def setUp(self):
out = run_train(
self, **{
'--generated-data': '',
'--dataset': 'ImageNet',
'--model-size': 50,
'--micro-batch-size': 8,
'--available-memory-proportion': 0.1,
'--iterations': 10,
'--BN-span': 2,
'--internal-exchange-optimisation-target': 'memory',
'--pipeline': '',
'--gradient-accumulation-count': 2,
'--pipeline-schedule': 'Sequential',
'--enable-recomputation': '',
'--pipeline-splits': 'b1/0/relu',
'--eight-bit': '',
'--replicas': 2,
'--enable-half-partials': '',
'--disable-variable-offloading': '',
'--batch-norm': '',
'--normalise-input': ''
})
self.validation = get_csv(out, 'validation.csv')
self.training = get_csv(out, 'training.csv')
def setUpClass(cls):
out = run_train(**{'--data-dir': cifar10_data_dir,
'--epochs': 10,
'--warmup-epochs': 0,
'--learning-rate-decay': '0.1',
'--learning-rate-schedule': '0.5,0.75,0.875'})
cls.validation = get_csv(out, 'validation.csv')
cls.training = get_csv(out, 'training.csv')
def setUpClass(cls):
out = run_train(**{'--generated-data': '',
'--dataset': 'ImageNet',
'--model-size': 50,
'--batch-size': 1,
'--available-memory-proportion': 0.1,
'--iterations': 10,
'--batches-per-step': 10})
cls.validation = get_csv(out, 'validation.csv')
cls.training = get_csv(out, 'training.csv')
def setUpClass(cls):
out = run_train(**{'--data-dir': cifar10_data_dir,
'--model': 'resnet',
'--lr-schedule': 'stepped',
'--learning-rate-decay': 0.5,
'--learning-rate-schedule': '0.5,0.9',
'--epochs': 20,
'--replicas': 2})
cls.validation = get_csv(out, 'validation.csv')
cls.training = get_csv(out, 'training.csv')
def setUpClass(cls):
out = run_train(**{'--data-dir': cifar10_data_dir,
'--model': "squeezenet",
'--epochs': 10,
'--use-bypass': '',
'--poly-lr-initial-lr': 0.1,
'--poly-lr-end-lr': 0.0001,
'--lr-schedule': "polynomial_decay_lr"})
cls.validation = get_csv(out, 'validation.csv')
cls.training = get_csv(out, 'training.csv')
def setUpClass(cls):
out = run_train(**{'--data-dir': cifar10_data_dir,
'--epochs': 50,
'--batch-size': 48,
'--warmup-epochs': 2,
'--lr-schedule': 'cosine',
'--label-smoothing': '0.05',
'--base-learning-rate': -5,
'--precision': '16.32'})
cls.validation = get_csv(out, 'validation.csv')
cls.training = get_csv(out, 'training.csv')
def setUpClass(cls):
out = run_train(**{'--data-dir': cifar10_data_dir,
'--name-suffix': 'penguin',
'--log-dir': 'logs/walrus',
'--iterations': 10,
'--batches-per-step': 10})
cls.logdir = None
for line in out.split('\n'):
if line.find('Saving to ') != -1:
cls.logdir = line[11:]
break
cls.validation = get_csv(out, 'validation.csv')
cls.training = get_csv(out, 'training.csv')
def setUp(self):
out = run_train(
self, **{
'--data-dir': cifar10_data_dir,
'--epochs': 50,
'--micro-batch-size': 48,
'--warmup-epochs': 2,
'--lr-schedule': 'cosine',
'--label-smoothing': '0.05',
'--base-learning-rate-exponent': -5,
'--precision': '16.32'
})
self.validation = get_csv(out, 'validation.csv')
self.training = get_csv(out, 'training.csv')
def setUp(self):
out = run_train(
self, **{
'--data-dir': cifar10_data_dir,
'--epochs': 10,
'--model': "resnext",
'--model-size': 29,
'--micro-batch-size': 8,
'--warmup-epochs': 0,
'--learning-rate-decay': '0.1',
'--learning-rate-schedule': '0.5,0.75,0.875'
})
self.validation = get_csv(out, 'validation.csv')
self.training = get_csv(out, 'training.csv')
def setUp(self):
out = run_train(
self, **{
'--data-dir': cifar10_data_dir,
'--model': 'resnext',
'--lr-schedule': 'stepped',
'--model-size': 29,
'--micro-batch-size': 4,
'--learning-rate-decay': 0.5,
'--learning-rate-schedule': '0.5,0.9',
'--epochs': 20,
'--replicas': 2
})
self.validation = get_csv(out, 'validation.csv')
self.training = get_csv(out, 'training.csv')
def setUpClass(cls):
out = run_train(
**{
'--data-dir': cifar10_data_dir,
'--epochs': 400,
'--model': "squeezenet",
'--use-bypass': '',
'--lr-schedule': 'polynomial_decay_lr',
'--label-smoothing': '0.05',
'--abs-learning-rate': 0.1,
'--abs-end-learning-rate': 0.0001,
'--warmup-epochs': 0,
'--precision': '16.32'
})
cls.validation = get_csv(out, 'validation.csv')
cls.training = get_csv(out, 'training.csv')
def setUp(self):
out = run_train(
self, **{
'--config': 'mk2_resnet8_test',
'--data-dir': cifar10_data_dir
})
self.training = get_csv(out, 'training.csv')
def setUpClass(cls):
out = run_train(**{'--dataset': 'cifar-10',
'--epochs': 10,
'--model-size': 14,
'--batch-norm': '',
'--pipeline-num-parallel': 8,
'--generated-data': '',
'--batch-size': 16,
'--base-learning-rate': -4,
'--precision': '32.32',
'--seed': 1234,
'--warmup-epochs': 0,
'--no-stochastic-rounding': '',
'--batches-per-step': 100
})
cls.validation = get_csv(out, 'validation.csv')
cls.training = get_csv(out, 'training.csv')
def setUpClass(cls):
out = run_train(**{'--config': 'resnet50_1_ipu_training',
'--model-size': 8,
'--epochs': 10,
'--data-dir': cifar10_data_dir,
'--dataset': 'cifar-10',
'--gradient-accumulation-count': 1
})
cls.training = get_csv(out, 'training.csv')
def setUp(self):
out = run_train(
self, **{
'--dataset': 'cifar-10',
'--model': 'resnext',
'--epochs': 2,
'--model-size': 29,
'--micro-batch-size': 4,
'--batch-norm': '',
'--pipeline-num-parallel': 8,
'--generated-data': '',
'--base-learning-rate-exponent': -4,
'--precision': '32.32',
'--seed': 1234,
'--warmup-epochs': 0,
'--no-stochastic-rounding': '',
'--batches-per-step': 100
})
self.validation = get_csv(out, 'validation.csv')
self.training = get_csv(out, 'training.csv')
def setUpClass(cls):
out = run_train(**{'--iterations': 100,
'--batches-per-step': 10,
'--dataset': 'cifar-10',
'--synthetic-data': '',
'--model': 'EfficientNet',
'--model-size': 'cifar',
'--batch-size': 10,
'--no-validation': '',
'--xla-recompute': '',
'--group-dim': 16,
'--expand-ratio': 4})
cls.training = get_csv(out, 'training.csv')
def setUpClass(cls):
out = run_train(
**{
'--synthetic-data': '',
'--dataset': 'ImageNet',
'--model': 'resnext',
'--model-size': 14,
'--shards': 2,
'--batch-size': 2,
'--iterations': '5000',
'--no-validation': ''
})
cls.training = get_csv(out, 'training.csv')
def setUpClass(cls):
out = run_train(**{'--synthetic-data': '',
'--dataset': 'ImageNet',
'--model': 'efficientnet',
'--shards': 2,
'--pipeline-depth': 256,
'--batch-size': 2,
'--no-validation': '',
'--xla-recompute': '',
'--available-memory-proportion': 0.2,
'--iterations': 10,
'--pipeline-splits': 'block3b'})
cls.out = out
cls.training = get_csv(out, 'training.csv')
def setUpClass(cls):
out = run_train(**{'--iterations': 10,
'--batches-per-step': 10,
'--dataset': 'imagenet',
'--generated-data': '',
'--model-size': 50,
'--shards': 2,
'--pipeline': '',
'--gradient-accumulation-count': 256,
'--batch-size': 2,
'--no-validation': '',
'--xla-recompute': '',
'--available-memory-proportion': 0.1,
'--pipeline-splits': 'b3/1/relu'})
cls.training = get_csv(out, 'training.csv')
def setUpClass(cls):
out = run_train(
**{
'--iterations': 1000,
'--dataset': 'imagenet',
'--model': 'resnext',
'--model-size': 14,
'--synthetic-data': '',
'--shards': 2,
'--pipeline-depth': 128,
'--batch-size': 1,
'--no-validation': '',
'--pipeline-splits': 'b2/0/relu'
})
cls.training = get_csv(out, 'training.csv')
print(cls.training)
def setUp(self):
out = run_train(
self, **{
'--iterations': 100,
'--batches-per-step': 10,
'--dataset': 'cifar-10',
'--generated-data': '',
'--model': 'EfficientNet',
'--model-size': 'cifar',
'--micro-batch-size': 10,
'--no-validation': '',
'--enable-recomputation': '',
'--group-dim': 16,
'--expand-ratio': 4
})
self.training = get_csv(out, 'training.csv')
def setUpClass(cls):
out = run_train(**{'--iterations': 10,
'--batches-per-step': 10,
'--dataset': 'imagenet',
'--synthetic-data': '',
'--model': 'EfficientNet',
'--model-size': 'B0',
'--shards': 2,
'--pipeline-depth': 128,
'--batch-size': 4,
'--no-validation': '',
'--xla-recompute': '',
'--pipeline-schedule': 'Grouped',
'--group-dim': 16,
'--expand-ratio': 4,
'--pipeline-splits': 'block3b'})
cls.training = get_csv(out, 'training.csv')
def setUpClass(cls):
out = run_train(**{'--generated-data': '',
'--dataset': 'ImageNet',
'--model': 'efficientnet',
'--model-size': 'B1',
'--shards': 4,
'--pipeline': '',
'--gradient-accumulation-count': 128,
'--batch-size': 4,
'--no-validation': '',
'--xla-recompute': '',
'--available-memory-proportion': 0.2,
'--pipeline-schedule': 'Grouped',
'--iterations': 10,
'--pipeline-splits': 'block2a/c', 'block4a': 'block5c'})
cls.out = out
cls.training = get_csv(out, 'training.csv')
def setUpClass(cls):
out = run_train(
**{
'--iterations': 500,
'--dataset': 'imagenet',
'--model': 'resnext',
'--model-size': 14,
'--generated-data': '',
'--shards': 2,
'--pipeline': '',
'--gradient-accumulation-count': 128,
'--batch-size': 1,
'--no-validation': '',
'--pipeline-splits': 'b2/0/relu'
})
cls.out = out
cls.training = get_csv(out, 'training.csv')
def setUp(self):
out = run_train(
self, **{
'--iterations': 500,
'--dataset': 'imagenet',
'--model': 'resnext',
'--model-size': 14,
'--generated-data': '',
'--shards': 2,
'--pipeline': '',
'--gradient-accumulation-count': 128,
'--micro-batch-size': 1,
'--no-validation': '',
'--pipeline-splits': 'b2/0/relu',
'--fused-preprocessing': ''
})
self.out = out
self.training = get_csv(out, 'training.csv')
def setUp(self):
out = run_train(
self, **{
'--generated-data': '',
'--dataset': 'ImageNet',
'--model': 'efficientnet',
'--shards': 2,
'--pipeline': '',
'--gradient-accumulation-count': 256,
'--micro-batch-size': 2,
'--no-validation': '',
'--enable-recomputation': '',
'--available-memory-proportion': 0.2,
'--iterations': 10,
'--pipeline-splits': 'block3b',
'--fused-preprocessing': ''
})
self.out = out
self.training = get_csv(out, 'training.csv')
def setUpClass(cls):
out = run_train(
**{
'--iterations': 10,
'--batches-per-step': 10,
'--dataset': 'imagenet',
'--synthetic-data': '',
'--model-size': 50,
'--shards': 2,
'--replicas': 2,
'--pipeline-depth': 128,
'--pipeline-schedule': 'Grouped',
'--batch-size': 2,
'--no-validation': '',
'--xla-recompute': '',
'--available-memory-proportion': 0.1,
'--pipeline-splits': 'b3/0/relu'
})
cls.training = get_csv(out, 'training.csv')
def setUp(self):
out = run_train(
self, **{
'--iterations': 10,
'--batches-per-step': 10,
'--dataset': 'imagenet',
'--generated-data': '',
'--model-size': 50,
'--shards': 2,
'--replicas': 2,
'--pipeline': '',
'--gradient-accumulation-count': 128,
'--pipeline-schedule': 'Grouped',
'--micro-batch-size': 2,
'--no-validation': '',
'--enable-recomputation': '',
'--available-memory-proportion': 0.1,
'--pipeline-splits': 'b3/0/relu'
})
self.training = get_csv(out, 'training.csv')
def setUp(self):
out = run_train(
self, **{
'--iterations': 10,
'--batches-per-step': 10,
'--dataset': 'imagenet',
'--generated-data': '',
'--model': 'EfficientNet',
'--model-size': 'B0',
'--shards': 2,
'--pipeline': '',
'--gradient-accumulation-count': 128,
'--micro-batch-size': 4,
'--no-validation': '',
'--enable-recomputation': '',
'--pipeline-schedule': 'Grouped',
'--group-dim': 16,
'--expand-ratio': 4,
'--pipeline-splits': 'block3b'
})
self.training = get_csv(out, 'training.csv')
def setUpClass(cls):
out = run_train(**{'--iterations': 10,
'--batches-per-step': 10,
'--dataset': 'imagenet',
'--synthetic-data': '',
'--model': 'EfficientNet',
'--model-size': 'B0',
'--shards': 2,
'--replicas': 2,
'--pipeline': '',
'--gradient-accumulation-count': 128,
'--pipeline-schedule': 'Grouped',
'--batch-size': 2,
'--no-validation': '',
'--xla-recompute': '',
'--group-dim': 16,
'--expand-ratio': 4,
'--use-relu': '',
'--available-memory-proportion': 0.2,
'--pipeline-splits': 'block3b'})
cls.training = get_csv(out, 'training.csv')
