Spaces:
Sleeping
Sleeping
Enhance model loading for prediction by integrating pre-trained BERT and refining checkpoint handling
Browse files- utils/prediction.py +19 -5
utils/prediction.py
CHANGED
@@ -1,6 +1,6 @@
|
|
1 |
from utils.model import BiLSTMAttentionBERT, BiLSTMConfig
|
2 |
import torch
|
3 |
-
from transformers import AutoTokenizer
|
4 |
from sklearn.preprocessing import LabelEncoder
|
5 |
import numpy as np
|
6 |
import streamlit as st
|
@@ -12,6 +12,11 @@ from huggingface_hub import hf_hub_download
|
|
12 |
def load_model_for_prediction():
|
13 |
try:
|
14 |
st.write("Starting model loading...")
|
|
|
|
|
|
|
|
|
|
|
15 |
config = BiLSTMConfig(
|
16 |
hidden_dim=128,
|
17 |
num_classes=22,
|
@@ -19,20 +24,29 @@ def load_model_for_prediction():
|
|
19 |
dropout=0.5
|
20 |
)
|
21 |
|
22 |
-
# Initialize model
|
23 |
model = BiLSTMAttentionBERT(config)
|
|
|
24 |
|
25 |
-
# Load checkpoint
|
26 |
model_path = hf_hub_download(
|
27 |
repo_id="joko333/BiLSTM_v01",
|
28 |
filename="model_epoch8_acc72.53.pt"
|
29 |
)
|
30 |
checkpoint = torch.load(model_path, map_location='cpu')
|
31 |
|
32 |
-
#
|
|
|
|
|
33 |
if 'model_state_dict' in checkpoint:
|
|
|
|
|
34 |
state_dict = checkpoint['model_state_dict']
|
35 |
-
|
|
|
|
|
|
|
|
|
|
|
36 |
st.write("Model loaded successfully")
|
37 |
else:
|
38 |
st.error("Invalid checkpoint format")
|
|
|
1 |
from utils.model import BiLSTMAttentionBERT, BiLSTMConfig
|
2 |
import torch
|
3 |
+
from transformers import AutoTokenizer, AutoModel
|
4 |
from sklearn.preprocessing import LabelEncoder
|
5 |
import numpy as np
|
6 |
import streamlit as st
|
|
|
12 |
def load_model_for_prediction():
|
13 |
try:
|
14 |
st.write("Starting model loading...")
|
15 |
+
|
16 |
+
# Initialize BERT first
|
17 |
+
bert = AutoModel.from_pretrained('dmis-lab/biobert-base-cased-v1.2')
|
18 |
+
|
19 |
+
# Initialize config and model
|
20 |
config = BiLSTMConfig(
|
21 |
hidden_dim=128,
|
22 |
num_classes=22,
|
|
|
24 |
dropout=0.5
|
25 |
)
|
26 |
|
|
|
27 |
model = BiLSTMAttentionBERT(config)
|
28 |
+
model.bert = bert # Set pre-trained BERT
|
29 |
|
30 |
+
# Load custom layers from checkpoint
|
31 |
model_path = hf_hub_download(
|
32 |
repo_id="joko333/BiLSTM_v01",
|
33 |
filename="model_epoch8_acc72.53.pt"
|
34 |
)
|
35 |
checkpoint = torch.load(model_path, map_location='cpu')
|
36 |
|
37 |
+
# Debug checkpoint structure
|
38 |
+
st.write("Checkpoint keys:", checkpoint.keys())
|
39 |
+
|
40 |
if 'model_state_dict' in checkpoint:
|
41 |
+
# Extract only custom layer weights
|
42 |
+
custom_state_dict = {}
|
43 |
state_dict = checkpoint['model_state_dict']
|
44 |
+
for key, value in state_dict.items():
|
45 |
+
if not key.startswith('bert.'):
|
46 |
+
custom_state_dict[key] = value
|
47 |
+
|
48 |
+
# Load custom layers
|
49 |
+
model.load_state_dict(custom_state_dict, strict=False)
|
50 |
st.write("Model loaded successfully")
|
51 |
else:
|
52 |
st.error("Invalid checkpoint format")
|