Add cleanup_torchscript_cache to test_quantize_conformance.py

[AlexanderDokuchaev]

Changes

Add cleanup_torchscript_cache function to test_quantize_conformance.py

Reason for changes

After run torch.jit.trace in convert_model, PyTorch does not clear the trace cache automatically.

Same function in torch test and openvino notebok:
https://github.com/pytorch/pytorch/blob/main/torch/testing/_internal/jit_utils.py#L59
https://github.com/openvinotoolkit/openvino_notebooks/blob/1932d4b4e99116bdedaa620c9dc92069fbb1f05e/notebooks/236-stable-diffusion-v2/implementation/conversion_helper_utils.py#L8

[codecov]

Codecov Report

Merging #2171 (c7deb33) into develop (bfefa7c) will decrease coverage by 0.03%.
Report is 7 commits behind head on develop.
The diff coverage is n/a.

Additional details and impacted files

@@             Coverage Diff             @@
##           develop    #2171      +/-   ##
===========================================
- Coverage    36.12%   36.09%   -0.03%     
===========================================
  Files          478      480       +2     
  Lines        42711    43243     +532     
===========================================
+ Hits         15429    15610     +181     
- Misses       27282    27633     +351     

see 43 files with indirect coverage changes

[AlexanderDokuchaev]

Reproducer:

import openvino as ov
import timm
import torch

dummy_tensor = torch.rand([1, 3, 224, 224])

def cleanup_torchscript_cache():
    torch._C._jit_clear_class_registry()
    torch.jit._recursive.concrete_type_store = torch.jit._recursive.ConcreteTypeStore()
    torch.jit._state._clear_class_state()

for i in range(10):
    timm_model = timm.create_model("hrnet_w18", num_classes=1000, in_chans=3, pretrained=True, checkpoint_path="")
    ov_model = ov.convert_model(timm_model, example_input=dummy_tensor)
    # cleanup_torchscript_cache()

Run script with mprof

pip install memory_profiler
mprof run --python python3 mem.py
mprof plot -o plot.png

Without cleanup_torchscript_cache:

With cleanup_torchscript_cache:

In test_quantize_conformance.py result looks like:

Model	RAM MiB
hf/bert-base-uncased	2198
timm/crossvit_9_240	4778
timm/darknet53	5264
timm/deit3_small_patch16_224	5259
timm/dla34	5345
timm/dpn68	5339
timm/efficientnet_b0	5476
timm/efficientnet_lite0	5626
timm/hrnet_w18	6224
timm/inception_resnet_v2	7009
timm/levit_128	7294
timm/mobilenetv2_050	7289
timm/mobilenetv3_small_050	7348
timm/regnetx_002	7482
timm/resnest14d	7631
timm/resnet18	7694
timm/swin_base_patch4_window7_224	8186
timm/tf_inception_v3	8440
timm/vgg11	8817
timm/visformer_small	8462
timm/wide_resnet50_2	8634

[vshampor]

Good job tracking this down to actual source.