Spaces:
Paused
Paused
Update text2vid_torch2.py
Browse files- text2vid_torch2.py +48 -1
text2vid_torch2.py
CHANGED
@@ -167,7 +167,8 @@ class AttnProcessor2_0:
|
|
167 |
hidden_states = hidden_states / attn.rescale_output_factor
|
168 |
|
169 |
return hidden_states
|
170 |
-
|
|
|
171 |
def get_qk(
|
172 |
self, query, key):
|
173 |
r"""
|
@@ -221,7 +222,53 @@ class AttnProcessor2_0:
|
|
221 |
|
222 |
|
223 |
return query, key, dynamic_lambda, key1
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
224 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
225 |
|
226 |
def init_attention_func(unet):
|
227 |
|
|
|
167 |
hidden_states = hidden_states / attn.rescale_output_factor
|
168 |
|
169 |
return hidden_states
|
170 |
+
|
171 |
+
"""
|
172 |
def get_qk(
|
173 |
self, query, key):
|
174 |
r"""
|
|
|
222 |
|
223 |
|
224 |
return query, key, dynamic_lambda, key1
|
225 |
+
"""
|
226 |
+
|
227 |
+
def get_qk(
|
228 |
+
self, query, key):
|
229 |
+
r"""
|
230 |
+
Compute the attention scores.
|
231 |
+
Args:
|
232 |
+
query (`torch.Tensor`): The query tensor.
|
233 |
+
key (`torch.Tensor`): The key tensor.
|
234 |
+
attention_mask (`torch.Tensor`, *optional*): The attention mask to use. If `None`, no mask is applied.
|
235 |
+
Returns:
|
236 |
+
`torch.Tensor`: The attention probabilities/scores.
|
237 |
+
"""
|
238 |
+
q_old = query.clone()
|
239 |
+
k_old = key.clone()
|
240 |
+
dynamic_lambda = None
|
241 |
+
key1 = None
|
242 |
+
|
243 |
+
if self.use_last_attn_slice:
|
244 |
+
if self.last_attn_slice is not None:
|
245 |
+
|
246 |
+
query_list = self.last_attn_slice[0]
|
247 |
+
key_list = self.last_attn_slice[1]
|
248 |
+
|
249 |
+
if query.shape[1] == self.num_frames and query.shape == key.shape:
|
250 |
+
key1 = key.clone()
|
251 |
+
key1[:,:1,:key_list.shape[2]] = key_list[:,:1]
|
252 |
+
dynamic_lambda = torch.tensor([1 + self.LAMBDA * (i/50) for i in range(self.num_frames)]).to(key.dtype).cuda()
|
253 |
|
254 |
+
if q_old.shape == k_old.shape and q_old.shape[1] != self.num_frames:
|
255 |
+
batch_dim = query_list.shape[0] // self.bs
|
256 |
+
all_dim = query.shape[0] // self.bs
|
257 |
+
for i in range(self.bs):
|
258 |
+
# Ensure slice dimensions match
|
259 |
+
target_size = min(query[i*all_dim:(i*all_dim) + batch_dim, :query_list.shape[1], :query_list.shape[2]].size(0),
|
260 |
+
query_list[i*batch_dim:(i+1)*batch_dim].size(0))
|
261 |
+
|
262 |
+
# Assign values from query_list to query
|
263 |
+
query[i*all_dim:(i*all_dim) + target_size, :query_list.shape[1], :query_list.shape[2]] = \
|
264 |
+
query_list[i*batch_dim:i*batch_dim + target_size]
|
265 |
+
|
266 |
+
if self.save_last_attn_slice:
|
267 |
+
self.last_attn_slice = [query, key]
|
268 |
+
self.save_last_attn_slice = False
|
269 |
+
|
270 |
+
return query, key, dynamic_lambda, key1
|
271 |
+
|
272 |
|
273 |
def init_attention_func(unet):
|
274 |
|