Spaces:
Runtime error
Runtime error
DimaKoshman
commited on
Commit
•
028951c
1
Parent(s):
88f8b47
trained model for a bit
Browse files- MakingGraphsAccessible.ipynb +0 -0
- app.py +48 -36
- checkpoint/added_tokens.json +11 -10
- checkpoint/config.json +5 -5
- checkpoint/generation_config.json +3 -4
- checkpoint/pytorch_model.bin +2 -2
- checkpoint/special_tokens_map.json +1 -1
- checkpoint/tokenizer.json +0 -0
- checkpoint/tokenizer_config.json +1 -1
- config.py +21 -0
- data.py +521 -0
- metrics.py +54 -0
- model.py +193 -0
- requirements.txt +0 -3
- train.py +114 -0
- utils.py +23 -0
MakingGraphsAccessible.ipynb
CHANGED
The diff for this file is too large to render.
See raw diff
|
|
app.py
CHANGED
@@ -1,44 +1,56 @@
|
|
1 |
import gradio
|
2 |
-
import
|
3 |
-
import
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
4 |
|
5 |
|
6 |
-
|
7 |
-
|
8 |
|
9 |
-
|
10 |
-
|
11 |
-
MODEL.encoder_decoder = transformers.VisionEncoderDecoderModel.from_pretrained(checkpoint_path)
|
12 |
-
MODEL.tokenizer = MODEL.donut_processor.tokenizer
|
13 |
|
|
|
|
|
|
|
14 |
|
15 |
-
|
16 |
-
decoder_output = MODEL.encoder_decoder.generate(
|
17 |
-
images,
|
18 |
-
max_length=MODEL.encoder_decoder.config.decoder.max_length,
|
19 |
-
eos_token_id=MODEL.tokenizer.eos_token_id,
|
20 |
-
return_dict_in_generate=True,
|
21 |
-
)
|
22 |
-
return MODEL.tokenizer.batch_decode(
|
23 |
-
decoder_output.sequences, skip_special_tokens=skip_special_tokens
|
24 |
-
)
|
25 |
|
26 |
-
|
27 |
-
|
28 |
-
|
29 |
-
|
30 |
-
|
31 |
-
|
32 |
-
|
33 |
-
|
34 |
-
|
35 |
-
|
36 |
-
|
37 |
-
|
38 |
-
|
39 |
-
|
40 |
-
|
41 |
-
|
42 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
43 |
|
44 |
-
interface.launch()
|
|
|
1 |
import gradio
|
2 |
+
import pandas as pd
|
3 |
+
from matplotlib import pyplot as plt
|
4 |
+
|
5 |
+
from config import CONFIG
|
6 |
+
from machine_learning.transformers.MakingGraphsAccessible.data import (
|
7 |
+
get_extra_tokens,
|
8 |
+
BenetechOutput,
|
9 |
+
ChartType,
|
10 |
+
)
|
11 |
+
from model import predict_string, build_model
|
12 |
|
13 |
|
14 |
+
def gradio_visualize_prediction(string):
|
15 |
+
string = string.removeprefix(get_extra_tokens().benetech_prompt)
|
16 |
|
17 |
+
if not BenetechOutput.does_string_match_expected_pattern(string):
|
18 |
+
return
|
|
|
|
|
19 |
|
20 |
+
benetech_output = BenetechOutput.from_string(string)
|
21 |
+
x = benetech_output.x_data[: len(benetech_output.y_data)]
|
22 |
+
y = benetech_output.y_data[: len(benetech_output.x_data)]
|
23 |
|
24 |
+
df = pd.DataFrame(dict(x=x, y=y))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
25 |
|
26 |
+
plt_plot = {
|
27 |
+
ChartType.line: plt.plot,
|
28 |
+
ChartType.scatter: plt.scatter,
|
29 |
+
ChartType.horizontal_bar: plt.barh,
|
30 |
+
ChartType.vertical_bar: plt.bar,
|
31 |
+
ChartType.dot: plt.scatter,
|
32 |
+
}
|
33 |
+
|
34 |
+
plt_plot[benetech_output.chart_type](x, y)
|
35 |
+
plt.xticks(rotation=30)
|
36 |
+
plt.savefig("plot.png")
|
37 |
+
|
38 |
+
...
|
39 |
+
|
40 |
+
|
41 |
+
def main():
|
42 |
+
config = CONFIG
|
43 |
+
config.pretrained_model_name = "checkpoint"
|
44 |
+
model = build_model(config)
|
45 |
+
|
46 |
+
interface = gradio.Interface(
|
47 |
+
title="Making graphs accessible",
|
48 |
+
description="Generate textual representation of a graph\n"
|
49 |
+
"https://www.kaggle.com/competitions/benetech-making-graphs-accessible",
|
50 |
+
fn=lambda image: predict_string(image, model),
|
51 |
+
inputs="image",
|
52 |
+
outputs="text",
|
53 |
+
examples="examples",
|
54 |
+
)
|
55 |
|
56 |
+
interface.launch()
|
checkpoint/added_tokens.json
CHANGED
@@ -1,18 +1,19 @@
|
|
1 |
{
|
2 |
"1": 57537,
|
3 |
"</benetech_prompt>": 57526,
|
4 |
-
"<;>":
|
5 |
"<benetech_prompt>": 57525,
|
6 |
-
"<categorical>":
|
7 |
-
"<dot>":
|
8 |
-
"<horizontal_bar>":
|
9 |
-
"<line>":
|
10 |
-
"<numerical>":
|
11 |
"<s_iitcdip>": 57523,
|
12 |
"<s_synthdog>": 57524,
|
13 |
-
"<scatter>":
|
14 |
"<sep/>": 57522,
|
15 |
-
"<vertical_bar>":
|
16 |
-
"<x_start>":
|
17 |
-
"<y_start>":
|
|
|
18 |
}
|
|
|
1 |
{
|
2 |
"1": 57537,
|
3 |
"</benetech_prompt>": 57526,
|
4 |
+
"<;>": 57529,
|
5 |
"<benetech_prompt>": 57525,
|
6 |
+
"<categorical>": 57535,
|
7 |
+
"<dot>": 57530,
|
8 |
+
"<horizontal_bar>": 57531,
|
9 |
+
"<line>": 57533,
|
10 |
+
"<numerical>": 57536,
|
11 |
"<s_iitcdip>": 57523,
|
12 |
"<s_synthdog>": 57524,
|
13 |
+
"<scatter>": 57534,
|
14 |
"<sep/>": 57522,
|
15 |
+
"<vertical_bar>": 57532,
|
16 |
+
"<x_start>": 57527,
|
17 |
+
"<y_start>": 57528,
|
18 |
+
"ދ": 57538
|
19 |
}
|
checkpoint/config.json
CHANGED
@@ -4,7 +4,7 @@
|
|
4 |
"architectures": [
|
5 |
"VisionEncoderDecoderModel"
|
6 |
],
|
7 |
-
"bos_token_id":
|
8 |
"decoder": {
|
9 |
"_name_or_path": "",
|
10 |
"activation_dropout": 0.0,
|
@@ -51,7 +51,7 @@
|
|
51 |
"LABEL_1": 1
|
52 |
},
|
53 |
"length_penalty": 1.0,
|
54 |
-
"max_length":
|
55 |
"max_position_embeddings": 1536,
|
56 |
"min_length": 0,
|
57 |
"model_type": "mbart",
|
@@ -88,9 +88,9 @@
|
|
88 |
"typical_p": 1.0,
|
89 |
"use_bfloat16": false,
|
90 |
"use_cache": true,
|
91 |
-
"vocab_size":
|
92 |
},
|
93 |
-
"decoder_start_token_id":
|
94 |
"encoder": {
|
95 |
"_name_or_path": "",
|
96 |
"add_cross_attention": false,
|
@@ -187,7 +187,7 @@
|
|
187 |
"use_bfloat16": false,
|
188 |
"window_size": 10
|
189 |
},
|
190 |
-
"eos_token_id":
|
191 |
"is_encoder_decoder": true,
|
192 |
"model_type": "vision-encoder-decoder",
|
193 |
"pad_token_id": 1,
|
|
|
4 |
"architectures": [
|
5 |
"VisionEncoderDecoderModel"
|
6 |
],
|
7 |
+
"bos_token_id": 3,
|
8 |
"decoder": {
|
9 |
"_name_or_path": "",
|
10 |
"activation_dropout": 0.0,
|
|
|
51 |
"LABEL_1": 1
|
52 |
},
|
53 |
"length_penalty": 1.0,
|
54 |
+
"max_length": 20,
|
55 |
"max_position_embeddings": 1536,
|
56 |
"min_length": 0,
|
57 |
"model_type": "mbart",
|
|
|
88 |
"typical_p": 1.0,
|
89 |
"use_bfloat16": false,
|
90 |
"use_cache": true,
|
91 |
+
"vocab_size": 57539
|
92 |
},
|
93 |
+
"decoder_start_token_id": 3,
|
94 |
"encoder": {
|
95 |
"_name_or_path": "",
|
96 |
"add_cross_attention": false,
|
|
|
187 |
"use_bfloat16": false,
|
188 |
"window_size": 10
|
189 |
},
|
190 |
+
"eos_token_id": 3,
|
191 |
"is_encoder_decoder": true,
|
192 |
"model_type": "vision-encoder-decoder",
|
193 |
"pad_token_id": 1,
|
checkpoint/generation_config.json
CHANGED
@@ -1,10 +1,9 @@
|
|
1 |
{
|
2 |
"_from_model_config": true,
|
3 |
-
"bos_token_id":
|
4 |
-
"decoder_start_token_id":
|
5 |
-
"eos_token_id":
|
6 |
"forced_eos_token_id": 2,
|
7 |
-
"max_length": 512,
|
8 |
"pad_token_id": 1,
|
9 |
"transformers_version": "4.26.1"
|
10 |
}
|
|
|
1 |
{
|
2 |
"_from_model_config": true,
|
3 |
+
"bos_token_id": 3,
|
4 |
+
"decoder_start_token_id": 3,
|
5 |
+
"eos_token_id": 3,
|
6 |
"forced_eos_token_id": 2,
|
|
|
7 |
"pad_token_id": 1,
|
8 |
"transformers_version": "4.26.1"
|
9 |
}
|
checkpoint/pytorch_model.bin
CHANGED
@@ -1,3 +1,3 @@
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:
|
3 |
-
size
|
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:7c9a42d9810580ea7d19acdfe533e97b4be48693c18c82c2b5f337eb879921ff
|
3 |
+
size 809236249
|
checkpoint/special_tokens_map.json
CHANGED
@@ -5,7 +5,7 @@
|
|
5 |
],
|
6 |
"bos_token": "<s>",
|
7 |
"cls_token": "<s>",
|
8 |
-
"eos_token": "
|
9 |
"mask_token": {
|
10 |
"content": "<mask>",
|
11 |
"lstrip": true,
|
|
|
5 |
],
|
6 |
"bos_token": "<s>",
|
7 |
"cls_token": "<s>",
|
8 |
+
"eos_token": "<unk>",
|
9 |
"mask_token": {
|
10 |
"content": "<mask>",
|
11 |
"lstrip": true,
|
checkpoint/tokenizer.json
CHANGED
The diff for this file is too large to render.
See raw diff
|
|
checkpoint/tokenizer_config.json
CHANGED
@@ -11,7 +11,7 @@
|
|
11 |
"single_word": false
|
12 |
},
|
13 |
"model_max_length": 1000000000000000019884624838656,
|
14 |
-
"name_or_path": "
|
15 |
"pad_token": "<pad>",
|
16 |
"processor_class": "DonutProcessor",
|
17 |
"sep_token": "</s>",
|
|
|
11 |
"single_word": false
|
12 |
},
|
13 |
"model_max_length": 1000000000000000019884624838656,
|
14 |
+
"name_or_path": "naver-clova-ix/donut-base",
|
15 |
"pad_token": "<pad>",
|
16 |
"processor_class": "DonutProcessor",
|
17 |
"sep_token": "</s>",
|
config.py
ADDED
@@ -0,0 +1,21 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
class CONFIG:
|
2 |
+
debug = False
|
3 |
+
accelerator = "cpu" if debug else "gpu"
|
4 |
+
devices = "auto" if accelerator == "cpu" else [1]
|
5 |
+
batch_size = 2 if debug else 1
|
6 |
+
limit_train_batches = 2 if debug else None
|
7 |
+
limit_val_batches = 2 if debug else 100
|
8 |
+
learning_rate = 3e-5
|
9 |
+
val_fraction = 0.1
|
10 |
+
seed = 42
|
11 |
+
train_val_indices_path = "data/train_val_indices.pickle"
|
12 |
+
float_scientific_notation_string_precision = 5
|
13 |
+
pretrained_model_name = "naver-clova-ix/donut-base"
|
14 |
+
image_width = 720
|
15 |
+
image_height = 512
|
16 |
+
unknown_tokens_for_tokenizer_path = "data/unknown_tokens_for_tokenizer.pickle"
|
17 |
+
decoder_sequence_max_length = 512
|
18 |
+
num_workers = 4
|
19 |
+
training_directory = "training"
|
20 |
+
save_top_k_checkpoints = 3
|
21 |
+
wandb_project_name = "MakingGraphsAccessible"
|
data.py
ADDED
@@ -0,0 +1,521 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import dataclasses
|
2 |
+
import enum
|
3 |
+
import functools
|
4 |
+
import json
|
5 |
+
import os
|
6 |
+
import re
|
7 |
+
import types
|
8 |
+
from typing import Callable
|
9 |
+
|
10 |
+
import einops
|
11 |
+
import imageio
|
12 |
+
import numpy as np
|
13 |
+
import torch.utils.data
|
14 |
+
import torchvision
|
15 |
+
import tqdm
|
16 |
+
|
17 |
+
from config import CONFIG
|
18 |
+
from utils import load_pickle_or_build_object_and_save
|
19 |
+
|
20 |
+
|
21 |
+
class Source(enum.Enum):
|
22 |
+
generated = "generated"
|
23 |
+
extracted = "extracted"
|
24 |
+
|
25 |
+
|
26 |
+
class ChartType(enum.Enum):
|
27 |
+
dot = "dot"
|
28 |
+
horizontal_bar = "horizontal_bar"
|
29 |
+
vertical_bar = "vertical_bar"
|
30 |
+
line = "line"
|
31 |
+
scatter = "scatter"
|
32 |
+
|
33 |
+
|
34 |
+
@dataclasses.dataclass
|
35 |
+
class PlotBoundingBox:
|
36 |
+
height: int
|
37 |
+
width: int
|
38 |
+
x0: int
|
39 |
+
y0: int
|
40 |
+
|
41 |
+
def get_bounds(self):
|
42 |
+
xs = [self.x0, self.x0 + self.width, self.x0 + self.width, self.x0, self.x0]
|
43 |
+
ys = [self.y0, self.y0, self.y0 + self.height, self.y0 + self.height, self.y0]
|
44 |
+
return xs, ys
|
45 |
+
|
46 |
+
|
47 |
+
@dataclasses.dataclass
|
48 |
+
class DataPoint:
|
49 |
+
x: float or str
|
50 |
+
y: float or str
|
51 |
+
|
52 |
+
|
53 |
+
class TextRole(enum.Enum):
|
54 |
+
axis_title = "axis_title"
|
55 |
+
chart_title = "chart_title"
|
56 |
+
legend_label = "legend_label"
|
57 |
+
tick_grouping = "tick_grouping"
|
58 |
+
tick_label = "tick_label"
|
59 |
+
other = "other"
|
60 |
+
|
61 |
+
|
62 |
+
@dataclasses.dataclass
|
63 |
+
class Polygon:
|
64 |
+
x0: int
|
65 |
+
x1: int
|
66 |
+
x2: int
|
67 |
+
x3: int
|
68 |
+
y0: int
|
69 |
+
y1: int
|
70 |
+
y2: int
|
71 |
+
y3: int
|
72 |
+
|
73 |
+
def get_bounds(self):
|
74 |
+
xs = [
|
75 |
+
self.x0,
|
76 |
+
self.x1,
|
77 |
+
self.x2,
|
78 |
+
self.x3,
|
79 |
+
self.x0,
|
80 |
+
]
|
81 |
+
ys = [
|
82 |
+
self.y0,
|
83 |
+
self.y1,
|
84 |
+
self.y2,
|
85 |
+
self.y3,
|
86 |
+
self.y0,
|
87 |
+
]
|
88 |
+
return xs, ys
|
89 |
+
|
90 |
+
|
91 |
+
@dataclasses.dataclass
|
92 |
+
class Text:
|
93 |
+
id: int
|
94 |
+
polygon: Polygon
|
95 |
+
role: TextRole
|
96 |
+
text: str
|
97 |
+
|
98 |
+
def __post_init__(self):
|
99 |
+
self.polygon = Polygon(**self.polygon)
|
100 |
+
self.role = TextRole(self.role)
|
101 |
+
|
102 |
+
|
103 |
+
class ValuesType(enum.Enum):
|
104 |
+
categorical = "categorical"
|
105 |
+
numerical = "numerical"
|
106 |
+
|
107 |
+
|
108 |
+
@dataclasses.dataclass
|
109 |
+
class Tick:
|
110 |
+
id: int
|
111 |
+
x: int
|
112 |
+
y: int
|
113 |
+
|
114 |
+
|
115 |
+
class TickType(enum.Enum):
|
116 |
+
markers = "markers"
|
117 |
+
separators = "separators"
|
118 |
+
|
119 |
+
|
120 |
+
@dataclasses.dataclass
|
121 |
+
class Axis:
|
122 |
+
values_type: ValuesType
|
123 |
+
tick_type: TickType
|
124 |
+
ticks: list[Tick]
|
125 |
+
|
126 |
+
def __post_init__(self):
|
127 |
+
self.values_type = ValuesType(self.values_type)
|
128 |
+
self.tick_type = TickType(self.tick_type)
|
129 |
+
self.ticks = [
|
130 |
+
Tick(id=kw["id"], x=kw["tick_pt"]["x"], y=kw["tick_pt"]["y"])
|
131 |
+
for kw in self.ticks
|
132 |
+
]
|
133 |
+
|
134 |
+
def get_bounds(self):
|
135 |
+
min_x = min(tick.x for tick in self.ticks)
|
136 |
+
max_x = max(tick.x for tick in self.ticks)
|
137 |
+
min_y = min(tick.y for tick in self.ticks)
|
138 |
+
max_y = max(tick.y for tick in self.ticks)
|
139 |
+
xs = [min_x, max_x, max_x, min_x, min_x]
|
140 |
+
ys = [min_y, min_y, max_y, max_y, min_y]
|
141 |
+
return xs, ys
|
142 |
+
|
143 |
+
|
144 |
+
def convert_dashes_to_underscores_in_key_names(dictionary):
|
145 |
+
return {k.replace("-", "_"): v for k, v in dictionary.items()}
|
146 |
+
|
147 |
+
|
148 |
+
@dataclasses.dataclass
|
149 |
+
class Axes:
|
150 |
+
x_axis: Axis
|
151 |
+
y_axis: Axis
|
152 |
+
|
153 |
+
def __post_init__(self):
|
154 |
+
self.x_axis = Axis(**convert_dashes_to_underscores_in_key_names(self.x_axis))
|
155 |
+
self.y_axis = Axis(**convert_dashes_to_underscores_in_key_names(self.y_axis))
|
156 |
+
|
157 |
+
|
158 |
+
def preprocess_numerical_value(value):
|
159 |
+
value = float(value)
|
160 |
+
value = 0 if np.isnan(value) else value
|
161 |
+
return value
|
162 |
+
|
163 |
+
|
164 |
+
def preprocess_value(value, value_type: ValuesType):
|
165 |
+
if value_type == ValuesType.numerical:
|
166 |
+
return preprocess_numerical_value(value)
|
167 |
+
else:
|
168 |
+
return str(value)
|
169 |
+
|
170 |
+
|
171 |
+
@dataclasses.dataclass
|
172 |
+
class Annotation:
|
173 |
+
source: Source
|
174 |
+
chart_type: ChartType
|
175 |
+
plot_bb: PlotBoundingBox
|
176 |
+
text: list[Text]
|
177 |
+
axes: Axes
|
178 |
+
data_series: list[DataPoint]
|
179 |
+
|
180 |
+
def __post_init__(self):
|
181 |
+
self.source = Source(self.source)
|
182 |
+
self.chart_type = ChartType(self.chart_type)
|
183 |
+
self.plot_bb = PlotBoundingBox(**self.plot_bb)
|
184 |
+
self.text = [Text(**kw) for kw in self.text]
|
185 |
+
self.axes = Axes(**convert_dashes_to_underscores_in_key_names(self.axes))
|
186 |
+
self.data_series = [DataPoint(**kw) for kw in self.data_series]
|
187 |
+
|
188 |
+
for i in range(len(self.data_series)):
|
189 |
+
self.data_series[i].x = preprocess_value(
|
190 |
+
self.data_series[i].x, self.axes.x_axis.values_type
|
191 |
+
)
|
192 |
+
self.data_series[i].y = preprocess_value(
|
193 |
+
self.data_series[i].y, self.axes.y_axis.values_type
|
194 |
+
)
|
195 |
+
|
196 |
+
@staticmethod
|
197 |
+
def from_dict_with_dashes(kwargs):
|
198 |
+
return Annotation(**convert_dashes_to_underscores_in_key_names(kwargs))
|
199 |
+
|
200 |
+
@staticmethod
|
201 |
+
def from_image_index(image_index: int):
|
202 |
+
image_id = load_train_image_ids()[image_index]
|
203 |
+
return Annotation.from_dict_with_dashes(load_image_annotation(image_id))
|
204 |
+
|
205 |
+
def get_text_by_role(self, text_role: TextRole) -> list[Text]:
|
206 |
+
return [t for t in self.text if t.role == text_role]
|
207 |
+
|
208 |
+
|
209 |
+
@dataclasses.dataclass
|
210 |
+
class AnnotatedImage:
|
211 |
+
id: str
|
212 |
+
image: np.ndarray
|
213 |
+
annotation: Annotation
|
214 |
+
|
215 |
+
@staticmethod
|
216 |
+
def from_image_id(image_id: str):
|
217 |
+
return AnnotatedImage(
|
218 |
+
id=image_id,
|
219 |
+
image=load_image(image_id),
|
220 |
+
annotation=Annotation.from_dict_with_dashes(
|
221 |
+
load_image_annotation(image_id)
|
222 |
+
),
|
223 |
+
)
|
224 |
+
|
225 |
+
@staticmethod
|
226 |
+
def from_image_index(image_index: int):
|
227 |
+
return AnnotatedImage.from_image_id(load_train_image_ids()[image_index])
|
228 |
+
|
229 |
+
|
230 |
+
def generate_annotated_images():
|
231 |
+
for image_id in tqdm.autonotebook.tqdm(
|
232 |
+
load_train_image_ids(), "Iterating over annotated images"
|
233 |
+
):
|
234 |
+
yield AnnotatedImage.from_image_id(image_id)
|
235 |
+
|
236 |
+
|
237 |
+
@functools.cache
|
238 |
+
def load_train_image_ids() -> list[str]:
|
239 |
+
train_image_ids = [i.replace(".jpg", "") for i in os.listdir("data/train/images")]
|
240 |
+
return train_image_ids[: 1000 if CONFIG.debug else None]
|
241 |
+
|
242 |
+
|
243 |
+
@functools.cache
|
244 |
+
def load_test_image_ids() -> list[str]:
|
245 |
+
return [i.replace(".jpg", "") for i in os.listdir("data/test/images")]
|
246 |
+
|
247 |
+
|
248 |
+
@functools.cache
|
249 |
+
def load_image_annotation(image_id: str) -> dict:
|
250 |
+
return json.load(open(f"data/train/annotations/{image_id}.json"))
|
251 |
+
|
252 |
+
|
253 |
+
def load_image(image_id: str) -> np.ndarray:
|
254 |
+
return imageio.v3.imread(open(f"data/train/images/{image_id}.jpg", "rb"))
|
255 |
+
|
256 |
+
|
257 |
+
@dataclasses.dataclass
|
258 |
+
class DataItem:
|
259 |
+
image: torch.FloatTensor
|
260 |
+
target_string: str
|
261 |
+
data_index: int
|
262 |
+
|
263 |
+
def __post_init__(self):
|
264 |
+
shape = einops.parse_shape(self.image, "channel height width")
|
265 |
+
assert shape["channel"] == 3, "Image is expected to have 3 channels."
|
266 |
+
|
267 |
+
|
268 |
+
def split_train_indices_by_source():
|
269 |
+
extracted_image_indices = []
|
270 |
+
generated_image_indices = []
|
271 |
+
for i, annotated_image in enumerate(generate_annotated_images()):
|
272 |
+
if annotated_image.annotation.source == Source.extracted:
|
273 |
+
extracted_image_indices.append(i)
|
274 |
+
else:
|
275 |
+
generated_image_indices.append(i)
|
276 |
+
return extracted_image_indices, generated_image_indices
|
277 |
+
|
278 |
+
|
279 |
+
def get_train_val_split_indices(val_fraction=0.1, seed=42):
|
280 |
+
np.random.seed(seed)
|
281 |
+
val_size = int(len(load_train_image_ids()) * val_fraction)
|
282 |
+
|
283 |
+
extracted_image_indices, generated_image_indices = split_train_indices_by_source()
|
284 |
+
extracted_image_indices = np.random.permutation(extracted_image_indices)
|
285 |
+
generated_image_indices = np.random.permutation(generated_image_indices)
|
286 |
+
|
287 |
+
val_indices = extracted_image_indices[:val_size]
|
288 |
+
n_generated_images_in_val = val_size - len(val_indices)
|
289 |
+
val_indices = np.concatenate(
|
290 |
+
[val_indices, generated_image_indices[:n_generated_images_in_val]]
|
291 |
+
)
|
292 |
+
|
293 |
+
train_indices = generated_image_indices[n_generated_images_in_val:]
|
294 |
+
|
295 |
+
assert len(set(train_indices) | set(val_indices)) == len(load_train_image_ids())
|
296 |
+
assert len(val_indices) == val_size
|
297 |
+
assert len(set(train_indices) & set(val_indices)) == 0
|
298 |
+
|
299 |
+
return train_indices, val_indices
|
300 |
+
|
301 |
+
|
302 |
+
def to_token_str(value: str or enum.Enum):
|
303 |
+
string = value.name if isinstance(value, enum.Enum) else value
|
304 |
+
if re.fullmatch("<.*>", string):
|
305 |
+
return string
|
306 |
+
else:
|
307 |
+
return f"<{string}>"
|
308 |
+
|
309 |
+
|
310 |
+
@functools.cache
|
311 |
+
def get_extra_tokens() -> types.SimpleNamespace:
|
312 |
+
token_ns = types.SimpleNamespace()
|
313 |
+
|
314 |
+
token_ns.benetech_prompt = to_token_str("benetech_prompt")
|
315 |
+
token_ns.benetech_prompt_end = to_token_str("/benetech_prompt")
|
316 |
+
token_ns.x_start = to_token_str("x_start")
|
317 |
+
token_ns.y_start = to_token_str("y_start")
|
318 |
+
token_ns.value_separator = to_token_str(";")
|
319 |
+
|
320 |
+
for chart_type in ChartType:
|
321 |
+
setattr(token_ns, chart_type.name, to_token_str(chart_type))
|
322 |
+
|
323 |
+
for values_type in ValuesType:
|
324 |
+
setattr(token_ns, values_type.name, to_token_str(values_type))
|
325 |
+
|
326 |
+
return token_ns
|
327 |
+
|
328 |
+
|
329 |
+
def convert_number_to_scientific_string(value: int or float) -> str:
|
330 |
+
return f"{value:.{CONFIG.float_scientific_notation_string_precision}e}"
|
331 |
+
|
332 |
+
|
333 |
+
def convert_axis_data_to_string(
|
334 |
+
axis_data: list[str or float], values_type: ValuesType
|
335 |
+
) -> str:
|
336 |
+
formatted_axis_data = []
|
337 |
+
for value in axis_data:
|
338 |
+
if values_type == ValuesType.numerical:
|
339 |
+
value = convert_number_to_scientific_string(value)
|
340 |
+
formatted_axis_data.append(value)
|
341 |
+
return get_extra_tokens().value_separator.join(formatted_axis_data)
|
342 |
+
|
343 |
+
|
344 |
+
def convert_string_to_axis_data(string, values_type: ValuesType):
|
345 |
+
data = string.split(get_extra_tokens().value_separator)
|
346 |
+
if values_type == ValuesType.numerical:
|
347 |
+
data = [float(i.replace(" ", "")) for i in data]
|
348 |
+
return data
|
349 |
+
|
350 |
+
|
351 |
+
@dataclasses.dataclass
|
352 |
+
class BenetechOutput:
|
353 |
+
chart_type: ChartType
|
354 |
+
x_values_type: ValuesType
|
355 |
+
y_values_type: ValuesType
|
356 |
+
x_data: list[str or float]
|
357 |
+
y_data: list[str or float]
|
358 |
+
|
359 |
+
def __post_init__(self):
|
360 |
+
self.chart_type = ChartType(self.chart_type)
|
361 |
+
self.x_values_type = ValuesType(self.x_values_type)
|
362 |
+
self.y_values_type = ValuesType(self.y_values_type)
|
363 |
+
assert isinstance(self.x_data, list)
|
364 |
+
assert isinstance(self.y_data, list)
|
365 |
+
|
366 |
+
def get_main_characteristics(self):
|
367 |
+
return (
|
368 |
+
self.chart_type,
|
369 |
+
self.x_values_type,
|
370 |
+
self.y_values_type,
|
371 |
+
len(self.x_data),
|
372 |
+
len(self.y_data),
|
373 |
+
)
|
374 |
+
|
375 |
+
@staticmethod
|
376 |
+
def from_annotation(annotation: Annotation):
|
377 |
+
return BenetechOutput(
|
378 |
+
chart_type=annotation.chart_type,
|
379 |
+
x_values_type=annotation.axes.x_axis.values_type,
|
380 |
+
y_values_type=annotation.axes.y_axis.values_type,
|
381 |
+
x_data=[dp.x for dp in annotation.data_series],
|
382 |
+
y_data=[dp.y for dp in annotation.data_series],
|
383 |
+
)
|
384 |
+
|
385 |
+
def to_string(self):
|
386 |
+
return self.format_strings(
|
387 |
+
chart_type=self.chart_type,
|
388 |
+
x_values_type=self.x_values_type,
|
389 |
+
y_values_type=self.y_values_type,
|
390 |
+
x_data=convert_axis_data_to_string(self.x_data, self.x_values_type),
|
391 |
+
y_data=convert_axis_data_to_string(self.y_data, self.y_values_type),
|
392 |
+
)
|
393 |
+
|
394 |
+
@staticmethod
|
395 |
+
def format_strings(*, chart_type, x_values_type, y_values_type, x_data, y_data):
|
396 |
+
chart_type = to_token_str(chart_type)
|
397 |
+
x_values_type = to_token_str(x_values_type)
|
398 |
+
y_values_type = to_token_str(y_values_type)
|
399 |
+
token_ns = get_extra_tokens()
|
400 |
+
return (
|
401 |
+
f"{token_ns.benetech_prompt}{chart_type}"
|
402 |
+
f"{token_ns.x_start}{x_values_type}{x_data}"
|
403 |
+
f"{token_ns.y_start}{y_values_type}{y_data}"
|
404 |
+
f"{token_ns.benetech_prompt_end}"
|
405 |
+
)
|
406 |
+
|
407 |
+
@staticmethod
|
408 |
+
def get_string_pattern():
|
409 |
+
field_names = [field.name for field in dataclasses.fields(BenetechOutput)]
|
410 |
+
pattern = BenetechOutput.format_strings(
|
411 |
+
**{field_name: f"(?P<{field_name}>.*?)" for field_name in field_names}
|
412 |
+
)
|
413 |
+
return pattern
|
414 |
+
|
415 |
+
@staticmethod
|
416 |
+
def does_string_match_expected_pattern(string):
|
417 |
+
try:
|
418 |
+
BenetechOutput.from_string(string)
|
419 |
+
return True
|
420 |
+
except:
|
421 |
+
return False
|
422 |
+
|
423 |
+
@staticmethod
|
424 |
+
def from_string(string):
|
425 |
+
fullmatch = re.fullmatch(BenetechOutput.get_string_pattern(), string)
|
426 |
+
benetech_kwargs = fullmatch.groupdict()
|
427 |
+
benetech_kwargs["chart_type"] = ChartType(benetech_kwargs["chart_type"])
|
428 |
+
benetech_kwargs["x_values_type"] = ValuesType(benetech_kwargs["x_values_type"])
|
429 |
+
benetech_kwargs["y_values_type"] = ValuesType(benetech_kwargs["y_values_type"])
|
430 |
+
benetech_kwargs["x_data"] = convert_string_to_axis_data(
|
431 |
+
benetech_kwargs["x_data"], benetech_kwargs["x_values_type"]
|
432 |
+
)
|
433 |
+
benetech_kwargs["y_data"] = convert_string_to_axis_data(
|
434 |
+
benetech_kwargs["y_data"], benetech_kwargs["y_values_type"]
|
435 |
+
)
|
436 |
+
return BenetechOutput(**benetech_kwargs)
|
437 |
+
|
438 |
+
|
439 |
+
def get_annotation_ground_truth_str(annotation: Annotation):
|
440 |
+
benetech_output = BenetechOutput(
|
441 |
+
chart_type=annotation.chart_type,
|
442 |
+
x_values_type=annotation.axes.x_axis.values_type,
|
443 |
+
x_data=[dp.x for dp in annotation.data_series],
|
444 |
+
y_values_type=annotation.axes.y_axis.values_type,
|
445 |
+
y_data=[dp.y for dp in annotation.data_series],
|
446 |
+
)
|
447 |
+
return benetech_output.to_string()
|
448 |
+
|
449 |
+
|
450 |
+
def get_annotation_ground_truth_str_from_image_index(image_index: int) -> str:
|
451 |
+
return get_annotation_ground_truth_str(Annotation.from_image_index(image_index))
|
452 |
+
|
453 |
+
|
454 |
+
class Dataset(torch.utils.data.Dataset):
|
455 |
+
def __init__(self, indices: list[int]):
|
456 |
+
super().__init__()
|
457 |
+
self.indices = indices
|
458 |
+
self.to_tensor = torchvision.transforms.ToTensor()
|
459 |
+
|
460 |
+
def __len__(self):
|
461 |
+
return len(self.indices)
|
462 |
+
|
463 |
+
def __getitem__(self, idx: int) -> DataItem:
|
464 |
+
data_index = self.indices[idx]
|
465 |
+
|
466 |
+
annotated_image = AnnotatedImage.from_image_index(data_index)
|
467 |
+
|
468 |
+
image = annotated_image.image
|
469 |
+
image = self.to_tensor(image)
|
470 |
+
|
471 |
+
target_string = get_annotation_ground_truth_str(annotated_image.annotation)
|
472 |
+
|
473 |
+
return DataItem(image=image, target_string=target_string, data_index=data_index)
|
474 |
+
|
475 |
+
|
476 |
+
def get_train_val_datasets():
|
477 |
+
train_indices, val_indices = load_pickle_or_build_object_and_save(
|
478 |
+
CONFIG.train_val_indices_path,
|
479 |
+
lambda: get_train_val_split_indices(CONFIG.val_fraction, CONFIG.seed),
|
480 |
+
)
|
481 |
+
return Dataset(train_indices), Dataset(val_indices)
|
482 |
+
|
483 |
+
|
484 |
+
def get_train_dataset():
|
485 |
+
return get_train_val_datasets()[0]
|
486 |
+
|
487 |
+
|
488 |
+
def get_val_dataset():
|
489 |
+
return get_train_val_datasets()[1]
|
490 |
+
|
491 |
+
|
492 |
+
@dataclasses.dataclass
|
493 |
+
class Batch:
|
494 |
+
images: torch.FloatTensor
|
495 |
+
labels: torch.IntTensor
|
496 |
+
data_indices: list[int]
|
497 |
+
|
498 |
+
def __post_init__(self):
|
499 |
+
if CONFIG.debug:
|
500 |
+
images_shape = einops.parse_shape(self.images, "batch channel height width")
|
501 |
+
labels_shape = einops.parse_shape(self.labels, "batch label")
|
502 |
+
assert images_shape["batch"] == labels_shape["batch"]
|
503 |
+
assert len(self.data_indices) == images_shape["batch"]
|
504 |
+
|
505 |
+
|
506 |
+
class Split(enum.Enum):
|
507 |
+
train = "train"
|
508 |
+
val = "val"
|
509 |
+
|
510 |
+
|
511 |
+
BatchCollateFunction = Callable[[list[DataItem], Split], Batch]
|
512 |
+
|
513 |
+
|
514 |
+
def build_dataloader(split: Split, batch_collate_function: BatchCollateFunction):
|
515 |
+
return torch.utils.data.DataLoader(
|
516 |
+
get_train_dataset() if split == Split.train else get_val_dataset(),
|
517 |
+
batch_size=CONFIG.batch_size,
|
518 |
+
shuffle=split == Split.train,
|
519 |
+
num_workers=CONFIG.num_workers,
|
520 |
+
collate_fn=functools.partial(batch_collate_function, split=split),
|
521 |
+
)
|
metrics.py
ADDED
@@ -0,0 +1,54 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import numpy as np
|
2 |
+
import rapidfuzz
|
3 |
+
import sklearn
|
4 |
+
|
5 |
+
from data import ValuesType, BenetechOutput, Annotation
|
6 |
+
|
7 |
+
|
8 |
+
def normalized_rmse(expected: list[float], predicted: list[float]) -> float:
|
9 |
+
return (1 - sklearn.metrics.r2_score(expected, predicted)) ** 0.5
|
10 |
+
|
11 |
+
|
12 |
+
def normalized_levenshtein_distance(expected: list[str], predicted: list[str]) -> float:
|
13 |
+
total_distance = 0
|
14 |
+
for e, p in zip(expected, predicted):
|
15 |
+
total_distance += rapidfuzz.distance.Levenshtein.distance(e, p)
|
16 |
+
total_length = np.sum([len(e) for e in expected])
|
17 |
+
return total_distance / total_length
|
18 |
+
|
19 |
+
|
20 |
+
def sigmoid(x):
|
21 |
+
return 1 / (1 + np.exp(-x))
|
22 |
+
|
23 |
+
|
24 |
+
def positive_loss_to_score(x):
|
25 |
+
return 2 * sigmoid(-x)
|
26 |
+
|
27 |
+
|
28 |
+
def score_axis_values(values_type, expected, predicted):
|
29 |
+
if values_type == ValuesType.numerical:
|
30 |
+
loss = normalized_rmse(expected, predicted)
|
31 |
+
else:
|
32 |
+
loss = normalized_levenshtein_distance(expected, predicted)
|
33 |
+
return positive_loss_to_score(loss)
|
34 |
+
|
35 |
+
|
36 |
+
def benetech_score(expected: BenetechOutput, predicted: BenetechOutput) -> float:
|
37 |
+
if expected.get_main_characteristics() != predicted.get_main_characteristics():
|
38 |
+
return 0
|
39 |
+
x_score = score_axis_values(
|
40 |
+
expected.x_values_type, expected.x_data, predicted.x_data
|
41 |
+
)
|
42 |
+
y_score = score_axis_values(
|
43 |
+
expected.y_values_type, expected.y_data, predicted.y_data
|
44 |
+
)
|
45 |
+
return (x_score + y_score) / 2
|
46 |
+
|
47 |
+
|
48 |
+
def benetech_score_string_prediction(expected_data_index: int, predicted_string: str):
|
49 |
+
if not BenetechOutput.does_string_match_expected_pattern(predicted_string):
|
50 |
+
return 0
|
51 |
+
expected_annotation = Annotation.from_image_index(expected_data_index)
|
52 |
+
expected_output = BenetechOutput.from_annotation(expected_annotation)
|
53 |
+
predicted_output = BenetechOutput.from_string(predicted_string)
|
54 |
+
return benetech_score(expected_output, predicted_output)
|
model.py
ADDED
@@ -0,0 +1,193 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import collections
|
2 |
+
import dataclasses
|
3 |
+
import types
|
4 |
+
|
5 |
+
import pytorch_lightning as pl
|
6 |
+
import torch.utils.data
|
7 |
+
import transformers
|
8 |
+
|
9 |
+
from data import (
|
10 |
+
generate_annotated_images,
|
11 |
+
get_annotation_ground_truth_str,
|
12 |
+
DataItem,
|
13 |
+
get_extra_tokens,
|
14 |
+
Batch,
|
15 |
+
Split,
|
16 |
+
BatchCollateFunction,
|
17 |
+
)
|
18 |
+
from utils import load_pickle_or_build_object_and_save
|
19 |
+
|
20 |
+
|
21 |
+
@dataclasses.dataclass
|
22 |
+
class Model:
|
23 |
+
processor: transformers.models.donut.processing_donut.DonutProcessor
|
24 |
+
tokenizer: transformers.models.xlm_roberta.tokenization_xlm_roberta_fast.XLMRobertaTokenizerFast
|
25 |
+
encoder_decoder: transformers.models.vision_encoder_decoder.modeling_vision_encoder_decoder.VisionEncoderDecoderModel
|
26 |
+
batch_collate_function: BatchCollateFunction
|
27 |
+
config: types.SimpleNamespace
|
28 |
+
|
29 |
+
|
30 |
+
def add_unknown_tokens_to_tokenizer(
|
31 |
+
tokenizer, encoder_decoder, unknown_tokens: list[str]
|
32 |
+
):
|
33 |
+
tokenizer.add_tokens(unknown_tokens)
|
34 |
+
encoder_decoder.decoder.resize_token_embeddings(len(tokenizer))
|
35 |
+
|
36 |
+
|
37 |
+
def find_unknown_tokens_for_tokenizer(tokenizer) -> collections.Counter:
|
38 |
+
unknown_tokens_counter = collections.Counter()
|
39 |
+
|
40 |
+
for annotated_image in generate_annotated_images():
|
41 |
+
ground_truth = get_annotation_ground_truth_str(annotated_image.annotation)
|
42 |
+
|
43 |
+
input_ids = tokenizer(ground_truth).input_ids
|
44 |
+
tokens = tokenizer.tokenize(ground_truth, add_special_tokens=True)
|
45 |
+
|
46 |
+
for token_id, token in zip(input_ids, tokens, strict=True):
|
47 |
+
if token_id == tokenizer.unk_token_id:
|
48 |
+
unknown_tokens_counter.update([token])
|
49 |
+
|
50 |
+
return unknown_tokens_counter
|
51 |
+
|
52 |
+
|
53 |
+
def replace_pad_token_id_with_negative_hundred_for_hf_transformers_automatic_batch_transformation(
|
54 |
+
tokenizer, token_ids
|
55 |
+
):
|
56 |
+
token_ids[token_ids == tokenizer.pad_token_id] = -100
|
57 |
+
return token_ids
|
58 |
+
|
59 |
+
|
60 |
+
@dataclasses.dataclass
|
61 |
+
class BatchCollateFunction:
|
62 |
+
processor: transformers.models.donut.processing_donut.DonutProcessor
|
63 |
+
tokenizer: transformers.models.xlm_roberta.tokenization_xlm_roberta_fast.XLMRobertaTokenizerFast
|
64 |
+
decoder_sequence_max_length: int
|
65 |
+
|
66 |
+
def __call__(self, batch: list[DataItem], split: Split) -> Batch:
|
67 |
+
images = [di.image for di in batch]
|
68 |
+
images = self.processor(
|
69 |
+
images, random_padding=split == Split.train, return_tensors="pt"
|
70 |
+
).pixel_values
|
71 |
+
|
72 |
+
target_token_ids = self.tokenizer(
|
73 |
+
[di.target_string for di in batch],
|
74 |
+
add_special_tokens=False,
|
75 |
+
max_length=self.decoder_sequence_max_length,
|
76 |
+
padding="max_length",
|
77 |
+
truncation=True,
|
78 |
+
return_tensors="pt",
|
79 |
+
).input_ids
|
80 |
+
labels = replace_pad_token_id_with_negative_hundred_for_hf_transformers_automatic_batch_transformation(
|
81 |
+
self.tokenizer, target_token_ids
|
82 |
+
)
|
83 |
+
|
84 |
+
data_indices = [di.data_index for di in batch]
|
85 |
+
|
86 |
+
return Batch(images=images, labels=labels, data_indices=data_indices)
|
87 |
+
|
88 |
+
|
89 |
+
def build_model(config: types.SimpleNamespace or object) -> Model:
|
90 |
+
donut_processor = transformers.DonutProcessor.from_pretrained(
|
91 |
+
config.pretrained_model_name
|
92 |
+
)
|
93 |
+
donut_processor.image_processor.size = dict(
|
94 |
+
width=config.image_width, height=config.image_height
|
95 |
+
)
|
96 |
+
donut_processor.image_processor.do_align_long_axis = False
|
97 |
+
|
98 |
+
tokenizer = donut_processor.tokenizer
|
99 |
+
|
100 |
+
encoder_decoder_config = transformers.VisionEncoderDecoderConfig.from_pretrained(
|
101 |
+
config.pretrained_model_name
|
102 |
+
)
|
103 |
+
encoder_decoder_config.encoder.image_size = (
|
104 |
+
config.image_width,
|
105 |
+
config.image_height,
|
106 |
+
)
|
107 |
+
|
108 |
+
encoder_decoder = transformers.VisionEncoderDecoderModel.from_pretrained(
|
109 |
+
config.pretrained_model_name, config=encoder_decoder_config
|
110 |
+
)
|
111 |
+
encoder_decoder_config.pad_token_id = tokenizer.pad_token_id
|
112 |
+
encoder_decoder_config.decoder_start_token_id = tokenizer.convert_tokens_to_ids(
|
113 |
+
get_extra_tokens().benetech_prompt
|
114 |
+
)
|
115 |
+
encoder_decoder_config.bos_token_id = encoder_decoder_config.decoder_start_token_id
|
116 |
+
encoder_decoder_config.eos_token_id = tokenizer.convert_tokens_to_ids(
|
117 |
+
get_extra_tokens().benetech_prompt_end
|
118 |
+
)
|
119 |
+
|
120 |
+
extra_tokens = list(get_extra_tokens().__dict__.values())
|
121 |
+
add_unknown_tokens_to_tokenizer(tokenizer, encoder_decoder, extra_tokens)
|
122 |
+
unknown_dataset_tokens = load_pickle_or_build_object_and_save(
|
123 |
+
config.unknown_tokens_for_tokenizer_path,
|
124 |
+
lambda: list(find_unknown_tokens_for_tokenizer(tokenizer).keys()),
|
125 |
+
)
|
126 |
+
add_unknown_tokens_to_tokenizer(tokenizer, encoder_decoder, unknown_dataset_tokens)
|
127 |
+
tokenizer.eos_token_id = encoder_decoder_config.eos_token_id
|
128 |
+
|
129 |
+
batch_collate_function = BatchCollateFunction(
|
130 |
+
processor=donut_processor,
|
131 |
+
tokenizer=tokenizer,
|
132 |
+
decoder_sequence_max_length=config.decoder_sequence_max_length,
|
133 |
+
)
|
134 |
+
|
135 |
+
return Model(
|
136 |
+
processor=donut_processor,
|
137 |
+
tokenizer=tokenizer,
|
138 |
+
encoder_decoder=encoder_decoder,
|
139 |
+
batch_collate_function=batch_collate_function,
|
140 |
+
config=config,
|
141 |
+
)
|
142 |
+
|
143 |
+
|
144 |
+
def generate_token_strings(
|
145 |
+
model: Model, images: torch.Tensor, skip_special_tokens=True
|
146 |
+
) -> list[str]:
|
147 |
+
decoder_output = model.encoder_decoder.generate(
|
148 |
+
images,
|
149 |
+
max_length=10
|
150 |
+
if model.config.debug
|
151 |
+
else model.config.decoder_sequence_max_length,
|
152 |
+
eos_token_id=model.tokenizer.eos_token_id,
|
153 |
+
return_dict_in_generate=True,
|
154 |
+
)
|
155 |
+
return model.tokenizer.batch_decode(
|
156 |
+
decoder_output.sequences, skip_special_tokens=skip_special_tokens
|
157 |
+
)
|
158 |
+
|
159 |
+
|
160 |
+
def predict_string(image, model: Model):
|
161 |
+
image = model.processor(
|
162 |
+
image, random_padding=False, return_tensors="pt"
|
163 |
+
).pixel_values
|
164 |
+
string = generate_token_strings(model, image)[0]
|
165 |
+
return string
|
166 |
+
|
167 |
+
|
168 |
+
class LightningModule(pl.LightningModule):
|
169 |
+
def __init__(self, config):
|
170 |
+
super().__init__()
|
171 |
+
self.save_hyperparameters()
|
172 |
+
self.model = build_model(config)
|
173 |
+
self.encoder_decoder = self.model.encoder_decoder
|
174 |
+
|
175 |
+
def training_step(self, batch: Batch, batch_idx: int) -> torch.Tensor:
|
176 |
+
loss = self.compute_loss(batch)
|
177 |
+
self.log("train_loss", loss)
|
178 |
+
return loss
|
179 |
+
|
180 |
+
def validation_step(self, batch: Batch, batch_idx: int, dataset_idx: int = 0):
|
181 |
+
loss = self.compute_loss(batch)
|
182 |
+
self.log("val_loss", loss)
|
183 |
+
|
184 |
+
def compute_loss(self, batch: Batch) -> torch.Tensor:
|
185 |
+
outputs = self.encoder_decoder(pixel_values=batch.images, labels=batch.labels)
|
186 |
+
loss = outputs.loss
|
187 |
+
return loss
|
188 |
+
|
189 |
+
def configure_optimizers(self) -> torch.optim.Optimizer:
|
190 |
+
optimizer = torch.optim.Adam(
|
191 |
+
self.parameters(), lr=self.hparams["config"].learning_rate
|
192 |
+
)
|
193 |
+
return optimizer
|
requirements.txt
DELETED
@@ -1,3 +0,0 @@
|
|
1 |
-
gradio==3.27.0
|
2 |
-
torch==2.0.0
|
3 |
-
transformers==4.26.1
|
|
|
|
|
|
|
|
train.py
ADDED
@@ -0,0 +1,114 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
|
3 |
+
import pandas as pd
|
4 |
+
import pytorch_lightning as pl
|
5 |
+
import transformers
|
6 |
+
import wandb
|
7 |
+
|
8 |
+
from config import CONFIG
|
9 |
+
from data import (
|
10 |
+
get_annotation_ground_truth_str_from_image_index,
|
11 |
+
load_train_image_ids,
|
12 |
+
build_dataloader,
|
13 |
+
Split,
|
14 |
+
Batch,
|
15 |
+
)
|
16 |
+
from metrics import benetech_score_string_prediction
|
17 |
+
from model import generate_token_strings, LightningModule
|
18 |
+
from utils import set_tokenizers_parallelism, set_torch_device_order_pci_bus
|
19 |
+
|
20 |
+
|
21 |
+
class MetricsCallback(pl.callbacks.Callback):
|
22 |
+
def on_validation_batch_start(
|
23 |
+
self, trainer, pl_module, batch: Batch, batch_idx, dataloader_idx=0
|
24 |
+
):
|
25 |
+
predicted_strings = generate_token_strings(pl_module.model, images=batch.images)
|
26 |
+
|
27 |
+
for expected_data_index, predicted_string in zip(
|
28 |
+
batch.data_indices, predicted_strings, strict=True
|
29 |
+
):
|
30 |
+
benetech_score = benetech_score_string_prediction(
|
31 |
+
expected_data_index=expected_data_index,
|
32 |
+
predicted_string=predicted_string,
|
33 |
+
)
|
34 |
+
wandb.log(dict(benetech_score=benetech_score))
|
35 |
+
|
36 |
+
ground_truth_strings = [
|
37 |
+
get_annotation_ground_truth_str_from_image_index(i)
|
38 |
+
for i in batch.data_indices
|
39 |
+
]
|
40 |
+
string_ids = [load_train_image_ids()[i] for i in batch.data_indices]
|
41 |
+
strings_dataframe = pd.DataFrame(
|
42 |
+
dict(
|
43 |
+
string_ids=string_ids,
|
44 |
+
ground_truth=ground_truth_strings,
|
45 |
+
predicted=predicted_strings,
|
46 |
+
)
|
47 |
+
)
|
48 |
+
wandb.log(dict(strings=wandb.Table(dataframe=strings_dataframe)))
|
49 |
+
|
50 |
+
|
51 |
+
class TransformersPreTrainedModelsCheckpointIO(pl.plugins.CheckpointIO):
|
52 |
+
def __init__(
|
53 |
+
self, pretrained_models: list[transformers.modeling_utils.PreTrainedModel]
|
54 |
+
):
|
55 |
+
super().__init__()
|
56 |
+
self.pretrained_models = pretrained_models
|
57 |
+
|
58 |
+
def save_checkpoint(self, checkpoint, path, storage_options=None):
|
59 |
+
for pretrained_model in self.pretrained_models:
|
60 |
+
pretrained_model.save_pretrained(path)
|
61 |
+
|
62 |
+
def load_checkpoint(self, path, storage_options=None):
|
63 |
+
self.pretrained_models = [
|
64 |
+
pm.from_pretrained(path) for pm in self.pretrained_models
|
65 |
+
]
|
66 |
+
|
67 |
+
def remove_checkpoint(self, path):
|
68 |
+
os.remove(path)
|
69 |
+
|
70 |
+
|
71 |
+
def train():
|
72 |
+
set_tokenizers_parallelism(False)
|
73 |
+
set_torch_device_order_pci_bus()
|
74 |
+
|
75 |
+
pl_module = LightningModule(CONFIG)
|
76 |
+
|
77 |
+
model_checkpoint = pl.callbacks.ModelCheckpoint(
|
78 |
+
dirpath=CONFIG.training_directory,
|
79 |
+
monitor="val_loss",
|
80 |
+
save_top_k=CONFIG.save_top_k_checkpoints,
|
81 |
+
)
|
82 |
+
metrics_callback = MetricsCallback()
|
83 |
+
|
84 |
+
logger = pl.loggers.WandbLogger(
|
85 |
+
project=CONFIG.wandb_project_name, save_dir=CONFIG.training_directory
|
86 |
+
)
|
87 |
+
|
88 |
+
plugin = TransformersPreTrainedModelsCheckpointIO(
|
89 |
+
[pl_module.model.processor, pl_module.model.encoder_decoder]
|
90 |
+
)
|
91 |
+
|
92 |
+
trainer = pl.Trainer(
|
93 |
+
accelerator=CONFIG.accelerator,
|
94 |
+
devices=CONFIG.devices,
|
95 |
+
plugins=[plugin],
|
96 |
+
callbacks=[model_checkpoint, metrics_callback],
|
97 |
+
logger=logger,
|
98 |
+
limit_train_batches=CONFIG.limit_train_batches,
|
99 |
+
limit_val_batches=CONFIG.limit_val_batches,
|
100 |
+
)
|
101 |
+
|
102 |
+
trainer.fit(
|
103 |
+
model=pl_module,
|
104 |
+
train_dataloaders=build_dataloader(
|
105 |
+
Split.train, pl_module.model.batch_collate_function
|
106 |
+
),
|
107 |
+
val_dataloaders=build_dataloader(
|
108 |
+
Split.val, pl_module.model.batch_collate_function
|
109 |
+
),
|
110 |
+
)
|
111 |
+
|
112 |
+
|
113 |
+
if __name__ == "__main__":
|
114 |
+
train()
|
utils.py
ADDED
@@ -0,0 +1,23 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import pickle
|
3 |
+
from typing import Callable, TypeVar
|
4 |
+
|
5 |
+
T = TypeVar("T")
|
6 |
+
|
7 |
+
|
8 |
+
def set_tokenizers_parallelism(enable: bool):
|
9 |
+
os.environ["TOKENIZERS_PARALLELISM"] = "true" if enable else "false"
|
10 |
+
|
11 |
+
|
12 |
+
def set_torch_device_order_pci_bus():
|
13 |
+
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
|
14 |
+
|
15 |
+
|
16 |
+
def load_pickle_or_build_object_and_save(
|
17 |
+
pickle_path: str, build_object: Callable[[], T], overwrite=False
|
18 |
+
) -> T:
|
19 |
+
if overwrite or not os.path.exists(pickle_path):
|
20 |
+
pickle.dump(build_object(), open(pickle_path, "wb"))
|
21 |
+
else:
|
22 |
+
print(f"Reusing object {pickle_path}.")
|
23 |
+
return pickle.load(open(pickle_path, "rb"))
|