Skip to content
This repository has been archived by the owner on Sep 18, 2024. It is now read-only.

fix checkpoint load error and stop updating paramters in evaluation stage #3124

Merged
merged 4 commits into from
Nov 30, 2020
Merged

fix checkpoint load error and stop updating paramters in evaluation stage #3124

Changes from 3 commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
05d2ef9
fix checkpoint load error and stop updating paramters in training stage
Nov 25, 2020
834d941
fix dtype mismatch problem
Nov 26, 2020
1ed0163
avoid modify tensor size during calculation
Nov 26, 2020
dc8185a
add cuda transformation before _quantize calculation
Nov 29, 2020
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
46 changes: 27 additions & 19 deletions nni/algorithms/compression/pytorch/quantization/quantizers.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -73,9 +73,9 @@ def update_quantization_param(bits, rmin, rmax):
----------
bits : int
quantization bits length
rmin : float
rmin : Tensor
min value of real value
rmax : float
rmax : Tensor
max value of real value

Returns
Expand All @@ -85,12 +85,17 @@ def update_quantization_param(bits, rmin, rmax):
# extend the [min, max] interval to ensure that it contains 0.
# Otherwise, we would not meet the requirement that 0 be an exactly
# representable value.
rmin = min(rmin, 0)
rmax = max(rmax, 0)
if rmin.is_cuda:
rmin = torch.min(rmin, torch.Tensor([0]).cuda())
rmax = torch.max(rmax, torch.Tensor([0]).cuda())
qmin = torch.Tensor([0]).cuda()
qmax = torch.Tensor([(1 << bits) - 1]).cuda()
else:
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

why this place use 1-dimension, it seems torch.min, torch.max and update_ema should return 0-dimension, although the call in line 277 will return 1-dimension due to use self.bound_model.steps. Maybe it would be better to unify these places.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

if we claim scale and zero-point as a 1-dimension tensor in QAT, but here changes scale and zero_point to 0-dimension, then we load checkpoint after saving, we would have type mismatch problem, cuz scale and zero_point are claimed as 1-dimension tensor, but in checkpoint, they are saved as a 0-dimension tensor

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

So what i do here is to keep the scale and zero_point's size consistent during all calculation

rmin = torch.min(rmin, torch.Tensor([0]))
rmax = torch.max(rmax, torch.Tensor([0]))
qmin = torch.Tensor([0])
qmax = torch.Tensor([(1 << bits) - 1])

# the min and max quantized values, as floating-point values
qmin = 0
qmax = (1 << bits) - 1
# First determine the scale.
scale = (rmax - rmin) / (qmax - qmin)

Expand Down Expand Up @@ -143,11 +148,11 @@ def __init__(self, model, config_list, optimizer=None):
types of nn.module you want to apply quantization, eg. 'Conv2d'
"""
super().__init__(model, config_list, optimizer)
self.steps = 1
modules_to_compress = self.get_modules_to_compress()
self.bound_model.register_buffer("steps", torch.Tensor([1]))
for layer, config in modules_to_compress:
layer.module.register_buffer("zero_point", None)
layer.module.register_buffer("scale", None)
layer.module.register_buffer("zero_point", torch.Tensor([0.0]))
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

If we just use 0-dimension definition here, would it be more simple and we don't need to add conversion?

Copy link
Contributor Author

@eedalong eedalong Nov 29, 2020 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

define 0-dimension does save some work,defining it to 1-dimension tensor is just for keeping consitent with other buffered variables. Because other buffered variables are also a real value, there's no reason why scale and zero_point, who are also just two real numbers, are defined as a 0-dimension tensor. It might confuse users

layer.module.register_buffer("scale", torch.Tensor([1.0]))
if "output" in config.get("quant_types", []):
layer.module.register_buffer('ema_decay', torch.Tensor([0.99]))
layer.module.register_buffer('tracked_min_biased', torch.zeros(1))
Expand Down Expand Up @@ -229,7 +234,8 @@ def quantize_weight(self, wrapper, **kwargs):
quant_start_step = config.get('quant_start_step', 0)
assert weight_bits >= 1, "quant bits length should be at least 1"

if quant_start_step > self.steps:
# we dont update weight in evaluation stage
if quant_start_step > self.bound_model.steps or not wrapper.training:
return weight

# if bias exists, quantize bias to uint32
Expand Down Expand Up @@ -258,15 +264,17 @@ def quantize_output(self, output, wrapper, **kwargs):
quant_start_step = config.get('quant_start_step', 0)
assert output_bits >= 1, "quant bits length should be at least 1"

if quant_start_step > self.steps:
if quant_start_step > self.bound_model.steps:
return output

current_min, current_max = torch.min(output), torch.max(output)
module.tracked_min_biased, module.tracked_min = update_ema(module.tracked_min_biased, current_min,
module.ema_decay, self.steps)
module.tracked_max_biased, module.tracked_max = update_ema(module.tracked_max_biased, current_max,
module.ema_decay, self.steps)
module.scale, module.zero_point = update_quantization_param(output_bits, module.tracked_min, module.tracked_max)
# we dont update output quantization parameters in evaluation stage
if wrapper.training:
current_min, current_max = torch.min(output), torch.max(output)
module.tracked_min_biased, module.tracked_min = update_ema(module.tracked_min_biased, current_min,
module.ema_decay, self.bound_model.steps)
module.tracked_max_biased, module.tracked_max = update_ema(module.tracked_max_biased, current_max,
module.ema_decay, self.bound_model.steps)
module.scale, module.zero_point = update_quantization_param(output_bits, module.tracked_min, module.tracked_max)
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

At present, we have to re-calculate the scale and zero_point of output and weight. Because the parameter of activation and weight all use the same scale and zero-point of such module. If quantize both activation and weight during testing without updating, parameters would be only one of them, and the result could be wrong. But this design that scale and zero_point using the same parameter doesn't make sense, we plan to refactor it in the following release.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

um, Indeed activation and weight use the same field, but once a user defined to quantize output of the layer, module.zero_point module.scale will always be activation's zero_point and scale , so it will not cause anything wrong

out = self._quantize(output_bits, module, output)
out = self._dequantize(module, out)
return out
Expand All @@ -279,7 +287,7 @@ def step_with_optimizer(self):
"""
override `compressor` `step` method, quantization only happens after certain number of steps
"""
self.steps += 1
self.bound_model.steps +=1


class DoReFaQuantizer(Quantizer):
Expand Down