File size: 10,319 Bytes
fbe096f
06c9264
e432586
fbe096f
 
 
06276f7
fbe096f
 
7ef202c
fbe096f
 
 
 
 
 
 
 
 
 
2894e20
fbe096f
 
 
 
 
 
2894e20
fbe096f
 
 
 
 
 
 
 
 
 
 
 
 
2894e20
fbe096f
 
 
 
 
 
 
 
 
 
2894e20
fbe096f
 
 
 
 
 
 
 
 
 
 
 
2894e20
fbe096f
 
2894e20
fbe096f
6d03810
e432586
fbe096f
 
 
 
 
 
 
 
 
 
 
 
2b183b5
fbe096f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7ef202c
fbe096f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2b183b5
fbe096f
 
 
2b183b5
 
 
fbe096f
 
 
 
 
 
 
2894e20
fbe096f
 
 
 
 
 
 
 
2894e20
fbe096f
 
 
 
 
 
 
 
 
 
 
2894e20
fbe096f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1675d25
6d03810
 
 
fbe096f
 
 
 
 
 
 
2b183b5
fbe096f
2b183b5
fbe096f
 
6d03810
1675d25
fbe096f
6d03810
fbe096f
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
import gradio as gr
import torch
import torch.nn as nn
import torch.nn.functional as F
from torchvision import transforms, models
import numpy as np
from PIL import Image
import matplotlib.pyplot as plt
import os

# Class Mapping for RAF-DB Dataset (7 classes)
class_mapping = {
    0: "Surprise",
    1: "Fear",
    2: "Disgust",
    3: "Happiness",
    4: "Sadness",
    5: "Anger",
    6: "Neutral"
}

# Transformations for inference (same as test transform)
transform = transforms.Compose([
    transforms.Resize((112, 112)),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])

# Feature Extraction Backbone
class IR50(nn.Module):
    def __init__(self):
        super(IR50, self).__init__()
        resnet = models.resnet50(weights='IMAGENET1K_V1')
        self.conv1 = resnet.conv1
        self.bn1 = resnet.bn1
        self.relu = resnet.relu
        self.maxpool = resnet.maxpool
        self.layer1 = resnet.layer1
        self.layer2 = resnet.layer2
        self.downsample = nn.Conv2d(512, 256, 1, stride=2)
        self.bn_downsample = nn.BatchNorm2d(256, eps=1e-5)

    def forward(self, x):
        x = self.conv1(x)
        x = self.bn1(x)
        x = self.relu(x)
        x = self.maxpool(x)
        x = self.layer1(x)
        x = self.layer2(x)
        x = self.downsample(x)
        x = self.bn_downsample(x)
        return x

# HLA Stream
class HLA(nn.Module):
    def __init__(self, in_channels=256, reduction=4):
        super(HLA, self).__init__()
        reduced_channels = in_channels // reduction
        self.spatial_branch1 = nn.Conv2d(in_channels, reduced_channels, 1)
        self.spatial_branch2 = nn.Conv2d(in_channels, reduced_channels, 1)
        self.sigmoid = nn.Sigmoid()
        self.channel_restore = nn.Conv2d(reduced_channels, in_channels, 1)
        self.channel_attention = nn.Sequential(
            nn.AdaptiveAvgPool2d(1),
            nn.Conv2d(in_channels, in_channels // reduction, 1, bias=False),
            nn.ReLU(),
            nn.Conv2d(in_channels // reduction, in_channels, 1, bias=False),
            nn.Sigmoid()
        )
        self.bn = nn.BatchNorm2d(in_channels, eps=1e-5)

    def forward(self, x):
        b1 = self.spatial_branch1(x)
        b2 = self.spatial_branch2(x)
        spatial_attn = self.sigmoid(torch.max(b1, b2))
        spatial_attn = self.channel_restore(spatial_attn)
        spatial_out = x * spatial_attn
        channel_attn = self.channel_attention(spatial_out)
        out = spatial_out * channel_attn
        out = self.bn(out)
        return out

# ViT Stream
class ViT(nn.Module):
    def __init__(self, in_channels=256, patch_size=1, embed_dim=768, num_layers=8, num_heads=12):  # 8 layers as in the 82.93% version
        super(ViT, self).__init__()
        self.patch_embed = nn.Conv2d(in_channels, embed_dim, kernel_size=patch_size, stride=patch_size)
        self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
        num_patches = (7 // patch_size) * (7 // patch_size)
        self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, embed_dim))
        self.transformer = nn.ModuleList([
            nn.TransformerEncoderLayer(embed_dim, num_heads, dim_feedforward=1536, activation="gelu")
            for _ in range(num_layers)
        ])
        self.ln = nn.LayerNorm(embed_dim)
        self.bn = nn.BatchNorm1d(embed_dim, eps=1e-5)

        # Initialize weights
        nn.init.xavier_uniform_(self.patch_embed.weight)
        nn.init.zeros_(self.patch_embed.bias)
        nn.init.normal_(self.cls_token, std=0.02)
        nn.init.normal_(self.pos_embed, std=0.02)

    def forward(self, x):
        x = self.patch_embed(x)
        x = x.flatten(2).transpose(1, 2)
        cls_tokens = self.cls_token.expand(x.size(0), -1, -1)
        x = torch.cat([cls_tokens, x], dim=1)
        x = x + self.pos_embed
        for layer in self.transformer:
            x = layer(x)
        x = x[:, 0]
        x = self.ln(x)
        x = self.bn(x)
        return x

# Intensity Stream
class IntensityStream(nn.Module):
    def __init__(self, in_channels=256):
        super(IntensityStream, self).__init__()
        sobel_x = torch.tensor([[-1, 0, 1], [-2, 0, 2], [-1, 0, 1]], dtype=torch.float32)
        sobel_y = torch.tensor([[-1, -2, -1], [0, 0, 0], [1, 2, 1]], dtype=torch.float32)
        self.sobel_x = nn.Conv2d(in_channels, in_channels, 3, padding=1, bias=False, groups=in_channels)
        self.sobel_y = nn.Conv2d(in_channels, in_channels, 3, padding=1, bias=False, groups=in_channels)
        self.sobel_x.weight.data = sobel_x.repeat(in_channels, 1, 1, 1)
        self.sobel_y.weight.data = sobel_y.repeat(in_channels, 1, 1, 1)
        self.conv = nn.Conv2d(in_channels, 128, 3, padding=1)
        self.bn = nn.BatchNorm2d(128, eps=1e-5)
        self.pool = nn.AdaptiveAvgPool2d(1)
        self.attention = nn.MultiheadAttention(embed_dim=128, num_heads=1)

        # Initialize weights
        nn.init.xavier_uniform_(self.conv.weight)
        nn.init.zeros_(self.conv.bias)

    def forward(self, x):
        gx = self.sobel_x(x)
        gy = self.sobel_y(x)
        grad_magnitude = torch.sqrt(gx**2 + gy**2 + 1e-8)
        variance = ((x - x.mean(dim=1, keepdim=True))**2).mean(dim=1).flatten(1)
        cnn_out = F.relu(self.conv(grad_magnitude))
        cnn_out = self.bn(cnn_out)
        texture_out = self.pool(cnn_out).squeeze(-1).squeeze(-1)
        attn_in = cnn_out.flatten(2).permute(2, 0, 1)
        attn_in = attn_in / (attn_in.norm(dim=-1, keepdim=True) + 1e-8)
        attn_out, _ = self.attention(attn_in, attn_in, attn_in)
        context_out = attn_out.mean(dim=0)
        out = torch.cat([texture_out, context_out], dim=1)
        return out, grad_magnitude, variance

# Full Model (Single-Label Prediction)
class TripleStreamHLAViT(nn.Module):
    def __init__(self, num_classes=7):
        super(TripleStreamHLAViT, self).__init__()
        self.backbone = IR50()
        self.hla = HLA()
        self.vit = ViT()
        self.intensity = IntensityStream()
        self.fc_hla = nn.Linear(256*7*7, 768)
        self.fc_intensity = nn.Linear(256, 768)
        self.fusion_fc = nn.Linear(768*3, 512)
        self.bn_fusion = nn.BatchNorm1d(512, eps=1e-5)
        self.dropout = nn.Dropout(0.5)
        self.classifier = nn.Linear(512, num_classes)

        # Initialize weights
        nn.init.xavier_uniform_(self.fc_hla.weight)
        nn.init.zeros_(self.fc_hla.bias)
        nn.init.xavier_uniform_(self.fc_intensity.weight)
        nn.init.zeros_(self.fc_intensity.bias)
        nn.init.xavier_uniform_(self.fusion_fc.weight)
        nn.init.zeros_(self.fusion_fc.bias)
        nn.init.xavier_uniform_(self.classifier.weight)
        nn.init.zeros_(self.classifier.bias)

    def forward(self, x):
        features = self.backbone(x)
        hla_out = self.hla(features)
        vit_out = self.vit(features)
        intensity_out, grad_magnitude, variance = self.intensity(features)
        hla_flat = self.fc_hla(hla_out.view(-1, 256*7*7))
        intensity_flat = self.fc_intensity(intensity_out)
        fused = torch.cat([hla_flat, vit_out, intensity_flat], dim=1)
        fused = F.relu(self.fusion_fc(fused))
        fused = self.bn_fusion(fused)
        fused = self.dropout(fused)
        logits = self.classifier(fused)
        return logits, hla_out, vit_out, grad_magnitude, variance

# Load the model
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"Using device: {device}")
model = TripleStreamHLAViT(num_classes=7).to(device)
model_path = "triple_stream_model_rafdb.pth"  # Ensure this file is in the Hugging Face Space repository
try:
    # Map the weights to the appropriate device
    map_location = torch.device('cpu') if not torch.cuda.is_available() else None
    model.load_state_dict(torch.load(model_path, map_location=map_location, weights_only=True))
    model.eval()
    print("Model loaded successfully")
except Exception as e:
    print(f"Error loading model: {e}")
    raise

# Inference and Visualization Function
def predict_emotion(image):
    # Convert the input image (from Gradio) to PIL Image
    if isinstance(image, np.ndarray):
        image = Image.fromarray(image)

    # Preprocess the image
    image_tensor = transform(image).unsqueeze(0).to(device)

    # Run inference
    with torch.no_grad():
        outputs, hla_out, _, grad_magnitude, _ = model(image_tensor)
        probs = F.softmax(outputs, dim=1)
        pred_label = torch.argmax(probs, dim=1).item()
        pred_label_name = class_mapping[pred_label]
        probabilities = probs.cpu().numpy()[0]

    # Create probability dictionary
    prob_dict = {class_mapping[i]: float(prob) for i, prob in enumerate(probabilities)}

    # Generate HLA heatmap
    heatmap = hla_out[0].mean(dim=0).detach().cpu().numpy()
    
    # Denormalize the image for visualization
    img = image_tensor[0].permute(1, 2, 0).detach().cpu().numpy()
    img = img * np.array([0.229, 0.224, 0.225]) + np.array([0.485, 0.456, 0.406])
    img = np.clip(img, 0, 1)

    # Plot the input image and heatmap
    fig, axs = plt.subplots(1, 2, figsize=(8, 4))
    axs[0].imshow(img)
    axs[0].set_title(f"Input Image\nPredicted: {pred_label_name}")
    axs[0].axis("off")
    axs[1].imshow(heatmap, cmap="jet")
    axs[1].set_title("HLA Heatmap")
    axs[1].axis("off")
    plt.tight_layout()
    
    # Save the plot to a temporary file
    plt.savefig("visualization.png")
    plt.close()

    return pred_label_name, prob_dict, "visualization.png"

# Gradio Interface
iface = gr.Interface(
    fn=predict_emotion,
    inputs=gr.Image(type="pil", label="Upload an Image"),
    outputs=[
        gr.Textbox(label="Predicted Emotion"),
        gr.Label(label="Probabilities"),
        gr.Image(label="Input Image and HLA Heatmap")
    ],
    title="Facial Emotion Recognition with TripleStreamHLAViT",
    description="Upload an image to predict the facial emotion (Surprise, Fear, Disgust, Happiness, Sadness, Anger, Neutral). This model achieves 82.93% test accuracy on the RAF-DB dataset. The HLA heatmap shows where the model focuses.",
    examples=[
        ["examples/surprise.jpg"],
        ["examples/sadness.jpg"]
    ]
)

# Launch the interface
if __name__ == "__main__":
    iface.launch(share=False)