def test_prune(self):
test = '''
model:
name: prune_yaml
framework: pytorch
device: cpu
pruning:
magnitude:
prune1:
weights: ['layer1.0.conv1.weight', 'layer1.0.conv2.weight']
target_sparsity: 0.3
end_epoch: 1
prune2:
weights: ['layer1.0.conv3.weight', 'layer1.0.conv4.weight']
target_sparsity: 0.2
start_epoch: 0
end_epoch: 20
frequency: 2
init_sparsity: 0.0
target_sparsity: 0.5
'''
helper(test)
config = conf.Conf('fake_conf.yaml')
def test_inputs_outputs(self):
test = '''
model:
name: inout_yaml
framework: mxnet
inputs: x, y
'''
helper(test)
config = conf.Conf('fake_conf.yaml')
self.assertEqual(config.usr_cfg.model.inputs, ['x', 'y'])
def test_prune(self):
test = '''
model:
name: imagenet_prune
framework: pytorch
pruning:
train:
start_epoch: 0
end_epoch: 4
dataloader:
batch_size: 30
dataset:
ImageFolder:
root: /path/to/training/dataset
optimizer:
SGD:
learning_rate: 0.1
momentum: 0.1
nesterov: True
weight_decay: 0.1
criterion:
CrossEntropyLoss:
reduction: sum
approach:
weight_compression:
initial_sparsity: 0.0
target_sparsity: 0.97
pruners:
- !Pruner
start_epoch: 1
end_epoch: 3
names: ['layer1.0.conv1.weight']
- !Pruner
start_epoch: 0
end_epoch: 4
target_sparsity: 0.6
update_frequency: 2
names: ['layer1.0.conv2.weight']
'''
helper(test)
config = conf.Conf('fake_conf.yaml')
def test_modelwise_conf_merge(self):
test = '''
model:
name: inout_yaml
framework: mxnet
quantization:
model_wise:
weight:
algorithm: minmax
activation:
algorithm: minmax
'''
helper(test)
config = conf.Conf('fake_conf.yaml')
framework_modelwise_capability = {
'CONV2D': {
'activation': {
'dtype': ['uint8', 'fp32'],
'scheme': ['asym', 'sym'],
'granularity': ['per_tensor'],
'algorithm': ['minmax', 'kl']
},
'weight': {
'dtype': ['int8', 'fp32'],
'scheme': [
'sym',
],
'granularity': ['per_channel', 'per_tensor'],
'algorithm': ['minmax']
},
},
}
tune_space = config.modelwise_tune_space(
framework_modelwise_capability)
self.assertEqual(tune_space['CONV2D']['activation']['algorithm'],
['minmax'])
def test_data_type(self):
test = '''
model:
name: test
framework: tensorflow
quantization:
calibration:
sampling_size: 20
dataloader:
batch_size: 1
dataset:
dummy:
shape: [[224,224], [256,256]]
high: [128., 127]
low: 1
dtype: ['float32', 'int8']
'''
helper(test)
cfg = conf.Conf('fake_conf.yaml').usr_cfg
dataset = cfg['quantization']['calibration']['dataloader']['dataset']['dummy']
self.assertTrue(isinstance(dataset['shape'][0], tuple))
self.assertTrue(isinstance(dataset['shape'], list))
self.assertTrue(isinstance(dataset['high'][1], float))
self.assertTrue(isinstance(dataset['high'][0], float))
self.assertTrue(isinstance(dataset['low'], float))
test = '''
model:
name: test
framework: tensorflow
quantization:
calibration:
sampling_size: 20
dataloader:
batch_size: 1
dataset:
dummy:
shape: [224,224]
high: 128
low: 0.1
dtype: ['float32', 'int8']
'''
helper(test)
cfg = conf.Conf('fake_conf.yaml').usr_cfg
dataset = cfg['quantization']['calibration']['dataloader']['dataset']['dummy']
self.assertTrue(isinstance(dataset['shape'], tuple))
self.assertTrue(isinstance(dataset['high'], float))
test = '''
model:
name: test
framework: tensorflow
quantization:
calibration:
sampling_size: 20
dataloader:
batch_size: 1
dataset:
style_transfer:
content_folder: test
style_folder: test
crop_ratio: 0.5
resize_shape: 10,10
transform:
RandomResizedCrop:
size: 10
scale: [0.07, 0.99]
ratio: [0.6, 0.8]
'''
helper(test)
cfg = conf.Conf('fake_conf.yaml').usr_cfg
shape_cfg = cfg['quantization']['calibration']['dataloader']['dataset']['style_transfer']['resize_shape']
self.assertTrue(isinstance(shape_cfg, list))
transform_cfg = cfg['quantization']['calibration']['dataloader']['transform']['RandomResizedCrop']
self.assertTrue(isinstance(transform_cfg['scale'], list))
self.assertTrue(isinstance(transform_cfg['ratio'], list))
test = '''
model:
name: test
framework: tensorflow
quantization:
calibration:
sampling_size: 20
dataloader:
batch_size: 1
dataset:
style_transfer:
content_folder: test
style_folder: test
crop_ratio: 0.5
resize_shape: [10,10]
'''
helper(test)
cfg = conf.Conf('fake_conf.yaml').usr_cfg
shape_cfg = cfg['quantization']['calibration']['dataloader']['dataset']['style_transfer']['resize_shape']
self.assertTrue(isinstance(shape_cfg, list))
test = '''
model:
name: test
framework: tensorflow
quantization:
calibration:
sampling_size: 20
dataloader:
batch_size: 1
dataset:
dummy:
shape: [224,224]
transform:
BilinearImagenet:
height: 224
width: 224
mean_value: 123.68 116.78 103.94
'''
helper(test)
cfg = conf.Conf('fake_conf.yaml').usr_cfg
shape_cfg = cfg['quantization']['calibration']['dataloader']['dataset']['dummy']['shape']
self.assertTrue(isinstance(shape_cfg, tuple))
transform_cfg = cfg['quantization']['calibration']['dataloader']['transform']['BilinearImagenet']
self.assertTrue(isinstance(transform_cfg['mean_value'], list))
def test_ops_override(self):
test = '''
model:
name: ops_override_yaml
framework: mxnet
quantization:
op_wise: {
'conv1': {
'activation': {'dtype': ['uint8', 'fp32'], 'algorithm': ['minmax'], 'scheme':['sym']},
'weight': {'dtype': ['int8', 'fp32'], 'algorithm': ['kl']}
},
'conv2': {
'activation': {'dtype': ['fp32']},
'weight': {'dtype': ['fp32']}
}
}
tuning:
accuracy_criterion:
relative: 0.01
objective: performance
'''
helper(test)
config = conf.Conf('fake_conf.yaml')
framework_modelwise_capability = {
'CONV2D': {
'activation': {
'dtype': ['uint8', 'fp32'],
'scheme': ['asym', 'sym'],
'granularity': ['per_tensor'],
'algorithm': ['minmax', 'kl']
},
'weight': {
'dtype': ['int8', 'fp32'],
'scheme': [
'sym',
],
'granularity': ['per_channel', 'per_tensor'],
'algorithm': ['minmax']
},
},
}
config.modelwise_tune_space(framework_modelwise_capability)
framework_opwise_capability = {
('conv1', 'CONV2D'): {
'activation': {
'dtype': ['uint8', 'fp32'],
'scheme': ['asym', 'sym'],
'granularity': ['per_tensor'],
'algorithm': ['minmax', 'kl']
},
'weight': {
'dtype': ['int8', 'fp32'],
'scheme': [
'sym',
],
'granularity': ['per_channel', 'per_tensor'],
'algorithm': ['minmax']
}},
('conv2', 'CONV2D'): {
'activation': {
'dtype': ['uint8', 'fp32'],
'scheme': ['asym', 'sym'],
'granularity': ['per_tensor'],
'algorithm': ['minmax', 'kl']
},
'weight': {
'dtype': ['int8', 'fp32'],
'scheme': [
'sym',
],
'granularity': ['per_channel', 'per_tensor'],
'algorithm': ['minmax']
}},
}
tune_space = config.opwise_tune_space(framework_opwise_capability)
self.assertEqual(tune_space[('conv1', 'CONV2D')]['weight']['algorithm'], ['minmax'])
self.assertEqual(tune_space[('conv2', 'CONV2D')]['activation']['dtype'], ['fp32'])
def test_quantization(self):
test = '''
model:
name: quant_yaml
framework: mxnet
quantization:
model_wise:
weights:
granularity: per_channel
'''
helper(test)
self.assertRaises(RuntimeError, conf.Conf, 'fake_conf.yaml')
test = '''
model:
name: quant_yaml
framework: mxnet
quantization:
model_wise:
approach:
'''
helper(test)
self.assertRaises(RuntimeError, conf.Conf, 'fake_conf.yaml')
test = '''
model:
name: quant_yaml
framework: mxnet
quantization:
approach: post_training_static_quant, quant_aware_training
'''
helper(test)
self.assertRaises(RuntimeError, conf.Conf, 'fake_conf.yaml')
test = '''
model:
name: quant_yaml
framework: mxnet
quantization:
model_wise:
activation:
scheme: asym
dtype: int8
weight:
scheme: asym
dtype: int8
'''
helper(test)
conf.Conf('fake_conf.yaml')
test = '''
model:
name: quant_yaml
framework: mxnet
quantization:
model_wise:
activation:
scheme:
dtype: int8
weight:
scheme: asym
dtype: int8
'''
helper(test)
self.assertRaises(RuntimeError, conf.Conf, 'fake_conf.yaml')