forked from jiangwx/DFQ
-
Notifications
You must be signed in to change notification settings - Fork 0
/
improve_dfq.py
371 lines (310 loc) · 14.8 KB
/
improve_dfq.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
import torch
import torch.nn as nn
import torch.nn.functional as F
from utils.quantize import QuantNConv2d, QuantNLinear, QuantConv2d, QuantLinear, QConv2d, QLinear
from PyTransformer.transformers.torchTransformer import TorchTransformer
from utils.layer_transform import set_quant_minmax, replace_op, restore_op
import numpy as np
import copy
from tensorboardX import SummaryWriter
import time
class GradHook():
def __init__(self, weight, scale=None, scale_prev=None, merge_scale=None, merge_scale_prev=None):
self.weight = weight
self.scale = scale
self.scale_prev = scale_prev
self.merge_scale = merge_scale
self.merge_scale_prev = merge_scale_prev
self.update_mask()
def update_mask(self):
weight = self.weight.detach()
if self.scale_prev is not None:
weight = self.merge_scale_prev(weight, self.scale_prev)
if self.scale is not None:
weight, _ = self.merge_scale(weight, None, self.scale)
# print("Weight max: {}, min: {}, range: {}".format(weight.max(), weight.min(), weight.max()-weight.min()))
std = torch.sqrt(torch.var(weight))
mean = weight.mean()
# mask = (weight > (mean + 2 * std)).float() + (weight < (mean - 2 * std)).float()
mask = (weight < (mean + 2 * std)).long() + (weight > (mean - 2 * std)).long()
weight[mask] = 0
self.mask = torch.abs(weight) / torch.abs(weight).max()
def get_weight_scaled(self):
weight = self.weight
if self.scale_prev is not None:
# weight = self.merge_scale_prev(weight, torch.clamp(self.scale_prev, max=1))
weight = self.merge_scale_prev(weight, self.scale_prev)
if self.scale is not None:
# weight, _ = self.merge_scale(weight, None, torch.clamp(self.scale, max=1))
weight, _ = self.merge_scale(weight, None, self.scale)
return weight
def hook_mask_grad_tensor(self, grad):
return grad
# print(grad.shape, self.mask.shape, np.count_nonzero(np.array(grad.shape) != np.array(self.mask.shape)))
if np.count_nonzero(np.array(grad.shape) != np.array(self.mask.shape)) != 0:
# mask = self.mask.view(self.mask.size(0), -1).max(-1)[0].view(-1, 1, 1, 1)
mask = self.mask.view(self.mask.size(0), -1).mean(-1).view(-1, 1, 1, 1)
else:
mask = self.mask
return grad * mask
def hook_mask_grad_input(self, m, grad_input, grad_output):
return grad_input
if type(m) == nn.Linear:
# mask = self.mask.max(0)[0].view(1, -1)
mask = self.mask.mean(0).view(1, -1)
grad_input = list(grad_input)
grad_input[1] *= mask
grad_input = tuple(grad_input)
elif type(m) == nn.Conv2d:
mask = self.mask.view(self.mask.size(1), -1).mean(-1).view(1, -1, 1, 1)
# mask = self.mask.view(self.mask.size(1), -1).max(-1)[0].view(1, -1, 1, 1)
grad_input = list(grad_input)
grad_input[0] *= mask
grad_input = tuple(grad_input)
else:
raise NotImplementedError
return grad_input
class ModuleHook(object):
"""
Forward_hook used to get the output and module object of the intermediate layer.
"""
def __init__(self):
super(ModuleHook, self).__init__()
self.input = None
self.outputs = None
self.module = None
def hook(self, module, input, output):
# if self.module is None:
self.module = module
self.inputs = input
self.outputs = output
def clear(self):
self.module = None
self.inputs = None
self.outputs = None
def set_scale(res, graph, bottoms, targ_layer):
# def _find_prev(graph, bottoms, layer_idx):
# bot = bottoms[layer_idx]
# last_bn = None
# prev_list = []
# while len(bot) == 1 and "Data" != bot[0]:
# if type(graph[bot[0]]) == nn.BatchNorm2d:
# last_bn = bot[0]
# if type(graph[bot[0]]) in targ_layer:
# prev_list.append((bot[0], last_bn))
# elif not(type(graph[bot[0]]) in [nn.BatchNorm2d, nn.ReLU] or
# (type(graph[bot[0]]) == str and ("F.pad" in bot[0] or "torch.mean" in bot[0]))):
# return None, None
# bot = bottoms[bot[0]]
# return None, None
layer_first_list = []
layer_second_list = []
for rr in res:
layer_first, layer_second, _ = rr.get_idxs()
scale = rr.get_scale_vec()
graph[layer_first].set_scale(scale=torch.ones(graph[layer_first].weight.shape[0]))
graph[layer_second].set_scale(scale_prev=graph[layer_first].scale)
layer_first_list.append(layer_first)
layer_second_list.append(layer_second)
# res_new = {}
# for idx in graph:
# if type(graph[idx]) in targ_layer:
# if idx not in layer_first_list:
# res_new[idx] = []
# elif idx not in layer_second_list:
# pass
def transform_quant_layer(model, graph, res, trainable=False):
for rr in res:
layer_first, layer_second, _ = rr.get_idxs()
graph[layer_first].merge_scale_to_weight()
graph[layer_second].merge_scale_to_weight()
if hasattr(graph[layer_first], 'scale'):
delattr(graph[layer_first], 'scale')
if hasattr(graph[layer_first], 'scale_prev'):
delattr(graph[layer_first], 'scale_prev')
if hasattr(graph[layer_second], 'scale'):
delattr(graph[layer_second], 'scale')
if hasattr(graph[layer_second], 'scale_prev'):
delattr(graph[layer_second], 'scale_prev')
transformer = TorchTransformer()
if trainable:
transformer.register(QConv2d, QuantConv2d)
transformer.register(QLinear, QuantLinear)
else:
transformer.register(QConv2d, QuantNConv2d)
transformer.register(QLinear, QuantNLinear)
model = transformer.trans_layers(model, update=True)
return model
def kl_categorical(p_logit, q_logit, dim=-1):
"""
https://blog.csdn.net/guotong1988/article/details/90262901
"""
p = F.softmax(p_logit, dim=dim)
_kl = torch.sum(p * (F.log_softmax(p_logit, dim=dim)
- F.log_softmax(q_logit, dim=dim)), dim)
return torch.mean(_kl)
def norm2(weight, idx, writer, step):
# print("Weight max: {}, min: {}, range: {}".format(round(float(weight.max()), 3), round(float(weight.min()), 3), round(float(weight.max()-weight.min()), 3)))
writer.add_scalar("{}/max".format(idx), weight.max().data, step)
writer.add_scalar("{}/min".format(idx), weight.min().data, step)
writer.add_scalar("{}/range".format(idx), (weight.max()-weight.min()).data, step)
mean = weight.mean()
std = torch.sqrt(torch.var(weight)+1e-8)
mask = ((weight < (mean - 2 * std)).float() + (weight > (mean + 2 * std)).float()) * (torch.abs(weight) > 2).float()
return torch.sqrt((torch.abs(weight*mask) ** 2).sum()+1e-8) * (weight.max() - weight.min())
def update_scale(qmodel, model, data_distill, graph, bottoms, res, targ_layer, num_epoch=1000):
"""
this function use data_distill to find optimized scale for DFQ
"""
print("Start updating scale")
writer = SummaryWriter("./tensorboard/exp_{}/".format(round(time.time())))
qmodel = qmodel.eval().cuda()
model = model.eval().cuda()
for idx in range(len(data_distill)):
data_distill[idx].requires_grad = False
graph_original = copy.deepcopy(graph)
optimizer = torch.optim.Adam([p for n, p in qmodel.named_parameters() if 'scale' in n], lr=0.001)
terminate = False
# hook params
hooks = []
hook_handle = []
for name, module in qmodel.named_modules():
if type(module) in targ_layer and hasattr(module, 'scale'):
# print("Add hook to scale of {} module".format(type(module)))
grad_hook = GradHook(module.weight, module.scale if hasattr(module, 'scale') else None,
module.scale_prev if hasattr(module, 'scale_prev') else None,
module.merge_scale if hasattr(module, 'scale') else None,
module.merge_scale_prev if hasattr(module, 'scale_prev') else None)
hooks.append(grad_hook)
# hook_handle.append(module.weight.register_hook(grad_hook.hook_mask_grad_tensor))
hook_handle.append(module.scale.register_hook(grad_hook.hook_mask_grad_tensor))
try:
"""
TODO: check if graph and model contains same module parameters!!!
"""
for epoch in range(num_epoch):
for it in range(len(data_distill)):
data = data_distill[it].cuda()
with torch.no_grad():
logit = model(data)
replace_op()
qlogit = qmodel(data)
restore_op()
klloss = kl_categorical(qlogit, logit) #+ kl_categorical(logit, qlogit)
normloss = 0
for idx, hook in enumerate(hooks):
normloss += norm2(hook.get_weight_scaled(), idx, writer, epoch*len(data_distill)+it+1)
loss = klloss
writer.add_scalar("loss", loss.data, epoch*len(data_distill)+it+1)
writer.add_scalar("norm", normloss.data, epoch*len(data_distill)+it+1)
writer.add_scalar("kldiv", klloss.data, epoch*len(data_distill)+it+1)
print("loss: {}, klloss: {}, norm: {}, iter: {}, epoch: {}".format(loss.data, klloss.data, normloss.data, it+1, epoch+1))
optimizer.zero_grad()
loss.backward()
optimizer.step()
for rr in res:
layer_first, _, bn_idx = rr.get_idxs()
# scale = torch.clamp(graph[layer_first].scale.detach().data.view(-1), max=1)
scale = graph[layer_first].scale.detach().data.view(-1)
graph[bn_idx].fake_weight.copy_(graph_original[bn_idx].fake_weight * scale)
graph[bn_idx].fake_bias.copy_(graph_original[bn_idx].fake_bias * scale)
set_quant_minmax(graph, bottoms, verbose=False)
# for hook in hooks:
# hook.update_mask()
# print("iter: {}, epoch: {}, mean: {}".format(it, epoch, hook.weight.mean()))
# print("="*150)
if loss.data < 0.02:
terminate = True
break
if terminate:
break
except KeyboardInterrupt:
for rr in res:
layer_first, _, bn_idx = rr.get_idxs()
scale = graph[layer_first].scale.detach().data.view(-1)
graph[bn_idx].fake_weight.copy_(graph_original[bn_idx].fake_weight * scale)
graph[bn_idx].fake_bias.copy_(graph_original[bn_idx].fake_bias * scale)
for handle in hook_handle:
handle.remove()
return qmodel
def update_quant_range(model, data, graph, bottoms, is_detection=False):
with torch.no_grad():
replace_op()
for batch_data in data:
batch_data = batch_data.cuda()
_ = model(batch_data)
restore_op()
for idx in graph:
if bottoms[idx] is None:
continue
if bottoms[idx][0] == "Data":
if not is_detection:
graph[idx].quant.running_max.fill_(2.64)
graph[idx].quant.running_min.fill_(-2.11790393)
else:
graph[idx].quant.running_max.fill_(1)
graph[idx].quant.running_min.fill_(-1)
return model
def set_update_stat(model, targ_type, update_stat):
"""!
this function turns on/off the update_stat flag in modules in targ_type
"""
for module_name in model._modules:
# has children
if len(model._modules[module_name]._modules) > 0 and type(getattr(model, module_name)) not in targ_type:
set_update_stat(model._modules[module_name], targ_type, update_stat)
else:
if type(getattr(model, module_name)) in targ_type:
getattr(model, module_name).set_update_stat(update_stat)
def bias_correction_distill(qmodel, model_original, data, targ_type, targ_type_original):
"""!
do bias correction based on distilled data
"""
qmodel = qmodel.cuda().eval()
model_original = model_original.cuda().eval()
hooks = []
hooks_original = []
hook_handles = []
for name, module in qmodel.named_modules():
if type(module) in targ_type:
hook = ModuleHook()
hooks.append(hook)
hook_handles.append(module.register_forward_hook(hook.hook))
for name, module in model_original.named_modules():
if type(module) in targ_type_original:
hook = ModuleHook()
hooks_original.append(hook)
hook_handles.append(module.register_forward_hook(hook.hook))
error_list = {}
assert len(hooks) == len(hooks_original), "len of hooks in 2 models must be the same"
with torch.no_grad():
for b, batch_data in enumerate(data):
for hook in hooks:
hook.clear()
for hook in hooks_original:
hook.clear()
batch_data = batch_data.cuda()
replace_op()
out = qmodel(batch_data)
restore_op()
out = model_original(batch_data)
for idx in range(len(hooks)):
# print("Hook {}, error mean: {}, error sum: {}".format(idx, (hooks_original[idx].outputs.mean(0) - hooks[idx].outputs.mean(0)).cpu().mean(), (hooks_original[idx].outputs.mean(0) - hooks[idx].outputs.mean(0)).cpu().sum()))
if b == 0:
error_list[idx] = [hooks[idx].outputs.mean(0).cpu(), hooks_original[idx].outputs.mean(0).cpu()]
else:
error_list[idx][0] += (hooks[idx].outputs.mean(0)).cpu()
error_list[idx][1] += (hooks_original[idx].outputs.mean(0)).cpu()
# error_list[idx].append((hooks[idx].outputs - hooks_original[idx].outputs).cpu())
for idx, hook in enumerate(hooks):
module = hook.module
error = (error_list[idx][0] - error_list[idx][1]) / len(data)
# print("Hook: {}, error_sum: {}, error_mean: {}".format(idx, error.sum(), error.mean()))
# for idx_error in range(1, len(error_list[idx])):
# error += error_list[idx][idx_error]
error = error.view(error.size(0), -1).sum(-1)
if not hasattr(module, "bias") or getattr(module, "bias") is None:
module.bias = torch.nn.Parameter(torch.zeros(error.size(0)), requires_grad=False)
module.bias.add_(-error.cuda())
for handle in hook_handles:
handle.remove()