File size: 13,927 Bytes
d8c12d2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
#!/usr/bin/env python3
"""Neural Storyteller – Gradio App for Hugging Face Spaces (Attention model)."""

import os, json, pickle
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from torchvision import models, transforms
from PIL import Image
import gradio as gr

# ── Device ──
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# ── Load config ──
with open("config.json", "r") as f:
    cfg = json.load(f)

EMBED_SIZE   = cfg["embed_size"]
HIDDEN_SIZE  = cfg["hidden_size"]
NUM_REGIONS  = cfg["num_regions"]
VOCAB_SIZE   = cfg["vocab_size"]
MAX_LEN      = cfg["max_len"]
DROPOUT      = cfg["dropout"]
BEAM_WIDTH   = cfg["beam_width"]
LENGTH_PEN   = cfg.get("length_penalty", 0.7)
REP_PEN      = cfg.get("repetition_penalty", 1.2)

# ── Vocabulary class (required for unpickling) ──
class Vocabulary:
    PAD, START, END, UNK = '<pad>', '<start>', '<end>', '<unk>'

    def __init__(self, freq_threshold=5):
        self.freq_threshold = freq_threshold
        self.word2idx = {}
        self.idx2word = {}
        self._idx = 0

    def __len__(self):
        return len(self.word2idx)

# ── Load vocabulary ──
with open("vocab.pkl", "rb") as f:
    vocab = pickle.load(f)


# ══════════════ Model Definitions (must match training) ══════════════

class Encoder(nn.Module):
    def __init__(self, feature_dim=2048, hidden_size=HIDDEN_SIZE,
                 num_regions=NUM_REGIONS, dropout=DROPOUT):
        super().__init__()
        self.num_regions = num_regions
        self.hidden_size = hidden_size
        self.project = nn.Linear(feature_dim, hidden_size * num_regions)
        self.bn      = nn.BatchNorm1d(hidden_size * num_regions)
        self.dropout = nn.Dropout(dropout)
        self.init_h  = nn.Linear(feature_dim, hidden_size)
        self.init_c  = nn.Linear(feature_dim, hidden_size)

    def forward(self, features):
        proj = self.dropout(F.relu(self.bn(self.project(features))))
        regions = proj.view(-1, self.num_regions, self.hidden_size)
        h0 = torch.tanh(self.init_h(features))
        c0 = torch.tanh(self.init_c(features))
        return regions, h0, c0


class BahdanauAttention(nn.Module):
    def __init__(self, hidden_size):
        super().__init__()
        self.W_enc = nn.Linear(hidden_size, hidden_size)
        self.W_dec = nn.Linear(hidden_size, hidden_size)
        self.V     = nn.Linear(hidden_size, 1)

    def forward(self, encoder_out, decoder_hidden):
        energy  = self.V(torch.tanh(
            self.W_enc(encoder_out) + self.W_dec(decoder_hidden).unsqueeze(1)
        ))
        weights = F.softmax(energy.squeeze(2), dim=1)
        context = (weights.unsqueeze(2) * encoder_out).sum(1)
        return context, weights


class AttentionDecoder(nn.Module):
    def __init__(self, vocab_size, embed_size=EMBED_SIZE,
                 hidden_size=HIDDEN_SIZE, dropout=DROPOUT):
        super().__init__()
        self.embed     = nn.Embedding(vocab_size, embed_size, padding_idx=0)
        self.attention = BahdanauAttention(hidden_size)
        self.lstm_cell = nn.LSTMCell(embed_size + hidden_size, hidden_size)
        self.fc_out    = nn.Linear(hidden_size + hidden_size, vocab_size)
        self.dropout   = nn.Dropout(dropout)

    def forward_step(self, word_idx, h, c, encoder_out):
        embed   = self.dropout(self.embed(word_idx))
        context, attn_w = self.attention(encoder_out, h)
        lstm_in = torch.cat([embed, context], dim=1)
        h, c    = self.lstm_cell(lstm_in, (h, c))
        logits  = self.fc_out(self.dropout(torch.cat([h, context], dim=1)))
        return logits, h, c, attn_w


class Seq2SeqCaptioner(nn.Module):
    def __init__(self, vocab_size, embed_size=EMBED_SIZE,
                 hidden_size=HIDDEN_SIZE, dropout=DROPOUT,
                 num_regions=NUM_REGIONS):
        super().__init__()
        self.encoder     = Encoder(2048, hidden_size, num_regions, dropout)
        self.decoder     = AttentionDecoder(vocab_size, embed_size, hidden_size, dropout)
        self.hidden_size = hidden_size

    def forward(self, features, captions, teacher_forcing_ratio=1.0):
        import random
        B = features.size(0)
        T = captions.size(1) - 1
        V = self.decoder.fc_out.out_features
        encoder_out, h, c = self.encoder(features)
        outputs = torch.zeros(B, T, V, device=features.device)
        inp = captions[:, 0]
        for t in range(T):
            logits, h, c, _ = self.decoder.forward_step(inp, h, c, encoder_out)
            outputs[:, t] = logits
            if random.random() < teacher_forcing_ratio:
                inp = captions[:, t + 1]
            else:
                inp = logits.argmax(dim=-1)
        return outputs


# ── Load trained weights ──
caption_model = Seq2SeqCaptioner(VOCAB_SIZE, EMBED_SIZE, HIDDEN_SIZE, DROPOUT, NUM_REGIONS).to(device)
caption_model.load_state_dict(torch.load("best_model.pth", map_location=device))
caption_model.eval()

# ── ResNet50 feature extractor ──
resnet = models.resnet50(weights=models.ResNet50_Weights.DEFAULT)
resnet = nn.Sequential(*list(resnet.children())[:-1])
resnet = resnet.to(device)
resnet.eval()

img_transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406],
                         std=[0.229, 0.224, 0.225]),
])


# ── Greedy Search (faster, simpler) ──
@torch.no_grad()
def greedy_search_inference(feature):
    feature = feature.unsqueeze(0).to(device)
    encoder_out, h, c = caption_model.encoder(feature)

    start_idx = vocab.word2idx[vocab.START]
    end_idx   = vocab.word2idx[vocab.END]
    
    sequence = [start_idx]
    inp = torch.tensor([start_idx], device=device)

    for _ in range(MAX_LEN):
        logits, h, c, _ = caption_model.decoder.forward_step(inp, h, c, encoder_out)
        predicted = logits.argmax(dim=-1).item()
        
        if predicted == end_idx:
            break
        
        sequence.append(predicted)
        inp = torch.tensor([predicted], device=device)

    words = [vocab.idx2word[i] for i in sequence
             if vocab.idx2word[i] not in (vocab.START, vocab.END, vocab.PAD)]
    return " ".join(words)


# ── Beam Search with penalties ──
@torch.no_grad()
def beam_search_inference(feature, beam_width=BEAM_WIDTH,
                          length_penalty=LENGTH_PEN,
                          repetition_penalty=REP_PEN):
    feature = feature.unsqueeze(0).to(device)
    encoder_out, h0, c0 = caption_model.encoder(feature)

    start_idx = vocab.word2idx[vocab.START]
    end_idx   = vocab.word2idx[vocab.END]
    pad_idx   = vocab.word2idx[vocab.PAD]

    beams     = [(0.0, [start_idx], h0, c0)]
    completed = []

    for _ in range(MAX_LEN):
        new_beams = []
        for log_prob, seq, h, c in beams:
            inp = torch.tensor([seq[-1]], device=device)
            logits, h_new, c_new, _ = caption_model.decoder.forward_step(
                inp, h, c, encoder_out)
            logits = logits.squeeze(0)

            for prev_tok in set(seq):
                if prev_tok not in (start_idx, end_idx, pad_idx):
                    logits[prev_tok] /= repetition_penalty

            log_probs = F.log_softmax(logits, dim=-1)
            topk_lp, topk_idx = log_probs.topk(beam_width)

            for k in range(beam_width):
                token   = topk_idx[k].item()
                new_lp  = log_prob + topk_lp[k].item()
                new_seq = seq + [token]
                if token == end_idx:
                    score = new_lp / (len(new_seq) ** length_penalty)
                    completed.append((score, new_seq))
                else:
                    new_beams.append((new_lp, new_seq, h_new, c_new))

        new_beams.sort(key=lambda x: x[0], reverse=True)
        beams = new_beams[:beam_width]
        if not beams or len(completed) >= beam_width:
            break

    if not completed and beams:
        for lp, seq, _, _ in beams:
            completed.append((lp / (len(seq) ** length_penalty), seq))

    completed.sort(key=lambda x: x[0], reverse=True)
    best_seq = completed[0][1] if completed else [start_idx]

    words = [vocab.idx2word[i] for i in best_seq
             if vocab.idx2word[i] not in (vocab.START, vocab.END, vocab.PAD)]
    return " ".join(words)


# ── Prediction function for Gradio ──
def predict(image, search_method, beam_width, length_penalty, repetition_penalty):
    """Take a PIL image -> return generated caption string."""
    if image is None:
        return """
        <div style="background: linear-gradient(135deg, #f093fb 0%, #f5576c 100%); padding: 30px; border-radius: 15px; text-align: center;">
            <p style="color: white; font-size: 20px; margin: 0;">⚠️ Please upload an image first</p>
        </div>
        """
    
    image = image.convert("RGB")
    img_tensor = img_transform(image).unsqueeze(0).to(device)

    with torch.no_grad():
        feature = resnet(img_tensor).view(1, -1).squeeze(0)

    if search_method == "Greedy Search (Fast)":
        caption = greedy_search_inference(feature)
        method_info = "πŸš€ Generated using Greedy Search"
    else:  # Beam Search
        caption = beam_search_inference(
            feature, 
            beam_width=int(beam_width),
            length_penalty=length_penalty,
            repetition_penalty=repetition_penalty
        )
        method_info = f"πŸ” Generated using Beam Search (width={int(beam_width)})"
    
    # Return beautiful HTML formatted caption
    return f"""
    <div style="background: linear-gradient(135deg, #667eea 0%, #764ba2 100%); padding: 40px; border-radius: 15px; box-shadow: 0 8px 32px rgba(0,0,0,0.1);">
        <p style="color: white; font-size: 28px; font-weight: 600; text-align: center; line-height: 1.6; margin: 0; text-shadow: 2px 2px 4px rgba(0,0,0,0.2);">
            "{caption}"
        </p>
        <p style="color: rgba(255,255,255,0.9); font-size: 14px; text-align: center; margin-top: 20px; font-style: italic;">
            {method_info}
        </p>
    </div>
    """


# ── Gradio Interface ──
with gr.Blocks(theme=gr.themes.Soft(), title="Neural Storyteller", css="""
    .caption-box {
        background: linear-gradient(135deg, #667eea 0%, #764ba2 100%);
        padding: 30px;
        border-radius: 15px;
        box-shadow: 0 8px 32px rgba(0,0,0,0.1);
        margin: 20px 0;
    }
    .caption-text {
        color: white;
        font-size: 24px;
        font-weight: 600;
        text-align: center;
        line-height: 1.6;
        text-shadow: 2px 2px 4px rgba(0,0,0,0.2);
    }
    .method-info {
        color: rgba(255,255,255,0.9);
        font-size: 14px;
        text-align: center;
        margin-top: 15px;
        font-style: italic;
    }
""") as demo:
    gr.Markdown("""
    # 🧠 Neural Storyteller – AI Image Captioning
    
    Upload any image and let the AI generate a natural language description using a **Seq2Seq model** 
    with ResNet50 encoder and Attention-based LSTM decoder, trained on Flickr30k dataset.
    """)
    
    with gr.Row():
        with gr.Column(scale=1):
            image_input = gr.Image(type="pil", label="πŸ“Έ Upload Your Image", height=400)
            
        with gr.Column(scale=1):
            gr.Markdown("### βš™οΈ Generation Settings")
            
            search_method = gr.Radio(
                choices=["Greedy Search (Fast)", "Beam Search (Better Quality)"],
                value="Beam Search (Better Quality)",
                label="🎯 Decoding Method",
                info="Greedy is faster, Beam produces better results"
            )
            
            with gr.Accordion("πŸ”§ Advanced Options (Beam Search Only)", open=False):
                beam_width = gr.Slider(
                    minimum=1, maximum=10, value=5, step=1,
                    label="Beam Width",
                    info="Number of candidates to explore (higher = better quality but slower)"
                )
                
                length_penalty = gr.Slider(
                    minimum=0.0, maximum=2.0, value=0.7, step=0.1,
                    label="Length Penalty",
                    info="Controls caption length (lower = shorter, higher = longer)"
                )
                
                repetition_penalty = gr.Slider(
                    minimum=1.0, maximum=2.0, value=1.2, step=0.1,
                    label="Repetition Penalty",
                    info="Reduces word repetition (higher = less repetition)"
                )
    
    generate_btn = gr.Button("✨ Generate Caption", variant="primary", size="lg", scale=1)
    
    # Beautiful caption display area
    gr.Markdown("## πŸ“ Generated Caption")
    output_text = gr.HTML(label="")
    
    with gr.Accordion("πŸ’‘ Tips & Model Details", open=False):
        gr.Markdown("""
        ### Tips:
        - Try both **Greedy** and **Beam** search to compare results
        - Increase **Beam Width** for more diverse captions
        - Adjust **Length Penalty** if captions are too short/long
        - Use **Repetition Penalty** to avoid repeated words
        
        ### Model Details:
        - **Encoder**: ResNet50 (pretrained on ImageNet)
        - **Decoder**: Attention-based LSTM
        - **Training Data**: Flickr30k dataset
        - **Vocabulary**: ~8000 words
        """)
    
    generate_btn.click(
        fn=predict,
        inputs=[image_input, search_method, beam_width, length_penalty, repetition_penalty],
        outputs=output_text
    )
    
    gr.Markdown("""
    ---
    <p style="text-align: center; color: #666;">
    Built with PyTorch, Gradio, and ❀️ | Model trained on Flickr30k
    </p>
    """)

if __name__ == "__main__":
    demo.launch()