def get_basic_sparsity_config(model_size=4,
input_sample_size=None,
sparsity_init=0.02,
sparsity_target=0.5,
sparsity_steps=2,
sparsity_training_steps=3):
if input_sample_size is None:
input_sample_size = [1, 1, 4, 4]
config = NNCFConfig()
config.update({
"model": "basic_sparse_conv",
"model_size": model_size,
"input_info": {
"sample_size": input_sample_size,
},
"compression": {
"algorithm": "rb_sparsity",
"params": {
"schedule": "polynomial",
"sparsity_init": sparsity_init,
"sparsity_target": sparsity_target,
"sparsity_steps": sparsity_steps,
"sparsity_training_steps": sparsity_training_steps
},
}
})
return config
def get_basic_pruning_config(input_sample_size=None) -> NNCFConfig:
if input_sample_size is None:
input_sample_size = [1, 1, 4, 4]
config = NNCFConfig()
config.update({
"model": "pruning_conv_model",
"input_info": {
"sample_size": input_sample_size,
},
"compression": {
"params": {}
}
})
return config
def get_basic_sparsity_plus_quantization_config(input_sample_size=None):
if input_sample_size is None:
input_sample_size = [1, 1, 4, 4]
config = NNCFConfig()
config.update({
"input_info": {
"sample_size": input_sample_size,
},
"compression": [{
"algorithm": "rb_sparsity",
}, {
"algorithm": "quantization"
}]
})
return config
def get_basic_magnitude_sparsity_config(input_sample_size=None):
if input_sample_size is None:
input_sample_size = [1, 1, 4, 4]
config = NNCFConfig()
config.update({
"model": "basic_sparse_conv",
"input_info": {
"sample_size": input_sample_size,
},
"compression": {
"algorithm": "magnitude_sparsity",
"params": {}
}
})
return config
def get_empty_config(model_size=4, input_sample_sizes: Union[Tuple[List[int]], List[int]] = None):
if input_sample_sizes is None:
input_sample_sizes = [1, 1, 4, 4]
def _create_input_info():
if isinstance(input_sample_sizes, tuple):
return [{"sample_size": sizes} for sizes in input_sample_sizes]
return [{"sample_size": input_sample_sizes}]
config = NNCFConfig()
config.update({
"model": "basic_sparse_conv",
"model_size": model_size,
"input_info": _create_input_info()
})
return config
def get_basic_quantization_config(model_size=4):
config = NNCFConfig()
config.update({
"model": "basic_quant_conv",
"model_size": model_size,
"input_info": {
"sample_size": [1, 1, model_size, model_size],
},
"compression": {
"algorithm": "quantization",
"initializer": {
"range": {
"num_init_steps": 0
}
}
}
})
return config
def test_get_default_weight_decay(algo, ref_weight_decay):
config = NNCFConfig()
config.update({"compression": {"algorithm": algo}})
assert ref_weight_decay == get_default_weight_decay(config)
def test_init_ranges_are_set(quantization_mode: str, per_channel: bool,
range_init_type_vs_ref_vals: Tuple[str, float,
float, float]):
class SyntheticDataset(torch.utils.data.Dataset):
def __init__(self):
super().__init__()
self._length = 1
def __getitem__(self, idx):
if idx >= self._length:
raise StopIteration
test_input_sample = torch.zeros([3, 100, 100])
for i in range(0, 100):
for j in range(0, 100):
test_input_sample[0][i][j] = i * 100 + j
test_input_sample[1] = test_input_sample[0]
test_input_sample[2] = test_input_sample[0]
return test_input_sample, test_input_sample
def __len__(self):
return self._length
data_loader = torch.utils.data.DataLoader(SyntheticDataset(),
batch_size=1,
drop_last=True)
range_init_type = range_init_type_vs_ref_vals[0]
config_with_init = NNCFConfig()
config_with_init.update({
"input_info": {
"sample_size": [1, 3, 100, 100]
},
"target_device": "TRIAL",
"compression": {
"algorithm": "quantization",
"activations": {
"mode": quantization_mode,
"per_channel": per_channel
},
"weights": {
"mode": quantization_mode,
"per_channel": per_channel
},
"initializer": {
"range": {
"num_init_samples": 1,
"type": range_init_type
}
}
}
})
if range_init_type == "percentile":
config_with_init["compression"]["initializer"]["range"]["params"] = {
"min_percentile": 32.10,
"max_percentile": 67.89
}
# Activations init check
id_model = SingleConv2dIdentityModel()
config_with_init.register_extra_structs(
[QuantizationRangeInitArgs(wrap_dataloader_for_init(data_loader))])
register_bn_adaptation_init_args(config_with_init)
_, compression_ctrl = create_compressed_model_and_algo_for_test(
id_model, config_with_init)
act_quantizer_info = next(
iter(compression_ctrl.non_weight_quantizers.values()))
ref_scale = range_init_type_vs_ref_vals[1]
ref_input_low = range_init_type_vs_ref_vals[2]
ref_input_high = range_init_type_vs_ref_vals[3]
def check_scales(quantizer: BaseQuantizer, per_channel: bool):
# Absolute tolerance is 1.0 due to percentile value interpolation
if quantization_mode == 'symmetric':
assert torch.allclose(quantizer.scale,
torch.ones_like(quantizer.scale) * ref_scale,
atol=1.0)
if per_channel:
assert quantizer.scale.numel() == 3
else:
assert quantizer.scale.numel() == 1
else:
assert torch.allclose(quantizer.input_low,
torch.ones_like(quantizer.input_low) *
ref_input_low,
atol=1.0)
assert torch.allclose(quantizer.input_range,
torch.ones_like(quantizer.input_low) *
ref_input_high,
atol=1.0)
if per_channel:
assert quantizer.input_low.numel() == 3
assert quantizer.input_range.numel() == 3
else:
assert quantizer.input_low.numel() == 1
assert quantizer.input_range.numel() == 1
check_scales(act_quantizer_info.quantizer_module_ref, per_channel)
# Weight init check
synth_weight_model = SingleConv2dSyntheticWeightModel()
_, compression_ctrl = create_compressed_model_and_algo_for_test(
synth_weight_model, config_with_init)
weight_quantizer_info = next(
iter(compression_ctrl.weight_quantizers.values()))
check_scales(weight_quantizer_info.quantizer_module_ref, per_channel)
def test_percentile_init(quantization_mode):
class SyntheticDataset(torch.utils.data.Dataset):
def __init__(self):
self._length = 1
def __getitem__(self, idx):
if idx >= self._length:
raise StopIteration
test_input_sample = torch.zeros([1, 100, 100])
for i in range(0, 100):
for j in range(0, 100):
test_input_sample[0][i][j] = i * 100 + j
return test_input_sample, test_input_sample
def __len__(self):
return self._length
data_loader = torch.utils.data.DataLoader(SyntheticDataset(), batch_size=1)
config_with_init = NNCFConfig()
config_with_init.update({
"input_info": {
"sample_size": [1, 1, 100, 100]
},
"compression": {
"algorithm": "quantization",
"activations": {
"mode": quantization_mode,
},
"weights": {
"mode": quantization_mode,
},
"initializer": {
"range": {
"num_init_steps": 1,
"type": "percentile",
"min_percentile": 32.10,
"max_percentile": 67.89
}
}
}
})
# Activations init check
id_model = SingleConv2dIdentityModel()
config_with_init.register_extra_structs(
[QuantizationRangeInitArgs(data_loader)])
_, compression_ctrl = create_compressed_model_and_algo_for_test(
id_model, config_with_init)
act_quantizer = next(iter(compression_ctrl.non_weight_quantizers.values()))
def assert_range(quantizer: BaseQuantizer):
# Absolute tolerance is 1.0 due to percentile value interpolation
if quantization_mode == 'symmetric':
assert quantizer.scale.item() == approx(6789, abs=1.0)
else:
assert quantizer.input_low.item() == approx(3210, abs=1.0)
assert quantizer.input_range.item() == approx(3578, abs=1.0)
assert_range(act_quantizer)
# Weight init check
synth_weight_model = SingleConv2dSyntheticWeightModel()
_, compression_ctrl = create_compressed_model_and_algo_for_test(
synth_weight_model, config_with_init)
weight_quantizer = next(
iter(compression_ctrl.non_weight_quantizers.values()))
assert_range(weight_quantizer)