Upload train.py
Browse files
train.py
ADDED
@@ -0,0 +1,189 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# %% Importing the dependencies we need
|
2 |
+
import numpy as np
|
3 |
+
import torch
|
4 |
+
from sklearn.datasets import fetch_20newsgroups
|
5 |
+
from sklearn.metrics import (accuracy_score, f1_score, confusion_matrix,
|
6 |
+
ConfusionMatrixDisplay, classification_report)
|
7 |
+
from sklearn.model_selection import train_test_split
|
8 |
+
from sklearn.pipeline import Pipeline
|
9 |
+
from skops import card, hub_utils
|
10 |
+
from skorch import NeuralNetClassifier
|
11 |
+
from skorch.callbacks import LRScheduler, ProgressBar
|
12 |
+
from skorch.hf import HuggingfacePretrainedTokenizer
|
13 |
+
from torch import nn
|
14 |
+
from torch.optim.lr_scheduler import LambdaLR
|
15 |
+
from transformers import AutoModelForSequenceClassification
|
16 |
+
from transformers import AutoTokenizer
|
17 |
+
# for model hosting and requirements
|
18 |
+
from pathlib import Path
|
19 |
+
import transformers
|
20 |
+
import skorch
|
21 |
+
import sklearn
|
22 |
+
import torch
|
23 |
+
|
24 |
+
# %%
|
25 |
+
# Choose a tokenizer and BERT model that work together
|
26 |
+
TOKENIZER = "distilbert-base-uncased"
|
27 |
+
PRETRAINED_MODEL = "distilbert-base-uncased"
|
28 |
+
|
29 |
+
# model hyper-parameters
|
30 |
+
OPTMIZER = torch.optim.AdamW
|
31 |
+
LR = 5e-5
|
32 |
+
MAX_EPOCHS = 3
|
33 |
+
CRITERION = nn.CrossEntropyLoss
|
34 |
+
BATCH_SIZE = 8
|
35 |
+
|
36 |
+
# device
|
37 |
+
DEVICE = 'cuda' if torch.cuda.is_available() else 'cpu'
|
38 |
+
|
39 |
+
# %% Load the dataset, define features & labels and split
|
40 |
+
dataset = fetch_20newsgroups()
|
41 |
+
|
42 |
+
print(dataset.DESCR.split('Usage')[0])
|
43 |
+
|
44 |
+
dataset.target_names
|
45 |
+
|
46 |
+
X = dataset.data
|
47 |
+
y = dataset.target
|
48 |
+
X_train, X_test, y_train, y_test, = train_test_split(X, y, stratify=y, random_state=0)
|
49 |
+
num_training_steps = MAX_EPOCHS * (len(X_train) // BATCH_SIZE + 1)
|
50 |
+
|
51 |
+
# %%
|
52 |
+
# Defining learning rate scheduler & BERT in nn.Module
|
53 |
+
|
54 |
+
def lr_schedule(current_step):
|
55 |
+
factor = float(num_training_steps - current_step) / float(max(1, num_training_steps))
|
56 |
+
assert factor > 0
|
57 |
+
return factor
|
58 |
+
|
59 |
+
class BertModule(nn.Module):
|
60 |
+
def __init__(self, name, num_labels):
|
61 |
+
super().__init__()
|
62 |
+
self.name = name
|
63 |
+
self.num_labels = num_labels
|
64 |
+
|
65 |
+
self.reset_weights()
|
66 |
+
|
67 |
+
def reset_weights(self):
|
68 |
+
self.bert = AutoModelForSequenceClassification.from_pretrained(
|
69 |
+
self.name, num_labels=self.num_labels
|
70 |
+
)
|
71 |
+
|
72 |
+
def forward(self, **kwargs):
|
73 |
+
pred = self.bert(**kwargs)
|
74 |
+
return pred.logits
|
75 |
+
|
76 |
+
# %% Chaining tokenizer and BERT in one pipeline
|
77 |
+
pipeline = Pipeline([
|
78 |
+
('tokenizer', HuggingfacePretrainedTokenizer(TOKENIZER)),
|
79 |
+
('net', NeuralNetClassifier(
|
80 |
+
BertModule,
|
81 |
+
module__name=PRETRAINED_MODEL,
|
82 |
+
module__num_labels=len(set(y_train)),
|
83 |
+
optimizer=OPTMIZER,
|
84 |
+
lr=LR,
|
85 |
+
max_epochs=MAX_EPOCHS,
|
86 |
+
criterion=CRITERION,
|
87 |
+
batch_size=BATCH_SIZE,
|
88 |
+
iterator_train__shuffle=True,
|
89 |
+
device=DEVICE,
|
90 |
+
callbacks=[
|
91 |
+
LRScheduler(LambdaLR, lr_lambda=lr_schedule, step_every='batch'),
|
92 |
+
ProgressBar(),
|
93 |
+
],
|
94 |
+
)),
|
95 |
+
])
|
96 |
+
|
97 |
+
torch.manual_seed(0)
|
98 |
+
torch.cuda.manual_seed(0)
|
99 |
+
torch.cuda.manual_seed_all(0)
|
100 |
+
np.random.seed(0)
|
101 |
+
|
102 |
+
# %% Training
|
103 |
+
%time pipeline.fit(X_train, y_train)
|
104 |
+
|
105 |
+
# %% Evaluate the model
|
106 |
+
%%time
|
107 |
+
with torch.inference_mode():
|
108 |
+
y_pred = pipeline.predict(X_test)
|
109 |
+
|
110 |
+
accuracy_score(y_test, y_pred)
|
111 |
+
|
112 |
+
# %% Save the model
|
113 |
+
import pickle
|
114 |
+
with open("model.pkl", mode="bw") as f:
|
115 |
+
pickle.dump(pipeline, file=f)
|
116 |
+
|
117 |
+
# %% Initialize the repository for Hub
|
118 |
+
local_repo = "model_repo"
|
119 |
+
hub_utils.init(
|
120 |
+
model="model.pkl",
|
121 |
+
requirements=[f"scikit-learn={sklearn.__version__}", f"transformers={transformers.__version__}",
|
122 |
+
f"torch={torch.__version__}", f"skorch={skorch.__version__}"],
|
123 |
+
dst=local_repo,
|
124 |
+
task="text-classification",
|
125 |
+
data=X_test,
|
126 |
+
)
|
127 |
+
|
128 |
+
# %% Create model card
|
129 |
+
model_card = card.Card(pipeline, metadata=card.metadata_from_config(Path("model_repo")))
|
130 |
+
|
131 |
+
# %% We will add information related to model
|
132 |
+
model_description = (
|
133 |
+
"This is a neural net classifier and distilbert model chained with sklearn Pipeline trained on 20 news groups dataset."
|
134 |
+
)
|
135 |
+
limitations = "This model is trained for a tutorial and is not ready to be used in production."
|
136 |
+
model_card.add(
|
137 |
+
model_description=model_description,
|
138 |
+
limitations=limitations
|
139 |
+
)
|
140 |
+
|
141 |
+
# %% We can add plots, evaluation results and more!
|
142 |
+
eval_descr = (
|
143 |
+
"The model is evaluated on validation data from 20 news group's test split,"
|
144 |
+
" using accuracy and F1-score with micro average."
|
145 |
+
)
|
146 |
+
model_card.add(eval_method=eval_descr)
|
147 |
+
|
148 |
+
accuracy = accuracy_score(y_test, y_pred)
|
149 |
+
f1 = f1_score(y_test, y_pred, average="micro")
|
150 |
+
model_card.add_metrics(**{"accuracy": accuracy, "f1 score": f1})
|
151 |
+
|
152 |
+
|
153 |
+
cm = confusion_matrix(y_test, y_pred, labels=pipeline.classes_)
|
154 |
+
disp = ConfusionMatrixDisplay(confusion_matrix=cm, display_labels=pipeline.classes_)
|
155 |
+
disp.plot()
|
156 |
+
|
157 |
+
disp.figure_.savefig(Path(local_repo) / "confusion_matrix.png")
|
158 |
+
model_card.add_plot(**{"Confusion matrix": "confusion_matrix.png"})
|
159 |
+
|
160 |
+
clf_report = classification_report(
|
161 |
+
y_test, y_pred, output_dict=True, target_names=dataset.target_names
|
162 |
+
)
|
163 |
+
# %% We can add classification report as a table
|
164 |
+
# We first need to convert classification report to DataFrame to add it as a table
|
165 |
+
import pandas as pd
|
166 |
+
del clf_report["accuracy"]
|
167 |
+
clf_report = pd.DataFrame(clf_report).T.reset_index()
|
168 |
+
model_card.add_table(
|
169 |
+
folded=True,
|
170 |
+
**{
|
171 |
+
"Classification Report": clf_report,
|
172 |
+
},
|
173 |
+
)
|
174 |
+
|
175 |
+
# %% We will save our model card
|
176 |
+
model_card.save(Path(local_repo) / "README.md")
|
177 |
+
|
178 |
+
# %% We will add the training script to our repository
|
179 |
+
hub_utils.add_files(__file__, dst=local_repo)
|
180 |
+
|
181 |
+
# %% Push to Hub! This requires us to authenticate ourselves first.
|
182 |
+
from huggingface_hub import notebook_login
|
183 |
+
notebook_login()
|
184 |
+
|
185 |
+
hub_utils.push(
|
186 |
+
repo_id="scikit-learn/skorch-text-classification",
|
187 |
+
source=local_repo,
|
188 |
+
create_remote=True,
|
189 |
+
)
|