Created by: wojtuss

PR types

Function optimization

PR changes

Others

Describe

In this PR several improvements to the Quant2 -> INT8 transformation process and tests have been added:

• process:

• the list of optimization passes is supplemented,
• improved logging,
• fixed conv2d quantization in case of ResidualData are present,

• tests:

• INT8 model: accuracy and performance results are reported,
• Quant model: accuracy results are reported,
• accuracy is compared between INT8 and Quant models,
• if an FP32 model is provided using the --fp32_model option, it will be optimized and both accuracy and performance are reported,
• Image Classification tests run without the --ops_to_quantize option, enabling quantization of all supported operators,
• NLP test run with the --ops_to_quantize option for code coverage purposes,
• removed option to save FP32 model obtained from the Quant model.