iicd/Graph-DiT
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

make data to use 1 rather than 2 size list

Browse Source
This commit is contained in:
mhz 2024-09-08 23:53:56 +02:00
parent 5dccf590e7
commit 297261d666
2 changed files with 18 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
graph_dit/models/conditions.py
View File

@ -76,6 +76,8 @@ class CategoricalEmbedder(nn.Module):
embeddings = embeddings + noise embeddings = embeddings + noise
return embeddings return embeddings
# 相似的condition cluster起来
# size
class ClusterContinuousEmbedder(nn.Module): class ClusterContinuousEmbedder(nn.Module):
def __init__(self, input_size, hidden_size, dropout_prob): def __init__(self, input_size, hidden_size, dropout_prob):
super().__init__() super().__init__()
@ -108,6 +110,8 @@ class ClusterContinuousEmbedder(nn.Module):
if drop_ids is not None: if drop_ids is not None:
embeddings = torch.zeros((labels.shape[0], self.hidden_size), device=labels.device) embeddings = torch.zeros((labels.shape[0], self.hidden_size), device=labels.device)
# print(labels[~drop_ids].shape)
# torch.Size([1200])
embeddings[~drop_ids] = self.mlp(labels[~drop_ids]) embeddings[~drop_ids] = self.mlp(labels[~drop_ids])
embeddings[drop_ids] += self.embedding_drop.weight[0] embeddings[drop_ids] += self.embedding_drop.weight[0]
else: else:

23
graph_dit/models/transformer.py
View File

@ -17,20 +17,22 @@ class Denoiser(nn.Module):
num_heads=16, num_heads=16,
mlp_ratio=4.0, mlp_ratio=4.0,
drop_condition=0.1, drop_condition=0.1,
Xdim=118, Xdim=7,
Edim=5, Edim=2,
ydim=3, ydim=1,
task_type='regression', task_type='regression',
): ):
super().__init__() super().__init__()
print(f"Denoiser, xdim: {Xdim}, edim: {Edim}, ydim: {ydim}, hidden_size: {hidden_size}, depth: {depth}, num_heads: {num_heads}, mlp_ratio: {mlp_ratio}, drop_condition: {drop_condition}")
self.num_heads = num_heads self.num_heads = num_heads
self.ydim = ydim self.ydim = ydim
self.x_embedder = nn.Linear(Xdim + max_n_nodes * Edim, hidden_size, bias=False) self.x_embedder = nn.Linear(Xdim + max_n_nodes * Edim, hidden_size, bias=False)
self.t_embedder = TimestepEmbedder(hidden_size) self.t_embedder = TimestepEmbedder(hidden_size)
#
self.y_embedding_list = torch.nn.ModuleList() self.y_embedding_list = torch.nn.ModuleList()
self.y_embedding_list.append(ClusterContinuousEmbedder(2, hidden_size, drop_condition)) self.y_embedding_list.append(ClusterContinuousEmbedder(1, hidden_size, drop_condition))
for i in range(ydim - 2): for i in range(ydim - 2):
if task_type == 'regression': if task_type == 'regression':
self.y_embedding_list.append(ClusterContinuousEmbedder(1, hidden_size, drop_condition)) self.y_embedding_list.append(ClusterContinuousEmbedder(1, hidden_size, drop_condition))
@ -88,6 +90,8 @@ class Denoiser(nn.Module):
# print("Denoiser Forward") # print("Denoiser Forward")
# print(x.shape, e.shape, y.shape, t.shape, unconditioned) # print(x.shape, e.shape, y.shape, t.shape, unconditioned)
# torch.Size([1200, 8, 7]) torch.Size([1200, 8, 8, 2]) torch.Size([1200, 2]) torch.Size([1200, 1]) False
# print(y)
force_drop_id = torch.zeros_like(y.sum(-1)) force_drop_id = torch.zeros_like(y.sum(-1))
# drop the nan values # drop the nan values
force_drop_id[torch.isnan(y.sum(-1))] = 1 force_drop_id[torch.isnan(y.sum(-1))] = 1
@ -109,11 +113,12 @@ class Denoiser(nn.Module):
c1 = self.t_embedder(t) c1 = self.t_embedder(t)
# print("C1 after t_embedder") # print("C1 after t_embedder")
# print(c1.shape) # print(c1.shape)
for i in range(1, self.ydim): c2 = self.y_embedding_list[0](y[:,0].unsqueeze(-1), self.training, force_drop_id, t)
if i == 1: # for i in range(1, self.ydim):
c2 = self.y_embedding_list[i-1](y[:, :2], self.training, force_drop_id, t) # if i == 1:
else: # c2 = self.y_embedding_list[i-1](y[:, :2], self.training, force_drop_id, t)
c2 = c2 + self.y_embedding_list[i-1](y[:, i:i+1], self.training, force_drop_id, t) # else:
# c2 = c2 + self.y_embedding_list[i-1](y[:, i:i+1], self.training, force_drop_id, t)
# print("C2 after y_embedding_list") # print("C2 after y_embedding_list")
# print(c2.shape) # print(c2.shape)
# print("C1 + C2") # print("C1 + C2")