def checkpoint_stage(stage, result):
checkpoint = utils.get_checkpoint().get(stage, {})
checkpoint.update({'result': result})
utils.update_checkpoint({
'stage': stages.index(stage),
stage: checkpoint,
})
def set_manual_stages(params):
start_stage_idx = len(stages) - 1
checkpoint = utils.get_checkpoint()['params']
# For each stage, check if any of the parameters have changed since the last run
changed_stages = set()
for stage,stage_params in params.items():
diff = {}
for key,value in stage_params.items():
if checkpoint.get(stage) and checkpoint.get(stage).get(key) != value:
# the value differs from the checkpoint
diff[key] = (value, checkpoint.get(stage).get(key))
if diff:
logger.info("parameters differ from checkpoint in stage %s: %s", stage, diff)
changed_stages.add(stage)
# For stage, mark is as changed if any of its dependencies changed
for stage in stages:
for dependency in dependencies.get(stage, []):
if stages[dependency] in changed_stages:
changed_stages.add(stage)
# Finally, add the required stages to the manual stages
if changed_stages:
start_stage_idx = min([ stages.index(stage) for stage in changed_stages ])
stage_required.manual_stages = range(start_stage_idx, len(stages))
def stage_required(stage):
# check that 'stage-1' has been checkpointed
stage_index = stages.index(stage)
required = (utils.get_checkpoint()['stage'] < stage_index)
if stage_required.manual_stages:
if stage_index > max(stage_required.manual_stages):
sys.exit(0)
elif stage_index in stage_required.manual_stages:
return True
else:
return False
return required
def do_linalg(parameters, sieve_result):
result = {}
# Check the rare case that factorization was completed in msieve's filtering
if utils.get_checkpoint().get('trial_division') != None:
result['duration'] = 0
elif stage_required('linalg'):
linalg_start = time.time()
linalg.run(parameters)
linalg_finish = time.time()
result['duration'] = linalg_finish - linalg_start
logger.info("\tLinalg in %s", utils.str_time(result['duration']))
checkpoint_stage('linalg', result)
else:
result = load_stage('linalg')
post_linalg = parameters.myparams({'post_linalg': None}, ['commands']).get('post_linalg')
if post_linalg != None:
logger.info('Post-linalg command %s', post_linalg)
utils.run_command(post_linalg, logger=logger)
return result
def load_stage(stage):
return utils.get_checkpoint().get(stage, {}).get('result')
def main():
# get args
parser = argparse.ArgumentParser(description="Im2Latex Training Program")
# parser.add_argument('--path', required=True, help='root of the model')
# model args
parser.add_argument("--emb_dim",
type=int,
default=80,
help="Embedding size")
parser.add_argument("--dec_rnn_h",
type=int,
default=512,
help="The hidden state of the decoder RNN")
parser.add_argument("--data_path",
type=str,
default="/root/private/im2latex/data/",
help="The dataset's dir")
parser.add_argument("--add_position_features",
action='store_true',
default=False,
help="Use position embeddings or not")
# training args
parser.add_argument("--max_len",
type=int,
default=150,
help="Max size of formula")
parser.add_argument("--dropout",
type=float,
default=0.,
help="Dropout probility")
parser.add_argument("--cuda",
action='store_true',
default=True,
help="Use cuda or not")
parser.add_argument("--batch_size", type=int, default=8)
parser.add_argument("--epoches", type=int, default=200)
parser.add_argument("--lr", type=float, default=3e-4, help="Learning Rate")
parser.add_argument("--min_lr",
type=float,
default=3e-5,
help="Learning Rate")
parser.add_argument("--sample_method",
type=str,
default="teacher_forcing",
choices=('teacher_forcing', 'exp', 'inv_sigmoid'),
help="The method to schedule sampling")
parser.add_argument("--decay_k", type=float, default=1.)
parser.add_argument("--lr_decay",
type=float,
default=0.5,
help="Learning Rate Decay Rate")
parser.add_argument("--lr_patience",
type=int,
default=3,
help="Learning Rate Decay Patience")
parser.add_argument("--clip",
type=float,
default=2.0,
help="The max gradient norm")
parser.add_argument("--save_dir",
type=str,
default="./ckpts",
help="The dir to save checkpoints")
parser.add_argument("--print_freq",
type=int,
default=100,
help="The frequency to print message")
parser.add_argument("--seed",
type=int,
default=2020,
help="The random seed for reproducing ")
parser.add_argument("--from_check_point",
action='store_true',
default=False,
help="Training from checkpoint or not")
parser.add_argument("--batch_size_per_gpu", type=int, default=16)
parser.add_argument("--gpu_num", type=int, default=4)
device_ids = [0, 1, 2, 3]
args = parser.parse_args()
max_epoch = args.epoches
from_check_point = args.from_check_point
if from_check_point:
checkpoint_path = get_checkpoint(args.save_dir)
checkpoint = torch.load(checkpoint_path)
args = checkpoint['args']
print("Training args:", args)
torch.manual_seed(args.seed)
torch.cuda.manual_seed(args.seed)
# Building vocab
print("Load vocab...")
vocab = load_vocab(args.data_path)
use_cuda = True if args.cuda and torch.cuda.is_available() else False
print(use_cuda)
device = torch.device("cuda" if use_cuda else "cpu")
# data loader
print("Construct data loader...")
# train_loader = DataLoader(
# Im2LatexDataset(args.data_path, 'train', args.max_len),
# batch_size=args.batch_size,
# collate_fn=partial(collate_fn, vocab.token2idx),
# pin_memory=True if use_cuda else False,
# num_workers=4)
train_loader = DataLoader(
Im2LatexDataset(args.data_path, 'train', args.max_len),
batch_size=args.batch_size_per_gpu * args.gpu_num,
collate_fn=partial(collate_fn, vocab.token2idx),
pin_memory=True if use_cuda else False,
num_workers=2)
# val_loader = DataLoader(
# Im2LatexDataset(args.data_path, 'validate', args.max_len),
# batch_size=args.batch_size,
# collate_fn=partial(collate_fn, vocab.token2idx),
# pin_memory=True if use_cuda else False,
# num_workers=4)
val_loader = DataLoader(Im2LatexDataset(args.data_path, 'validate',
args.max_len),
batch_size=args.batch_size_per_gpu * args.gpu_num,
collate_fn=partial(collate_fn, vocab.token2idx),
pin_memory=True if use_cuda else False,
num_workers=2)
# construct model
print("Construct model")
vocab_size = len(vocab)
model = Im2LatexModel(vocab_size,
args.emb_dim,
args.dec_rnn_h,
add_pos_feat=args.add_position_features,
dropout=args.dropout)
model = nn.DataParallel(model, device_ids=device_ids)
# model = model.
model = model.cuda()
print("Model Settings:")
print(model)
# construct optimizer
optimizer = optim.Adam(model.parameters(), lr=args.lr)
lr_scheduler = ReduceLROnPlateau(optimizer,
"min",
factor=args.lr_decay,
patience=args.lr_patience,
verbose=True,
min_lr=args.min_lr)
if from_check_point:
model.load_state_dict(checkpoint['model_state_dict'])
optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
epoch = checkpoint['epoch']
lr_scheduler.load_state_dict(checkpoint['lr_sche'])
# init trainer from checkpoint
trainer = Trainer(optimizer,
model,
lr_scheduler,
train_loader,
val_loader,
args,
use_cuda=use_cuda,
init_epoch=epoch,
last_epoch=max_epoch)
else:
trainer = Trainer(optimizer,
model,
lr_scheduler,
train_loader,
val_loader,
args,
use_cuda=use_cuda,
init_epoch=1,
last_epoch=args.epoches)
# begin training
trainer.train()
def run_rnn_rcnn(args):
dataset_dir = utils.choose_data(args)
with tf.device('/cpu:0'):
dataset = utils.load_dataset(dataset_dir)
print "Using checkpoint directory: {0}".format(args.ckpt_dir)
reward_output_file = os.path.join(args.ckpt_dir, "rewards")
accuracy_output_file = os.path.join(args.ckpt_dir, "accuracies")
model = utils.choose_model(args) # pass in necessary model parameters
print "Running {0} model for {1} epochs.".format(args.model, args.num_epochs)
global_step = tf.Variable(0, trainable=False, name='global_step')
saver = tf.train.Saver(max_to_keep=args.num_epochs)
with tf.Session(config=GPU_CONFIG) as session:
print "Inititialized TF Session!"
# Checkpoint
i_stopped, found_ckpt = utils.get_checkpoint(args, session, saver)
# Summary Writer
file_writer = tf.summary.FileWriter(args.ckpt_dir, graph=session.graph, max_queue=10, flush_secs=30)
# Val or Test set accuracie
# Make computational graph
if args.train == "train" and not found_ckpt:
init_op = tf.global_variables_initializer()
init_op.run()
else:
if not found_ckpt:
print "No checkpoint found for test or validation!"
return
if "pretrained" in args.model:
init_fn = tf.contrib.framework.assign_from_checkpoint_fn(
model_path='/data/yolo/YOLO_small.ckpt',
var_list=model.variables_to_restore)
init_fn(session)
if "neg_l1" in args.model:
model.add_loss_op('negative_l1_dist')
elif "iou" in args.model:
model.add_loss_op('iou')
model.add_summary_op()
# if "pretrained" in args.model:
if args.train == 'train':
for i in xrange(i_stopped, args.num_epochs):
print "Running epoch ({0})...".format(i)
with tf.device('/cpu:0'):
batched_data, batched_labels, batched_seq_lens, batched_bbox = utils.make_batches(dataset, batch_size=BATCH_SIZE)
rewards, accuracies = run_epoch(args, model, session, batched_data, batched_labels, batched_seq_lens, batched_bbox,saver,
file_writer, i)
with open(reward_output_file, "a") as reward_file:
reward_file.write("\n".join([r.astype('|S6') for r in rewards]) + "\n")
with open(accuracy_output_file, "a") as accuracy_file:
accuracy_file.write("\n".join([a.astype('|S6') for a in accuracies]) + "\n")
if args.train == 'test':
# model.add_error_op()
# model.add_summary_op()
batched_data, batched_labels, batched_seq_lens, batched_bbox = utils.make_batches(dataset, batch_size=BATCH_SIZE)
run_epoch(args, model, session, batched_data, batched_labels, batched_seq_lens, batched_bbox,saver,
file_writer, 1)
print "Output weights are same: " + str(output)
print "Output bias weights are same: " + str(output_bias)
print "-------------------------------------------------------------------"
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument("--nature", default=True, type=bool)
args = parser.parse_args()
epoch = 0
everywhere_all_the_time = True
while checkpoints_all_exist(epoch):
nets = []
for checkpoint_dir in CHECKPOINT_FOLDERS:
checkpoint = utils.get_checkpoint(checkpoint_dir, epoch)
nets.append(checkpoint['state_dict'])
'''
Print results
'''
print "EPOCH: " + str(epoch)
conv1 = True
conv2 = True
fc1 = True
output = True
conv1_bias = True
conv2_bias = True
fc1_bias = True
output_bias = True
if args.nature:
conv3 = True
