提交 49bed8cf 编写于 作者: W Webbley

add ogb graph classification task

上级 b0a20434
......@@ -16,7 +16,11 @@ python setup.py install
```
### How to run
For example, use GPU to train model on ogbg-molhiv dataset.
For example, use GPU to train model on ogbg-molhiv dataset and ogb-molpcba dataset.
```
python main_pgl.py --use_cuda --dataset ogbg-molhiv
export CUDA_VISIBLE_DEVICES=1
python -u main.py --config hiv_config.yaml
export CUDA_VISIBLE_DEVICES=2
python -u main.py --config pcba_config.yaml
```
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from __future__ import unicode_literals
from __future__ import absolute_import
import os
import time
import argparse
from utils.args import ArgumentGroup
# yapf: disable
parser = argparse.ArgumentParser(__doc__)
model_g = ArgumentGroup(parser, "model", "model configuration and paths.")
model_g.add_arg("init_checkpoint", str, None, "Init checkpoint to resume training from.")
model_g.add_arg("init_pretraining_params", str, None,
"Init pre-training params which preforms fine-tuning from. If the "
"arg 'init_checkpoint' has been set, this argument wouldn't be valid.")
model_g.add_arg("./save_dir", str, "./checkpoints", "Path to save checkpoints.")
model_g.add_arg("hidden_size", int, 128, "hidden size.")
train_g = ArgumentGroup(parser, "training", "training options.")
train_g.add_arg("epoch", int, 3, "Number of epoches for fine-tuning.")
train_g.add_arg("learning_rate", float, 5e-5, "Learning rate used to train with warmup.")
train_g.add_arg("lr_scheduler", str, "linear_warmup_decay",
"scheduler of learning rate.", choices=['linear_warmup_decay', 'noam_decay'])
train_g.add_arg("weight_decay", float, 0.01, "Weight decay rate for L2 regularizer.")
train_g.add_arg("warmup_proportion", float, 0.1,
"Proportion of training steps to perform linear learning rate warmup for.")
train_g.add_arg("save_steps", int, 10000, "The steps interval to save checkpoints.")
train_g.add_arg("validation_steps", int, 1000, "The steps interval to evaluate model performance.")
train_g.add_arg("use_dynamic_loss_scaling", bool, True, "Whether to use dynamic loss scaling.")
train_g.add_arg("init_loss_scaling", float, 102400,
"Loss scaling factor for mixed precision training, only valid when use_fp16 is enabled.")
train_g.add_arg("test_save", str, "./checkpoints/test_result", "test_save")
train_g.add_arg("metric", str, "simple_accuracy", "metric")
train_g.add_arg("incr_every_n_steps", int, 100, "Increases loss scaling every n consecutive.")
train_g.add_arg("decr_every_n_nan_or_inf", int, 2,
"Decreases loss scaling every n accumulated steps with nan or inf gradients.")
train_g.add_arg("incr_ratio", float, 2.0,
"The multiplier to use when increasing the loss scaling.")
train_g.add_arg("decr_ratio", float, 0.8,
"The less-than-one-multiplier to use when decreasing.")
log_g = ArgumentGroup(parser, "logging", "logging related.")
log_g.add_arg("skip_steps", int, 10, "The steps interval to print loss.")
log_g.add_arg("verbose", bool, False, "Whether to output verbose log.")
log_g.add_arg("log_dir", str, './logs/', "Whether to output verbose log.")
data_g = ArgumentGroup(parser, "data", "Data paths, vocab paths and data processing options")
data_g.add_arg("tokenizer", str, "FullTokenizer",
"ATTENTION: the INPUT must be splited by Word with blank while using SentencepieceTokenizer or WordsegTokenizer")
data_g.add_arg("train_set", str, None, "Path to training data.")
data_g.add_arg("test_set", str, None, "Path to test data.")
data_g.add_arg("dev_set", str, None, "Path to validation data.")
data_g.add_arg("aug1_type", str, "scheme1", "augment type")
data_g.add_arg("aug2_type", str, "scheme1", "augment type")
data_g.add_arg("batch_size", int, 32, "Total examples' number in batch for training. see also --in_tokens.")
data_g.add_arg("predict_batch_size", int, None, "Total examples' number in batch for predict. see also --in_tokens.")
data_g.add_arg("random_seed", int, None, "Random seed.")
data_g.add_arg("buf_size", int, 1000, "Random seed.")
run_type_g = ArgumentGroup(parser, "run_type", "running type options.")
run_type_g.add_arg("use_cuda", bool, False, "If set, use GPU for training.")
run_type_g.add_arg("num_iteration_per_drop_scope", int, 10, "Iteration intervals to drop scope.")
run_type_g.add_arg("do_train", bool, True, "Whether to perform training.")
run_type_g.add_arg("do_val", bool, True, "Whether to perform evaluation on dev data set.")
run_type_g.add_arg("do_test", bool, True, "Whether to perform evaluation on test data set.")
run_type_g.add_arg("metrics", bool, True, "Whether to perform evaluation on test data set.")
run_type_g.add_arg("shuffle", bool, True, "")
run_type_g.add_arg("for_cn", bool, True, "model train for cn or for other langs.")
run_type_g.add_arg("num_workers", int, 1, "use multiprocess to generate graph")
run_type_g.add_arg("output_dir", str, None, "path to save model")
run_type_g.add_arg("config", str, None, "configure yaml file")
run_type_g.add_arg("n", str, None, "task name")
run_type_g.add_arg("task_name", str, None, "task name")
run_type_g.add_arg("pretrain", bool, False, "Whether do pretrian")
run_type_g.add_arg("pretrain_name", str, None, "pretrain task name")
run_type_g.add_arg("pretrain_config", str, None, "pretrain config.yaml file")
run_type_g.add_arg("pretrain_model_step", str, None, "pretrain model step")
run_type_g.add_arg("model_type", str, "BaseLineModel", "pretrain model step")
run_type_g.add_arg("num_class", int, 1, "number class")
run_type_g.add_arg("dataset_name", str, None, "finetune dataset name")
run_type_g.add_arg("eval_metrics", str, None, "evaluate metrics")
run_type_g.add_arg("task_type", str, None, "regression or classification")
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import sys
import os
from ogb.graphproppred import GraphPropPredDataset
import pgl
from pgl.utils.logger import log
class BaseDataset(object):
def __init__(self):
pass
def __getitem__(self, idx):
raise NotImplementedError
def __len__(self):
raise NotImplementedError
class Subset(BaseDataset):
r"""
Subset of a dataset at specified indices.
Arguments:
dataset (Dataset): The whole Dataset
indices (sequence): Indices in the whole set selected for subset
"""
def __init__(self, dataset, indices):
self.dataset = dataset
self.indices = indices
def __getitem__(self, idx):
return self.dataset[self.indices[idx]]
def __len__(self):
return len(self.indices)
class Dataset(BaseDataset):
def __init__(self, args):
self.args = args
self.raw_dataset = GraphPropPredDataset(name=args.dataset_name)
self.num_tasks = self.raw_dataset.num_tasks
self.eval_metrics = self.raw_dataset.eval_metric
self.task_type = self.raw_dataset.task_type
self.pgl_graph_list = []
self.graph_label_list = []
for i in range(len(self.raw_dataset)):
graph, label = self.raw_dataset[i]
edges = list(zip(graph["edge_index"][0], graph["edge_index"][1]))
g = pgl.graph.Graph(num_nodes=graph["num_nodes"], edges=edges)
if graph["edge_feat"] is not None:
g.edge_feat["feat"] = graph["edge_feat"]
if graph["node_feat"] is not None:
g.node_feat["feat"] = graph["node_feat"]
self.pgl_graph_list.append(g)
self.graph_label_list.append(label)
def __getitem__(self, idx):
return self.pgl_graph_list[idx], self.graph_label_list[idx]
def __len__(self):
return len(slef.pgl_graph_list)
def get_idx_split(self):
return self.raw_dataset.get_idx_split()
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
This file implement the graph dataloader.
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from __future__ import unicode_literals
from __future__ import absolute_import
import ssl
ssl._create_default_https_context = ssl._create_unverified_context
# SSL
import torch
import sys
import six
from io import open
import collections
from collections import namedtuple
import numpy as np
import tqdm
import time
import paddle
import paddle.fluid as fluid
import paddle.fluid.layers as fl
import pgl
from pgl.utils import mp_reader
from pgl.utils.logger import log
from ogb.graphproppred import GraphPropPredDataset
def batch_iter(data, batch_size, fid, num_workers):
"""node_batch_iter
"""
size = len(data)
perm = np.arange(size)
np.random.shuffle(perm)
start = 0
cc = 0
while start < size:
index = perm[start:start + batch_size]
start += batch_size
cc += 1
if cc % num_workers != fid:
continue
yield data[index]
def scan_batch_iter(data, batch_size, fid, num_workers):
"""scan_batch_iter
"""
batch = []
cc = 0
for line_example in data.scan():
cc += 1
if cc % num_workers != fid:
continue
batch.append(line_example)
if len(batch) == batch_size:
yield batch
batch = []
if len(batch) > 0:
yield batch
class GraphDataloader(object):
"""Graph Dataloader
"""
def __init__(self,
dataset,
graph_wrapper,
batch_size,
seed=0,
num_workers=1,
buf_size=1000,
shuffle=True):
self.shuffle = shuffle
self.seed = seed
self.num_workers = num_workers
self.buf_size = buf_size
self.batch_size = batch_size
self.dataset = dataset
self.graph_wrapper = graph_wrapper
def batch_fn(self, batch_examples):
""" batch_fn batch producer"""
graphs = [b[0] for b in batch_examples]
labels = [b[1] for b in batch_examples]
join_graph = pgl.graph.MultiGraph(graphs)
labels = np.array(labels)
feed_dict = self.graph_wrapper.to_feed(join_graph)
batch_valid = (labels == labels).astype("float32")
labels = np.nan_to_num(labels).astype("float32")
feed_dict['labels'] = labels
feed_dict['unmask'] = batch_valid
return feed_dict
def batch_iter(self, fid):
"""batch_iter"""
if self.shuffle:
for batch in batch_iter(self, self.batch_size, fid,
self.num_workers):
yield batch
else:
for batch in scan_batch_iter(self, self.batch_size, fid,
self.num_workers):
yield batch
def __len__(self):
"""__len__"""
return len(self.dataset)
def __getitem__(self, idx):
"""__getitem__"""
if isinstance(idx, collections.Iterable):
return [self[bidx] for bidx in idx]
else:
return self.dataset[idx]
def __iter__(self):
"""__iter__"""
def worker(filter_id):
def func_run():
for batch_examples in self.batch_iter(filter_id):
batch_dict = self.batch_fn(batch_examples)
yield batch_dict
return func_run
if self.num_workers == 1:
r = paddle.reader.buffered(worker(0), self.buf_size)
else:
worker_pool = [worker(wid) for wid in range(self.num_workers)]
worker = mp_reader.multiprocess_reader(
worker_pool, use_pipe=True, queue_size=1000)
r = paddle.reader.buffered(worker, self.buf_size)
for batch in r():
yield batch
def scan(self):
"""scan"""
for example in self.dataset:
yield example
if __name__ == "__main__":
from base_dataset import BaseDataset, Subset
dataset = GraphPropPredDataset(name="ogbg-molhiv")
splitted_index = dataset.get_idx_split()
train_dataset = Subset(dataset, splitted_index['train'])
valid_dataset = Subset(dataset, splitted_index['valid'])
test_dataset = Subset(dataset, splitted_index['test'])
log.info("Train Examples: %s" % len(train_dataset))
log.info("Val Examples: %s" % len(valid_dataset))
log.info("Test Examples: %s" % len(test_dataset))
# train_loader = GraphDataloader(train_dataset, batch_size=3)
# for batch_data in train_loader:
# graphs, labels = batch_data
# print(labels.shape)
# time.sleep(4)
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import sys
import os
import logging
from random import random
import pandas as pd
import numpy as np
from itertools import compress
import scipy.sparse as sp
from sklearn.model_selection import StratifiedKFold
from sklearn.preprocessing import StandardScaler
from rdkit.Chem.Scaffolds import MurckoScaffold
import pgl
from pgl.utils import paddle_helper
try:
from dataset.Dataset import Subset
from dataset.Dataset import ChemDataset
except:
from Dataset import Subset
from Dataset import ChemDataset
log = logging.getLogger("logger")
def random_split(dataset, args):
total_precent = args.frac_train + args.frac_valid + args.frac_test
np.testing.assert_almost_equal(total_precent, 1.0)
length = len(dataset)
perm = list(range(length))
np.random.shuffle(perm)
num_train = int(args.frac_train * length)
num_valid = int(args.frac_valid * length)
num_test = int(args.frac_test * length)
train_indices = perm[0:num_train]
valid_indices = perm[num_train:(num_train + num_valid)]
test_indices = perm[(num_train + num_valid):]
assert (len(train_indices) + len(valid_indices) + len(test_indices)
) == length
train_dataset = Subset(dataset, train_indices)
valid_dataset = Subset(dataset, valid_indices)
test_dataset = Subset(dataset, test_indices)
return train_dataset, valid_dataset, test_dataset
def scaffold_split(dataset, args, return_smiles=False):
total_precent = args.frac_train + args.frac_valid + args.frac_test
np.testing.assert_almost_equal(total_precent, 1.0)
smiles_list_file = os.path.join(args.data_dir, "smiles.csv")
smiles_list = pd.read_csv(smiles_list_file, header=None)[0].tolist()
non_null = np.ones(len(dataset)) == 1
smiles_list = list(compress(enumerate(smiles_list), non_null))
# create dict of the form {scaffold_i: [idx1, idx....]}
all_scaffolds = {}
for i, smiles in smiles_list:
scaffold = MurckoScaffold.MurckoScaffoldSmiles(
smiles=smiles, includeChirality=True)
# scaffold = generate_scaffold(smiles, include_chirality=True)
if scaffold not in all_scaffolds:
all_scaffolds[scaffold] = [i]
else:
all_scaffolds[scaffold].append(i)
# sort from largest to smallest sets
all_scaffolds = {
key: sorted(value)
for key, value in all_scaffolds.items()
}
all_scaffold_sets = [
scaffold_set
for (scaffold, scaffold_set) in sorted(
all_scaffolds.items(),
key=lambda x: (len(x[1]), x[1][0]),
reverse=True)
]
# get train, valid test indices
train_cutoff = args.frac_train * len(smiles_list)
valid_cutoff = (args.frac_train + args.frac_valid) * len(smiles_list)
train_idx, valid_idx, test_idx = [], [], []
for scaffold_set in all_scaffold_sets:
if len(train_idx) + len(scaffold_set) > train_cutoff:
if len(train_idx) + len(valid_idx) + len(
scaffold_set) > valid_cutoff:
test_idx.extend(scaffold_set)
else:
valid_idx.extend(scaffold_set)
else:
train_idx.extend(scaffold_set)
assert len(set(train_idx).intersection(set(valid_idx))) == 0
assert len(set(test_idx).intersection(set(valid_idx))) == 0
# log.info(len(scaffold_set))
# log.info(["train_idx", train_idx])
# log.info(["valid_idx", valid_idx])
# log.info(["test_idx", test_idx])
train_dataset = Subset(dataset, train_idx)
valid_dataset = Subset(dataset, valid_idx)
test_dataset = Subset(dataset, test_idx)
if return_smiles:
train_smiles = [smiles_list[i][1] for i in train_idx]
valid_smiles = [smiles_list[i][1] for i in valid_idx]
test_smiles = [smiles_list[i][1] for i in test_idx]
return train_dataset, valid_dataset, test_dataset, (
train_smiles, valid_smiles, test_smiles)
return train_dataset, valid_dataset, test_dataset
if __name__ == "__main__":
file_path = os.path.dirname(os.path.realpath(__file__))
proj_path = os.path.join(file_path, '../')
sys.path.append(proj_path)
from utils.config import Config
from dataset.Dataset import Subset
from dataset.Dataset import ChemDataset
config_file = "./finetune_config.yaml"
args = Config(config_file)
log.info("loading dataset")
dataset = ChemDataset(args)
train_dataset, valid_dataset, test_dataset = scaffold_split(dataset, args)
log.info("Train Examples: %s" % len(train_dataset))
log.info("Val Examples: %s" % len(valid_dataset))
log.info("Test Examples: %s" % len(test_dataset))
import ipdb
ipdb.set_trace()
log.info("preprocess finish")
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import ssl
ssl._create_default_https_context = ssl._create_unverified_context
# SSL
import torch
import os
import re
import time
from random import random
from functools import reduce, partial
import numpy as np
import multiprocessing
from ogb.graphproppred import Evaluator
import paddle
import paddle.fluid as F
import paddle.fluid.layers as L
import pgl
from pgl.utils import paddle_helper
from pgl.utils.logger import log
from utils.args import print_arguments, check_cuda, prepare_logger
from utils.init import init_checkpoint, init_pretraining_params
from utils.config import Config
from optimization import optimization
from monitor.train_monitor import train_and_evaluate
from args import parser
import model as Model
from data.base_dataset import Subset, Dataset
from data.dataloader import GraphDataloader
def main(args):
log.info('loading data')
dataset = Dataset(args)
args.num_class = dataset.num_tasks
args.eval_metrics = dataset.eval_metrics
args.task_type = dataset.task_type
splitted_index = dataset.get_idx_split()
train_dataset = Subset(dataset, splitted_index['train'])
valid_dataset = Subset(dataset, splitted_index['valid'])
test_dataset = Subset(dataset, splitted_index['test'])
log.info("preprocess finish")
log.info("Train Examples: %s" % len(train_dataset))
log.info("Val Examples: %s" % len(valid_dataset))
log.info("Test Examples: %s" % len(test_dataset))
train_prog = F.Program()
startup_prog = F.Program()
if args.use_cuda:
dev_list = F.cuda_places()
place = dev_list[0]
dev_count = len(dev_list)
else:
place = F.CPUPlace()
dev_count = int(os.environ.get('CPU_NUM', multiprocessing.cpu_count()))
# dev_count = args.cpu_num
log.info("building model")
with F.program_guard(train_prog, startup_prog):
with F.unique_name.guard():
graph_model = getattr(Model, args.model_type)(args, dataset)
train_ds = GraphDataloader(
train_dataset,
graph_model.graph_wrapper,
batch_size=args.batch_size)
num_train_examples = len(train_dataset)
max_train_steps = args.epoch * num_train_examples // args.batch_size // dev_count
warmup_steps = int(max_train_steps * args.warmup_proportion)
scheduled_lr, loss_scaling = optimization(
loss=graph_model.loss,
warmup_steps=warmup_steps,
num_train_steps=max_train_steps,
learning_rate=args.learning_rate,
train_program=train_prog,
startup_prog=startup_prog,
weight_decay=args.weight_decay,
scheduler=args.lr_scheduler,
use_fp16=False,
use_dynamic_loss_scaling=args.use_dynamic_loss_scaling,
init_loss_scaling=args.init_loss_scaling,
incr_every_n_steps=args.incr_every_n_steps,
decr_every_n_nan_or_inf=args.decr_every_n_nan_or_inf,
incr_ratio=args.incr_ratio,
decr_ratio=args.decr_ratio)
test_prog = F.Program()
with F.program_guard(test_prog, startup_prog):
with F.unique_name.guard():
_graph_model = getattr(Model, args.model_type)(args, dataset)
test_prog = test_prog.clone(for_test=True)
valid_ds = GraphDataloader(
valid_dataset,
graph_model.graph_wrapper,
batch_size=args.batch_size,
shuffle=False)
test_ds = GraphDataloader(
test_dataset,
graph_model.graph_wrapper,
batch_size=args.batch_size,
shuffle=False)
exe = F.Executor(place)
exe.run(startup_prog)
for init in graph_model.init_vars:
init(place)
for init in _graph_model.init_vars:
init(place)
if args.init_pretraining_params is not None:
init_pretraining_params(
exe, args.init_pretraining_params, main_program=startup_prog)
nccl2_num_trainers = 1
nccl2_trainer_id = 0
if dev_count > 1:
exec_strategy = F.ExecutionStrategy()
exec_strategy.num_threads = dev_count
train_exe = F.ParallelExecutor(
use_cuda=args.use_cuda,
loss_name=graph_model.loss.name,
exec_strategy=exec_strategy,
main_program=train_prog,
num_trainers=nccl2_num_trainers,
trainer_id=nccl2_trainer_id)
test_exe = exe
else:
train_exe, test_exe = exe, exe
evaluator = Evaluator(args.dataset_name)
train_and_evaluate(
exe=exe,
train_exe=train_exe,
valid_exe=test_exe,
train_ds=train_ds,
valid_ds=valid_ds,
test_ds=test_ds,
train_prog=train_prog,
valid_prog=test_prog,
args=args,
dev_count=dev_count,
evaluator=evaluator,
model=graph_model)
if __name__ == "__main__":
args = parser.parse_args()
if args.config is not None:
args = Config(args.config, isCreate=True, isSave=True)
log.info(args)
main(args)
#-*- coding: utf-8 -*-
import os
import re
import time
import logging
from random import random
from functools import reduce, partial
import numpy as np
import multiprocessing
import paddle
import paddle.fluid as F
import paddle.fluid.layers as L
import pgl
from pgl.graph_wrapper import GraphWrapper
from pgl.layers.conv import gcn, gat
from pgl.utils import paddle_helper
from pgl.utils.logger import log
from utils.args import print_arguments, check_cuda, prepare_logger
from utils.init import init_checkpoint, init_pretraining_params
from mol_encoder import AtomEncoder, BondEncoder
def copy_send(src_feat, dst_feat, edge_feat):
return src_feat["h"]
def mean_recv(feat):
return L.sequence_pool(feat, pool_type="average")
def sum_recv(feat):
return L.sequence_pool(feat, pool_type="sum")
def max_recv(feat):
return L.sequence_pool(feat, pool_type="max")
def unsqueeze(tensor):
tensor = L.unsqueeze(tensor, axes=-1)
tensor.stop_gradient = True
return tensor
class Metric:
def __init__(self, **args):
self.args = args
@property
def vars(self):
values = [self.args[k] for k in self.args.keys()]
return values
def parse(self, fetch_list):
tup = list(zip(self.args.keys(), [float(v[0]) for v in fetch_list]))
return dict(tup)
def gin_layer(gw, node_features, edge_features, train_eps, name):
def send_func(src_feat, dst_feat, edge_feat):
"""Send"""
return src_feat["h"] + edge_feat["h"]
epsilon = L.create_parameter(
shape=[1, 1],
dtype="float32",
attr=F.ParamAttr(name="%s_eps" % name),
default_initializer=F.initializer.ConstantInitializer(value=0.0))
if not train_eps:
epsilon.stop_gradient = True
msg = gw.send(
send_func,
nfeat_list=[("h", node_features)],
efeat_list=[("h", edge_features)])
node_feat = gw.recv(msg, "sum") + node_features * (epsilon + 1.0)
# if apply_func is not None:
# node_feat = apply_func(node_feat, name)
return node_feat
class GNNModel(object):
def __init__(self, args, dataset):
self.args = args
self.dataset = dataset
self.hidden_size = self.args.hidden_size
self.embed_dim = self.args.embed_dim
self.dropout_prob = self.args.dropout_rate
self.pool_type = self.args.pool_type
self._init_vars = []
graph_data = []
g, label = self.dataset[0]
graph_data.append(g)
g, label = self.dataset[1]
graph_data.append(g)
batch_graph = pgl.graph.MultiGraph(graph_data)
graph_data = batch_graph
graph_data.edge_feat["feat"] = graph_data.edge_feat["feat"].astype(
"int64")
graph_data.node_feat["feat"] = graph_data.node_feat["feat"].astype(
"int64")
self.graph_wrapper = GraphWrapper(
name="graph",
place=F.CPUPlace(),
node_feat=graph_data.node_feat_info(),
edge_feat=graph_data.edge_feat_info())
self.atom_encoder = AtomEncoder(name="atom", emb_dim=self.embed_dim)
self.bond_encoder = BondEncoder(name="bond", emb_dim=self.embed_dim)
self.labels = L.data(
"labels",
shape=[None, self.args.num_class],
dtype="float32",
append_batch_size=False)
self.unmask = L.data(
"unmask",
shape=[None, self.args.num_class],
dtype="float32",
append_batch_size=False)
self.build_model()
def build_model(self):
node_features = self.atom_encoder(self.graph_wrapper.node_feat['feat'])
edge_features = self.bond_encoder(self.graph_wrapper.edge_feat['feat'])
self._enc_out = self.node_repr_encode(node_features, edge_features)
logits = L.fc(self._enc_out,
self.args.num_class,
act=None,
param_attr=F.ParamAttr(name="final_fc"))
# L.Print(self.labels, message="labels")
# L.Print(self.unmask, message="unmask")
loss = L.sigmoid_cross_entropy_with_logits(x=logits, label=self.labels)
loss = loss * self.unmask
self.loss = L.reduce_sum(loss) / L.reduce_sum(self.unmask)
self.pred = L.sigmoid(logits)
self._metrics = Metric(loss=self.loss)
def node_repr_encode(self, node_features, edge_features):
features_list = [node_features]
for layer in range(self.args.num_layers):
feat = gin_layer(
self.graph_wrapper,
features_list[layer],
edge_features,
train_eps=self.args.train_eps,
name="gin_%s" % layer, )
feat = self.mlp(feat, name="mlp_%s" % layer)
feat = feat + features_list[layer] # residual
features_list.append(feat)
output = pgl.layers.graph_pooling(
self.graph_wrapper, features_list[-1], self.args.pool_type)
return output
def mlp(self, features, name):
h = features
dim = features.shape[-1]
dim_list = [dim * 2, dim]
for i in range(2):
h = L.fc(h,
size=dim_list[i],
name="%s_fc_%s" % (name, i),
act=None)
if self.args.norm_type == "layer_norm":
log.info("norm_type is %s" % self.args.norm_type)
h = L.layer_norm(
h,
begin_norm_axis=1,
param_attr=F.ParamAttr(
name="norm_scale_%s_%s" % (name, i),
initializer=F.initializer.Constant(1.0)),
bias_attr=F.ParamAttr(
name="norm_bias_%s_%s" % (name, i),
initializer=F.initializer.Constant(0.0)), )
else:
log.info("using batch_norm")
h = L.batch_norm(h)
h = pgl.layers.graph_norm(self.graph_wrapper, h)
h = L.relu(h)
return h
def get_enc_output(self):
return self._enc_out
@property
def init_vars(self):
return self._init_vars
@property
def metrics(self):
return self._metrics
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""MolEncoder for ogb
"""
import paddle.fluid as fluid
from ogb.utils.features import get_atom_feature_dims, get_bond_feature_dims
class AtomEncoder(object):
"""AtomEncoder for encoding node features"""
def __init__(self, name, emb_dim):
self.emb_dim = emb_dim
self.name = name
def __call__(self, x):
atom_feature = get_atom_feature_dims()
atom_input = fluid.layers.split(
x, num_or_sections=len(atom_feature), dim=-1)
outputs = None
count = 0
for _x, _atom_input_dim in zip(atom_input, atom_feature):
count += 1
emb = fluid.layers.embedding(
_x,
size=(_atom_input_dim, self.emb_dim),
param_attr=fluid.ParamAttr(
name=self.name + '_atom_feat_%s' % count))
if outputs is None:
outputs = emb
else:
outputs = outputs + emb
return outputs
class BondEncoder(object):
"""Bond for encoding edge features"""
def __init__(self, name, emb_dim):
self.emb_dim = emb_dim
self.name = name
def __call__(self, x):
bond_feature = get_bond_feature_dims()
bond_input = fluid.layers.split(
x, num_or_sections=len(bond_feature), dim=-1)
outputs = None
count = 0
for _x, _bond_input_dim in zip(bond_input, bond_feature):
count += 1
emb = fluid.layers.embedding(
_x,
size=(_bond_input_dim, self.emb_dim),
param_attr=fluid.ParamAttr(
name=self.name + '_bond_feat_%s' % count))
if outputs is None:
outputs = emb
else:
outputs = outputs + emb
return outputs
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import tqdm
import json
import numpy as np
import os
from datetime import datetime
import logging
from collections import defaultdict
from tensorboardX import SummaryWriter
import paddle.fluid as F
from pgl.utils.logger import log
def multi_device(reader, dev_count):
if dev_count == 1:
for batch in reader:
yield batch
else:
batches = []
for batch in reader:
batches.append(batch)
if len(batches) == dev_count:
yield batches
batches = []
def evaluate(exe, loader, prog, model, evaluator):
total_labels = []
for i in range(len(loader.dataset)):
g, l = loader.dataset[i]
total_labels.append(l)
total_labels = np.vstack(total_labels)
pred_output = []
for feed_dict in loader:
ret = exe.run(prog, feed=feed_dict, fetch_list=model.pred)
pred_output.append(ret[0])
pred_output = np.vstack(pred_output)
result = evaluator.eval({"y_true": total_labels, "y_pred": pred_output})
return result
def _create_if_not_exist(path):
basedir = os.path.dirname(path)
if not os.path.exists(basedir):
os.makedirs(basedir)
def train_and_evaluate(exe,
train_exe,
valid_exe,
train_ds,
valid_ds,
test_ds,
train_prog,
valid_prog,
args,
model,
evaluator,
dev_count=1):
global_step = 0
timestamp = datetime.now().strftime("%Hh%Mm%Ss")
log_path = os.path.join(args.log_dir, "tensorboard_log_%s" % timestamp)
_create_if_not_exist(log_path)
writer = SummaryWriter(log_path)
best_valid_score = 0.0
for e in range(args.epoch):
for feed_dict in multi_device(train_ds, dev_count):
if dev_count > 1:
ret = train_exe.run(feed=feed_dict,
fetch_list=model.metrics.vars)
ret = [[np.mean(v)] for v in ret]
else:
ret = train_exe.run(train_prog,
feed=feed_dict,
fetch_list=model.metrics.vars)
ret = model.metrics.parse(ret)
if global_step % args.train_log_step == 0:
writer.add_scalar(
"batch_loss", ret['loss'], global_step=global_step)
log.info("epoch: %d | step: %d | loss: %.4f " %
(e, global_step, ret['loss']))
global_step += 1
if global_step % args.eval_step == 0:
valid_ret = evaluate(exe, valid_ds, valid_prog, model,
evaluator)
message = "valid: "
for key, value in valid_ret.items():
message += "%s %.4f | " % (key, value)
writer.add_scalar(
"eval_%s" % key, value, global_step=global_step)
log.info(message)
# testing
test_ret = evaluate(exe, test_ds, valid_prog, model, evaluator)
message = "test: "
for key, value in test_ret.items():
message += "%s %.4f | " % (key, value)
writer.add_scalar(
"test_%s" % key, value, global_step=global_step)
log.info(message)
# evaluate after one epoch
valid_ret = evaluate(exe, valid_ds, valid_prog, model, evaluator)
message = "epoch %s valid: " % e
for key, value in valid_ret.items():
message += "%s %.4f | " % (key, value)
writer.add_scalar("eval_%s" % key, value, global_step=global_step)
log.info(message)
# testing
test_ret = evaluate(exe, test_ds, valid_prog, model, evaluator)
message = "epoch %s test: " % e
for key, value in test_ret.items():
message += "%s %.4f | " % (key, value)
writer.add_scalar("test_%s" % key, value, global_step=global_step)
log.info(message)
message = "epoch %s best %s result | " % (e, args.eval_metrics)
if valid_ret[args.eval_metrics] > best_valid_score:
best_valid_score = valid_ret[args.eval_metrics]
best_test_score = test_ret[args.eval_metrics]
message += "valid %.4f | test %.4f" % (best_valid_score,
best_test_score)
log.info(message)
# if global_step % args.save_step == 0:
# F.io.save_persistables(exe, os.path.join(args.save_dir, "%s" % global_step), train_prog)
writer.close()
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Optimization and learning rate scheduling."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from __future__ import unicode_literals
from __future__ import absolute_import
import numpy as np
import paddle.fluid as fluid
from utils.fp16 import create_master_params_grads, master_param_to_train_param, apply_dynamic_loss_scaling
def linear_warmup_decay(learning_rate, warmup_steps, num_train_steps):
""" Applies linear warmup of learning rate from 0 and decay to 0."""
with fluid.default_main_program()._lr_schedule_guard():
lr = fluid.layers.tensor.create_global_var(
shape=[1],
value=0.0,
dtype='float32',
persistable=True,
name="scheduled_learning_rate")
global_step = fluid.layers.learning_rate_scheduler._decay_step_counter(
)
with fluid.layers.control_flow.Switch() as switch:
with switch.case(global_step < warmup_steps):
warmup_lr = learning_rate * (global_step / warmup_steps)
fluid.layers.tensor.assign(warmup_lr, lr)
with switch.default():
decayed_lr = fluid.layers.learning_rate_scheduler.polynomial_decay(
learning_rate=learning_rate,
decay_steps=num_train_steps,
end_learning_rate=0.0,
power=1.0,
cycle=False)
fluid.layers.tensor.assign(decayed_lr, lr)
return lr
def optimization(loss,
warmup_steps,
num_train_steps,
learning_rate,
train_program,
startup_prog,
weight_decay,
scheduler='linear_warmup_decay',
use_fp16=False,
use_dynamic_loss_scaling=False,
init_loss_scaling=1.0,
incr_every_n_steps=1000,
decr_every_n_nan_or_inf=2,
incr_ratio=2.0,
decr_ratio=0.8):
if warmup_steps > 0:
if scheduler == 'noam_decay':
scheduled_lr = fluid.layers.learning_rate_scheduler\
.noam_decay(1/(warmup_steps *(learning_rate ** 2)),
warmup_steps)
elif scheduler == 'linear_warmup_decay':
scheduled_lr = linear_warmup_decay(learning_rate, warmup_steps,
num_train_steps)
else:
raise ValueError("Unkown learning rate scheduler, should be "
"'noam_decay' or 'linear_warmup_decay'")
optimizer = fluid.optimizer.Adam(learning_rate=scheduled_lr)
else:
scheduled_lr = fluid.layers.create_global_var(
name=fluid.unique_name.generate("learning_rate"),
shape=[1],
value=learning_rate,
dtype='float32',
persistable=True)
optimizer = fluid.optimizer.Adam(learning_rate=scheduled_lr)
optimizer._learning_rate_map[fluid.default_main_program(
)] = scheduled_lr
fluid.clip.set_gradient_clip(
clip=fluid.clip.GradientClipByGlobalNorm(clip_norm=1.0))
def exclude_from_weight_decay(name):
if name.find("layer_norm") > -1:
return True
bias_suffix = ["_bias", "_b", ".b_0"]
for suffix in bias_suffix:
if name.endswith(suffix):
return True
return False
param_list = dict()
loss_scaling = fluid.layers.create_global_var(
name=fluid.unique_name.generate("loss_scaling"),
shape=[1],
value=init_loss_scaling,
dtype='float32',
persistable=True)
if use_fp16:
loss *= loss_scaling
param_grads = optimizer.backward(loss)
master_param_grads = create_master_params_grads(
param_grads, train_program, startup_prog, loss_scaling)
for param, _ in master_param_grads:
param_list[param.name] = param * 1.0
param_list[param.name].stop_gradient = True
if use_dynamic_loss_scaling:
apply_dynamic_loss_scaling(
loss_scaling, master_param_grads, incr_every_n_steps,
decr_every_n_nan_or_inf, incr_ratio, decr_ratio)
optimizer.apply_gradients(master_param_grads)
if weight_decay > 0:
for param, grad in master_param_grads:
if exclude_from_weight_decay(param.name.rstrip(".master")):
continue
with param.block.program._optimized_guard(
[param, grad]), fluid.framework.name_scope("weight_decay"):
updated_param = param - param_list[
param.name] * weight_decay * scheduled_lr
fluid.layers.assign(output=param, input=updated_param)
master_param_to_train_param(master_param_grads, param_grads,
train_program)
else:
for param in train_program.global_block().all_parameters():
param_list[param.name] = param * 1.0
param_list[param.name].stop_gradient = True
_, param_grads = optimizer.minimize(loss)
if weight_decay > 0:
for param, grad in param_grads:
if exclude_from_weight_decay(param.name):
continue
with param.block.program._optimized_guard(
[param, grad]), fluid.framework.name_scope("weight_decay"):
updated_param = param - param_list[
param.name] * weight_decay * scheduled_lr
fluid.layers.assign(output=param, input=updated_param)
return scheduled_lr, loss_scaling
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Arguments for configuration."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from __future__ import unicode_literals
from __future__ import absolute_import
import six
import os
import sys
import argparse
import logging
import paddle.fluid as fluid
log = logging.getLogger("logger")
def prepare_logger(logger, debug=False, save_to_file=None):
formatter = logging.Formatter(
fmt='[%(levelname)s] %(asctime)s [%(filename)12s:%(lineno)5d]:\t%(message)s'
)
# console_hdl = logging.StreamHandler()
# console_hdl.setFormatter(formatter)
# logger.addHandler(console_hdl)
if save_to_file is not None: #and not os.path.exists(save_to_file):
if os.path.isdir(save_to_file):
file_hdl = logging.FileHandler(
os.path.join(save_to_file, 'log.txt'))
else:
file_hdl = logging.FileHandler(save_to_file)
file_hdl.setFormatter(formatter)
logger.addHandler(file_hdl)
logger.setLevel(logging.DEBUG)
logger.propagate = False
def str2bool(v):
# because argparse does not support to parse "true, False" as python
# boolean directly
return v.lower() in ("true", "t", "1")
class ArgumentGroup(object):
def __init__(self, parser, title, des):
self._group = parser.add_argument_group(title=title, description=des)
def add_arg(self,
name,
type,
default,
help,
positional_arg=False,
**kwargs):
prefix = "" if positional_arg else "--"
type = str2bool if type == bool else type
self._group.add_argument(
prefix + name,
default=default,
type=type,
help=help + ' Default: %(default)s.',
**kwargs)
def print_arguments(args):
log.info('----------- Configuration Arguments -----------')
for arg, value in sorted(six.iteritems(vars(args))):
log.info('%s: %s' % (arg, value))
log.info('------------------------------------------------')
def check_cuda(use_cuda, err = \
"\nYou can not set use_cuda = True in the model because you are using paddlepaddle-cpu.\n \
Please: 1. Install paddlepaddle-gpu to run your models on GPU or 2. Set use_cuda = False to run models on CPU.\n"
):
try:
if use_cuda == True and fluid.is_compiled_with_cuda() == False:
log.error(err)
sys.exit(1)
except Exception as e:
pass
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from __future__ import unicode_literals
from __future__ import absolute_import
import os
def get_cards():
"""
get gpu cards number
"""
num = 0
cards = os.environ.get('CUDA_VISIBLE_DEVICES', '')
if cards != '':
num = len(cards.split(","))
return num
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
This file implement a class for model configure.
"""
import datetime
import os
import yaml
import random
import shutil
import six
import logging
log = logging.getLogger("logger")
class AttrDict(dict):
"""Attr dict
"""
def __init__(self, d):
self.dict = d
def __getattr__(self, attr):
value = self.dict[attr]
if isinstance(value, dict):
return AttrDict(value)
else:
return value
def __str__(self):
return str(self.dict)
class Config(object):
"""Implementation of Config class for model configure.
Args:
config_file(str): configure filename, which is a yaml file.
isCreate(bool): if true, create some neccessary directories to save models, log file and other outputs.
isSave(bool): if true, save config_file in order to record the configure message.
"""
def __init__(self, config_file, isCreate=False, isSave=False):
self.config_file = config_file
# self.config = self.get_config_from_yaml(config_file)
self.config = self.load_config(config_file)
if isCreate:
self.create_necessary_dirs()
if isSave:
self.save_config_file()
def load_config(self, config_file):
"""Load config file"""
with open(config_file) as f:
if hasattr(yaml, 'FullLoader'):
config = yaml.load(f, Loader=yaml.FullLoader)
else:
config = yaml.load(f)
return config
def create_necessary_dirs(self):
"""Create some necessary directories to save some important files.
"""
self.config['log_dir'] = os.path.join(self.config['log_dir'],
self.config['task_name'])
self.config['save_dir'] = os.path.join(self.config['save_dir'],
self.config['task_name'])
self.config['output_dir'] = os.path.join(self.config['output_dir'],
self.config['task_name'])
self.make_dir(self.config['log_dir'])
self.make_dir(self.config['save_dir'])
self.make_dir(self.config['output_dir'])
def save_config_file(self):
"""Save config file so that we can know the config when we look back
"""
filename = self.config_file.split('/')[-1]
targetpath = os.path.join(self.config['save_dir'], filename)
try:
shutil.copyfile(self.config_file, targetpath)
except shutil.SameFileError:
log.info("%s and %s are the same file, did not copy by shutil"\
% (self.config_file, targetpath))
def make_dir(self, path):
"""Build directory"""
if not os.path.exists(path):
os.makedirs(path)
def __getitem__(self, key):
return self.config[key]
def __call__(self):
"""__call__"""
return self.config
def __getattr__(self, attr):
try:
result = self.config[attr]
except KeyError:
log.warn("%s attribute is not existed, return None" % attr)
result = None
return result
def __setitem__(self, key, value):
self.config[key] = value
def __str__(self):
return str(self.config)
def pretty_print(self):
log.info(
"-----------------------------------------------------------------")
log.info("config file: %s" % self.config_file)
for key, value in sorted(
self.config.items(), key=lambda item: item[0]):
log.info("%s: %s" % (key, value))
log.info(
"-----------------------------------------------------------------")
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import print_function
import paddle
import paddle.fluid as fluid
def append_cast_op(i, o, prog):
"""
Append a cast op in a given Program to cast input `i` to data type `o.dtype`.
Args:
i (Variable): The input Variable.
o (Variable): The output Variable.
prog (Program): The Program to append cast op.
"""
prog.global_block().append_op(
type="cast",
inputs={"X": i},
outputs={"Out": o},
attrs={"in_dtype": i.dtype,
"out_dtype": o.dtype})
def copy_to_master_param(p, block):
v = block.vars.get(p.name, None)
if v is None:
raise ValueError("no param name %s found!" % p.name)
new_p = fluid.framework.Parameter(
block=block,
shape=v.shape,
dtype=fluid.core.VarDesc.VarType.FP32,
type=v.type,
lod_level=v.lod_level,
stop_gradient=p.stop_gradient,
trainable=p.trainable,
optimize_attr=p.optimize_attr,
regularizer=p.regularizer,
gradient_clip_attr=p.gradient_clip_attr,
error_clip=p.error_clip,
name=v.name + ".master")
return new_p
def apply_dynamic_loss_scaling(loss_scaling, master_params_grads,
incr_every_n_steps, decr_every_n_nan_or_inf,
incr_ratio, decr_ratio):
_incr_every_n_steps = fluid.layers.fill_constant(
shape=[1], dtype='int32', value=incr_every_n_steps)
_decr_every_n_nan_or_inf = fluid.layers.fill_constant(
shape=[1], dtype='int32', value=decr_every_n_nan_or_inf)
_num_good_steps = fluid.layers.create_global_var(
name=fluid.unique_name.generate("num_good_steps"),
shape=[1],
value=0,
dtype='int32',
persistable=True)
_num_bad_steps = fluid.layers.create_global_var(
name=fluid.unique_name.generate("num_bad_steps"),
shape=[1],
value=0,
dtype='int32',
persistable=True)
grads = [fluid.layers.reduce_sum(g) for [_, g] in master_params_grads]
all_grads = fluid.layers.concat(grads)
all_grads_sum = fluid.layers.reduce_sum(all_grads)
is_overall_finite = fluid.layers.isfinite(all_grads_sum)
update_loss_scaling(is_overall_finite, loss_scaling, _num_good_steps,
_num_bad_steps, _incr_every_n_steps,
_decr_every_n_nan_or_inf, incr_ratio, decr_ratio)
# apply_gradient append all ops in global block, thus we shouldn't
# apply gradient in the switch branch.
with fluid.layers.Switch() as switch:
with switch.case(is_overall_finite):
pass
with switch.default():
for _, g in master_params_grads:
fluid.layers.assign(fluid.layers.zeros_like(g), g)
def create_master_params_grads(params_grads, main_prog, startup_prog,
loss_scaling):
master_params_grads = []
for p, g in params_grads:
with main_prog._optimized_guard([p, g]):
# create master parameters
master_param = copy_to_master_param(p, main_prog.global_block())
startup_master_param = startup_prog.global_block()._clone_variable(
master_param)
startup_p = startup_prog.global_block().var(p.name)
append_cast_op(startup_p, startup_master_param, startup_prog)
# cast fp16 gradients to fp32 before apply gradients
if g.name.find("layer_norm") > -1:
scaled_g = g / loss_scaling
master_params_grads.append([p, scaled_g])
continue
master_grad = fluid.layers.cast(g, "float32")
master_grad = master_grad / loss_scaling
master_params_grads.append([master_param, master_grad])
return master_params_grads
def master_param_to_train_param(master_params_grads, params_grads, main_prog):
for idx, m_p_g in enumerate(master_params_grads):
train_p, _ = params_grads[idx]
if train_p.name.find("layer_norm") > -1:
continue
with main_prog._optimized_guard([m_p_g[0], m_p_g[1]]):
append_cast_op(m_p_g[0], train_p, main_prog)
def update_loss_scaling(is_overall_finite, prev_loss_scaling, num_good_steps,
num_bad_steps, incr_every_n_steps,
decr_every_n_nan_or_inf, incr_ratio, decr_ratio):
"""
Update loss scaling according to overall gradients. If all gradients is
finite after incr_every_n_steps, loss scaling will increase by incr_ratio.
Otherwisw, loss scaling will decrease by decr_ratio after
decr_every_n_nan_or_inf steps and each step some gradients are infinite.
Args:
is_overall_finite (Variable): A boolean variable indicates whether
all gradients are finite.
prev_loss_scaling (Variable): Previous loss scaling.
num_good_steps (Variable): A variable accumulates good steps in which
all gradients are finite.
num_bad_steps (Variable): A variable accumulates bad steps in which
some gradients are infinite.
incr_every_n_steps (Variable): A variable represents increasing loss
scaling every n consecutive steps with
finite gradients.
decr_every_n_nan_or_inf (Variable): A variable represents decreasing
loss scaling every n accumulated
steps with nan or inf gradients.
incr_ratio(float): The multiplier to use when increasing the loss
scaling.
decr_ratio(float): The less-than-one-multiplier to use when decreasing
loss scaling.
"""
zero_steps = fluid.layers.fill_constant(shape=[1], dtype='int32', value=0)
with fluid.layers.Switch() as switch:
with switch.case(is_overall_finite):
should_incr_loss_scaling = fluid.layers.less_than(
incr_every_n_steps, num_good_steps + 1)
with fluid.layers.Switch() as switch1:
with switch1.case(should_incr_loss_scaling):
new_loss_scaling = prev_loss_scaling * incr_ratio
loss_scaling_is_finite = fluid.layers.isfinite(
new_loss_scaling)
with fluid.layers.Switch() as switch2:
with switch2.case(loss_scaling_is_finite):
fluid.layers.assign(new_loss_scaling,
prev_loss_scaling)
with switch2.default():
pass
fluid.layers.assign(zero_steps, num_good_steps)
fluid.layers.assign(zero_steps, num_bad_steps)
with switch1.default():
fluid.layers.increment(num_good_steps)
fluid.layers.assign(zero_steps, num_bad_steps)
with switch.default():
should_decr_loss_scaling = fluid.layers.less_than(
decr_every_n_nan_or_inf, num_bad_steps + 1)
with fluid.layers.Switch() as switch3:
with switch3.case(should_decr_loss_scaling):
new_loss_scaling = prev_loss_scaling * decr_ratio
static_loss_scaling = \
fluid.layers.fill_constant(shape=[1],
dtype='float32',
value=1.0)
less_than_one = fluid.layers.less_than(new_loss_scaling,
static_loss_scaling)
with fluid.layers.Switch() as switch4:
with switch4.case(less_than_one):
fluid.layers.assign(static_loss_scaling,
prev_loss_scaling)
with switch4.default():
fluid.layers.assign(new_loss_scaling,
prev_loss_scaling)
fluid.layers.assign(zero_steps, num_good_steps)
fluid.layers.assign(zero_steps, num_bad_steps)
with switch3.default():
fluid.layers.assign(zero_steps, num_good_steps)
fluid.layers.increment(num_bad_steps)
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from __future__ import unicode_literals
from __future__ import absolute_import
import os
import six
import ast
import copy
import logging
import numpy as np
import paddle.fluid as fluid
log = logging.getLogger("logger")
def cast_fp32_to_fp16(exe, main_program):
log.info("Cast parameters to float16 data format.")
for param in main_program.global_block().all_parameters():
if not param.name.endswith(".master"):
param_t = fluid.global_scope().find_var(param.name).get_tensor()
data = np.array(param_t)
if param.name.startswith("encoder_layer") \
and "layer_norm" not in param.name:
param_t.set(np.float16(data).view(np.uint16), exe.place)
#load fp32
master_param_var = fluid.global_scope().find_var(param.name +
".master")
if master_param_var is not None:
master_param_var.get_tensor().set(data, exe.place)
def init_checkpoint(exe, init_checkpoint_path, main_program, use_fp16=False):
assert os.path.exists(
init_checkpoint_path), "[%s] cann't be found." % init_checkpoint_path
def existed_persitables(var):
if not fluid.io.is_persistable(var):
return False
return os.path.exists(os.path.join(init_checkpoint_path, var.name))
fluid.io.load_vars(
exe,
init_checkpoint_path,
main_program=main_program,
predicate=existed_persitables)
log.info("Load model from {}".format(init_checkpoint_path))
if use_fp16:
cast_fp32_to_fp16(exe, main_program)
def init_pretraining_params(exe,
pretraining_params_path,
main_program,
use_fp16=False):
assert os.path.exists(pretraining_params_path
), "[%s] cann't be found." % pretraining_params_path
def existed_params(var):
if not isinstance(var, fluid.framework.Parameter):
return False
return os.path.exists(os.path.join(pretraining_params_path, var.name))
fluid.io.load_vars(
exe,
pretraining_params_path,
main_program=main_program,
predicate=existed_params)
log.info("Load pretraining parameters from {}.".format(
pretraining_params_path))
if use_fp16:
cast_fp32_to_fp16(exe, main_program)
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册