How to compare error between int8 and fp16 seg model?

I have an onnx model, and through the onnx2trt function, I obtained the fp16 engine segmentation model and the int8 engine segmentation model. The int8 model has been calibrated with over 1000 data points, but on the F1 metric, the int8 model is 10 points lower than the fp16 model.

`
def onnx2trt(
model,
log_level='ERROR',
max_batch_size=1,
min_input_shapes=None,
max_input_shapes=None,
max_workspace_size=4,
fp16_mode=True,
strict_type_constraints=False,
int8_mode=False,
int8_calibrator=None,
):

logger = trt.Logger(getattr(trt.Logger, log_level))
builder = trt.Builder(logger)

network = builder.create_network(
    1 << (int)(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH))
parser = trt.OnnxParser(network, logger)
if isinstance(model, str):
    with open(model, 'rb') as f:
        flag = parser.parse(f.read())
else:
    flag = parser.parse(model.read())
if not flag:
    for error in range(parser.num_errors):
        print(parser.get_error(error))

# re-order output tensor
output_tensors = [network.get_output(i)
                  for i in range(network.num_outputs)]

[network.unmark_output(tensor) for tensor in output_tensors]
for tensor in output_tensors:
    identity_out_tensor = network.add_identity(tensor).get_output(0)
    identity_out_tensor.name = 'identity_{}'.format(tensor.name)
    network.mark_output(tensor=identity_out_tensor)

builder.max_batch_size = max_batch_size

config = builder.create_builder_config()
config.max_workspace_size = max_workspace_size * (1 << 32)
if fp16_mode:
    config.set_flag(trt.BuilderFlag.FP16)
if strict_type_constraints:
    config.set_flag(trt.BuilderFlag.STRICT_TYPES)
if int8_mode:
    config.set_flag(trt.BuilderFlag.INT8)
    if int8_calibrator is None:
        shapes = [(1,) + network.get_input(i).shape[1:]
                  for i in range(network.num_inputs)]
        dummy_data = gen_ones(shapes)
        int8_calibrator = EntropyCalibrator2(CustomDataset(dummy_data))
    config.int8_calibrator = int8_calibrator


# set dynamic shape profile
assert not (bool(min_input_shapes) ^ bool(max_input_shapes))

profile = builder.create_optimization_profile()

input_shapes = [network.get_input(i).shape[1:]
                for i in range(network.num_inputs)]
if not min_input_shapes:
    min_input_shapes = input_shapes
if not max_input_shapes:
    max_input_shapes = input_shapes

assert len(min_input_shapes) == len(max_input_shapes) == len(input_shapes)

for i in range(network.num_inputs):
    tensor = network.get_input(i)
    name = tensor.name
    min_shape = (1,) + min_input_shapes[i]
    max_shape = (max_batch_size,) + max_input_shapes[i]
    opt_shape = [(min_ + max_) // 2
                 for min_, max_ in zip(min_shape, max_shape)]
    profile.set_shape(name, min_shape, opt_shape, max_shape)
config.add_optimization_profile(profile)
engine = builder.build_engine(network, config)
return engine

`

I want to know:

1. Is there any tools to compare the errors of each layer, find the layer with larger errors, and set thses layers to fp32 precision?

2. I tried using polygraphy to analyze, but how should I parse the resulting JSON file?polygraphy run 20251023_174037_fp16.engine --trt --trt-outputs mark all --save-outputs fp16_all_outputs.json
polygraphy run 20251023_174037_int8.engine --trt --trt-outputs mark all --save-outputs int8_all_outputs.json

3. I noticed that trtexec is used for calibration in issues, but how can I obtain the calibration cache file using my own data?

Start with the onnx2trt configuration in the issue and the two Polygraphy commands that save fp16_all_outputs.json and int8_all_outputs.json. Read the TensorRT calibration and precision-analysis entry points, including trtexec, to determine how the custom dataset and calibration cache relate to the observed F1 difference. Done means the layer-comparison and calibration workflow is documented for this setup.