def fwd_graph(popart_model, torch_model, mapping=None, transform=None):
# ------------------- PopART --------------------
config = popart_model.config
builder = popart_model.builder
sequence_info = popart.TensorInfo(
"INT32", [config.batch_size * config.sequence_length])
indices = builder.addInputTensor(sequence_info)
positions = builder.addInputTensor(sequence_info)
segments = builder.addInputTensor(sequence_info)
data = {
indices:
np.random.randint(0, config.vocab_length,
(config.batch_size * config.sequence_length)).astype(
np.int32),
positions:
np.random.randint(0, config.sequence_length,
(config.batch_size * config.sequence_length)).astype(
np.int32),
segments:
np.random.randint(0, 2,
(config.batch_size * config.sequence_length)).astype(
np.int32)
}
output = popart_model.build_graph(indices, positions, segments)
proto = builder.getModelProto()
outputs, post_proto = run_py(
proto,
data,
output,
ipus=math.ceil(config.num_layers / config.layers_per_ipu) +
popart_model.layer_offset)
# ----------------- PopART -> PyTorch ----------------
proto = onnx.load_model_from_string(proto)
inputs = {
"input_ids":
data[indices].reshape(config.batch_size, config.sequence_length),
"position_ids":
data[positions].reshape(config.batch_size, config.sequence_length),
"token_type_ids":
data[segments].reshape(config.batch_size, config.sequence_length)
}
torch_to_onnx = get_mapping(config, init=mapping)
transform_weights = get_transform(config, init=transform)
# ------------------- PyTorch -------------------------
# Turn off dropout
torch_model.eval()
copy_weights_to_torch(torch_model, proto, torch_to_onnx, transform_weights)
torch_outputs = run_fwd_model(inputs, torch_model)
check_tensors(torch_outputs, outputs)
def pytorch_result_and_model(torch_config, inputs, popart_proto, is_bwd=False):
# Conversion of the popart model to onnx
proto = onnx.load_model_from_string(popart_proto)
torch_model = BertFCN(torch_config)
# Turn off dropout
torch_model.eval()
copy_weights_to_torch(torch_model,
proto,
TORCH_TO_ONNX,
transform=TRANSPOSE_WEIGHTS)
result = run_fwd_model(inputs, torch_model)
if is_bwd:
l1_lambda = 0.1
optim = torch.optim.SGD(torch_model.parameters(),
0.01,
weight_decay=0.0,
momentum=0.0)
result = torch_model(*[torch.from_numpy(t).float() for t in inputs])[0]
torch_loss = l1_lambda * torch.norm(result, 1)
torch_loss.backward()
optim.step()
result = result.detach().numpy()
return result, torch_model
def pytorch_result_and_model(torch_config,
inputs,
popart_proto,
mode,
is_bwd=False,
momentum=0.0):
# Conversion of the popart model to onnx
proto = onnx.load_model_from_string(popart_proto)
torch_model = BertFCN(torch_config)
# Turn off dropout
torch_model.eval()
copy_weights_to_torch(torch_model,
proto,
TORCH_TO_ONNX[mode],
transform=TRANSPOSE_WEIGHTS)
result = run_fwd_model(inputs, torch_model)
if is_bwd:
l1_lambda = 0.1
optim = torch.optim.SGD(torch_model.parameters(),
lr,
weight_decay=0.0,
momentum=momentum)
if momentum > 0.0:
for group in optim.param_groups:
for p in group['params']:
optim.state[p]['momentum_buffer'] = p.data * 0.
optim.state[p]['exp_avg'] = p.data * 0.
optim.state[p]['exp_avg_sq'] = p.data * 0.
optim.state[p]['step'] = 0
for _ in range(num_reps_bwd):
result = torch_model(
*[torch.from_numpy(t).float() for t in inputs])[0]
torch_loss = l1_lambda * torch.norm(result, 1)
torch_loss.backward()
optim.step()
optim.zero_grad()
result = [result.detach().numpy()]
return result, torch_model
def pytorch_result_and_model(torch_config,
inputs,
popart_proto,
weight_decay=0.0,
lr=0.0,
l1_lambda=0.0):
proto = onnx.load_model_from_string(popart_proto)
torch_model = BertFCN(torch_config)
torch_model.eval() # Turn off dropout
copy_weights_to_torch(torch_model,
proto,
TORCH_TO_ONNX,
transform=TRANSPOSE_WEIGHTS)
run_fwd_model(inputs, torch_model)
decay = []
no_decay = []
for name, param in torch_model.named_parameters():
if "bias" in name or "LayerNorm" in name:
no_decay.append(param)
else:
decay.append(param)
params = [{
'params': no_decay,
'weight_decay': 0.
}, {
'params': decay,
'weight_decay': weight_decay
}]
optim = torch.optim.SGD(params, lr, momentum=0.0)
result = torch_model(*[torch.from_numpy(t).float() for t in inputs])[0]
torch_loss = l1_lambda * torch.norm(result, 1)
torch_loss.backward()
optim.step()
result = result.detach().numpy()
return result, torch_model
def test_embedding_projection_bwd(custom_ops):
l1_lambda = 0.1
# ------------------- PopART --------------------
builder = popart.Builder(opsets={
"ai.onnx": 9,
"ai.onnx.ml": 1,
"ai.graphcore": 1
})
config = BertConfig(vocab_length=9728,
embedding_serialization_vocab_steps=4,
micro_batch_size=1,
hidden_size=288,
sequence_length=128,
activation_type='relu',
popart_dtype="FLOAT",
no_dropout=True,
no_cls_layer=False,
# Currently updating embedding dict with projection is only
# available with momentum. And PopART != Pytorch momentum
# due to a bootstrapping step on iter 0.
update_embedding_dict=False)
popart_model = Bert(config, builder=builder)
sequence_info = popart.TensorInfo(
"UINT32", [config.micro_batch_size * config.sequence_length])
indices = builder.addInputTensor(sequence_info)
data = {
indices:
np.random.randint(0, config.vocab_length,
(config.micro_batch_size * config.sequence_length)).astype(
np.uint32)
}
x = popart_model.gather(
indices, config.vocab_length, "Embedding_Dict")
x = popart_model.norm(x)
x = popart_model.dropout(x)
with popart_model.device_scope(nameScope="CLS"):
x = popart_model.lm_prediction_head(x)
output = popart_model.projection(x)
l1 = builder.aiGraphcore.l1loss(
[output], l1_lambda, debugPrefix="l1LossVal", reduction=popart.ReductionType.Sum)
proto = builder.getModelProto()
optimizer = popart.ConstSGD(0.01)
outputs, post_proto = run_py(proto,
data, output,
loss=l1,
optimizer=optimizer)
# ----------------- PopART -> PyTorch ----------------
proto = onnx.load_model_from_string(proto)
inputs = [data[indices].reshape(
config.micro_batch_size, config.sequence_length).astype(np.int32)]
# ------------------- PyTorch -------------------------
torch_model = EmbeddingProjectionModel(
TorchBertConfig(config.vocab_length,
config.hidden_size,
max_position_embeddings=config.max_positional_length,
layer_norm_eps=config.layer_norm_eps,
no_cls_layer=config.no_cls_layer,
update_embedding_dict=config.update_embedding_dict))
# Turn off dropout
torch_model.eval()
copy_weights_to_torch(torch_model,
proto,
TORCH_TO_ONNX,
transform=TRANSPOSE_WEIGHTS)
optim = torch.optim.SGD(torch_model.parameters(),
0.01,
weight_decay=0.0,
momentum=0.0)
torch_output = torch_model(*[torch.from_numpy(t).long() for t in inputs])
torch_loss = l1_lambda * torch.norm(torch_output, 1)
torch_loss.backward()
optim.step()
check_tensors([torch_output.detach().numpy()], outputs, margin=1e-5)
check_model(torch_model,
post_proto,
TORCH_TO_ONNX,
transform=TRANSPOSE_WEIGHTS)
def test_embedding_bwd(custom_ops, mode):
# ------------------- PopART --------------------
config = BertConfig(task="SQUAD",
vocab_length=9728,
batch_size=1,
hidden_size=768,
sequence_length=128,
activation_type='relu',
popart_dtype="FLOAT",
no_dropout=True,
update_embedding_dict=False,
embedding_serialization_vocab_steps=1)
popart_model = get_model(config, mode, 'embedding')
# Prevent virtualGraph attributes being added to the ops.
sequence_info = popart.TensorInfo(
"UINT32", [config.batch_size * config.sequence_length])
indices = popart_model.builder.addInputTensor(sequence_info)
positions = popart_model.builder.addInputTensor(sequence_info)
segments = popart_model.builder.addInputTensor(sequence_info)
data = {
indices:
np.random.randint(0, config.vocab_length,
(config.batch_size * config.sequence_length)).astype(
np.uint32),
positions:
np.random.randint(0, config.max_positional_length,
(config.batch_size * config.sequence_length)).astype(
np.uint32),
segments:
np.random.randint(0, 2,
(config.batch_size * config.sequence_length)).astype(
np.uint32)
}
optimizer = popart.ConstSGD(0.01)
l1_lambda = 0.1
if mode == ExecutionMode.PHASED:
user_options = {
"batchSerializationFactor": 1,
"executionPhases": popart_model.total_execution_phases,
}
output = popart_model(indices, positions, segments)
with popart_model.scope_provider(popart_model.builder,
popart_model.norm.scope):
l1 = popart_model.builder.aiGraphcore.l1loss(
[output],
l1_lambda,
debugPrefix="l1LossVal",
reduction=popart.ReductionType.Sum)
else:
user_options = {"enableStochasticRounding": True}
output = popart_model.embedding(indices, positions, segments)
l1 = popart_model.builder.aiGraphcore.l1loss(
[output],
l1_lambda,
debugPrefix="l1LossVal",
reduction=popart.ReductionType.Sum)
proto = popart_model.builder.getModelProto()
outputs, post_proto = run_py(proto,
data,
output,
ipus=1,
loss=l1,
optimizer=optimizer,
user_options=user_options,
execution_mode=mode)
# ----------------- PopART -> PyTorch ----------------
proto = onnx.load_model_from_string(proto)
inputs = [
data[t].reshape(config.batch_size,
config.sequence_length).astype(np.int32)
for t in [indices, positions, segments]
]
# ------------------- PyTorch -------------------------
torch_model = BertEmbeddings(
TorchBertConfig(config.vocab_length,
config.hidden_size,
max_position_embeddings=config.max_positional_length,
layer_norm_eps=config.layer_norm_eps))
# Turn off dropout
torch_model.eval()
copy_weights_to_torch(torch_model, proto, TORCH_TO_ONNX[mode], {})
optim = torch.optim.SGD(torch_model.parameters(),
0.01,
weight_decay=0.0,
momentum=0.0)
torch_output = torch_model(*[torch.from_numpy(t).long() for t in inputs])
torch_loss = l1_lambda * torch.norm(torch_output, 1)
torch_loss.backward()
optim.step()
torch_outputs = [torch_output.detach().numpy()]
check_tensors(torch_outputs, outputs, margin=1e-06)
check_model(torch_model, post_proto, TORCH_TO_ONNX[mode], {}, margin=1e-06)
def fwd_graph(popart_model,
torch_model,
mapping=None,
transform=None,
replication_factor=1,
replicated_tensor_sharding=False):
# ------------------- PopART --------------------
config = popart_model.config
builder = popart_model.builder
sequence_info = popart.TensorInfo(
"UINT32", [config.micro_batch_size * config.sequence_length])
indices = builder.addInputTensor(sequence_info)
positions = builder.addInputTensor(sequence_info)
segments = builder.addInputTensor(sequence_info)
data = {
indices:
np.random.randint(0, config.vocab_length,
(replication_factor, config.micro_batch_size *
config.sequence_length)).astype(np.uint32),
positions:
np.random.randint(0, config.sequence_length,
(replication_factor, config.micro_batch_size *
config.sequence_length)).astype(np.uint32),
segments:
np.random.randint(0, 2, (replication_factor, config.micro_batch_size *
config.sequence_length)).astype(np.uint32)
}
output = popart_model.build_graph(indices, positions, segments)
ipus = popart_model.total_ipus
proto = builder.getModelProto()
outputs, _ = run_py(proto,
data,
output,
replication_factor=replication_factor,
replicated_tensor_sharding=replicated_tensor_sharding,
ipus=ipus)
# ----------------- PopART -> PyTorch ----------------
proto = onnx.load_model_from_string(proto)
inputs = {
"input_ids":
data[indices].reshape(replication_factor * config.micro_batch_size,
config.sequence_length).astype(np.int32),
"position_ids":
data[positions].reshape(replication_factor * config.micro_batch_size,
config.sequence_length).astype(np.int32),
"token_type_ids":
data[segments].reshape(replication_factor * config.micro_batch_size,
config.sequence_length).astype(np.int32)
}
torch_to_onnx = get_mapping(config, init=mapping)
transform_weights = get_transform(config, init=transform)
# ------------------- PyTorch -------------------------
# Turn off dropout
torch_model.eval()
copy_weights_to_torch(torch_model, proto, torch_to_onnx, transform_weights)
torch_outputs = run_fwd_model(inputs, torch_model)
check_tensors(torch_outputs, outputs)
def test_embedding_fwd(custom_ops):
# ------------------- PopART --------------------
config = BertConfig(task="SQUAD",
vocab_length=9728,
micro_batch_size=1,
hidden_size=768,
sequence_length=128,
activation_type='relu',
popart_dtype="FLOAT",
no_dropout=True,
inference=True)
popart_model = Bert(config)
sequence_info = popart.TensorInfo(
"UINT32", [config.micro_batch_size * config.sequence_length])
indices = popart_model.builder.addInputTensor(sequence_info)
positions = popart_model.builder.addInputTensor(sequence_info)
segments = popart_model.builder.addInputTensor(sequence_info)
data = {
indices:
np.random.randint(
0, config.vocab_length,
(config.micro_batch_size * config.sequence_length)).astype(
np.uint32),
positions:
np.random.randint(
0, config.max_positional_length,
(config.micro_batch_size * config.sequence_length)).astype(
np.uint32),
segments:
np.random.randint(
0, 2, (config.micro_batch_size * config.sequence_length)).astype(
np.uint32)
}
user_options = {"enableStochasticRounding": True}
with popart_model.builder.nameScope("Embedding"):
output = popart_model.embedding(indices, positions, segments)
proto = popart_model.builder.getModelProto()
outputs, post_proto = run_py(proto,
data,
output,
user_options=user_options)
# ----------------- PopART -> PyTorch ----------------
proto = onnx.load_model_from_string(proto)
inputs = [
data[t].reshape(config.micro_batch_size,
config.sequence_length).astype(np.int32)
for t in [indices, positions, segments]
]
# ------------------- PyTorch -------------------------
torch_model = BertEmbeddings(
TorchBertConfig(config.vocab_length,
config.hidden_size,
max_position_embeddings=config.max_positional_length,
layer_norm_eps=config.layer_norm_eps))
torch_model.eval()
copy_weights_to_torch(torch_model, proto, TORCH_TO_ONNX, {})
torch_outputs = run_fwd_model(inputs, torch_model)
check_tensors(torch_outputs, outputs, margin=5e-7)
def test_embedding_fwd(custom_ops):
# ------------------- PopART --------------------
builder = popart.Builder(opsets={
"ai.onnx": 9,
"ai.onnx.ml": 1,
"ai.graphcore": 1
})
config = BertConfig(vocab_length=9728,
batch_size=1,
hidden_size=768,
sequence_length=128,
activation_type='relu',
popart_dtype="FLOAT",
no_dropout=True,
custom_ops=['gather'],
inference=True)
popart_model = Bert(config, builder=builder)
# Prevent virtualGraph attributes being added to the ops.
popart_model.embedding_scope = popart_model.device_scope(None, None)
popart_model.embedding_split_scope = popart_model.embedding_scope
sequence_info = popart.TensorInfo(
"UINT32", [config.batch_size * config.sequence_length])
indices = builder.addInputTensor(sequence_info)
positions = builder.addInputTensor(sequence_info)
segments = builder.addInputTensor(sequence_info)
data = {
indices:
np.random.randint(0, config.vocab_length,
(config.batch_size * config.sequence_length)).astype(
np.uint32),
positions:
np.random.randint(0, config.max_positional_length,
(config.batch_size * config.sequence_length)).astype(
np.uint32),
segments:
np.random.randint(0, 2,
(config.batch_size * config.sequence_length)).astype(
np.uint32)
}
# Use the custom embedding for layout
output = popart_model.embedding(indices, positions, segments)
proto = builder.getModelProto()
outputs, post_proto = run_py(
proto, data, output, user_options={"enableStochasticRounding": True})
# ----------------- PopART -> PyTorch ----------------
proto = onnx.load_model_from_string(proto)
inputs = [
data[t].reshape(config.batch_size,
config.sequence_length).astype(np.int32)
for t in [indices, positions, segments]
]
torch_to_onnx = {
"word_embeddings.weight": "Embedding_Dict",
"position_embeddings.weight": "Positional_Dict",
"token_type_embeddings.weight": "Segment_Dict",
"LayerNorm.weight": "Gamma",
"LayerNorm.bias": "Beta"
}
transposed_weights = {
"word_embeddings.weight": np.transpose,
"position_embeddings.weight": np.transpose,
}
# ------------------- PyTorch -------------------------
torch_model = BertEmbeddings(
TorchBertConfig(config.vocab_length,
config.hidden_size,
max_position_embeddings=config.max_positional_length,
layer_norm_eps=config.layer_norm_eps))
torch_model.eval()
copy_weights_to_torch(torch_model, proto, torch_to_onnx,
transposed_weights)
torch_outputs = run_fwd_model(inputs, torch_model)
check_tensors(torch_outputs, outputs)
def test_embedding_projection_bwd(custom_ops):
l1_lambda = 0.1
# ------------------- PopART --------------------
builder = popart.Builder(opsets={
"ai.onnx": 9,
"ai.onnx.ml": 1,
"ai.graphcore": 1
})
config = BertConfig(vocab_length=9728,
batch_size=1,
hidden_size=768,
sequence_length=128,
activation_type='relu',
popart_dtype="FLOAT",
no_dropout=True,
custom_ops=['gather'])
popart_model = Bert(config, builder=builder)
sequence_info = popart.TensorInfo(
"INT32", [config.batch_size * config.sequence_length])
indices = builder.addInputTensor(sequence_info)
data = {
indices:
np.random.randint(0, config.vocab_length,
(config.batch_size * config.sequence_length)).astype(
np.int32)
}
x = popart_model.embedding_custom(
indices, config.vocab_length, "Embedding_Dict", detach=True)
x = popart_model.norm(x)
x = popart_model.dropout(x)
with popart_model.device_scope(nameScope="CLS"):
x = popart_model.lm_prediction_head(x)
output = popart_model.projection(x)
proto = builder.getModelProto()
l1 = popart.L1Loss(output, "l1LossVal", l1_lambda)
optimizer = popart.ConstSGD(0.01)
outputs, post_proto = run_py(proto,
data, output,
loss=l1,
optimizer=optimizer,
user_options={"enableStochasticRounding": True})
# ----------------- PopART -> PyTorch ----------------
proto = onnx.load_model_from_string(proto)
inputs = [data[indices].reshape(config.batch_size, config.sequence_length)]
# ------------------- PyTorch -------------------------
torch_model = EmbeddingProjectionModel(
TorchBertConfig(config.vocab_length,
config.hidden_size,
max_position_embeddings=config.max_positional_length,
layer_norm_eps=config.layer_norm_eps))
# Turn off dropout
torch_model.eval()
copy_weights_to_torch(torch_model,
proto,
torch_to_onnx,
transform=transposed_weights)
optim = torch.optim.SGD(torch_model.parameters(),
0.01,
weight_decay=0.0,
momentum=0.0)
torch_output = torch_model(*[torch.from_numpy(t).long() for t in inputs])
torch_loss = l1_lambda * torch.norm(torch_output, 1)
torch_loss.backward()
optim.step()
check_tensors([torch_output.detach().numpy()], outputs, margin=1e-5)
check_model(torch_model,
post_proto,
torch_to_onnx,
transform=transposed_weights)
def test_attention_fwd(custom_ops):
# ------------------- PopART --------------------
builder = popart.Builder()
config = BertConfig(task="PRETRAINING",
vocab_length=9728,
batch_size=1,
hidden_size=768,
sequence_length=128,
activation_type='relu',
popart_dtype="FLOAT",
no_dropout=True,
custom_ops=['attention'],
inference=True)
popart_model = Bert(config, builder=builder)
input_info = popart.TensorInfo(
config.popart_dtype,
[config.batch_size * config.sequence_length, config.hidden_size])
input_tensor = builder.addInputTensor(input_info)
mask_info = popart.TensorInfo("INT32", [config.batch_size])
mmask_tensor = builder.addInputTensor(mask_info)
smask_tensor = builder.addInputTensor(mask_info)
data = {
input_tensor:
np.random.normal(0, 0.02, input_info.shape()).astype(config.dtype),
mmask_tensor:
np.random.randint(0, config.mask_tokens + 1,
(config.batch_size, )).astype(np.int32),
smask_tensor:
np.random.randint(config.mask_tokens, config.sequence_length + 1,
(config.batch_size, )).astype(np.int32)
}
output = popart_model.attention(input_tensor, [mmask_tensor, smask_tensor])
proto = builder.getModelProto()
outputs, post_proto = run_py(proto, data, output)
# ----------------- PopART -> PyTorch ----------------
proto = onnx.load_model_from_string(proto)
inputs = [
data[input_tensor].reshape(config.batch_size, config.sequence_length,
config.hidden_size).astype(np.float32),
get_torch_mask(config, [data[mmask_tensor], data[smask_tensor]])
]
torch_to_onnx = {
"self.query.weight": "QKV",
"self.key.weight": "QKV",
"self.value.weight": "QKV",
"output.dense.weight": "Out",
"output.LayerNorm.weight": "Gamma",
"output.LayerNorm.bias": "Beta"
}
split_qkv = {
"self.query.weight":
lambda arr: arr[:, 0:config.hidden_size].T,
"self.key.weight":
lambda arr: arr[:, config.hidden_size:config.hidden_size * 2].T,
"self.value.weight":
lambda arr: arr[:, config.hidden_size * 2:config.hidden_size * 3].T,
"output.dense.weight":
np.transpose
}
# ------------------- PyTorch -------------------------
torch_model = BertAttention(
TorchBertConfig(config.vocab_length,
config.hidden_size,
config.num_layers,
config.attention_heads,
layer_norm_eps=config.layer_norm_eps))
# Turn off dropout
torch_model.eval()
copy_weights_to_torch(torch_model,
proto,
torch_to_onnx,
transform=split_qkv)
# Model to test against
torch_outputs = run_fwd_model(inputs, torch_model)
check_tensors(torch_outputs, outputs)
def test_attention_bwd(custom_ops):
l1_lambda = 0.1
# ------------------- PopART --------------------
builder = popart.Builder()
config = BertConfig(task="PRETRAINING",
vocab_length=9728,
batch_size=1,
hidden_size=768,
sequence_length=128,
activation_type='relu',
popart_dtype="FLOAT",
no_dropout=True,
custom_ops=['attention'])
popart_model = Bert(config, builder=builder)
input_info = popart.TensorInfo(
config.popart_dtype,
[config.batch_size * config.sequence_length, config.hidden_size])
input_tensor = builder.addInputTensor(input_info)
mask_info = popart.TensorInfo("INT32", [config.batch_size])
mmask_tensor = builder.addInputTensor(mask_info)
smask_tensor = builder.addInputTensor(mask_info)
data = {
input_tensor:
np.random.normal(0, 0.02, input_info.shape()).astype(config.dtype),
mmask_tensor:
np.random.randint(0, config.mask_tokens + 1,
(config.batch_size, )).astype(np.int32),
smask_tensor:
np.random.randint(config.mask_tokens, config.sequence_length + 1,
(config.batch_size, )).astype(np.int32)
}
output = popart_model.attention(input_tensor, [mmask_tensor, smask_tensor])
proto = builder.getModelProto()
l1 = popart.L1Loss(output, "l1LossVal", l1_lambda)
optimizer = popart.ConstSGD(0.01)
outputs, post_proto = run_py(proto,
data, (output, l1.output(0)),
loss=l1,
optimizer=optimizer)
# ----------------- PopART -> PyTorch ----------------
proto = onnx.load_model_from_string(proto)
inputs = [
data[input_tensor].reshape(config.batch_size, config.sequence_length,
config.hidden_size),
get_torch_mask(config, [data[mmask_tensor], data[smask_tensor]])
]
torch_to_onnx = {
"self.query.weight": "QKV",
"self.key.weight": "QKV",
"self.value.weight": "QKV",
"output.dense.weight": "Out",
"output.LayerNorm.weight": "Gamma",
"output.LayerNorm.bias": "Beta"
}
split_qkv = {
"self.query.weight":
lambda arr: arr[:, 0:config.hidden_size].T,
"self.key.weight":
lambda arr: arr[:, config.hidden_size:config.hidden_size * 2].T,
"self.value.weight":
lambda arr: arr[:, config.hidden_size * 2:config.hidden_size * 3].T,
"output.dense.weight":
np.transpose
}
# ------------------- PyTorch -------------------------
torch_model = BertAttention(
TorchBertConfig(config.vocab_length,
config.hidden_size,
config.num_layers,
config.attention_heads,
layer_norm_eps=config.layer_norm_eps))
# Turn off dropout
torch_model.eval()
copy_weights_to_torch(torch_model,
proto,
torch_to_onnx,
transform=split_qkv)
optim = torch.optim.SGD(torch_model.parameters(),
0.01,
weight_decay=0.0,
momentum=0.0)
torch_output = torch_model(*[torch.from_numpy(t).float()
for t in inputs])[0]
torch_loss = l1_lambda * torch.norm(torch_output, 1)
torch_loss.backward()
optim.step()
check_tensors([torch_output.detach().numpy()], outputs)
check_model(torch_model, post_proto, torch_to_onnx, transform=split_qkv)
def test_embedding_fwd(custom_ops):
# ------------------- PopART --------------------
builder = popart.Builder(opsets={
"ai.onnx": 9,
"ai.onnx.ml": 1,
"ai.graphcore": 1
})
config = BertConfig(task="SQUAD",
vocab_length=9728,
batch_size=1,
hidden_size=768,
sequence_length=128,
activation_type='relu',
popart_dtype="FLOAT",
no_dropout=True,
inference=True)
popart_model = Bert(config, builder=builder)
# Prevent virtualGraph attributes being added to the ops.
popart_model.embedding_scope = popart_model.device_scope(None, None)
popart_model.embedding_split_scope = popart_model.embedding_scope
sequence_info = popart.TensorInfo(
"UINT32", [config.batch_size * config.sequence_length])
indices = builder.addInputTensor(sequence_info)
positions = builder.addInputTensor(sequence_info)
segments = builder.addInputTensor(sequence_info)
data = {
indices:
np.random.randint(0, config.vocab_length,
(config.batch_size * config.sequence_length)).astype(
np.uint32),
positions:
np.random.randint(0, config.max_positional_length,
(config.batch_size * config.sequence_length)).astype(
np.uint32),
segments:
np.random.randint(0, 2,
(config.batch_size * config.sequence_length)).astype(
np.uint32)
}
output = popart_model.embedding(indices, positions, segments)
proto = builder.getModelProto()
outputs, post_proto = run_py(proto, data, output)
# ----------------- PopART -> PyTorch ----------------
proto = onnx.load_model_from_string(proto)
inputs = [
data[t].reshape(config.batch_size,
config.sequence_length).astype(np.int32)
for t in [indices, positions, segments]
]
# ------------------- PyTorch -------------------------
torch_model = BertEmbeddings(
TorchBertConfig(config.vocab_length,
config.hidden_size,
max_position_embeddings=config.max_positional_length,
layer_norm_eps=config.layer_norm_eps))
torch_model.eval()
copy_weights_to_torch(torch_model, proto, TORCH_TO_ONNX, {})
torch_outputs = run_fwd_model(inputs, torch_model)
check_tensors(torch_outputs, outputs)
def test_attention_bwd(mode):
l1_lambda = 0.1
# ------------------- PopART --------------------
config = BertConfig(task="PRETRAINING",
vocab_length=9728,
batch_size=1,
hidden_size=768,
sequence_length=128,
activation_type='relu',
popart_dtype="FLOAT",
no_dropout=True,
no_attn_dropout=True)
popart_model = get_model(config, mode, 'attention')
input_info = popart.TensorInfo(
config.popart_dtype,
[config.batch_size * config.sequence_length, config.hidden_size])
input_tensor = popart_model.builder.addInputTensor(input_info)
mask_info = popart.TensorInfo("UINT32", [config.batch_size])
mmask_tensor = popart_model.builder.addInputTensor(mask_info)
smask_tensor = popart_model.builder.addInputTensor(mask_info)
data = {
input_tensor:
np.random.normal(0, 0.02, input_info.shape()).astype(config.dtype),
mmask_tensor:
np.random.randint(0, config.mask_tokens + 1,
(config.batch_size, )).astype(np.uint32),
smask_tensor:
np.random.randint(config.mask_tokens, config.sequence_length + 1,
(config.batch_size, )).astype(np.uint32)
}
user_options = {}
if mode == ExecutionMode.PHASED:
user_options = {
"batchSerializationFactor": 1,
"executionPhases": popart_model.total_execution_phases
}
output = popart_model(input_tensor, [mmask_tensor, smask_tensor])
with popart_model.scope_provider(popart_model.builder,
popart_model.norm.scope):
l1 = popart_model.builder.aiGraphcore.l1loss(
[output],
l1_lambda,
debugPrefix="l1LossVal",
reduction=popart.ReductionType.Sum)
else:
user_options = {"enableStochasticRounding": True}
output = popart_model.attention(input_tensor,
[mmask_tensor, smask_tensor])
l1 = popart_model.builder.aiGraphcore.l1loss(
[output], l1_lambda, reduction=popart.ReductionType.Sum)
proto = popart_model.builder.getModelProto()
optimizer = popart.ConstSGD(0.01)
outputs, post_proto = run_py(proto,
data, (output, l1),
loss=l1,
optimizer=optimizer,
user_options=user_options,
execution_mode=mode)
# ----------------- PopART -> PyTorch ----------------
proto = onnx.load_model_from_string(proto)
inputs = [
data[input_tensor].reshape(config.batch_size, config.sequence_length,
config.hidden_size),
get_torch_mask(config, [data[mmask_tensor], data[smask_tensor]])
]
split_qkv = {
"self.query.weight":
lambda arr: arr[:, 0:config.hidden_size].T,
"self.key.weight":
lambda arr: arr[:, config.hidden_size:config.hidden_size * 2].T,
"self.value.weight":
lambda arr: arr[:, config.hidden_size * 2:config.hidden_size * 3].T,
"output.dense.weight":
np.transpose
}
# ------------------- PyTorch -------------------------
torch_model = BertAttention(
TorchBertConfig(config.vocab_length,
config.hidden_size,
config.num_layers,
config.attention_heads,
layer_norm_eps=config.layer_norm_eps))
# Turn off dropout
torch_model.eval()
copy_weights_to_torch(torch_model,
proto,
TORCH_TO_ONNX[mode],
transform=split_qkv)
optim = torch.optim.SGD(torch_model.parameters(),
0.01,
weight_decay=0.0,
momentum=0.0)
torch_output = torch_model(*[torch.from_numpy(t).float()
for t in inputs])[0]
torch_loss = l1_lambda * torch.norm(torch_output, 1)
torch_loss.backward()
optim.step()
check_tensors([torch_output.detach().numpy()], outputs)
check_model(torch_model,
post_proto,
TORCH_TO_ONNX[mode],
transform=split_qkv)
def test_embedding_projection_fwd(custom_ops):
# ------------------- PopART --------------------
builder = popart.Builder(opsets={
"ai.onnx": 9,
"ai.onnx.ml": 1,
"ai.graphcore": 1
})
config = BertConfig(vocab_length=9728,
embedding_serialization_vocab_steps=4,
micro_batch_size=1,
hidden_size=768,
sequence_length=128,
activation_type='relu',
popart_dtype="FLOAT",
no_dropout=True,
no_cls_layer=False,
inference=True)
popart_model = Bert(config, builder=builder)
sequence_info = popart.TensorInfo(
"UINT32", [config.micro_batch_size * config.sequence_length])
indices = builder.addInputTensor(sequence_info)
data = {
indices:
np.random.randint(0, config.vocab_length,
(config.micro_batch_size * config.sequence_length)).astype(
np.uint32)
}
x = popart_model.gather(
indices, config.vocab_length, "Embedding_Dict")
x = popart_model.norm(x)
x = popart_model.dropout(x)
with popart_model.builder.nameScope("CLS"):
x = popart_model.lm_prediction_head(x)
output = popart_model.projection(x)
proto = builder.getModelProto()
outputs, post_proto = run_py(proto, data, output)
# ----------------- PopART -> PyTorch ----------------
proto = onnx.load_model_from_string(proto)
inputs = [data[indices].reshape(
config.micro_batch_size, config.sequence_length).astype(np.int32)]
# ------------------- PyTorch -------------------------
torch_model = EmbeddingProjectionModel(
TorchBertConfig(config.vocab_length,
config.hidden_size,
max_position_embeddings=config.max_positional_length,
layer_norm_eps=config.layer_norm_eps,
no_cls_layer=config.no_cls_layer))
torch_model.eval()
copy_weights_to_torch(torch_model, proto, TORCH_TO_ONNX,
TRANSPOSE_WEIGHTS)
torch_model.tie_weights()
torch_outputs = run_fwd_model(inputs, torch_model)
check_tensors(torch_outputs, outputs)
def test_embedding_bwd(custom_ops):
l1_lambda = 0.1
# ------------------- PopART --------------------
builder = popart.Builder(opsets={
"ai.onnx": 9,
"ai.onnx.ml": 1,
"ai.graphcore": 1
})
config = BertConfig(vocab_length=9728,
batch_size=1,
hidden_size=768,
sequence_length=128,
activation_type='relu',
popart_dtype="FLOAT",
no_dropout=True,
custom_ops=['gather'])
popart_model = Bert(config, builder=builder)
# Prevent virtualGraph attributes being added to the ops.
popart_model.embedding_scope = popart_model.device_scope(None, None)
popart_model.embedding_split_scope = popart_model.embedding_scope
sequence_info = popart.TensorInfo(
"UINT32", [config.batch_size * config.sequence_length])
indices = builder.addInputTensor(sequence_info)
positions = builder.addInputTensor(sequence_info)
segments = builder.addInputTensor(sequence_info)
data = {
indices:
np.random.randint(0, config.vocab_length,
(config.batch_size * config.sequence_length)).astype(
np.uint32),
positions:
np.random.randint(0, config.max_positional_length,
(config.batch_size * config.sequence_length)).astype(
np.uint32),
segments:
np.random.randint(0, 2,
(config.batch_size * config.sequence_length)).astype(
np.uint32)
}
output = popart_model.embedding(indices, positions, segments)
proto = builder.getModelProto()
l1 = popart.L1Loss(output, "l1LossVal", l1_lambda)
optimizer = popart.ConstSGD(0.01)
outputs, post_proto = run_py(
proto,
data,
output,
loss=l1,
optimizer=optimizer,
user_options={"enableStochasticRounding": True})
# ----------------- PopART -> PyTorch ----------------
proto = onnx.load_model_from_string(proto)
inputs = [
data[t].reshape(config.batch_size,
config.sequence_length).astype(np.int32)
for t in [indices, positions, segments]
]
torch_to_onnx = {
"word_embeddings.weight": "Embedding_Dict",
"position_embeddings.weight": "Positional_Dict",
"token_type_embeddings.weight": "Segment_Dict",
"LayerNorm.weight": "Gamma",
"LayerNorm.bias": "Beta"
}
transposed_weights = {
"word_embeddings.weight": np.transpose,
"position_embeddings.weight": np.transpose,
}
# ------------------- PyTorch -------------------------
torch_model = BertEmbeddings(
TorchBertConfig(config.vocab_length,
config.hidden_size,
max_position_embeddings=config.max_positional_length,
layer_norm_eps=config.layer_norm_eps))
# Turn off dropout
torch_model.eval()
copy_weights_to_torch(torch_model,
proto,
torch_to_onnx,
transform=transposed_weights)
optim = torch.optim.SGD(torch_model.parameters(),
0.01,
weight_decay=0.0,
momentum=0.0)
torch_output = torch_model(*[torch.from_numpy(t).long() for t in inputs])
torch_loss = l1_lambda * torch.norm(torch_output, 1)
torch_loss.backward()
optim.step()
torch_outputs = [torch_output.detach().numpy()]
check_tensors(torch_outputs, outputs)
check_model(torch_model,
post_proto,
torch_to_onnx,
transform=transposed_weights)
def test_embedding_projection_fwd(custom_ops):
# ------------------- PopART --------------------
builder = popart.Builder(opsets={
"ai.onnx": 9,
"ai.onnx.ml": 1,
"ai.graphcore": 1
})
config = BertConfig(vocab_length=9728,
batch_size=1,
hidden_size=768,
sequence_length=128,
activation_type='relu',
popart_dtype="FLOAT",
no_dropout=True,
custom_ops=['gather'],
inference=True)
popart_model = Bert(config, builder=builder)
sequence_info = popart.TensorInfo(
"INT32", [config.batch_size * config.sequence_length])
indices = builder.addInputTensor(sequence_info)
data = {
indices:
np.random.randint(0, config.vocab_length,
(config.batch_size * config.sequence_length)).astype(
np.int32)
}
x = popart_model.embedding_custom(
indices, config.vocab_length, "Embedding_Dict", detach=True)
x = popart_model.norm(x)
x = popart_model.dropout(x)
with popart_model.device_scope(nameScope="CLS"):
x = popart_model.lm_prediction_head(x)
output = popart_model.projection(x)
proto = builder.getModelProto()
outputs, post_proto = run_py(proto, data, output,
user_options={"enableStochasticRounding": True})
# ----------------- PopART -> PyTorch ----------------
proto = onnx.load_model_from_string(proto)
inputs = [data[indices].reshape(config.batch_size, config.sequence_length)]
# ------------------- PyTorch -------------------------
torch_model = EmbeddingProjectionModel(
TorchBertConfig(config.vocab_length,
config.hidden_size,
max_position_embeddings=config.max_positional_length,
layer_norm_eps=config.layer_norm_eps))
torch_model.eval()
copy_weights_to_torch(torch_model, proto, torch_to_onnx,
transposed_weights)
torch_model.tie_weights()
torch_outputs = run_fwd_model(inputs, torch_model)
check_tensors(torch_outputs, outputs)
def test_embedding_bwd(custom_ops):
# ------------------- PopART --------------------
config = BertConfig(task="SQUAD",
vocab_length=9728,
micro_batch_size=1,
hidden_size=768,
sequence_length=128,
activation_type='relu',
popart_dtype="FLOAT",
no_dropout=True,
update_embedding_dict=True)
popart_model = Bert(config)
# Prevent virtualGraph attributes being added to the ops
sequence_info = popart.TensorInfo(
"UINT32", [config.micro_batch_size * config.sequence_length])
indices = popart_model.builder.addInputTensor(sequence_info)
positions = popart_model.builder.addInputTensor(sequence_info)
segments = popart_model.builder.addInputTensor(sequence_info)
data = {
indices:
np.random.randint(
0, config.vocab_length,
(config.micro_batch_size * config.sequence_length)).astype(
np.uint32),
positions:
np.random.randint(
0, config.max_positional_length,
(config.micro_batch_size * config.sequence_length)).astype(
np.uint32),
segments:
np.random.randint(
0, 2, (config.micro_batch_size * config.sequence_length)).astype(
np.uint32)
}
optimizer = popart.ConstSGD(0.01)
l1_lambda = 0.1
with popart_model.builder.nameScope("Embedding"):
output = popart_model.embedding(indices, positions, segments)
l1 = popart_model.builder.aiGraphcore.l1loss(
[output],
l1_lambda,
debugContext="l1LossVal",
reduction=popart.ReductionType.Sum)
num_reps = 5
proto = popart_model.builder.getModelProto()
outputs, post_proto = run_py(proto,
data,
output,
ipus=1,
loss=l1,
num_reps=num_reps,
optimizer=optimizer)
# ----------------- PopART -> PyTorch ----------------
proto = onnx.load_model_from_string(proto)
inputs = [
data[t].reshape(config.micro_batch_size,
config.sequence_length).astype(np.int32)
for t in [indices, positions, segments]
]
# ------------------- PyTorch -------------------------
torch_model = BertEmbeddings(
TorchBertConfig(config.vocab_length,
config.hidden_size,
max_position_embeddings=config.max_positional_length,
layer_norm_eps=config.layer_norm_eps,
update_embedding_dict=config.update_embedding_dict))
# Turn off dropout
torch_model.eval()
copy_weights_to_torch(torch_model, proto, TORCH_TO_ONNX, {})
optim = torch.optim.SGD(torch_model.parameters(), 0.01)
for _ in range(num_reps):
torch_output = torch_model(
*[torch.from_numpy(t).long() for t in inputs])
torch_loss = l1_lambda * torch.norm(torch_output, 1)
torch_loss.backward()
optim.step()
optim.zero_grad()
torch_outputs = [torch_output.detach().numpy()]
check_tensors(torch_outputs, outputs, margin=7e-6)
check_model(torch_model, post_proto, TORCH_TO_ONNX, {}, margin=7e-06)
def test_embedding_fwd(custom_ops, mode, batch_size,
batch_serialization_factor,
embedding_serialization_vocab_steps):
# ------------------- PopART --------------------
config = BertConfig(
task="SQUAD",
vocab_length=9728,
batch_size=batch_size,
hidden_size=768,
sequence_length=128,
activation_type='relu',
popart_dtype="FLOAT",
no_dropout=True,
inference=True,
embedding_serialization_vocab_steps=embedding_serialization_vocab_steps
)
popart_model = get_model(config, mode, 'embedding')
sequence_info = popart.TensorInfo(
"UINT32", [config.batch_size * config.sequence_length])
indices = popart_model.builder.addInputTensor(sequence_info)
positions = popart_model.builder.addInputTensor(sequence_info)
segments = popart_model.builder.addInputTensor(sequence_info)
data = {
indices:
np.random.randint(0, config.vocab_length,
(config.batch_size * config.sequence_length)).astype(
np.uint32),
positions:
np.random.randint(0, config.max_positional_length,
(config.batch_size * config.sequence_length)).astype(
np.uint32),
segments:
np.random.randint(0, 2,
(config.batch_size * config.sequence_length)).astype(
np.uint32)
}
user_options = {}
if mode == ExecutionMode.PHASED:
user_options = {
"batchSerializationFactor": batch_serialization_factor,
"executionPhases": popart_model.total_execution_phases
}
output = popart_model(indices, positions, segments)
else:
user_options = {"enableStochasticRounding": True}
with popart_model.builder.nameScope("Embedding"):
output = popart_model.embedding(indices, positions, segments)
proto = popart_model.builder.getModelProto()
outputs, post_proto = run_py(proto,
data,
output,
user_options=user_options,
execution_mode=mode)
# ----------------- PopART -> PyTorch ----------------
proto = onnx.load_model_from_string(proto)
inputs = [
data[t].reshape(config.batch_size,
config.sequence_length).astype(np.int32)
for t in [indices, positions, segments]
]
# ------------------- PyTorch -------------------------
torch_model = BertEmbeddings(
TorchBertConfig(config.vocab_length,
config.hidden_size,
max_position_embeddings=config.max_positional_length,
layer_norm_eps=config.layer_norm_eps))
torch_model.eval()
expanded_name_map, remapped_transform_map = expand_torch_to_onnx_map(
TORCH_TO_ONNX[mode], config, mode)
copy_weights_to_torch(torch_model, proto, expanded_name_map,
remapped_transform_map)
torch_outputs = run_fwd_model(inputs, torch_model)
check_tensors(torch_outputs, outputs, margin=5e-7)
def embedding_bwd(custom_ops,
mode,
momentum,
batch_size,
batch_serialization_factor,
embedding_serialization_vocab_steps,
vocab_length=9728,
hidden_size=768):
# ------------------- PopART --------------------
config = BertConfig(
task="SQUAD",
vocab_length=vocab_length,
batch_size=batch_size,
hidden_size=hidden_size,
sequence_length=128,
activation_type='relu',
popart_dtype="FLOAT",
no_dropout=True,
update_embedding_dict=True,
embedding_serialization_vocab_steps=embedding_serialization_vocab_steps
)
popart_model = get_model(config, mode, 'embedding')
# Prevent virtualGraph attributes being added to the ops
sequence_info = popart.TensorInfo(
"UINT32", [config.batch_size * config.sequence_length])
indices = popart_model.builder.addInputTensor(sequence_info)
positions = popart_model.builder.addInputTensor(sequence_info)
segments = popart_model.builder.addInputTensor(sequence_info)
data = {
indices:
np.random.randint(0, config.vocab_length,
(config.batch_size * config.sequence_length)).astype(
np.uint32),
positions:
np.random.randint(0, config.max_positional_length,
(config.batch_size * config.sequence_length)).astype(
np.uint32),
segments:
np.random.randint(0, 2,
(config.batch_size * config.sequence_length)).astype(
np.uint32)
}
if momentum:
optimizer = popart.SGD({
"defaultLearningRate": (0.01, True),
"defaultMomentum": (momentum, True),
"defaultDampening": (0.0, True),
"defaultVelocityScaling": (1.0, True),
"lossScaling": (1.0, True),
"defaultWeightDecay": (0.0, True)
})
else:
optimizer = popart.ConstSGD(0.01)
l1_lambda = 0.1
if mode == ExecutionMode.PHASED:
user_options = {
"batchSerializationFactor": batch_serialization_factor,
"executionPhases": popart_model.total_execution_phases,
}
output = popart_model(indices, positions, segments)
with popart_model.scope_provider(popart_model.builder,
popart_model.norm.scope):
l1 = popart_model.builder.aiGraphcore.l1loss(
[output],
l1_lambda,
debugPrefix="l1LossVal",
reduction=popart.ReductionType.Sum)
else:
user_options = {"enableStochasticRounding": True}
with popart_model.builder.nameScope("Embedding"):
output = popart_model.embedding(indices, positions, segments)
l1 = popart_model.builder.aiGraphcore.l1loss(
[output],
l1_lambda,
debugPrefix="l1LossVal",
reduction=popart.ReductionType.Sum)
num_reps = 5
proto = popart_model.builder.getModelProto()
outputs, post_proto = run_py(proto,
data,
output,
ipus=1,
loss=l1,
num_reps=num_reps,
optimizer=optimizer,
user_options=user_options,
execution_mode=mode)
# ----------------- PopART -> PyTorch ----------------
proto = onnx.load_model_from_string(proto)
inputs = [
data[t].reshape(config.batch_size,
config.sequence_length).astype(np.int32)
for t in [indices, positions, segments]
]
# ------------------- PyTorch -------------------------
torch_model = BertEmbeddings(
TorchBertConfig(config.vocab_length,
config.hidden_size,
max_position_embeddings=config.max_positional_length,
layer_norm_eps=config.layer_norm_eps,
update_embedding_dict=config.update_embedding_dict))
# Turn off dropout
torch_model.eval()
expanded_name_map, remapped_transform_map = expand_torch_to_onnx_map(
TORCH_TO_ONNX[mode], config, mode)
copy_weights_to_torch(torch_model, proto, expanded_name_map,
remapped_transform_map)
optim = torch.optim.SGD(torch_model.parameters(),
0.01,
weight_decay=0.0,
dampening=0.0,
momentum=momentum)
if momentum > 0.:
for group in optim.param_groups:
for p in group['params']:
optim.state[p]['momentum_buffer'] = p.data * 0
optim.state[p]['exp_avg'] = p.data * 0
optim.state[p]['exp_avg_sq'] = p.data * 0
optim.state[p]['step'] = 0
for _ in range(num_reps):
torch_output = torch_model(
*[torch.from_numpy(t).long() for t in inputs])
torch_loss = l1_lambda * torch.norm(torch_output, 1)
torch_loss.backward()
optim.step()
optim.zero_grad()
torch_outputs = [torch_output.detach().numpy()]
check_tensors(torch_outputs, outputs, margin=7e-6)
expanded_name_map, remapped_transform_map = expand_torch_to_onnx_map(
TORCH_TO_ONNX[mode], config, mode)
check_model(torch_model,
post_proto,
expanded_name_map,
remapped_transform_map,
margin=7e-06)
def test_attention_fwd(mode, micro_batch_size, batch_serialisation_factor,
number_attention_splits, attention_bias, split_qkv):
# ------------------- PopART --------------------
config = BertConfig(task="PRETRAINING",
vocab_length=9728,
micro_batch_size=micro_batch_size,
hidden_size=768,
attention_heads=4,
sequence_length=128,
activation_type='relu',
popart_dtype="FLOAT",
no_dropout=True,
no_attn_dropout=True,
inference=True,
split_qkv=split_qkv,
attention_bias=attention_bias,
num_attention_splits=number_attention_splits)
popart_model = get_model(config, mode, 'attention')
input_info = popart.TensorInfo(
config.popart_dtype,
[config.micro_batch_size * config.sequence_length, config.hidden_size])
input_tensor = popart_model.builder.addInputTensor(input_info)
mask_info = popart.TensorInfo(
"UINT32", [config.micro_batch_size, config.sequence_length])
mmask_tensor = popart_model.builder.addInputTensor(mask_info)
smask_tensor = popart_model.builder.addInputTensor(mask_info)
data = {
input_tensor:
np.random.normal(0, 0.02, input_info.shape()).astype(config.dtype),
mmask_tensor:
np.random.randint(0, config.mask_tokens + 1, (
config.micro_batch_size,
config.sequence_length,
)).astype(np.uint32),
smask_tensor:
np.random.randint(config.mask_tokens, config.sequence_length + 1, (
config.micro_batch_size,
config.sequence_length,
)).astype(np.uint32)
}
user_options = {}
if mode == ExecutionMode.PHASED:
user_options = {
"batchSerializationFactor": batch_serialisation_factor,
"executionPhases": popart_model.total_execution_phases
}
output = popart_model(input_tensor, [mmask_tensor, smask_tensor])
else:
user_options = {"enableStochasticRounding": True}
output = popart_model.attention(input_tensor,
[mmask_tensor, smask_tensor])
proto = popart_model.builder.getModelProto()
outputs, post_proto = run_py(proto,
data,
output,
user_options=user_options,
execution_mode=mode)
# ----------------- PopART -> PyTorch ----------------
proto = onnx.load_model_from_string(proto)
inputs = [
data[input_tensor].reshape(config.micro_batch_size,
config.sequence_length,
config.hidden_size).astype(np.float32),
get_torch_mask(config, [data[mmask_tensor], data[smask_tensor]])
]
# ------------------- PyTorch -------------------------
torch_model = BertAttention(
TorchBertConfig(config.vocab_length,
config.hidden_size,
config.num_layers,
config.attention_heads,
attention_bias=config.attention_bias,
layer_norm_eps=config.layer_norm_eps))
# Turn off dropout
torch_model.eval()
mapping = TORCH_TO_ONNX[mode]
if split_qkv:
mapping = TORCH_TO_ONNX_SPLIT_QKV[mode]
copy_weights_to_torch(torch_model,
proto,
mapping,
transform=get_transform(split_qkv,
config.hidden_size))
# Model to test against
torch_outputs = run_fwd_model(inputs, torch_model)
check_tensors(torch_outputs, outputs)
def bwd_graph(popart_model,
torch_model,
popart_loss_fn,
torch_loss_fn,
mapping=None,
transform=None):
np.random.seed(1984)
random.seed(1984)
torch.manual_seed(1984)
# ------------------- PopART --------------------
config = popart_model.config
builder = popart_model.builder
sequence_info = popart.TensorInfo(
"UINT32", [config.batch_size * config.sequence_length])
indices = builder.addInputTensor(sequence_info)
positions = builder.addInputTensor(sequence_info)
segments = builder.addInputTensor(sequence_info)
data = {
indices: np.random.randint(
0, config.vocab_length, (config.batch_size * config.sequence_length)).astype(np.uint32),
positions: np.random.randint(
0, config.sequence_length, (config.batch_size * config.sequence_length)).astype(np.uint32),
segments: np.random.randint(
0, 2, (config.batch_size * config.sequence_length)).astype(np.uint32)
}
output = popart_model.build_graph(indices, positions, segments)
proto = builder.getModelProto()
losses = popart_loss_fn(output)
optimizer = popart.ConstSGD(0.01)
outputs, post_proto = run_py(
proto, data, output, loss=losses, optimizer=optimizer,
ipus=math.ceil(config.num_layers / config.layers_per_ipu) + popart_model.layer_offset)
# ----------------- PopART -> PyTorch ----------------
proto = onnx.load_model_from_string(proto)
inputs = {
"input_ids": data[indices].reshape(config.batch_size, config.sequence_length).astype(np.int32),
"position_ids": data[positions].reshape(config.batch_size, config.sequence_length).astype(np.int32),
"token_type_ids": data[segments].reshape(config.batch_size, config.sequence_length).astype(np.int32)
}
torch_to_onnx = get_mapping(config, init=mapping)
transform_weights = get_transform(config, init=transform)
# ------------------- PyTorch -------------------------
# Turn off dropout
torch_model.eval()
copy_weights_to_torch(torch_model, proto,
torch_to_onnx, transform_weights)
optim = torch.optim.SGD(torch_model.parameters(), 0.01,
weight_decay=0.0, momentum=0.0)
torch_outputs = torch_model(
**{k: torch.from_numpy(t).long() for k, t in inputs.items()})
torch_loss = torch_loss_fn(torch_outputs)
torch_loss.backward()
optim.step()
check_tensors([output.detach().numpy() for output in torch_outputs], outputs)
check_model(torch_model, post_proto,
torch_to_onnx, transform_weights,
margin=6e-7)
def test_attention_bwd(mode, momentum, micro_batch_size,
batch_serialisation_factor, number_attention_splits,
attention_bias):
l1_lambda = 0.1
num_reps = 5
np.random.seed(1984)
torch.manual_seed(1984)
split_qkv = False
# ------------------- PopART --------------------
config = BertConfig(task="PRETRAINING",
vocab_length=9728,
micro_batch_size=micro_batch_size,
hidden_size=768,
sequence_length=128,
activation_type='relu',
popart_dtype="FLOAT",
no_dropout=True,
no_attn_dropout=True,
split_qkv=split_qkv,
attention_bias=attention_bias,
num_attention_splits=number_attention_splits)
popart_model = get_model(config, mode, 'attention')
input_info = popart.TensorInfo(
config.popart_dtype,
[config.micro_batch_size * config.sequence_length, config.hidden_size])
input_tensor = popart_model.builder.addInputTensor(input_info)
mask_info = popart.TensorInfo(
"UINT32", [config.micro_batch_size, config.sequence_length])
mmask_tensor = popart_model.builder.addInputTensor(mask_info)
smask_tensor = popart_model.builder.addInputTensor(mask_info)
data = {
input_tensor:
np.random.normal(0, 0.02, input_info.shape()).astype(config.dtype),
mmask_tensor:
np.random.randint(0, config.mask_tokens + 1, (
config.micro_batch_size,
config.sequence_length,
)).astype(np.uint32),
smask_tensor:
np.random.randint(config.mask_tokens, config.sequence_length + 1, (
config.micro_batch_size,
config.sequence_length,
)).astype(np.uint32)
}
user_options = {}
if mode == ExecutionMode.PHASED:
user_options = {
"batchSerializationFactor": batch_serialisation_factor,
"executionPhases": popart_model.total_execution_phases
}
output = popart_model(input_tensor, [mmask_tensor, smask_tensor])
with popart_model.scope_provider(popart_model.builder,
popart_model.norm.scope):
l1 = popart_model.builder.aiGraphcore.l1loss(
[output],
l1_lambda,
debugPrefix="l1LossVal",
reduction=popart.ReductionType.Sum)
else:
user_options = {}
output = popart_model.attention(input_tensor,
[mmask_tensor, smask_tensor])
l1 = popart_model.builder.aiGraphcore.l1loss(
[output], l1_lambda, reduction=popart.ReductionType.Sum)
proto = popart_model.builder.getModelProto()
if momentum:
optimizer = popart.SGD({
"defaultLearningRate": (0.01, True),
"defaultMomentum": (momentum, True)
})
else:
optimizer = popart.ConstSGD(0.01)
outputs, post_proto = run_py(proto,
data, (output, l1),
loss=l1,
optimizer=optimizer,
num_reps=num_reps,
user_options=user_options,
execution_mode=mode)
# ----------------- PopART -> PyTorch ----------------
proto = onnx.load_model_from_string(proto)
inputs = [
data[input_tensor].reshape(config.micro_batch_size,
config.sequence_length, config.hidden_size),
get_torch_mask(config, [data[mmask_tensor], data[smask_tensor]])
]
# ------------------- PyTorch -------------------------
torch_model = BertAttention(
TorchBertConfig(config.vocab_length,
config.hidden_size,
config.num_layers,
config.attention_heads,
attention_bias=config.attention_bias,
layer_norm_eps=config.layer_norm_eps))
# Turn off dropout
torch_model.eval()
mapping = TORCH_TO_ONNX[mode]
if split_qkv:
mapping = TORCH_TO_ONNX_SPLIT_QKV[mode]
copy_weights_to_torch(torch_model,
proto,
mapping,
transform=get_transform(split_qkv,
config.hidden_size))
optim = torch.optim.SGD(torch_model.parameters(),
0.01,
weight_decay=0.0,
momentum=momentum)
if momentum:
for group in optim.param_groups:
for p in group['params']:
optim.state[p]['momentum_buffer'] = p.data * 0
optim.state[p]['exp_avg'] = p.data * 0
optim.state[p]['exp_avg_sq'] = p.data * 0
optim.state[p]['step'] = 0
for _ in range(num_reps):
torch_output = torch_model(
*[torch.from_numpy(t).float() for t in inputs])[0]
torch_loss = l1_lambda * torch.norm(torch_output, 1)
torch_loss.backward()
optim.step()
optim.zero_grad()
check_tensors([torch_output.detach().numpy()], outputs, margin=6e-07)
check_model(torch_model,
post_proto,
mapping,
transform=get_transform(split_qkv, config.hidden_size),
margin=2e-7)
def bwd_graph(popart_model,
torch_model,
popart_loss_fn,
torch_loss_fn,
mapping=None,
transform=None,
replication_factor=1,
replicated_tensor_sharding=False,
opt_type="SGD"):
np.random.seed(1984)
random.seed(1984)
torch.manual_seed(1984)
# ------------------- PopART --------------------
config = popart_model.config
builder = popart_model.builder
sequence_info = popart.TensorInfo(
"UINT32", [config.micro_batch_size * config.sequence_length])
indices = builder.addInputTensor(sequence_info)
positions = builder.addInputTensor(sequence_info)
segments = builder.addInputTensor(sequence_info)
data = {
indices:
np.random.randint(0, config.vocab_length,
(replication_factor, config.micro_batch_size *
config.sequence_length)).astype(np.uint32),
positions:
np.random.randint(0, config.sequence_length,
(replication_factor, config.micro_batch_size *
config.sequence_length)).astype(np.uint32),
segments:
np.random.randint(0, 2, (replication_factor, config.micro_batch_size *
config.sequence_length)).astype(np.uint32)
}
num_reps = 5
output = popart_model.build_graph(indices, positions, segments)
ipus = popart_model.total_ipus
loss = popart_loss_fn(output)
proto = builder.getModelProto()
if opt_type == "SGD":
optimizer = popart.ConstSGD(1e-3)
elif opt_type == "LAMB":
optMap = {
"defaultLearningRate": (1e-3, True),
"defaultBeta1": (0.9, True),
"defaultBeta2": (0.999, True),
"defaultWeightDecay": (0.0, True),
"maxWeightNorm": (10.0, True),
"defaultEps": (1e-8, True),
"lossScaling": (1.0, True),
}
optimizer = popart.Adam(optMap, mode=popart.AdamMode.Lamb)
elif opt_type == "LAMB_NO_BIAS":
optMap = {
"defaultLearningRate": (1, False),
"defaultBeta1": (0, False),
"defaultBeta2": (0, False),
"defaultWeightDecay": (0.0, False),
"defaultEps": (1e-8, False),
"lossScaling": (1.0, False),
}
optimizer = popart.Adam(optMap, mode=popart.AdamMode.LambNoBias)
else:
raise ValueError(f"Unknown opt_type={opt_type}")
outputs, post_proto = run_py(
proto,
data,
output,
loss=loss,
optimizer=optimizer,
replication_factor=replication_factor,
replicated_tensor_sharding=replicated_tensor_sharding,
ipus=ipus,
num_reps=num_reps)
# ----------------- PopART -> PyTorch ----------------
proto = onnx.load_model_from_string(proto)
inputs = {
"input_ids":
data[indices].reshape(replication_factor * config.micro_batch_size,
config.sequence_length).astype(np.int32),
"position_ids":
data[positions].reshape(replication_factor * config.micro_batch_size,
config.sequence_length).astype(np.int32),
"token_type_ids":
data[segments].reshape(replication_factor * config.micro_batch_size,
config.sequence_length).astype(np.int32)
}
torch_to_onnx = get_mapping(config, init=mapping)
transform_weights = get_transform(config, init=transform)
# ------------------- PyTorch -------------------------
# Turn off dropout
torch_model.eval()
copy_weights_to_torch(torch_model, proto, torch_to_onnx, transform_weights)
if opt_type == "SGD":
optim = torch.optim.SGD(torch_model.parameters(),
1e-3,
weight_decay=0.0,
momentum=0.0)
elif opt_type == "LAMB":
optim = torch_lamb.Lamb(torch_model.parameters(),
lr=1e-3,
weight_decay=0.0,
biasCorrection=True)
for _ in range(num_reps):
torch_outputs = torch_model(
**{k: torch.from_numpy(t).long()
for k, t in inputs.items()})
torch_loss = torch_loss_fn(torch_outputs)
torch_loss.backward()
optim.step()
optim.zero_grad()
check_tensors([output.detach().numpy() for output in torch_outputs],
outputs,
margin=1.5e-06)
check_model(torch_model,
post_proto,
torch_to_onnx,
transform_weights,
margin=5e-5)
def test_attention_bwd(attention_bias, split_qkv):
l1_lambda = 0.1
num_reps = 5
np.random.seed(1984)
torch.manual_seed(1984)
# ------------------- PopART --------------------
config = BertConfig(task="PRETRAINING",
vocab_length=9728,
micro_batch_size=1,
hidden_size=768,
sequence_length=128,
activation_type='relu',
popart_dtype="FLOAT",
no_dropout=True,
no_attn_dropout=True,
split_qkv=split_qkv,
attention_bias=attention_bias)
popart_model = Bert(config, pipeline=True)
input_info = popart.TensorInfo(
config.popart_dtype,
[config.micro_batch_size * config.sequence_length, config.hidden_size])
input_tensor = popart_model.builder.addInputTensor(input_info)
mask_info = popart.TensorInfo(
"UINT32", [config.micro_batch_size, config.sequence_length])
mmask_tensor = popart_model.builder.addInputTensor(mask_info)
smask_tensor = popart_model.builder.addInputTensor(mask_info)
data = {
input_tensor:
np.random.normal(0, 0.02, input_info.shape()).astype(config.dtype),
mmask_tensor:
np.random.randint(0, config.mask_tokens + 1, (
config.micro_batch_size,
config.sequence_length,
)).astype(np.uint32),
smask_tensor:
np.random.randint(config.mask_tokens, config.sequence_length + 1, (
config.micro_batch_size,
config.sequence_length,
)).astype(np.uint32)
}
user_options = {}
output = popart_model.attention(input_tensor, [mmask_tensor, smask_tensor])
l1 = popart_model.builder.aiGraphcore.l1loss(
[output], l1_lambda, reduction=popart.ReductionType.Sum)
proto = popart_model.builder.getModelProto()
optimizer = popart.ConstSGD(0.01)
outputs, post_proto = run_py(proto,
data, (output, l1),
loss=l1,
optimizer=optimizer,
num_reps=num_reps,
user_options=user_options,
pipeline=popart_model.pipeline)
# ----------------- PopART -> PyTorch ----------------
proto = onnx.load_model_from_string(proto)
inputs = [
data[input_tensor].reshape(config.micro_batch_size,
config.sequence_length, config.hidden_size),
get_torch_mask(config, [data[mmask_tensor], data[smask_tensor]])
]
# ------------------- PyTorch -------------------------
torch_model = BertAttention(
TorchBertConfig(config.vocab_length,
config.hidden_size,
config.num_layers,
config.attention_heads,
attention_bias=config.attention_bias,
layer_norm_eps=config.layer_norm_eps))
# Turn off dropout
torch_model.eval()
mapping = TORCH_TO_ONNX_SPLIT_QKV if split_qkv else TORCH_TO_ONNX
copy_weights_to_torch(torch_model,
proto,
mapping,
transform=get_transform(split_qkv,
config.hidden_size))
optim = torch.optim.SGD(torch_model.parameters(), 0.01)
for _ in range(num_reps):
torch_output = torch_model(
*[torch.from_numpy(t).float() for t in inputs])[0]
torch_loss = l1_lambda * torch.norm(torch_output, 1)
torch_loss.backward()
optim.step()
optim.zero_grad()
check_tensors([torch_output.detach().numpy()], outputs, margin=6e-07)
check_model(torch_model,
post_proto,
mapping,
transform=get_transform(split_qkv, config.hidden_size),
margin=2e-7)