-
Notifications
You must be signed in to change notification settings - Fork 0
/
runt5.py
157 lines (135 loc) · 5.44 KB
/
runt5.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
# 3. Load wandb
# 1. Load data using templates
# 2. split in to train and test
# ----
# 4. train
# 5. upload model to wandb
""" Data Processing
{
"version": "v0",
"target": "prochoice_PCTS",
"description": "Prochoice dataset with parent child toxicity summary",
"wandb-project": "knoxcs/detoxify",
"base-dataset": "prochoice.summarized.json:v0",
"dataset": {
"path": "./artifacts/detoxify/prochoice.summarized.json:v0/prochoice.summarized.json",
"preprocess": "get_parent_child_toxic_summary"
},
"split": {
"train": 0.5,
"eval": 0.2
}
}
"""
import wandb, json
from templates.templates import process_data
from sklearn.model_selection import train_test_split
import pandas as pd
from models import t5sicon
def get_dataset(config):
ds = wandb.use_artifact(f'{config["wandb-project"]}/{config["base-dataset"]}:{config["base-version"]}', type="dataset")
ds.download()
return process_data(config)
def split_dataset(config, ds):
train_frac = config["split"]["train"]
eval_frac = config["split"]["eval"]
create_split(train_frac, eval_frac, ds)
def create_split(ds, train_frac, eval_frac):
if type(ds) == str:
ds = pd.read_json(ds)
train, test = train_test_split(ds, train_size=train_frac, random_state=42, shuffle=True)
eval_frac = eval_frac/(1-train_frac)
eval_df, test = train_test_split(test, train_size=eval_frac, random_state=42, shuffle=True)
name = "{}.json"
train.to_json("split/" + name.format("train"))
eval_df.to_json("split/" + name.format("eval"))
test.to_json("split/" + name.format("test"))
def load_split(config):
train = pd.read_json("train.json")
eval_df = pd.read_json("eval.json")
test = pd.read_json("test.json")
return train, eval_df, test
def build_t5_dataset(config):
print("running config:", config)
config = json.load(open(config))
entity, project = config["wandb-project"].split("/")
wandb_logger = wandb.init(project=project, entity=entity, config=config)
print(json.dumps(config))
ds = get_dataset(config)
for k, v in ds.items():
v.to_json(f'{k}.json')
#split_dataset(config, ds)
artifact = wandb.Artifact(config["target"], type="dataset", description=config["description"])
artifact.add_file("train.json")
artifact.add_file("test.json")
artifact.add_file("eval.json")
wandb_logger.log_artifact(artifact)
from pytorch_lightning.loggers import WandbLogger
# def train(train_df, eval_df, prototype="t5", base_model="t5-large", output_dir="outputs", logger="default"):
def train_t5(config):
config = json.load(open(config))
entity, project = config["wandb-project"].split("/")
experiment = wandb.init(name=config["name"], project=project, entity=entity, group="hyperion")
dataset = wandb.use_artifact(config["wandb-project"] + "/" + config["dataset"])
dataset = dataset.download()
train = pd.read_json(dataset+"/train.json")
eval_df = pd.read_json(dataset+"/eval.json")
wandb_logger = WandbLogger(name=config["name"], experiment=experiment, log_model=False, project=project, entity=entity, config=config, group="hyperion", tags=["model", config["prototype"], config["base_model"]])
print("starting training")
t5sicon.train(train,
eval_df,
prototype=config["prototype"],
base_model=config["base_model"],
logger=wandb_logger,
args=config.get("args", {}))
artifact = wandb.Artifact(config["name"], type="model")
artifact.add_dir(config["args"]["output_dir"]+"_model")
experiment.log_artifact(artifact)
def test_t5(config, n=-1):
from simplet5 import SimpleT5
config = json.load(open(config))
entity, project = config["wandb-project"].split("/")
experiment = wandb.init(project=project, entity=entity, group="hyperion")
print(config["dataset"], config["name"])
dataset = wandb.use_artifact(config["wandb-project"] + "/" + config["dataset"])
dataset = dataset.download()
model = wandb.use_artifact(config["wandb-project"] + "/" + config["name"]+":latest")
model_path = model.download()
test = pd.read_json(dataset+"/test.json")
if n > -1:
test = test.head(n)
model = SimpleT5()
# load (supports t5, mt5, byT5 models)
model.from_pretrained(config["prototype"], model_path)
def get_predictions(x):
h = model.predict(x.replace("A low toxicity reply:", "A high toxicity reply:"))
hp = x.replace("A low toxicity reply:", "A high toxicity reply:")
l = model.predict(x.replace("A high toxicity reply:", "A low toxicity reply:"))
lp = x.replace("A high toxicity reply:", "A low toxicity reply:")
return h, hp, l, lp
print("starting test")
test["high"], test["high_prompt"], test["low"], test["low_prompt"] = zip(*test["source_text"].progress_map(lambda x: get_predictions(x)))
test.to_csv("result.csv")
artifact = wandb.Artifact(config["name"]+"-test", type="dataset")
artifact.add_file("result.csv")
experiment.log_artifact(artifact)
"""
{
"dataset": "prochoice_PCTS:v0",
"wandb-project": "knoxcs/detoxify",
"prototype": "t5",
"base_model": "t5-large",
"args": {
"source_max_token_len": 512,
"target_max_token_len": 512,
"batch_size": 8,
"max_epochs": 3,
"use_gpu": 3,
"output_dir": "outputs",
}
}
"""
if __name__ == "__main__":
import os
r = os.environ["T5RUN"]
train_t5(r)