文章目录
1. The number of sizes provided (0) must be greater or equal to the number of dimensions in the tensor(1)
I’m trying to convert a CPU model to GPU using Pytorch, but I’m running into issues. I’m running this on Colab and I’m sure that Pytorch detects a GPU. This is a deep Q network (RL).
I declare my network as: Q = Q_Network(input_size, hidden_size, output_size).to(device)
I ran into an issue when I tried to pass arguments through the network (It expected type cuda but got type cpu) so I add .to(device):
batch = np.array(shuffled_memory[i:i+batch_size])
b_pobs = np.array(batch[:, 0].tolist(), dtype=np.float32).reshape(batch_size, -1)
b_pact = np.array(batch[:, 1].tolist(), dtype=np.int32)
b_reward = np.array(batch[:, 2].tolist(), dtype=np.int32)
b_obs = np.array(batch[:, 3].tolist(), dtype=np.float32).reshape(batch_size, -1)
b_done = np.array(batch[:, 4].tolist(), dtype=np.bool)
q = Q(torch.from_numpy(b_pobs).to(device))
q_ = Q_ast(torch.from_numpy(b_obs).to(device))
maxq = torch.max(q_.data,axis=1)
target = copy.deepcopy(q.data)
for j in range(batch_size):
print(target[j, b_pact[j]].shape) # torch.Size([])
target[j, b_pact[j]] = b_reward[j]+gamma*maxq[j]*(not b_done[j]) #I run into issues here
Here is the error:
RuntimeError: expand(torch.cuda.FloatTensor{[50]}, size=[]): the number of sizes provided (0) must be greater or equal to the number of dimensions in the tensor (1)
target[j, b_pact[j]] is a single element of the tensor (a scalar, hence size of torch.Size([])). If you want to assign anything to it, the right hand side can only be a scalar. That is not the case, as one of the terms is a tensor with 1 dimension (a vector), namely your maxq[j].
由于target[j, b_pact[j]]的值是单个元素,也就是标量,如果要进行赋值的话,=右边也应该是一个标量,但是事实却是右边是一个矢量;例如 maxq[j]这就是一个矢量;
When specifying a dimension dim (axis is treated as a synonym) to torch.max, it will return a named tuple of (values, indices), where values contains the maximum values and indices the location of each of the maximum values (equivalent to argmax).
maxq[j] is not indexing into the maximum values, but rather the tuple of (values, indices). If you only want the values you can use one of the following to get the values out of the tuple (all of them are equivalent, you can use whichever you prefer):
根据需求,要使得 对 maxq[j] 提取到的是一个标量,这个时候可以使用一些几种方法
1. Destructure/unpack and ignore the indices
maxq, _ = torch.max(q_.data,axis=1)
2. Access first element of the tuple
maxq = torch.max(q_.data,axis=1)[0]
3. Access `values` of the named tuple
maxq = torch.max(q_.data,axis=1).values
建议
for循环和计算结果不适合 GPU,因为它们可以执行大量并行计算,所以它们大多速度非常快。您需要避免 for 循环,因为它们无法被 PyTorch 优化,并用矢量化计算替换它们;
替换方法:
for j in range(batch_size):
print(target[j, b_pact[j]].shape) # torch.Size([])
target[j, b_pact[j]] = b_reward[j]+gamma*maxq[j]*(not b_done[j]) #I run into issues here
这两种方法等价
target[torch.arange(batch_size), b_pact] = b_reward + gamma * maxq * (~b_done).