parser.add_argument("--folder", type=str, nargs="?", default=None)
args = vars(parser.parse_args())
ds_shape = d5ds.dataset_shape('ucf101')
ds_classes, sample_shape = ds_shape[0], ds_shape[1:]
train_set, validation_set = d5ds.load_ucf101('0', 'label', folder=args['folder'],
normalize=True, max_length=700, skip_frames=10)
seq_lengths = [av.open(path).streams.video[0].frames for path in train_set.data]
train_sampler = d5.BucketSampler(train_set, BATCH_SIZE, seq_lengths, max_length=500,
transformations=[d5ref.Crop((224, 224)),])
validation_sampler = d5.OrderedSampler(validation_set, BATCH_SIZE,
transformations=[d5ref.Crop((224, 224)),])
model = ResNet50LSTM(num_classes=ds_classes, pretrained=True)
loss = torch.nn.CrossEntropyLoss()
executor = d5fw.PyTorchNativeGraphExecutor(model, loss, device=d5.GPUDevice())
optimizer = d5fw.GradientDescent(executor, 'loss')
METRICS = [
d5.TestAccuracy(),
d5.WallclockTime(reruns=0, avg_over=1)
]
results = d5.test_training(executor, train_sampler, validation_sampler,
optimizer, EPOCHS, BATCH_SIZE, 'output',
metrics=[m for m in METRICS])
def run_recipe(fixed: Dict[str, Any], mutable: Dict[str, Any],
metrics: List[Tuple[d5.TestMetric, Any]]) -> bool:
""" Runs a Deep500 recipe (see file documentation). Returns True on success
and False on failure, printing the unacceptable metrics. """
# Argument validation
if any(k in mutable for k in fixed.keys()):
raise RuntimeError('Fixed and mutable components cannot overlap')
# Create unified dictionary
comps = dict(fixed, **mutable)
# Add missing arguments and keyword arguments
old_keys = list(comps.keys())
for k in old_keys:
if (k not in ['batch_size', 'epochs', 'events']
and not (k.endswith('_args') or k.endswith('_kwargs'))):
if ('%s_args' % k) not in comps:
comps['%s_args' % k] = tuple()
if ('%s_kwargs' % k) not in comps:
comps['%s_kwargs' % k] = {}
########################################################################
# Obtain dataset metadata
if 'dataset' not in comps:
raise SyntaxError('Dataset must be specified in training recipe')
if isinstance(comps['dataset'], str):
loss_op = d5ds.dataset_loss(comps['dataset'])
ds_shape = d5ds.dataset_shape(comps['dataset'])
else:
loss_op = comps['dataset'].loss
ds_shape = comps['dataset'].shape
ds_classes, sample_shape = ds_shape[0], ds_shape[1:]
# Construct network
if 'model' not in comps:
raise SyntaxError('Model must be specified in recipe')
if 'batch_size' not in comps:
raise SyntaxError('Batch size must be specified in training recipe')
batch = comps['batch_size']
if isinstance(comps['model'], str):
# ONNX file
if os.path.isfile(comps['model']):
network = d5.parser.load_and_parse_model(comps['model'])
input_node = network.get_input_nodes()[0].name
output_node = network.get_output_nodes()[0].name
else: # Standard model
network, input_node, output_node = \
d5nt.create_model(comps['model'], batch, *comps['model_args'],
classes=ds_classes, shape=sample_shape,
**comps['model_kwargs'])
else: # Callable
network, input_node, output_node = comps['model'](
batch,
*comps['model_args'],
classes=ds_classes,
shape=sample_shape,
**comps['model_kwargs'])
# Add loss function to model
network.add_operation(loss_op([output_node, 'label'], 'loss'))
# Construct dataset
if isinstance(comps['dataset'], str):
train_set, validation_set = d5ds.load_dataset(
comps['dataset'], input_node, 'label', *comps['dataset_args'],
**comps['dataset_kwargs'])
else:
train_set, validation_set = comps['dataset'](input_node, 'label',
*comps['dataset_args'],
**comps['dataset_kwargs'])
# Construct samplers
if 'train_sampler' in comps:
if isinstance(comps['train_sampler'], d5.Sampler):
train_sampler = comps['train_sampler']
else:
train_sampler = comps['train_sampler'](
train_set, batch, *comps['train_sampler_args'],
**comps['train_sampler_kwargs'])
else:
train_sampler = train_set
if 'validation_sampler' in comps:
if isinstance(comps['validation_sampler'], d5.Sampler):
validation_sampler = comps['validation_sampler']
else:
validation_sampler = comps['validation_sampler'](
validation_set, batch, *comps['validation_sampler_args'],
**comps['validation_sampler_kwargs'])
else:
validation_sampler = validation_set
# Construct executor
if 'executor' not in comps:
raise SyntaxError('Executor must be specified in recipe')
if isinstance(comps['executor'], d5.GraphExecutor):
executor = comps['executor']
else:
executor = comps['executor'](network, *comps['executor_args'],
**comps['executor_kwargs'])
# Construct optimizer
if 'optimizer' not in comps:
raise SyntaxError('Optimizer must be specified in training recipe')
optimizer = comps['optimizer'](executor, 'loss', *comps['optimizer_args'],
**comps['optimizer_kwargs'])
# Add total time to metrics
metrics.append((d5.WallclockTime(reruns=0, avg_over=1), None))
########################################################################
# Create trainer and run
if 'epochs' not in comps:
raise SyntaxError('Epochs must be specified in training recipe')
if 'events' not in comps:
comps['events'] = None
results = d5.test_training(executor,
train_sampler,
validation_sampler,
optimizer,
comps['epochs'],
batch,
output_node,
metrics=[m[0] for m in metrics],
events=comps['events'])
# Verify results
ok = True
for (metric, acceptable), result in zip(metrics, results):
if acceptable is not None:
if result < acceptable:
print('FAIL %s: %s (Acceptable: %s)' %
(type(metric).__name__, result, acceptable))
ok = False
if not ok:
return False
else:
print('PASSED')
return True
# Create distributed samplers
if PARTITION_DATASET:
train_sampler = d5.PartitionedDistributedSampler(
d5.ShuffleSampler(train_set, BATCH_SIZE), comm)
else:
train_sampler = d5.DistributedSampler(
d5.ShuffleSampler(train_set, BATCH_SIZE), comm)
if comm is None or comm.rank == 0:
# No need to distribute test_set
test_sampler = d5.ShuffleSampler(test_set, BATCH_SIZE)
else:
# No need to test if not rank 0
test_sampler = None
#############################
# Events: Only print progress on rank 0
events = d5.DefaultRunnerEvents(MAX_EPOCHS) if comm is None or comm.rank == 0 else []
# Metrics: Add communication volume
metrics = d5.DefaultTrainingMetrics() + [d5.CommunicationVolume()]
# Run distributed training
d5.test_training(executor, train_sampler, test_sampler, optimizer,
MAX_EPOCHS, BATCH_SIZE, OUTPUT_NODE, events=events,
metrics=metrics)
# Wait for everyone to finish and finalize MPI if necessary
d5.mpi_end_barrier()
z_t = self.z[param_name]
z_t2 = z_t - self.alpha_t * eta_t * grad
y_t2 = old_param - eta_t * grad
self.z[param_name] = z_t2
self.y[param_name] = y_t2
self.squares[param_name] = squared_grad
adjusted_lr = self.lr / (self.eps + np.sqrt(squared_grad))
self.init = False
return old_param - adjusted_lr * grad
if __name__ == '__main__':
from deep500 import networks as d5net, datasets as d5ds
from deep500.frameworks import tensorflow as d5tf
from deep500.frameworks import reference as d5ref
batch_size = 1024
# Create network and dataset
net, innode, outnode = d5net.create_model('simple_cnn', batch_size)
net.add_operation(d5.ops.LabelCrossEntropy([outnode, 'label'], 'loss'))
train, test = d5ds.load_dataset('mnist', innode, 'label')
# Create executor and optimizer
executor = d5tf.from_model(net)
opt = AcceleGradOptimizer(executor)
# Run training
d5.test_training(executor, train, test, opt, 5, batch_size, outnode)