2024-09-18 21:42:09 +00:00
|
|
|
import unittest
|
2024-09-25 09:05:11 -07:00
|
|
|
from functools import partial
|
|
|
|
|
2025-02-14 20:42:19 +00:00
|
|
|
import pytest
|
2025-01-29 15:25:10 -08:00
|
|
|
from torch.utils.data import DataLoader
|
|
|
|
from tqdm import tqdm
|
2024-09-25 09:05:11 -07:00
|
|
|
from transformers import AutoProcessor
|
2024-09-18 21:42:09 +00:00
|
|
|
|
2025-01-27 18:30:41 +00:00
|
|
|
from olmocr.train.dataloader import (
|
2024-10-16 18:06:27 +00:00
|
|
|
build_finetuning_dataset,
|
2024-09-18 22:52:42 +00:00
|
|
|
extract_openai_batch_response,
|
2025-01-29 15:25:10 -08:00
|
|
|
list_dataset_files,
|
2024-10-08 22:10:18 +00:00
|
|
|
load_jsonl_into_ds,
|
2024-09-18 22:52:42 +00:00
|
|
|
)
|
2025-01-27 18:30:41 +00:00
|
|
|
from olmocr.train.dataprep import batch_prepare_data_for_qwen2_training
|
2024-09-25 09:05:11 -07:00
|
|
|
|
2024-09-18 21:42:09 +00:00
|
|
|
|
2025-02-14 20:42:19 +00:00
|
|
|
@pytest.mark.nonci
|
2024-09-18 21:42:09 +00:00
|
|
|
class TestBatchQueryResponseDataset(unittest.TestCase):
|
|
|
|
def testLoadS3(self):
|
2024-10-07 07:49:16 -07:00
|
|
|
ds = load_jsonl_into_ds("s3://ai2-oe-data/jakep/openai_batch_data_v2/*.jsonl", first_n_files=3)
|
2024-09-18 21:42:09 +00:00
|
|
|
|
|
|
|
print(f"Loaded {len(ds)} entries")
|
|
|
|
print(ds)
|
|
|
|
print(ds["train"])
|
2024-09-18 22:52:42 +00:00
|
|
|
|
2024-10-16 18:06:27 +00:00
|
|
|
def testFinetuningDS(self):
|
|
|
|
ds = build_finetuning_dataset(
|
2024-10-08 22:10:18 +00:00
|
|
|
response_glob_path="s3://ai2-oe-data/jakep/pdfdata/openai_batch_done_v5_1_eval/*.json",
|
2024-09-18 22:52:42 +00:00
|
|
|
)
|
2024-09-18 21:42:09 +00:00
|
|
|
|
|
|
|
print(ds)
|
|
|
|
|
2024-10-02 22:45:40 +00:00
|
|
|
processor = AutoProcessor.from_pretrained("Qwen/Qwen2-VL-2B-Instruct")
|
2024-10-16 18:26:25 +00:00
|
|
|
|
|
|
|
ds = ds.with_transform(partial(batch_prepare_data_for_qwen2_training, processor=processor, target_longest_image_dim=1024, target_anchor_text_len=6000))
|
2024-10-02 22:45:40 +00:00
|
|
|
|
|
|
|
print(ds[0])
|
|
|
|
|
2024-09-25 09:05:11 -07:00
|
|
|
def testPlotSequenceLengthHistogram(self):
|
2025-01-29 15:30:39 -08:00
|
|
|
import plotly.express as px
|
2024-09-25 09:05:11 -07:00
|
|
|
|
2024-10-16 18:26:25 +00:00
|
|
|
ds = build_finetuning_dataset(
|
2024-10-02 22:45:40 +00:00
|
|
|
response_glob_path="s3://ai2-oe-data/jakep/pdfdata/openai_batch_done_v5_1_eval/*.json",
|
2024-09-25 09:05:11 -07:00
|
|
|
)
|
2024-10-16 18:26:25 +00:00
|
|
|
|
2024-09-25 09:05:11 -07:00
|
|
|
processor = AutoProcessor.from_pretrained("Qwen/Qwen2-VL-2B-Instruct")
|
|
|
|
|
2024-10-16 18:26:25 +00:00
|
|
|
ds = ds.with_transform(partial(batch_prepare_data_for_qwen2_training, processor=processor, target_longest_image_dim=1024, target_anchor_text_len=6000))
|
2024-10-02 22:45:40 +00:00
|
|
|
|
2024-10-16 18:26:25 +00:00
|
|
|
processor = AutoProcessor.from_pretrained("Qwen/Qwen2-VL-2B-Instruct")
|
2024-09-27 15:48:56 +00:00
|
|
|
|
2024-10-16 18:26:25 +00:00
|
|
|
initial_len = len(ds)
|
|
|
|
|
|
|
|
train_dataloader = DataLoader(ds, batch_size=1, num_workers=30, shuffle=False)
|
2024-09-25 09:05:11 -07:00
|
|
|
|
|
|
|
max_seen_len = 0
|
|
|
|
steps = 0
|
|
|
|
sequence_lengths = [] # List to store sequence lengths
|
|
|
|
for entry in tqdm(train_dataloader):
|
|
|
|
num_input_tokens = entry["input_ids"].shape[1]
|
|
|
|
max_seen_len = max(max_seen_len, num_input_tokens)
|
|
|
|
sequence_lengths.append(num_input_tokens) # Collecting sequence lengths
|
|
|
|
|
|
|
|
if steps % 100 == 0:
|
|
|
|
print(f"Max input len {max_seen_len}")
|
|
|
|
|
|
|
|
steps += 1
|
|
|
|
|
|
|
|
# model.forward(**{k: v.to("cuda:0") for (k,v) in entry.items()})
|
|
|
|
print(f"Max input len {max_seen_len}")
|
2024-10-02 22:45:40 +00:00
|
|
|
print(f"Total elements before filtering: {initial_len}")
|
|
|
|
print(f"Total elements after filtering: {steps}")
|
2024-09-25 09:05:11 -07:00
|
|
|
|
|
|
|
# Plotting the histogram using Plotly
|
|
|
|
fig = px.histogram(
|
2025-01-29 15:30:39 -08:00
|
|
|
sequence_lengths, nbins=100, title="Distribution of Input Sequence Lengths", labels={"value": "Sequence Length", "count": "Frequency"}
|
2024-09-25 09:05:11 -07:00
|
|
|
)
|
|
|
|
|
|
|
|
fig.write_image("sequence_lengths_histogram.png")
|