PromptEngineering/main.py at master · Pennelli02/PromptEngineering · GitHub

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
from tqdm import tqdm

import classifier
import dataset
import metrics
import prompt

# VALORI PER CARICARE IL DATASET
installDataset = False  # serve nel caso si perdesse tutto
startMiniDt = False
NAME = "test_OS"
MAX_IMAGES = 300
SHUFFLE = False
# ======================================

# VALORI PER IL PROMPT
INDEX_PROMPT = 6  # (0-6)
IS_ITALIAN = False
SHOW_IMAGES = False
ONESHOT = True
UNCERTAIN_EN = False  # abilitare l'opzione al modello di rispondere incerto
# ===================================
# VALORI PER IL MODELLO

MODEL_NAME = "gemma3:4b"
# ===================================
# MODALITA' AUTOMATICA
AUTO_ON = False

# =============================
# GEMMA3 ONESHOT
imagePath = "photoEx/real/woman3-4.jpg"
description = ("This is a close-up, head-and-shoulders portrait of a young woman. She has shoulder-length, dark brown "
               "hair that is parted on her left side, with some strands falling forward around her face. Her face is "
               "a smooth, light tone. She has large, brown eyes that are looking directly at the camera, "
               "a small nose, and full, light-toned lips. Her expression is calm and neutral. The lighting is soft, "
               "evenly illuminating her face and hair. The background is blurred, but you can make out the vague "
               "shapes and colors of what appear to be party balloons, including a reddish one on the upper left and "
               "a light-colored one on the upper right with a faint, darker pattern on it.")
labelImage = "real face"
reason = ("The image shows a person with natural skin texture, including visible pores and slight unevenness in tone, "
          "particularly around the nose. The eyes have realistic reflections and subtle imperfections. The hair has a "
          "natural variation in strand thickness and highlights, and there are realistic, soft focus gradients in the "
          "background, consistent with a photograph taken with a camera.")
# ho usato la spiegazione fornita da gemini (stessa casa di produzione di gemma3)
# ======================================
# dataset section
if startMiniDt:
    images_with_labels, fakes, reals = dataset.loadDataset(MAX_IMAGES)
    dataset.saveDataset(images_with_labels, NAME)
else:
    images_with_labels, fakes, reals = dataset.loadExistingDataset(NAME)

if SHUFFLE:
    dataset.shuffleDataset(images_with_labels)

# prompt section
oneShotMessage = None

if AUTO_ON:
    # DATASETS = ["test_2", "test_3"]
    # for data in DATASETS:
    #     print(f"\n=== Running dataset: {data} ===")
    #
    #     images_with_labels, fakes, reals = dataset.loadExistingDataset(data)
    #     if SHUFFLE:
    #         dataset.shuffleDataset(images_with_labels)
    #     if data == "test_2":
    #         startPoint = 2
    #     else:
    #         startPoint = 0
    for INDEX_PROMPT in range(7):  # Prompt da 0 a 6
        for IS_ITALIAN in [False, True]:  # Prima inglese, poi italiano
            userPrompt = prompt.chooseAPrompt(INDEX_PROMPT, IS_ITALIAN)
            systemPrompt = prompt.getSystemPrompt(IS_ITALIAN, UNCERTAIN_EN)

            print(f"\n--- Prompt {INDEX_PROMPT} | Language: {'Italian' if IS_ITALIAN else 'English'} ---")
            print("Prompt:", userPrompt)

            # Inizializza le metriche
            counters = metrics.initMetrics()
            results = []

            # Ciclo principale di analisi immagini
            for img_path, label in tqdm(images_with_labels,
                                        desc=f"Analyzing images (Prompt {INDEX_PROMPT}, {'IT' if IS_ITALIAN else 'EN'})"):
                print(img_path)
                result, counters = classifier.analyze_image(
                    img_path, label, userPrompt, MODEL_NAME,
                    ONESHOT, oneShotMessage, systemPrompt,
                    counters, SHOW_IMAGES
                )
                results.append(result)

            # Salvataggio dei risultati
            jsonINFO = metrics.analyzeMetrics(
                counters, images_with_labels, userPrompt,
                systemPrompt, ONESHOT, oneShotMessage,
                reals, fakes
            )

            metrics.saveAllJson(jsonINFO, results, IS_ITALIAN, MODEL_NAME, INDEX_PROMPT)

else:
    userPrompt = prompt.chooseAPrompt(INDEX_PROMPT, IS_ITALIAN)
    systemPrompt = prompt.getSystemPrompt(IS_ITALIAN, UNCERTAIN_EN)
    if ONESHOT:
        oneShotMessage = prompt.usingOneShot(imagePath, description, labelImage, userPrompt, IS_ITALIAN, reason)

    print("you choose " + userPrompt)
    # initialize metrics
    counters = metrics.initMetrics()
    results = []
    # Main analysis loop
    for img_path, label in tqdm(images_with_labels, desc=" Analyzing images"):
        print(img_path)
        result, counters = classifier.analyze_image(img_path, label, userPrompt, MODEL_NAME, ONESHOT, oneShotMessage,
                                                    systemPrompt, counters, SHOW_IMAGES)
        results.append(result)
    jsonINFO = metrics.analyzeMetrics(counters, images_with_labels, userPrompt, systemPrompt, ONESHOT, oneShotMessage,
                                      reals, fakes)
    metrics.saveAllJson(jsonINFO, results, IS_ITALIAN, MODEL_NAME, INDEX_PROMPT)