[LAB 06 · 颜色干预:视图、投影与读出](/course/notebooks/7c4f531113d7473abc174e2e014e1340.html) # 从一张图到两个视图 · 实现代码 [独立阅读](/course/notebook-blocks/c62130014946402d998cecf03bd50f05.html) 主线阅读与优先实验 真实RGB对比训练、唯一指定增强变化、h/z形状与颜色读出 展开完整实现代码 ``` def make_data(seed=SEED): rng=np.random.default_rng(seed) yy,xx=np.mgrid[:16,:16] images=[]; shape=[]; color=[]; split=[] # Stratified before view construction. One independently generated source per ID. for s in range(3): for c in range(2): for i in range(140): cx,cy=rng.uniform(6,10,2); r=rng.uniform(3.0,4.7) dx,dy=xx-cx,yy-cy if s==0: mask=dx*dx+dy*dy<=r*r elif s==1: mask=np.maximum(abs(dx),abs(dy))<=r else: mask=abs(dx)+abs(dy)<=r*1.35 # Red and blue have equal channel sums: grayscale erases hue by design. rgb=np.array([.9,.15,.15] if c==0 else [.15,.15,.9]) bg=rng.uniform(.015,.075,(16,16,1)) im=np.where(mask[...,None],rgb*rng.uniform(.8,1),bg) im+=rng.normal(0,.012,im.shape) images.append(np.clip(im,0,1).transpose(2,0,1)) shape.append(s); color.append(c); split.append(0 if i<100 else 1) x=np.asarray(images,dtype='float32'); split=np.asarray(split) ids=np.arange(len(x)); train=ids[split==0]; test=ids[split==1] assert set(train).isdisjoint(test) assert len({hashlib.sha256(v.tobytes()).hexdigest() for v in x})==len(x) return torch.from_numpy(x),np.asarray(shape),np.asarray(color),train,test def views(x, destroy_color=False): # Independent transforms of each source; no labels/scene attributes are read. a=x.clone(); n=len(a) shifts=torch.randint(-1,2,(n,2)) for i,(dy,dx) in enumerate(shifts.tolist()): a[i]=torch.roll(a[i],(dy,dx),(-2,-1)) a=a*torch.empty(n,1,1,1).uniform_(.8,1.2) a=a+torch.randn_like(a)*.025 # The ablation changes only the hue swap. Every run consumes this RNG draw. swap=torch.rand(n)<.5 if destroy_color: a[swap]=a[swap][:,[2,1,0]] return a.clamp(0,1) ```