\u4e00\u3001\u7cfb\u7edf\u67b6\u6784\u8bbe\u8ba1<\/p>\n
A[\u6570\u636e\u91c7\u96c6] –> B[\u9884\u5904\u7406\u6a21\u5757]<\/p>\n
B –> C[\u7279\u5f81\u63d0\u53d6]<\/p>\n
C –> D[\u591a\u6a21\u6001\u878d\u5408]<\/p>\n
D –> E[\u60c5\u7eea\u5206\u7c7b]<\/p>\n
E –> F[\u7cfb\u7edf\u90e8\u7f72]<\/p>\n
F –> G[\u7528\u6237\u754c\u9762]<\/p>\n
\u4e8c\u3001\u6570\u636e\u51c6\u5907\u4e0e\u5904\u7406<\/p>\n
1. \u6570\u636e\u6536\u96c6<\/p>\n
– \u89c6\u9891\u6570\u636e：FER2013（\u9759\u6001\u56fe\u50cf）\u3001RAVDESS（\u52a8\u6001\u89c6\u9891）<\/p>\n
– \u97f3\u9891\u6570\u636e：CREMA-D\u3001IEMOCAP<\/p>\n
– \u81ea\u5b9a\u4e49\u91c7\u96c6：\u4f7f\u7528OpenCV+PyAudio\u5b9e\u73b0\u540c\u6b65\u91c7\u96c6<\/p>\n
2. \u6570\u636e\u9884\u5904\u7406<\/p>\n
\u89c6\u9891\u5904\u7406：<\/p>\n
import cv2<\/p>\n
def process_video(video_path):<\/p>\n
\u00a0 \u00a0 cap = cv2.VideoCapture(video_path)<\/p>\n
\u00a0 \u00a0 frames = []<\/p>\n
\u00a0 \u00a0 while cap.isOpened():<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 ret, frame = cap.read()<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 if not ret: break<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 # \u4eba\u8138\u68c0\u6d4b<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 face = cv2.CascadeClassifier('haarcascade_frontalface_default.xml')<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 faces = face.detectMultiScale(gray, 1.3, 5)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 # \u88c1\u526a\u548c\u5f52\u4e00\u5316<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 if len(faces) > 0:<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 (x,y,w,h) = faces[0]<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 roi = cv2.resize(gray[y:y+h, x:x+w], (128,128))<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 frames.append(roi)<\/p>\n
\u00a0 \u00a0 return np.array(frames)<\/p>\n
\u00a0<\/p>\n
\u97f3\u9891\u5904\u7406：<\/p>\n
import librosa<\/p>\n
def extract_audio_features(audio_path):<\/p>\n
\u00a0 \u00a0 y, sr = librosa.load(audio_path, sr=16000)<\/p>\n
\u00a0 \u00a0 # \u5206\u5e27\u5904\u7406（30ms\u7a97\u53e3）<\/p>\n
\u00a0 \u00a0 frames = librosa.util.frame(y, frame_length=480, hop_length=160)<\/p>\n
\u00a0 \u00a0 # \u63d0\u53d6MFCC\u7279\u5f81<\/p>\n
\u00a0 \u00a0 mfcc = librosa.feature.mfcc(y=y, sr=sr, n_mfcc=40)<\/p>\n
\u00a0 \u00a0 # \u52a8\u6001\u7279\u5f81\u62fc\u63a5<\/p>\n
\u00a0 \u00a0 delta = librosa.feature.delta(mfcc)<\/p>\n
\u00a0 \u00a0 ddelta = librosa.feature.delta(mfcc, order=2)<\/p>\n
\u00a0 \u00a0 return np.concatenate([mfcc, delta, ddelta], axis=0)<\/p>\n
\u00a0<\/p>\n
3. \u6570\u636e\u540c\u6b65\u7b56\u7565<\/p>\n
– \u4f7f\u7528FFmpeg\u63d0\u53d6\u89c6\u9891\u65f6\u95f4\u6233<\/p>\n
– \u52a8\u6001\u65f6\u95f4\u89c4\u6574（DTW）\u5bf9\u9f50\u97f3\u89c6\u9891\u5e8f\u5217<\/p>\n
– \u521b\u5efa\u65f6\u95f4\u5bf9\u9f50\u7684\u5143\u6570\u636e\u6587\u4ef6<\/p>\n
\u00a0<\/p>\n
\u4e09\u3001\u6a21\u578b\u8bbe\u8ba1\u4e0e\u8bad\u7ec3<\/p>\n
1. \u89c6\u89c9\u5206\u652f（PyTorch\u5b9e\u73b0）<\/p>\n
import torch<\/p>\n
from torchvision.models import resnet34<\/p>\n
\u00a0<\/p>\n
class VisualNet(nn.Module):<\/p>\n
\u00a0 \u00a0 def __init__(self):<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 super().__init__()<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 self.base = resnet34(pretrained=True)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 self.base.fc = nn.Identity() # \u79fb\u9664\u5168\u8fde\u63a5\u5c42<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 self.temporal = nn.LSTM(512, 256, bidirectional=True)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0\u00a0<\/p>\n
\u00a0 \u00a0 def forward(self, x):<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 # x: (B, T, C, H, W)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 B, T = x.shape[:2]<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 x = x.view(B*T, *x.shape[2:])<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 features = self.base(x) # (B*T, 512)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 features = features.view(B, T, -1)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 out, _ = self.temporal(features)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 return out[:, -1] # \u53d6\u6700\u540e\u65f6\u523b\u8f93\u51fa<\/p>\n
\u00a0<\/p>\n
2. \u97f3\u9891\u5206\u652f<\/p>\n
class AudioNet(nn.Module):<\/p>\n
\u00a0 \u00a0 def __init__(self):<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 super().__init__()<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 self.conv = nn.Sequential(<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 nn.Conv1d(120, 64, 3, padding=1),<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 nn.BatchNorm1d(64),<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 nn.ReLU(),<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 nn.MaxPool1d(2))<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 self.lstm = nn.LSTM(64, 128, bidirectional=True)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0\u00a0<\/p>\n
\u00a0 \u00a0 def forward(self, x):<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 # x: (B, T, Features)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 x = x.permute(0,2,1) # (B, Features, T)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 x = self.conv(x)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 x = x.permute(2,0,1) # (T, B, Features)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 out, _ = self.lstm(x)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 return out[-1]<\/p>\n
3. \u591a\u6a21\u6001\u878d\u5408<\/p>\n
\u6ce8\u610f\u529b\u878d\u5408\u5c42：<\/p>\n
class FusionModule(nn.Module):<\/p>\n
\u00a0 \u00a0 def __init__(self, v_dim, a_dim):<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 super().__init__()<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 self.v_proj = nn.Linear(v_dim, 256)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 self.a_proj = nn.Linear(a_dim, 256)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 self.attention = nn.MultiheadAttention(256, 4)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0\u00a0<\/p>\n
\u00a0 \u00a0 def forward(self, v_feat, a_feat):<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 v = self.v_proj(v_feat).unsqueeze(0) # (1,B,256)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 a = self.a_proj(a_feat).unsqueeze(0)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 combined = torch.cat([v, a], dim=0) # (2,B,256)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 attn_out, _ = self.attention(combined, combined, combined)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 return attn_out.mean(dim=0)<\/p>\n
\u00a0<\/p>\n
\u00a0\u56db\u3001\u8bad\u7ec3\u7b56\u7565<\/p>\n
1. \u635f\u5931\u51fd\u6570\u8bbe\u8ba1<\/p>\n
class MultimodalLoss(nn.Module):<\/p>\n
\u00a0 \u00a0 def __init__(self):<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 super().__init__()<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 self.ce = nn.CrossEntropyLoss()<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 self.kl = nn.KLDivLoss()<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0\u00a0<\/p>\n
\u00a0 \u00a0 def forward(self, pred, label, v_out, a_out):<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 # \u4e3b\u635f\u5931<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 main_loss = self.ce(pred, label)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 # \u6a21\u6001\u4e00\u81f4\u6027\u635f\u5931<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 p_v = F.log_softmax(v_out, dim=1)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 p_a = F.softmax(a_out, dim=1)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 consistency_loss = self.kl(p_v, p_a.detach())<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 return main_loss + 0.5 * consistency_loss<\/p>\n
\u00a0<\/p>\n
2. \u8bad\u7ec3\u6280\u5de7<\/p>\n
– \u5206\u9636\u6bb5\u8bad\u7ec3：\u5148\u5355\u6a21\u6001\u9884\u8bad\u7ec3，\u518d\u8054\u5408\u5fae\u8c03<\/p>\n
– \u6570\u636e\u589e\u5f3a\u7b56\u7565：<\/p>\n
\u00a0 – \u89c6\u89c9：\u968f\u673a\u906e\u6321\u3001\u8272\u5f69\u6296\u52a8<\/p>\n
\u00a0 – \u97f3\u9891：\u6dfb\u52a0\u566a\u58f0\u3001\u65f6\u79fb\u53d8\u6362<\/p>\n
– \u4f18\u5316\u5668\u914d\u7f6e：<\/p>\n
\u00a0 optimizer = torch.optim.AdamW([<\/p>\n
\u00a0 \u00a0 \u00a0 {'params': visual_net.parameters(), 'lr': 1e-4},<\/p>\n
\u00a0 \u00a0 \u00a0 {'params': audio_net.parameters(), 'lr': 3e-4},<\/p>\n
\u00a0 \u00a0 \u00a0 {'params': fusion_module.parameters(), 'lr': 5e-4}<\/p>\n
\u00a0 ], weight_decay=1e-5)<\/p>\n
\u4e94\u3001\u5b9e\u65f6\u5904\u7406\u4e0e\u90e8\u7f72<\/p>\n
1. \u5b9e\u65f6\u5904\u7406\u67b6\u6784<\/p>\n
import queue<\/p>\n
from threading import Thread<\/p>\n
\u00a0<\/p>\n
class RealTimeProcessor:<\/p>\n
\u00a0 \u00a0 def __init__(self):<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 self.video_queue = queue.Queue(maxsize=30)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 self.audio_queue = queue.Queue(maxsize=100)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0\u00a0<\/p>\n
\u00a0 \u00a0 def video_capture(self):<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 cap = cv2.VideoCapture(0)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 while True:<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 ret, frame = cap.read()<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 processed = process_frame(frame)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 self.video_queue.put(processed)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u00a0<\/p>\n
\u00a0 \u00a0 def audio_capture(self):<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 p = pyaudio.PyAudio()<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 stream = p.open(format=pyaudio.paInt16, channels=1,<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 rate=16000, input=True,<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 frames_per_buffer=1024)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 while True:<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 data = stream.read(1024)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 features = extract_features(data)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 self.audio_queue.put(features)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u00a0<\/p>\n
\u00a0 \u00a0 def sync_processor(self):<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 while True:<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 # \u52a8\u6001\u65f6\u95f4\u5bf9\u9f50\u7b97\u6cd5<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 video_batch = self.get_video_window()<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 audio_batch = self.get_audio_window()<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 aligned_data = dtw_align(video_batch, audio_batch)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 yield aligned_data<\/p>\n
\u00a0<\/p>\n
2. \u90e8\u7f72\u4f18\u5316\u65b9\u6848<\/p>\n
– \u4f7f\u7528TensorRT\u8fdb\u884c\u6a21\u578b\u91cf\u5316：<\/p>\n
\u00a0 trtexec –onnx=model.onnx –saveEngine=model.engine \\\\<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0–fp16 –workspace=2048<\/p>\n
– \u8fb9\u7f18\u8bbe\u5907\u4f18\u5316：<\/p>\n
\u00a0 import torch_tensorrt<\/p>\n
\u00a0 traced_model = torch.jit.trace(model, example_input)<\/p>\n
\u00a0 trt_model = torch_tensorrt.compile(traced_model,<\/p>\n
\u00a0 \u00a0 \u00a0 inputs= [torch_tensorrt.Input((1, 3, 128, 128),\u00a0<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0torch_tensorrt.Input((1, 100, 120))],<\/p>\n
\u00a0 \u00a0 \u00a0 enabled_precisions= {torch.float16})<\/p>\n
\u00a0<\/p>\n
\u516d\u3001\u8bc4\u4f30\u4e0e\u8c03\u4f18<\/p>\n
\u00a01. \u8bc4\u4f30\u6307\u6807<\/p>\n
from sklearn.metrics import f1_score, confusion_matrix<\/p>\n
\u00a0<\/p>\n
def evaluate(y_true, y_pred):<\/p>\n
\u00a0 \u00a0 acc = (y_true == y_pred).mean()<\/p>\n
\u00a0 \u00a0 f1 = f1_score(y_true, y_pred, average='macro')<\/p>\n
\u00a0 \u00a0 cm = confusion_matrix(y_true, y_pred)<\/p>\n
\u00a0 \u00a0 return {'accuracy': acc, 'f1': f1, 'confusion_matrix': cm}<\/p>\n
\u00a0<\/p>\n
2. \u6a21\u578b\u5206\u6790\u5de5\u5177<\/p>\n
import shap<\/p>\n
\u00a0<\/p>\n
def explain_sample(video, audio):<\/p>\n
\u00a0 \u00a0 explainer = shap.DeepExplainer(model)<\/p>\n
\u00a0 \u00a0 shap_values = explainer.shap_values([video, audio])<\/p>\n
\u00a0 \u00a0\u00a0<\/p>\n
\u00a0 \u00a0 # \u53ef\u89c6\u5316\u5404\u6a21\u6001\u8d21\u732e\u5ea6<\/p>\n
\u00a0 \u00a0 shap.image_plot(shap_values[0], video)<\/p>\n
\u00a0 \u00a0 shap.summary_plot(shap_values[1], audio)<\/p>\n
\u00a0<\/p>\n
\u4e03\u3001\u7cfb\u7edf\u96c6\u6210\u65b9\u6848<\/p>\n
1. \u670d\u52a1\u7aef\u67b6\u6784<\/p>\n
from fastapi import FastAPI<\/p>\n
from pydantic import BaseModel<\/p>\n
app = FastAPI()<\/p>\n
class Request(BaseModel):<\/p>\n
\u00a0 \u00a0 video_url: str<\/p>\n
\u00a0 \u00a0 audio_url: str<\/p>\n
@app.post("\/analyze")<\/p>\n
async def analyze(data: Request):<\/p>\n
\u00a0 \u00a0 video = download_and_process(data.video_url)<\/p>\n
\u00a0 \u00a0 audio = process_audio(data.audio_url)<\/p>\n
\u00a0 \u00a0 with torch.no_grad():<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 prediction = model(video, audio)<\/p>\n
\u00a0 \u00a0 return {"emotion": class_names[prediction.argmax()]}<\/p>\n
\u00a02. \u524d\u7aef\u754c\u9762\u793a\u4f8b<\/p>\n
\/\/ React\u7ec4\u4ef6\u793a\u4f8b<\/p>\n
function EmotionDetector() {<\/p>\n
\u00a0 const [result, setResult] = useState(null);<\/p>\n
\u00a0 const handleUpload = async (files) => {<\/p>\n
\u00a0 \u00a0 const formData = new FormData();<\/p>\n
\u00a0 \u00a0 formData.append('video', files[0]);<\/p>\n
\u00a0 \u00a0 formData.append('audio', files[1]);<\/p>\n
\u00a0 \u00a0 const res = await fetch('\/analyze', {<\/p>\n
\u00a0 \u00a0 \u00a0 method: 'POST',<\/p>\n
\u00a0 \u00a0 \u00a0 body: formData<\/p>\n
\u00a0 \u00a0 });<\/p>\n
\u00a0 \u00a0 setResult(await res.json());<\/p>\n
\u00a0 };<\/p>\n
\u00a0 return (<\/p>\n
\u00a0 \u00a0 <div><\/p>\n
\u00a0 \u00a0 \u00a0 <input type="file" onChange={e => handleUpload(e.target.files)} \/><\/p>\n
\u00a0 \u00a0 \u00a0 {result && <EmotionChart data={result}\/>}<\/p>\n
\u00a0 \u00a0 <\/div><\/p>\n
\u00a0 );<\/p>\n
}<\/p>\n
\u516b\u3001\u6311\u6218\u89e3\u51b3\u65b9\u6848<\/p>\n
1. \u6a21\u6001\u5f02\u6b65\u95ee\u9898：<\/p>\n
\u00a0 \u00a0– \u91c7\u7528\u53cc\u7f13\u51b2\u961f\u5217+\u52a8\u6001\u65f6\u95f4\u89c4\u6574<\/p>\n
\u00a0 \u00a0– \u8bbe\u7f6e\u6700\u5927\u7b49\u5f85\u65f6\u5ef6（200ms），\u8d85\u65f6\u4f7f\u7528\u63d2\u503c\u8865\u507f<\/p>\n
2. \u566a\u58f0\u5904\u7406：<\/p>\n
\u00a0 \u00a0def denoise_audio(audio):<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0return nr.reduce_noise(y=audio, sr=16000,\u00a0<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0stationary=True,<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0prop_decrease=0.8)\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u00a0<\/p>\n
\u00a0 \u00a0def enhance_video(frame):<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0clahe = cv2.createCLAHE(clipLimit=3.0, tileGridSize=(8,8))<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0return clahe.apply(frame)<\/p>\n
3. \u8d44\u6e90\u4f18\u5316：<\/p>\n
\u00a0 \u00a0– \u4f7f\u7528\u6a21\u578b\u84b8\u998f\u6280\u672f：<\/p>\n
\u00a0 \u00a0distiller = Distiller(teacher=teacher_model, student=student_model)<\/p>\n
\u00a0 \u00a0distiller.train_with_distillation(train_loader,\u00a0<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0alpha=0.3,\u00a0<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0temperature=4)<\/p>\n
\u603b\u7ed3：<\/p>\n
\u8be5\u65b9\u6848\u5b8c\u6574\u8986\u76d6\u4e86\u4ece\u6570\u636e\u91c7\u96c6\u5230\u90e8\u7f72\u7684\u5168\u6d41\u7a0b，\u91cd\u70b9\u89e3\u51b3\u4e86\u591a\u6a21\u6001\u7cfb\u7edf\u4e2d\u7684\u5173\u952e\u6311\u6218\u3002\u5b9e\u9645\u90e8\u7f72\u65f6\u53ef\u6839\u636e\u786c\u4ef6\u8d44\u6e90\u8c03\u6574\u6a21\u578b\u590d\u6742\u5ea6，\u63a8\u8350\u4f7f\u7528NVIDIA Jetson\u7cfb\u5217\u8bbe\u5907\u8fdb\u884c\u8fb9\u7f18\u90e8\u7f72\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
\u6587\u7ae0\u6d4f\u89c8\u9605\u8bfb1.8k\u6b21\uff0c\u70b9\u8d5e10\u6b21\uff0c\u6536\u85cf14\u6b21\u3002return out[:, -1] # \u53d6\u6700\u540e\u65f6\u523b\u8f93\u51fa\u3002- \u89c6\u9891\u6570\u636e\uff1aFER2013\uff08\u9759\u6001\u56fe\u50cf\uff09\u3001RAVDESS\uff08\u52a8\u6001\u89c6\u9891\uff09- \u81ea\u5b9a\u4e49\u91c7\u96c6\uff1a\u4f7f\u7528OpenCV+PyAudio\u5b9e\u73b0\u540c\u6b65\u91c7\u96c6\u3002- \u8bbe\u7f6e\u6700\u5927\u7b49\u5f85\u65f6\u5ef6\uff08200ms\uff09\uff0c\u8d85\u65f6\u4f7f\u7528\u63d2\u503c\u8865\u507f\u3002- \u97f3\u9891\u6570\u636e\uff1aCREMA-D\u3001IEMOCAP\u3002- \u5206\u9636\u6bb5\u8bad\u7ec3\uff1a\u5148\u5355\u6a21\u6001\u9884\u8bad\u7ec3\uff0c\u518d\u8054\u5408\u5fae\u8c03\u3002A[\u6570\u636e\u91c7\u96c6] –> B[\u9884\u5904\u7406\u6a21\u5757]- \u52a8\u6001\u65f6\u95f4\u89c4\u6574\uff08DTW\uff09\u5bf9\u9f50\u97f3\u89c6\u9891\u5e8f\u5217\u3002- \u91c7\u7528\u53cc\u7f13\u51b2\u961f\u5217+\u52a8\u6001\u65f6\u95f4\u89c4\u6574\u3002- \u89c6\u89c9\uff1a\u968f\u673a\u906e\u6321\u3001\u8272\u5f69\u6296\u52a8\u3002- \u97f3\u9891\uff1a\u6dfb\u52a0\u566a\u58f0\u3001\u65f6\u79fb\u53d8\u6362\u3002_python \u591a\u6a21\u6001\u8bc6\u522b\u7684\u7b56\u7565<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[81,1395,214],"topic":[],"class_list":["post-31016","post","type-post","status-publish","format-standard","hentry","category-server","tag-python","tag-1395","tag-214"],"yoast_head":"\n