-
Notifications
You must be signed in to change notification settings - Fork 5.6k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
add new post-quant methods #32208
add new post-quant methods #32208
Conversation
Thanks for your contribution! |
请为新增代码补充下单测。 |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
请添加相应单侧,要不然代码覆盖率达不到,参考../tests/下面的单侧示例。
不同方法的实验数据最好贴到pr上,包括量化模型精度、量化过程的时间等。
@@ -138,8 +138,10 @@ def __init__(self, | |||
batch_size=10, | |||
batch_nums=None, | |||
algo="KL", | |||
hist_perc=0.99999, |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
建议用完整的hist_percent
quantizable_op_type=["conv2d", "depthwise_conv2d", "mul"], | ||
is_full_quantize=False, | ||
bias_correct=False, |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
bias_correction?
python/paddle/fluid/contrib/slim/quantization/post_training_quantization.py
Outdated
Show resolved
Hide resolved
@@ -373,14 +389,17 @@ def quantize(self): | |||
batch_id += 1 | |||
if self._batch_nums and batch_id >= self._batch_nums: | |||
break | |||
|
|||
if self._algo == 'avg': |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
这是获取阈值的逻辑,后面计算阈值的部分。
if self._algo == "abs_max": | ||
self._sample_abs_max() | ||
if self._algo in ["avg", "abs_max"]: | ||
self._sample_abs_max_avg() |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
两个不相同的采样方式,分开成两个函数。
if mse_loss <= self._best_mse_loss[var_name]: | ||
self._best_mse_loss[var_name] = mse_loss | ||
best_scale = scale | ||
if best_scale > 0.0: |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
这个判断没有必要,self._quantized_threshold[var_name] = best_scale 可以放到if mse_loss <= self._best_mse_loss[var_name]:中
save_info( | ||
op_node, out_var_name, self._quantized_threshold, | ||
argname_index[0] + str(argname_index[1]) + "_threshold", | ||
"post_absmax") |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
三种方法有点点差别,建议区分开来
if self._algo == 'avg': | ||
if (var_name not in self._quantized_var_avg): | ||
self._quantized_var_avg[var_name] = [] | ||
abs_avg_value = float(np.mean(np.max(np.abs(var_tensor.reshape(var_tensor.shape[0], -1)), axis=(1)))) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
注意代码每行的长度
@@ -1154,7 +1158,35 @@ def apply(self, graph): | |||
else: | |||
quant_axis = 0 | |||
quantized_param_v = self._quant( | |||
param_v, scale_v, self._weight_bits, quant_axis) | |||
param_v.copy(), scale_v, self._weight_bits, quant_axis) | |||
if self._bias_correct == True: |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
将bias_correction功能独立为一个函数实现
if isinstance(scale_v, list): | ||
if quant_axis == 0: | ||
for i, s in enumerate(scale_v): | ||
quantized_param_v[i] = quantized_param_v[i] * s / bnt |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
这是dequantized_param_v了
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
LGTM
PR types
Performance optimization
PR changes
APIs
Describe
1.Add new methods: 'mse', 'hist', 'avg' of getting threshold values of activations for post-training quantization in 'post_training_quantization.py'.
2.Add bias correction method in of https://arxiv.org/abs/1810.05723 for post-training quantization in 'quantization_pass.py'. Bias correction changes the quantized weights by the following formulation:
Experimental results on 6-bit MobileNetV1 which calibrated by a batch of 32 images:
abs_max: 44.99 abs_max+bias_correction: 47.91
avg: 56.34 avg+bias_correction: 56.17
mse: 61.32 mse+bias_correction: 61.42
hist(0.9999): 60.72 hist+bias_correction: 62.44
KL: 53.11 KL+bias_correction: 58.46
Time cost: Abs_max and avg cost about 1 minute; hist and KL cost about 6 minutes; mse cost about 10 minutes. Bias correction cost little time.