Upload modeling_mimi.py with huggingface_hub

modeling_mimi.py

@@ -0,0 +1,73 @@
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import math
+
+class PositionalEncoding(nn.Module):
+    def __init__(self, d_model, max_len=15000):
+        super().__init__()
+        pe = torch.zeros(max_len, d_model)
+        position = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1)
+        div_term = torch.exp(torch.arange(0, d_model, 2).float() * (-math.log(10000.0) / d_model))
+        pe[:, 0::2] = torch.sin(position * div_term)
+        pe[:, 1::2] = torch.cos(position * div_term)
+        self.register_buffer('pe', pe.unsqueeze(0))
+
+    def forward(self, x):
+        seq_len = x.size(1)
+        if seq_len > self.pe.size(1):
+             x = x[:, :self.pe.size(1), :]
+             seq_len = x.size(1)
+        return x + self.pe[:, :seq_len, :]
+
+class MemoryModule(nn.Module):
+    def __init__(self, input_dim, mem_dim=64, num_slots=20):
+        super().__init__()
+        self.mem_dim = mem_dim
+        self.num_slots = num_slots
+        self.query_proj = nn.Linear(input_dim, mem_dim)
+        self.memory = nn.Parameter(torch.randn(num_slots, mem_dim))
+        nn.init.kaiming_uniform_(self.memory)
+
+    def forward(self, x):
+        q = self.query_proj(x)
+        att_logits = torch.matmul(q, self.memory.t())
+        att_weights = F.softmax(att_logits, dim=-1)
+        read_content = torch.matmul(att_weights, self.memory)
+        return read_content, att_weights
+
+class MemoryTransformerDetector(nn.Module):
+    def __init__(self, rgb_dim=384, flow_dim=1024, audio_dim=768, d_model=256, nhead=4, num_layers=2):
+        super().__init__()
+        part_dim = d_model // 3
+        self.rgb_proj = nn.Linear(rgb_dim, part_dim)
+        self.flow_proj = nn.Linear(flow_dim, part_dim)
+        self.audio_proj = nn.Linear(audio_dim, d_model - (2 * part_dim))
+        self.pos_encoder = PositionalEncoding(d_model, max_len=15000)
+        encoder_layer = nn.TransformerEncoderLayer(d_model=d_model, nhead=nhead,
+                                                   dim_feedforward=512, dropout=0.3,
+                                                   batch_first=True)
+        self.transformer = nn.TransformerEncoder(encoder_layer, num_layers=num_layers)
+        self.memory = MemoryModule(input_dim=d_model, mem_dim=d_model, num_slots=50)
+        self.classifier = nn.Sequential(
+            nn.Linear(d_model * 2, 128),
+            nn.ReLU(),
+            nn.Dropout(0.3),
+            nn.Linear(128, 1),
+            nn.Sigmoid()
+        )
+
+    def forward(self, rgb, flow, audio):
+        if rgb.dim() == 2:
+            rgb, flow, audio = rgb.unsqueeze(0), flow.unsqueeze(0), audio.unsqueeze(0)
+        x_rgb = self.rgb_proj(rgb)
+        x_flow = self.flow_proj(flow)
+        x_audio = self.audio_proj(audio)
+        x = torch.cat((x_rgb, x_flow, x_audio), dim=2)
+        x = self.pos_encoder(x)
+        x_trans = self.transformer(x)
+        x_mem, att_weights = self.memory(x_trans)
+        x_final = torch.cat((x_trans, x_mem), dim=2)
+        logits = self.classifier(x_final)
+        return logits.squeeze(2), att_weights