mirror of
https://github.com/Vision-CAIR/MiniGPT-4.git
synced 2025-04-21 21:20:46 +00:00
393 lines
17 KiB
Python
393 lines
17 KiB
Python
import copy
|
||
import pickle
|
||
import torch
|
||
import torch.nn as nn
|
||
import torch.nn.functional as F
|
||
|
||
|
||
class MoELayer(nn.Module):
|
||
def __init__(self, hidden_size, expert, gate, num_experts, route_method, topk=1, use_balance_loss=True, weight_type='l2_norm'):
|
||
# remove hash list
|
||
nn.Module.__init__(self)
|
||
self.num_experts = num_experts
|
||
self.experts = nn.ModuleList([copy.deepcopy(expert) for i in range(num_experts)])
|
||
self.route_method = route_method
|
||
self.topk = topk
|
||
self.use_balance_loss = use_balance_loss
|
||
self.weight_type = weight_type
|
||
|
||
if route_method in ["gate-token", "gate-sentence"]:
|
||
self.gate = gate
|
||
else:
|
||
raise KeyError("Routing method not supported.")
|
||
|
||
|
||
def _forward_gate_sentence(self, x, attention_mask):
|
||
"""
|
||
x: query_attention_output , torch.Size([bz, 32, 768])
|
||
attention_mask: torch.ones([bz, 32])
|
||
|
||
### Notice:
|
||
the raw version of expert_attention_mask is the extended_attention_mask,
|
||
which will be add to attention_score directly
|
||
the values of extended_attention_mask are -0.0 or -10000
|
||
it should be adjust to 1/0 version to be processed by experts
|
||
"""
|
||
attention_mask = torch.ones(attention_mask.shape[0], attention_mask.shape[1]).to(x.device)
|
||
x_masked = x * attention_mask.unsqueeze(-1) # torch.Size([bz, 32, 768])
|
||
x_average = x_masked.sum(1) / attention_mask.unsqueeze(-1).sum(1) # torch.Size([bz, 768])
|
||
logits_gate = self.gate(x_average) # torch.Size([bz, num_experts])
|
||
prob_gate = F.softmax(logits_gate, dim=-1) # torch.Size([bz, num_experts])
|
||
select_prob_gate, gate = torch.topk(prob_gate, self.topk, dim=1) # gate, 每个样本被分配的expert: torch.Size([bz, topk])
|
||
|
||
# 这里用l2 norm 去加权
|
||
if self.weight_type == 'l2_norm':
|
||
# normalized_tensor = torch.nn.functional.normalize(select_prob_gate, p=2, dim=0) # L2 Normalization torch.Size([bz, topk])
|
||
normalized_tensor = select_prob_gate
|
||
|
||
num_sentences = F.one_hot(gate, self.num_experts).sum(1).gt(0).sum(0) # 每个expert被分配的样本数 torch.Size([num_expert])
|
||
gate_load = num_sentences.clone()
|
||
|
||
# forward experts
|
||
def forward_expert(input_x, expert_idx):
|
||
input_x = self.experts[expert_idx].forward(input_x)
|
||
return input_x
|
||
|
||
result_lst = list()
|
||
for i in range(self.topk):
|
||
# top1、top2... 分别为一组，进行gate分组之后过expert，然后乘以概率后相加
|
||
tmp_gate = gate[:,i]
|
||
tmp_prob = normalized_tensor[:,i].unsqueeze(-1).unsqueeze(-1)
|
||
order = tmp_gate.argsort(0)
|
||
num_sentences_t = F.one_hot(tmp_gate, self.num_experts).gt(0).sum(0)
|
||
x1 = x[order] # reorder according to expert number
|
||
x1 = x1.split(num_sentences_t.tolist(), dim=0) # a list of length self.num_experts
|
||
|
||
result = []
|
||
for i in range(self.num_experts):
|
||
if x1[i].size(0) > 0:
|
||
result.append(forward_expert(x1[i], i))
|
||
result = torch.vstack(result)
|
||
result = result[order.argsort(0)] # restore original order
|
||
|
||
# result_lst.append(result * tmp_prob) # result * prob
|
||
result_lst.append(result) # result * prob
|
||
|
||
moe_result = sum(result_lst)
|
||
print('Layer Qformer MoE: \n',prob_gate)
|
||
return moe_result, select_prob_gate, gate
|
||
|
||
|
||
def forward(self, x, attention_mask):
|
||
if self.route_method == "gate-token":
|
||
x, balance_loss, gate_load = self._forward_gate_token(x)
|
||
elif self.route_method == "gate-sentence":
|
||
if x.size(0) == 1:
|
||
x, balance_loss, gate_load = self._forward_sentence_single_expert(x, attention_mask)
|
||
else:
|
||
x, balance_loss, gate_load = self._forward_gate_sentence(x, attention_mask)
|
||
elif self.route_method == "gate-sentence-post":
|
||
x, balance_loss, gate_load = self._forward_gate_sentence_post(x, attention_mask)
|
||
else:
|
||
raise KeyError("Routing method not supported.")
|
||
|
||
return x, balance_loss, gate_load
|
||
|
||
|
||
class RouteMoELayer(nn.Module):
|
||
def __init__(self, hidden_size, expert, gate, num_experts, num_beams=2, layer_judge=None, route_method="pre-route"):
|
||
# remove hash list
|
||
nn.Module.__init__(self)
|
||
self.num_experts = num_experts
|
||
self.experts = nn.ModuleList([copy.deepcopy(expert) for i in range(num_experts)])
|
||
self.num_beams = num_beams
|
||
self.hidden_size = hidden_size
|
||
self.layer_judge = layer_judge
|
||
|
||
self.route_method = route_method
|
||
if self.route_method == "pre-route":
|
||
self.gate = nn.Linear(hidden_size, num_experts, bias=False).float()
|
||
elif self.route_method == "post-route":
|
||
# gate = nn.Linear(hidden_size, 1, bias=False).float()
|
||
self.gates = nn.ModuleList([copy.deepcopy(gate) for i in range(num_experts)])
|
||
|
||
def forward_gate(self, x):
|
||
"""
|
||
x : torch.Size([bz*num_beams, 32, 768]) or torch.Size([bz, 32, 768])
|
||
prob_gate : torch.Size([bz*num_beams, num_experts]) or torch.Size([bz, num_experts])
|
||
"""
|
||
attention_mask = torch.ones(x.shape[0], x.shape[1]).to(x.device)
|
||
x_masked = x * attention_mask.unsqueeze(-1) # torch.Size([bz*num_beams, 32, 768])
|
||
x_average = x_masked.sum(1) / attention_mask.unsqueeze(-1).sum(1) # torch.Size([bz*num_beams, 768])
|
||
logits_gate = self.gate(x_average) # torch.Size([bz*num_beams, num_experts])
|
||
prob_gate = F.softmax(logits_gate, dim=-1) # torch.Size([bz*num_beams, num_experts])
|
||
return prob_gate
|
||
|
||
def beam_search(self, current_scores_log, beam_scores, expert_route, batch_size):
|
||
import pdb;pdb.set_trace()
|
||
if self.layer_judge=='first' and self.route_method=='pre-route':
|
||
assert beam_scores==None and expert_route==None
|
||
current_scores = torch.exp(current_scores_log)
|
||
topk_values, gate = torch.topk(current_scores, self.num_beams, dim=1) # gate, 每个样本被分配的expert: torch.Size([bz, topk])
|
||
beam_scores = topk_values.view(self.num_beams * batch_size) # torch.Size([bz * num_beams])
|
||
expert_route = gate.view(self.num_beams * batch_size).unsqueeze(1) # torch.Size([bz * num_beams,1])
|
||
beam_idx = None
|
||
else:
|
||
if self.layer_judge=='first' and self.route_method == 'post-route':
|
||
batch_size = batch_size
|
||
next_scores_raw1 = torch.exp(current_scores_log) # torch.Size([bz, num_experts])
|
||
else:
|
||
batch_size = int(batch_size // self.num_beams)
|
||
next_scores_raw = current_scores_log + torch.log(beam_scores).unsqueeze(1) # torch.Size([4*3, 5]) # 取log 之后，可以直接相加概率
|
||
next_scores_exp = torch.exp(next_scores_raw)
|
||
next_scores_raw1 = next_scores_exp.view(
|
||
batch_size, self.num_beams * self.num_experts
|
||
) # torch.Size([bz, num_beams*num_experts])
|
||
|
||
next_scores, next_experts = torch.topk(next_scores_raw1, self.num_beams, dim=1, largest=True, sorted=True)
|
||
# next_scores torch.Size([bz, num_beams])
|
||
# next_tokens torch.Size([bz, num_beams])
|
||
print(next_scores_raw1)
|
||
print(next_scores)
|
||
print(next_experts)
|
||
|
||
next_batch_beam = list()
|
||
for batch_idx in range(batch_size):
|
||
next_sent_beam = list()
|
||
for rank, (expert_id, expert_score) in enumerate(
|
||
zip(next_experts[batch_idx], next_scores[batch_idx])
|
||
):
|
||
expert_id = expert_id.item()
|
||
beam_id = expert_id // self.num_experts
|
||
ex_id = expert_id % self.num_experts
|
||
effective_beam_id = batch_idx*self.num_beams + beam_id
|
||
|
||
next_sent_beam.append((expert_score, ex_id, effective_beam_id))
|
||
next_batch_beam.extend(next_sent_beam)
|
||
|
||
import pdb;pdb.set_trace()
|
||
|
||
if self.layer_judge=='first' and self.route_method == 'post-route':
|
||
beam_scores = next_scores.view(self.num_beams * batch_size) # torch.Size([bz * num_beams])
|
||
expert_route = next_experts.view(self.num_beams * batch_size)
|
||
beam_scores = beam_scores.new([x[0] for x in next_batch_beam])
|
||
beam_experts = expert_route.new([x[1] for x in next_batch_beam]).unsqueeze(-1)
|
||
beam_idx = expert_route.new([int(x[2]/self.num_beams) for x in next_batch_beam])
|
||
expert_route = beam_experts
|
||
else:
|
||
beam_scores = beam_scores.new([x[0] for x in next_batch_beam])
|
||
beam_experts = expert_route[:,-1].new([x[1] for x in next_batch_beam])
|
||
beam_idx = expert_route[:,-1].new([x[2] for x in next_batch_beam])
|
||
pre_route = expert_route[beam_idx,:]
|
||
expert_route = torch.cat([pre_route, beam_experts.unsqueeze(1)], dim=-1)
|
||
|
||
import pdb;pdb.set_trace()
|
||
|
||
return beam_scores, expert_route, beam_idx
|
||
|
||
|
||
def forward_expert_ffn(self, x, expert_select, beam_scores):
|
||
"""
|
||
x_repeat : [bz*num_beams, 32,768]
|
||
expert_select : [bz*num_beams]
|
||
"""
|
||
# add_1212 l2_normalization
|
||
# normalized_tensor = torch.nn.functional.normalize(beam_scores, p=2, dim=0) # L2 Normalization torch.Size([bz, topk])
|
||
# tmp_prob = normalized_tensor.unsqueeze(-1).unsqueeze(-1)
|
||
|
||
outputs = list()
|
||
for i in range(x.shape[0]):
|
||
output_x = self.experts[expert_select[i]].forward(x[i])
|
||
outputs.append(output_x.unsqueeze(0))
|
||
candidate_output = torch.cat(outputs)
|
||
|
||
# candidate_output = candidate_output * tmp_prob
|
||
return candidate_output # torch.Size([bz*num_beams, 32, 768])
|
||
|
||
|
||
def forward_pre_route(self, x, beam_scores, expert_route, use_log=True):
|
||
|
||
current_scores = self.forward_gate(x) # [bz*num_beams, 5]
|
||
|
||
if use_log:
|
||
current_scores_log = torch.log(current_scores) # 取log之后可以直接相加
|
||
else:
|
||
current_scores_log = current_scores
|
||
|
||
batch_size, num_tokens = x.shape[0], x.shape[1]
|
||
beam_scores, expert_route, _ = self.beam_search(current_scores_log, beam_scores, expert_route, batch_size)
|
||
|
||
current_expert_select = expert_route[:,-1]
|
||
|
||
if self.layer_judge=='first': # expand first dim to batch_size * num_beams
|
||
replicated_tensor = x.unsqueeze(1).expand(batch_size, self.num_beams, num_tokens, self.hidden_size)
|
||
x = replicated_tensor.contiguous().view(-1, num_tokens, self.hidden_size) # [bz*num_beams, 32,768]
|
||
|
||
candidate_output = self.forward_expert_ffn(x, current_expert_select, beam_scores) # [bz*num_beams, 32,768]
|
||
|
||
return candidate_output, beam_scores, expert_route
|
||
|
||
|
||
def forward_post_route(self, x, beam_scores, expert_route, use_log=True):
|
||
|
||
# if self.layer_judge=='first': # expand first dim to batch_size * num_beams
|
||
# batch_size, num_tokens = x.shape[0], x.shape[1]
|
||
# replicated_tensor = x.unsqueeze(1).expand(batch_size, self.num_beams, num_tokens, self.hidden_size)
|
||
# x = replicated_tensor.contiguous().view(-1, num_tokens, self.hidden_size) # [bz*num_beams, 32,768]
|
||
|
||
attention_mask = torch.ones(x.shape[0], x.shape[1]).to(x.device)
|
||
x_masked = x * attention_mask.unsqueeze(-1) # torch.Size([bz, 32, 768])
|
||
|
||
def forward_expert(input_x, expert_idx):
|
||
output_x = self.experts[expert_idx].forward(input_x)
|
||
return output_x
|
||
|
||
outputs = list()
|
||
logits_gate_lst = list()
|
||
for expert_idx in range(self.num_experts):
|
||
output_x = forward_expert(x_masked, expert_idx)
|
||
outputs.append(output_x.unsqueeze(0))
|
||
output_x_aver = output_x.sum(1) / attention_mask.unsqueeze(-1).sum(1) # torch.Size([bz*num_beam, 768])
|
||
gate_acore = self.gates[expert_idx](output_x_aver)
|
||
logits_gate_lst.append(gate_acore)
|
||
candidate_output = torch.cat(outputs) # torch.Size([num_expert, bz*num_beam, 32, 768])
|
||
logits_gate = torch.cat(logits_gate_lst,dim=1)# torch.Size([bz*num_beam, num_expert])
|
||
current_scores = F.softmax(logits_gate, dim=-1) # torch.Size([bz*num_beam, num_experts])
|
||
|
||
if use_log:
|
||
current_scores_log = torch.log(current_scores) # 取log之后可以直接相加
|
||
else:
|
||
current_scores_log = current_scores
|
||
|
||
import pdb;pdb.set_trace()
|
||
|
||
batch_size = x.shape[0] # bz*num_beam
|
||
beam_scores, expert_route, beam_idx = self.beam_search(current_scores_log, beam_scores, expert_route, batch_size)
|
||
# beam_scores torch.Size([bz*num_beam])
|
||
# expert_route torch.Size([bz*num_beam, layer_n])
|
||
current_select_expert = expert_route[:,-1]
|
||
|
||
output = list()
|
||
for i in range(beam_idx.shape[0]):
|
||
b_idx = beam_idx[i]
|
||
ex_idx = current_select_expert[i]
|
||
ex_out = candidate_output[ex_idx, b_idx, :,:]
|
||
output.append(ex_out.unsqueeze(0))
|
||
|
||
final_output = torch.concat(output, dim=0)
|
||
|
||
return final_output, beam_scores, expert_route, beam_idx
|
||
|
||
|
||
def forward(self, x, attention_mask, beam_scores, expert_route, use_log=True):
|
||
"""
|
||
if first_layer: x [bz, 32, 768]
|
||
else: x [bz*num_beams, 32, 768]
|
||
|
||
"""
|
||
if self.route_method == 'pre-route':
|
||
candidate_output, beam_scores, expert_route, _ = self.forward_pre_route(x, beam_scores, expert_route, use_log=True)
|
||
elif self.route_method == "post-route":
|
||
candidate_output, beam_scores, expert_route, beam_idx = self.forward_post_route(x, beam_scores, expert_route, use_log=True)
|
||
|
||
return candidate_output, beam_scores, expert_route, beam_idx
|
||
|
||
|
||
if __name__ == '__main__':
|
||
|
||
import sys
|
||
sys.path.append("/mnt/pfs-guan-ssai/nlu/wanghanzi/multimodal/PromptMoE")
|
||
from minigpt4.models.QformerRouteMoE import BertConfig
|
||
from minigpt4.models.QformerRouteMoE import FeedForward
|
||
|
||
from minigpt4.models.moe.utils import (
|
||
use_experts,
|
||
moe_layer_judge,
|
||
)
|
||
vision_width = 1408
|
||
cross_attention_freq = 2
|
||
num_query_token = 32
|
||
# init_QformerMoE
|
||
config = BertConfig.from_pretrained("/mnt/pfs-guan-ssai/nlu/wanghanzi/models/bert-base-uncased")
|
||
config.encoder_width = vision_width
|
||
# insert cross-attention layer every other block
|
||
config.add_cross_attention = True
|
||
config.cross_attention_freq = cross_attention_freq
|
||
config.query_length = num_query_token
|
||
config.moebert_expert_num = 3
|
||
config.moebert_num_beams = 3
|
||
config.moebert_route_method = 'gate-sentence'
|
||
config.moe_topk = 2
|
||
config.use_balance_loss = False
|
||
config.moe_weight_type = 'l2_norm'
|
||
|
||
batch_size = 4
|
||
x = torch.randn(batch_size, 32, 768)
|
||
beam_scores, expert_route = None, None
|
||
|
||
x1 = x
|
||
x2 = x
|
||
beam_scores1, expert_route1 = None, None
|
||
|
||
for layer_num in [6, 8, 10]:
|
||
layer_judge = moe_layer_judge(layer_num)
|
||
ffn = FeedForward(config)
|
||
gate = nn.Linear(768, config.moebert_expert_num, bias=False).float()
|
||
|
||
# experts = RouteMoELayer(
|
||
# hidden_size=768,
|
||
# expert=ffn,
|
||
# gate = gate,
|
||
# num_experts=config.moebert_expert_num,
|
||
# num_beams=config.moebert_num_beams,
|
||
# layer_judge = layer_judge,
|
||
# route_method = "pre-route"
|
||
# )
|
||
# layer_output = experts(x, None, beam_scores, expert_route)
|
||
# hidden_states1, beam_scores, expert_route,_ = layer_output
|
||
|
||
# print(beam_scores)
|
||
# print(expert_route)
|
||
|
||
gate1 = nn.Linear(768, 1, bias=False).float()
|
||
experts_post = RouteMoELayer(
|
||
hidden_size=768,
|
||
expert=ffn,
|
||
gate = gate1,
|
||
num_experts=config.moebert_expert_num,
|
||
num_beams=config.moebert_num_beams,
|
||
layer_judge = layer_judge,
|
||
route_method = "post-route"
|
||
)
|
||
layer_output = experts_post(x1, None, beam_scores1, expert_route1, False)
|
||
hidden_states2, beam_scores1, expert_route1, beam_idx = layer_output
|
||
|
||
print(beam_scores1)
|
||
print(expert_route1)
|
||
print(beam_idx)
|
||
|
||
# experts_moe = MoELayer(
|
||
# hidden_size=config.hidden_size,
|
||
# expert=ffn,
|
||
# gate=gate,
|
||
# num_experts=config.moebert_expert_num,
|
||
# route_method=config.moebert_route_method,
|
||
# topk=config.moe_topk,
|
||
# use_balance_loss=config.use_balance_loss,
|
||
# weight_type=config.moe_weight_type,
|
||
# )
|
||
# attn_mask = torch.ones([batch_size, 32])
|
||
# layer_output = experts_moe(x2, attn_mask)
|
||
# hidden_states3, select_prob_gate, gate_load,_ = layer_output
|
||
|
||
# print(select_prob_gate)
|
||
# print(gate_load)
|
||
|
||
|
||
# x = hidden_states1
|
||
x1 = hidden_states2
|
||
# x2 = hidden_states3
|
||
|
||
print("------------------------------------")
|
||
|