STAT] Tensor::conv takes 10.501 seconds.
[RT_STAT] Tensor::conv takes 10.498 seconds.
[RT_STAT] Tensor::add takes 0.007 seconds.
[RT_STAT] Tensor::conv takes 25.199 seconds.
[RT_STAT] Tensor::conv takes 20.107 seconds.
[RT_STAT] Tensor::conv takes 19.462 seconds.
[RT_STAT] Tensor::add takes 0.000 seconds.
[RT_STAT] Tensor::conv takes 20.124 seconds.
[RT_STAT] Tensor::conv takes 20.239 seconds.
[RT_STAT] Tensor::add takes 0.007 seconds.
[RT_STAT] Tensor::conv takes 20.240 seconds.
[RT_STAT] Tensor::conv takes 20.260 seconds.
[RT_STAT] Tensor::add takes 0.007 seconds.
[[RT_STAT] Tensor::global_avg_pool takes 1.605 seconds.
[RT_STAT] Tensor::flatten takes 0.000 seconds.
[RT_STAT] Tensor::gemm takes 1.840 seconds.
Result: [-2.5673, -2.3168, -2.5635, 21.8516, -3.2558, 2.0432, -3.0956, -2.7133, -5.0904, -2.2925], exp
ected=3.
[INFO] infer image 0 success
[RESULT] infer 1 images, pass 1 1.000, fail 0 0.000.
Total memory size for keys: rot_key_cnt = 227, rot_key_size = 32847769864 bytes, total_key_size = 3307
6359968 bytes
Total memory size for weight plain: cnt = 6044, size = 12337456096 bytes

How can I print an execution trace of CKKS primitives (HEAdd, HEMult, Rescale, Rotate, Bootstrap) that are executed for this sample resnet20 test? Is it possible to get a list of all CKKS operations for the workload without actually running it?

[ywdyang]

Sorry. I do not know about this. I have just tried this repo few days ago.

[laijx-antresearch]

You may also turn on runtime statistics by environment variables RTLIB_TIMING_OUTPUT to display timing info for low level polynomial operations.
$ export RTLIB_TIMING_OUTPUT=stdout

[Sirius-sketch]

I also have the same problem, have you solved it?