ViT 可解释代码无法运行 #74
Description
尝试用可解释代码运行自己训练的ViT_small_patch16_224, 遇到了一些问题，请问有没有什么建议：
1） sg = it.SmoothGradInterpreter(paddle_model, device='cpu')
gradients = sg.interpret(
img_path,
noise_amount=0.1,
visual=True,
labels=None,
save_path=None)
该程序运行出的结果有斑块状噪声
2） 运行以下程序会报错：
tam = it.TAMInterpreter(paddle_model, device='cpu')
heatmap = tam.interpret(
img_path,
start_layer=4,
label=None, # cat
visual=True,
save_path=None)
错误如下：
TAMInterpreter.interpret(self, inputs, start_layer, steps, label, resize_to, crop_to, visual, save_path)
96 data_scaled = data * alpha
97 _, gradients, _ = self.predict_fn(data_scaled, label=label)
---> 99 total_gradients += gradients
101 W_state = np.mean((total_gradients / steps).clip(min=0), axis=1)[:, 0, :].reshape((b, 1, s))
103 tam_explanation = (states * W_state)[:, 0, 1:].reshape((-1, 14, 14))
UFuncTypeError: Cannot cast ufunc 'add' output from dtype('O') to dtype('float64') with casting rule 'same_kind'
3）运行以下程序会报错：
bt = it.BTCVInterpreter(paddle_model, device='gpu:0')
heatmap = bt.interpret(
img_path,
ap_mode = 'head',
start_layer=4,
label=None, # cat
visual=True,
save_path=None)
错误如下：
495 raise ValueError("gradient_of should be one of [logits, probability].")
497 for i, attn in enumerate(block_attns):
--> 498 grad = attn.grad.numpy()
499 block_attns_grads.append(grad)
500 block_attns[i] = attn.numpy()
AttributeError: 'NoneType' object has no attribute 'numpy'