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_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 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_embedding(config, phase):
# define input
indices = np.random.randint(
0, test_config.vocab_size,
(test_config.batch_size, test_config.sequence_length)).astype(np.int32)
positions = np.reshape(
np.arange(test_config.sequence_length),
(test_config.batch_size, test_config.sequence_length)).astype(np.int32)
segments = np.random.randint(
0, 2,
(test_config.batch_size, test_config.sequence_length)).astype(np.int32)
inputs = [d for d in [indices, positions, segments]]
# build model
# PyTorch model
torch_config = TorchBertConfig(
vocab_size_or_config_json_file=test_config.vocab_size,
hidden_size=test_config.hidden_size,
hidden_act=test_config.hidden_act,
num_attention_heads=test_config.num_attention_heads,
hidden_dropout_prob=test_config.hidden_dropout_prob,
max_position_embeddings=test_config.max_position_embeddings,
type_vocab_size=test_config.type_vocab_size,
update_embedding_dict=True,
layer_norm_eps=test_config.layer_norm_eps)
torch_model = TorchBertEmbeddings(torch_config)
torch_model.eval()
# TF model
tf_config = TFBertConfig(
vocab_size=test_config.vocab_size,
hidden_size=test_config.hidden_size,
hidden_act=test_config.hidden_act,
num_attention_heads=test_config.num_attention_heads,
max_position_embeddings=test_config.max_position_embeddings,
max_predictions_per_seq=test_config.max_predictions_per_seq,
hidden_dropout_prob=test_config.hidden_dropout_prob,
type_vocab_size=test_config.type_vocab_size,
initializer_range=test_config.initializer_range,
dtype=test_config.dtype,
matmul_serialize_factor=test_config.matmul_serialize_factor,
static_mask=False)
# farward check
if phase == "fwd":
torch_outputs = run_fwd_model(inputs, torch_model)
with tf.Graph().as_default():
tf_model = TFBertModel(tf_config, is_training=True)
with ops.device('cpu'):
input_ids = tf.placeholder(shape=[
test_config.batch_size, test_config.sequence_length
],
dtype=tf.int32)
position_ids = tf.placeholder(shape=[
test_config.batch_size, test_config.sequence_length
],
dtype=tf.int32)
segment_ids = tf.placeholder(shape=[
test_config.batch_size, test_config.sequence_length
],
dtype=tf.int32)
cfg = utils.create_ipu_config()
cfg = utils.auto_select_ipus(cfg, 1)
utils.configure_ipu_system(cfg)
utils.move_variable_initialization_to_cpu()
with ops.device("/device:IPU:0"):
opt = ipu_compiler.compile(
tf_model.embeddings_layer,
inputs=[input_ids, position_ids, segment_ids])
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
# copy pytorch weight to tf
var_and_init = copy_torch_weights_to_tf(
torch_model, tf_model, TF_TO_TORCH, {}, sess)
sess.run(var_and_init)
# run tf feed feed farward
tf_outputs = sess.run(
opt, {
input_ids: indices,
position_ids: positions,
segment_ids: segments
})
# compare tf output with pytorch output
check_tensors(tf_outputs, torch_outputs, margin=1.5e-8)
# backward check
elif phase == "bwd":
l1_lambda = 0.1
base_lr = 0.01
optim = torch.optim.SGD(torch_model.parameters(),
base_lr,
weight_decay=0.0,
momentum=0.0)
torch_output = torch_model(
*[torch.from_numpy(t).long() for t in inputs])
# pytorch backward
torch_loss = l1_lambda * torch.norm(torch_output, 1)
torch_loss.backward() # calculate gradients
optim.step() # update gradients
torch_outputs = [torch_output.detach().numpy()]
# TF
with tf.Graph().as_default():
tf_model = TFBertModel(tf_config, is_training=True)
with ops.device('cpu'):
input_ids = tf.placeholder(shape=[
test_config.batch_size, test_config.sequence_length
],
dtype=tf.int32)
position_ids = tf.placeholder(shape=[
test_config.batch_size, test_config.sequence_length
],
dtype=tf.int32)
segment_ids = tf.placeholder(shape=[
test_config.batch_size, test_config.sequence_length
],
dtype=tf.int32)
cfg = utils.create_ipu_config()
cfg = utils.auto_select_ipus(cfg, 1)
utils.configure_ipu_system(cfg)
utils.move_variable_initialization_to_cpu()
def embedding_graph(input_ids, position_ids, segment_ids):
embedding_output = tf_model.embeddings_layer(
input_ids, position_ids, segment_ids)
l1_loss = l1_lambda * tf.norm(embedding_output, 1)
optimizer = tf.train.GradientDescentOptimizer(base_lr)
train_step = optimizer.minimize(l1_loss)
return embedding_output, l1_loss, train_step
with ops.device("/device:IPU:0"):
opt = ipu_compiler.compile(
embedding_graph,
inputs=[input_ids, position_ids, segment_ids])
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
var_and_init = copy_torch_weights_to_tf(
torch_model, tf_model, TF_TO_TORCH, {}, sess)
sess.run(var_and_init)
tvars = sess.run({v.name: v for v in tf.trainable_variables()})
print(tvars)
tf_outputs, tf_loss = sess.run(
opt, {
input_ids: indices,
position_ids: positions,
segment_ids: segments
})
# sess.run(opt, {input_ids: indices, position_ids: positions, segment_ids: segments})
# Compare the farward output
check_tf_torch_model(sess,
torch_model,
TF_TO_TORCH,
margin=5e-7)
check_tensors(torch_outputs, tf_outputs, margin=5e-7)
else:
raise ValueError(
f"`phase` only can be set to [`fwd`, `bwd`] which mean farward or backward respectively."
)
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)