Update codeparrot_training.py
Browse files- codeparrot_training.py +11 -6
codeparrot_training.py
CHANGED
@@ -15,7 +15,7 @@ import wandb
|
|
15 |
|
16 |
class ConstantLengthDataset(IterableDataset):
|
17 |
|
18 |
-
def __init__(self, tokenizer, dataset, seq_length=1024,
|
19 |
num_of_sequences=1024, chars_per_token=3.6):
|
20 |
self.tokenizer = tokenizer
|
21 |
self.concat_token_id = tokenizer.bos_token_id
|
@@ -23,6 +23,7 @@ class ConstantLengthDataset(IterableDataset):
|
|
23 |
self.seq_length = seq_length
|
24 |
self.input_characters = seq_length * chars_per_token * num_of_sequences
|
25 |
self.epoch = 0
|
|
|
26 |
|
27 |
def __iter__(self):
|
28 |
iterator = iter(self.dataset)
|
@@ -36,9 +37,13 @@ class ConstantLengthDataset(IterableDataset):
|
|
36 |
buffer.append(next(iterator)['content'])
|
37 |
buffer_len += len(buffer[-1])
|
38 |
except StopIteration:
|
39 |
-
|
40 |
-
|
41 |
-
|
|
|
|
|
|
|
|
|
42 |
tokenized_inputs = tokenizer(buffer, truncation=False)['input_ids']
|
43 |
all_token_ids = []
|
44 |
for tokenized_input in tokenized_inputs:
|
@@ -77,9 +82,9 @@ def create_dataloaders(dataset_name, args):
|
|
77 |
train_data = train_data.shuffle(buffer_size=args.shuffle_buffer,
|
78 |
seed=args.seed)
|
79 |
valid_data = load_dataset(dataset_name+'-valid', split="train", **ds_kwargs)
|
80 |
-
train_dataset = ConstantLengthDataset(tokenizer, train_data,
|
81 |
seq_length=args.seq_length)
|
82 |
-
valid_dataset = ConstantLengthDataset(tokenizer, valid_data,
|
83 |
seq_length=args.seq_length)
|
84 |
train_dataloader=DataLoader(train_dataset, batch_size=args.train_batch_size)
|
85 |
eval_dataloader=DataLoader(valid_dataset, batch_size=args.valid_batch_size)
|
|
|
15 |
|
16 |
class ConstantLengthDataset(IterableDataset):
|
17 |
|
18 |
+
def __init__(self, tokenizer, dataset, infinite=False, seq_length=1024,
|
19 |
num_of_sequences=1024, chars_per_token=3.6):
|
20 |
self.tokenizer = tokenizer
|
21 |
self.concat_token_id = tokenizer.bos_token_id
|
|
|
23 |
self.seq_length = seq_length
|
24 |
self.input_characters = seq_length * chars_per_token * num_of_sequences
|
25 |
self.epoch = 0
|
26 |
+
self.infinite = infinite
|
27 |
|
28 |
def __iter__(self):
|
29 |
iterator = iter(self.dataset)
|
|
|
37 |
buffer.append(next(iterator)['content'])
|
38 |
buffer_len += len(buffer[-1])
|
39 |
except StopIteration:
|
40 |
+
if self.infinite:
|
41 |
+
iterator = iter(self.dataset)
|
42 |
+
self.epoch += 1
|
43 |
+
logger.info(f"Dataset epoch: {self.epoch}")
|
44 |
+
else:
|
45 |
+
more_examples = False
|
46 |
+
break
|
47 |
tokenized_inputs = tokenizer(buffer, truncation=False)['input_ids']
|
48 |
all_token_ids = []
|
49 |
for tokenized_input in tokenized_inputs:
|
|
|
82 |
train_data = train_data.shuffle(buffer_size=args.shuffle_buffer,
|
83 |
seed=args.seed)
|
84 |
valid_data = load_dataset(dataset_name+'-valid', split="train", **ds_kwargs)
|
85 |
+
train_dataset = ConstantLengthDataset(tokenizer, train_data, infinite=True,
|
86 |
seq_length=args.seq_length)
|
87 |
+
valid_dataset = ConstantLengthDataset(tokenizer, valid_data, infinite=False,
|
88 |
seq_length=args.seq_length)
|
89 |
train_dataloader=DataLoader(train_dataset, batch_size=args.train_batch_size)
|
90 |
eval_dataloader=DataLoader(valid_dataset, batch_size=args.valid_batch_size)
|