I would like to find where are the parameters quant_max, quant_min, min_val, max_val stored in QuantizedConv2d block. I was able to locate them using the following code in the observers
from torch.quantization.observer import MinMaxObserver, MovingAverageMinMaxObserver, HistogramObserver
C, L = 3, 4
normal = torch.distributions.normal.Normal(0,1)
inputs = [normal.sample((C, L)), normal.sample((C, L))]
observers = [MinMaxObserver(), MovingAverageMinMaxObserver(), HistogramObserver()]
for obs in observers:
for x in inputs:
obs(x)
print(obs.__class__.__name__, obs.calculate_qparams(), obs.quant_min, obs.quant_max, obs.min_val, obs.max_val)
The output being
MinMaxObserver (tensor([0.0113]), tensor([136], dtype=torch.int32)) 0 255 tensor(-1.5460) tensor(1.3433)
MovingAverageMinMaxObserver (tensor([0.0111]), tensor([134], dtype=torch.int32)) 0 255 tensor(-1.4766) tensor(1.3414)
HistogramObserver (tensor([0.0082]), tensor([143], dtype=torch.int32)) 0 255 tensor(-1.5460) tensor(1.3612)
However I am not able to locate the same after converting a Conv2d block into QuantizedConv2d block
class M(nn.Module):
def __init__(self):
super(M, self).__init__()
self.quant = torch.quantization.QuantStub()
self.conv = torch.nn.Conv2d(in_channels=1, out_channels=1, kernel_size=3, stride=1, padding=1)
self.dequant = torch.quantization.DeQuantStub()
def forward(self, x):
x = self.quant(x)
x = self.conv(x)
x = self.dequant(x)
return x
model_fp32 = M()
model_fp32.eval()
model_fp32.qconfig = torch.quantization.get_default_qconfig('fbgemm')
model_fp32_prepared = torch.quantization.prepare(model_fp32)
model_fp32_converted = torch.quantization.convert(model_fp32_prepared, inplace=True)
The weights of the model after passing the inputs through the model
model_fp32_converted(inputs)
model_fp32_converted.conv.weight()
tensor([[[[ 0.03294463083148002625, -0.13431271910667419434,
-0.21540719270706176758],
[ 0.14191532135009765625, -0.14191532135009765625,
-0.24835182726383209229],
[ 0.32184368371963500977, -0.14444953203201293945,
-0.21287299692630767822]]]], size=(1, 1, 3, 3), dtype=torch.qint8,
quantization_scheme=torch.per_channel_affine,
scale=tensor([0.00253420230001211166], dtype=torch.float64),
zero_point=tensor([0]), axis=0)
I would like to know
-
How to get quant_max, quant_min, min_val, max_val from this model_fp32_converted.conv block?
-
Point the location in the github code where these parameters from the observers get stored in the QuantizedConv2d block