File size: 6,451 Bytes
2bdd84f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e37cfd0
2bdd84f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e37cfd0
2bdd84f
 
 
 
e37cfd0
2bdd84f
e37cfd0
2bdd84f
 
e37cfd0
2bdd84f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e37cfd0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2bdd84f
 
e37cfd0
 
2bdd84f
 
 
 
 
 
 
 
 
 
 
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
158
159
160
161
162
163
164
165
import streamlit as st
import pandas as pd
import json
import os
from datetime import datetime
from utils import (
    load_model, 
    get_hf_token, 
    simulate_training, 
    plot_training_metrics, 
    load_finetuned_model, 
    save_model
)

st.title("πŸ”₯ Fine-tune the Gemma Model")

# -------------------------------
# Finetuning Option Selection
# -------------------------------
finetune_option = st.radio("Select Finetuning Option", ["Fine-tune from scratch", "Refinetune existing model"])

# -------------------------------
# Model Selection Logic
# -------------------------------
selected_model = None
saved_model_path = None

if finetune_option == "Fine-tune from scratch":
    # Display Hugging Face model list
    model_list = [
        "google/gemma-3-1b-pt",
        "google/gemma-3-1b-it", 
        "google/gemma-3-4b-pt", 
        "google/gemma-3-4b-it",
        "google/gemma-3-12b-pt", 
        "google/gemma-3-12b-it", 
        "google/gemma-3-27b-pt", 
        "google/gemma-3-27b-it"
    ]
    selected_model = st.selectbox("πŸ› οΈ Select Gemma Model to Fine-tune", model_list)

elif finetune_option == "Refinetune existing model":
    # Dynamically list all saved models from the /models folder
    model_dir = "models"
    
    if os.path.exists(model_dir):
        saved_models = [f for f in os.listdir(model_dir) if f.endswith(".pt")]
    else:
        saved_models = []

    if saved_models:
        saved_model_path = st.selectbox("Select a saved model to re-finetune", saved_models)
        saved_model_path = os.path.join(model_dir, saved_model_path)
        st.success(f"βœ… Selected model for refinement: `{saved_model_path}`")
    else:
        st.warning("⚠️ No saved models found! Switching to fine-tuning from scratch.")
        finetune_option = "Fine-tune from scratch"

# -------------------------------
# Dataset Selection
# -------------------------------
st.subheader("πŸ“š Dataset Selection")
dataset_option = st.radio("Choose dataset:", ["Upload New Dataset", "Use Existing Dataset (`train_data.csv`)"])
dataset_path = "datasets/train_data.csv"

if dataset_option == "Upload New Dataset":
    uploaded_file = st.file_uploader("πŸ“€ Upload Dataset (CSV or JSON)", type=["csv", "json"])
    if uploaded_file is not None:
        if uploaded_file.name.endswith(".csv"):
            new_data = pd.read_csv(uploaded_file)
        elif uploaded_file.name.endswith(".json"):
            json_data = json.load(uploaded_file)
            new_data = pd.json_normalize(json_data)
        else:
            st.error("❌ Unsupported file format. Please upload CSV or JSON.")
            st.stop()

        if os.path.exists(dataset_path):
            new_data.to_csv(dataset_path, mode='a', index=False, header=False)
            st.success(f"βœ… Data appended to `{dataset_path}`!")
        else:
            new_data.to_csv(dataset_path, index=False)
            st.success(f"βœ… Dataset saved as `{dataset_path}`!")
elif dataset_option == "Use Existing Dataset (`train_data.csv`)":
    if os.path.exists(dataset_path):
        st.success("βœ… Using existing `train_data.csv` for fine-tuning.")
    else:
        st.error("❌ `train_data.csv` not found! Please upload a new dataset.")
        st.stop()

# -------------------------------
# Hyperparameters Configuration
# -------------------------------
st.subheader("πŸ”§ Hyperparameter Configuration")
learning_rate = st.number_input("πŸ“Š Learning Rate", value=1e-4, format="%.5f")
batch_size = st.number_input("πŸ› οΈ Batch Size", value=16, step=1)
epochs = st.number_input("⏱️ Epochs", value=3, step=1)


# -------------------------------
# Fine-tuning Execution with Real-Time Visualization
# -------------------------------
if st.button("πŸš€ Start Fine-tuning"):
    st.info("Fine-tuning process initiated...")
    hf_token = get_hf_token()

    # Model loading logic
    if finetune_option == "Refinetune existing model" and saved_model_path:
        tokenizer, model = load_model("google/gemma-3-1b-it", hf_token)
        model = load_finetuned_model(model, saved_model_path)
        if model:
            st.success(f"βœ… Loaded saved model: `{saved_model_path}` for refinement!")
        else:
            st.error("❌ Failed to load the saved model. Aborting.")
            st.stop()
    else:
        if not selected_model:
            st.error("❌ Please select a model to fine-tune.")
            st.stop()
        tokenizer, model = load_model(selected_model, hf_token)
        if model:
            st.success(f"βœ… Base model loaded: `{selected_model}`")
        else:
            st.error("❌ Failed to load the base model. Aborting.")
            st.stop()

    # Create placeholders for training progress
    loss_chart = st.line_chart()  # Loss curve
    acc_chart = st.line_chart()   # Accuracy curve
    progress_text = st.empty()
    
    # Simulate training loop with real-time visualization
    losses_over_epochs = []
    accuracies_over_epochs = []
    
    for epoch, losses, accs in simulate_training(epochs, learning_rate, batch_size):
        # Update training text
        progress_text.text(f"Epoch {epoch}/{epochs} in progress...")
        
        # Assume simulate_training returns overall average loss and accuracy per epoch
        losses_over_epochs.append(losses)  # e.g., average loss of the epoch
        accuracies_over_epochs.append(accs)  # e.g., average accuracy of the epoch
        
        # Update real-time charts
        loss_chart.add_rows(pd.DataFrame({"Loss": [losses]}))
        acc_chart.add_rows(pd.DataFrame({"Accuracy": [accs]}))
    
    progress_text.text("Fine-tuning completed!")
    
    # Save fine-tuned model with timestamp
    timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
    model_identifier = selected_model if selected_model else os.path.basename(saved_model_path)
    new_model_name = f"models/fine_tuned_model_{model_identifier.replace('/', '_')}_{timestamp}.pt"
    
    saved_model_path = save_model(model, new_model_name)
    if saved_model_path:
        st.success(f"βœ… Fine-tuning completed! Model saved as `{saved_model_path}`")
        model = load_finetuned_model(model, saved_model_path)
        if model:
            st.success("πŸ› οΈ Fine-tuned model loaded and ready for inference!")
        else:
            st.error("❌ Failed to load the fine-tuned model for inference.")
    else:
        st.error("❌ Failed to save the fine-tuned model.")