def get_basic_sparsity_config(model_size=4,
input_sample_size=None,
sparsity_init=0.02,
sparsity_target=0.5,
sparsity_target_epoch=2,
sparsity_freeze_epoch=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',
'sparsity_init': sparsity_init,
'params': {
'schedule': 'polynomial',
'sparsity_target': sparsity_target,
'sparsity_target_epoch': sparsity_target_epoch,
'sparsity_freeze_epoch': sparsity_freeze_epoch
},
}
})
return config
def worker(rank: int, world_size: int) -> None:
torch.distributed.init_process_group(backend="nccl",
init_method='tcp://127.0.0.1:8999',
world_size=world_size,
rank=rank)
model = TestModelWithChangedTrain(freezing_stages=1)
model.cuda()
model.to(rank)
nncf_config = NNCFConfig()
nncf_config.update({
"input_info": {
"sample_size": [1, 1, 30, 30]
},
"compression": {
"algorithm": "quantization",
"initializer": {
"range": {
"num_init_samples": 10
},
"batchnorm_adaptation": {
"num_bn_adaptation_samples": 10
}
}
}
})
dataloader = create_random_mock_dataloader(nncf_config, num_samples=10)
register_default_init_args(nncf_config, dataloader)
_, compressed_model = create_compressed_model(model, nncf_config)
# At this part the additional processes may be freezing
_ = torch.nn.parallel.DistributedDataParallel(compressed_model,
device_ids=[rank])
def test_model_device_before_create_compressed_model(device_placing,
inference_type):
if not torch.cuda.is_available() and not inference_type == 'cpu':
pytest.skip("Skipping CUDA test cases for CPU only setups")
input_size = [1, 1, 8, 8]
config = NNCFConfig()
config = get_kd_config(config)
config.update({
"input_info": {
"sample_size": input_size,
},
})
if inference_type == 'DDP':
ngpus_per_node = torch.cuda.device_count()
config.world_size = ngpus_per_node
torch.multiprocessing.spawn(run_training_for_device_testing,
nprocs=ngpus_per_node,
args=(config, inference_type,
ngpus_per_node, device_placing),
join=True)
else:
run_training_for_device_testing(None,
config,
inference_type,
None,
device_placing=device_placing)
def get_basic_sparsity_config(model_size=4,
input_sample_size=None,
sparsity_init=0.02,
sparsity_target=0.5,
sparsity_target_epoch=2,
sparsity_freeze_epoch=3,
scheduler='polinomial'):
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",
"sparsity_init": sparsity_init,
"params": {
"schedule": scheduler,
"sparsity_target": sparsity_target,
"sparsity_target_epoch": sparsity_target_epoch,
"sparsity_freeze_epoch": sparsity_freeze_epoch
},
}
})
return config
def get_config_for_logarithm_scale(logarithm_scale: bool,
quantization_type: str) -> NNCFConfig:
nncf_config = NNCFConfig()
nncf_config.update({
"input_info": {
"sample_size": SAMPLE_SIZE
},
"target_device": 'TRIAL',
"compression": {
"algorithm": "quantization",
"initializer": {
"range": {
"num_init_samples": 4,
"type": "percentile",
"params": {
"min_percentile": 0.001,
"max_percentile": 99.999
}
}
},
"activations": {
"mode": quantization_type,
"logarithm_scale": logarithm_scale
},
"weights": {
"mode": quantization_type,
"signed": True,
"logarithm_scale": logarithm_scale
}
}
})
class RandDatasetMock:
def __getitem__(self, index):
return torch.rand(*SAMPLE_SIZE)
def __len__(self):
return 4
data_loader = torch.utils.data.DataLoader(RandDatasetMock(),
batch_size=1,
shuffle=False,
drop_last=True)
class SquadInitializingDataloader(
nncf.torch.initialization.PTInitializingDataLoader):
def get_inputs(self, batch):
return batch, {}
def get_target(self, batch):
return None
initializing_data_loader = SquadInitializingDataloader(data_loader)
init_range = nncf.config.structures.QuantizationRangeInitArgs(
initializing_data_loader)
nncf_config.register_extra_structs([init_range])
register_bn_adaptation_init_args(nncf_config)
return nncf_config
def get_config_for_export_mode(should_be_onnx_standard: bool) -> NNCFConfig:
nncf_config = NNCFConfig()
nncf_config.update({
"input_info": {
"sample_size": [1, 1, 4, 4]
},
"compression": {
"algorithm": "quantization",
"export_to_onnx_standard_ops": should_be_onnx_standard
}
})
return nncf_config
def get_basic_quantization_config(model_size=4):
config = NNCFConfig()
config.update(
Dict({
'model': 'basic_quant_conv',
'input_info': {
'sample_size': [1, model_size, model_size, 1],
},
'compression': {
'algorithm': 'quantization',
}
}))
return config
def get_quantization_config_without_range_init(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_samples": 0
}
}
}
})
return config
def get_binarization_config() -> NNCFConfig:
config = NNCFConfig()
config.update({
"model":
"resnet18",
"input_info": {
"sample_size": [1, *LeNet.INPUT_SIZE]
},
"compression": [{
"algorithm": "binarization",
"mode": "xnor",
"params": {
"activations_quant_start_epoch": 0,
"weights_quant_start_epoch": 0
}
}]
})
return config
def get_config_for_test(batch_size=10, num_bn_adaptation_samples=100):
config = NNCFConfig()
config.update(
Dict({
"compression": {
"algorithm": "quantization",
"initializer": {
"batchnorm_adaptation": {
"num_bn_adaptation_samples": num_bn_adaptation_samples,
}
}
}
}))
dataset = get_dataset_for_test()
config = register_default_init_args(config, dataset, batch_size)
return config
def get_basic_pruning_config(model_size=8):
config = NNCFConfig()
config.update(Dict({
"model": "basic",
"input_info":
{
"sample_size": [1, model_size, model_size, 1],
},
"compression":
{
"algorithm": "filter_pruning",
"pruning_init": 0.5,
"params": {
"prune_first_conv": True,
}
}
}))
return config
def get_basic_quantization_config():
config = NNCFConfig()
config.update({
"model": "AlexNet",
"input_info": {
"sample_size": [1, 3, 32, 32],
},
"compression": {
"algorithm": "quantization",
"quantize_inputs": True,
"initializer": {
"range": {
"num_init_samples": 0
}
}
}
})
return config
def get_basic_quantization_config():
config = NNCFConfig()
config.update({
'model': 'AlexNet',
'input_info': {
'sample_size': [1, 3, 32, 32],
},
'compression': {
'algorithm': 'quantization',
'quantize_inputs': True,
'initializer': {
'range': {
'num_init_samples': 0
}
}
}
})
register_bn_adaptation_init_args(config)
return config