Quantizer

Quantize your network

---

Quantizer

 Quantizer (backend='x86')

Initialize self. See help(type(self)) for accurate signature.

path = untar_data(URLs.PETS)
files = get_image_files(path/"images")

def label_func(f): return f[0].isupper()

dls = ImageDataLoaders.from_name_func(path, files, label_func, item_tfms=Resize(64))

import timm
pretrained_resnet_34 = timm.create_model('resnet34', pretrained=True)
qt = Quantizer()

q_model = qt.quantize(pretrained_resnet_34, dls); q_model

GraphModule(
  (conv1): QuantizedConvReLU2d(3, 64, kernel_size=(7, 7), stride=(2, 2), scale=0.057870179414749146, zero_point=0, padding=(3, 3))
  (maxpool): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False)
  (layer1): Module(
    (0): Module(
      (conv1): QuantizedConv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), scale=0.06179201602935791, zero_point=77, padding=(1, 1))
      (drop_block): Identity()
      (act1): ReLU(inplace=True)
      (aa): Identity()
      (conv2): QuantizedConv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), scale=0.18166492879390717, zero_point=39, padding=(1, 1))
    )
    (1): Module(
      (conv1): QuantizedConv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), scale=0.12881271541118622, zero_point=68, padding=(1, 1))
      (drop_block): Identity()
      (act1): ReLU(inplace=True)
      (aa): Identity()
      (conv2): QuantizedConv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), scale=0.21964868903160095, zero_point=87, padding=(1, 1))
    )
    (2): Module(
      (conv1): QuantizedConv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), scale=0.14362940192222595, zero_point=74, padding=(1, 1))
      (drop_block): Identity()
      (act1): ReLU(inplace=True)
      (aa): Identity()
      (conv2): QuantizedConv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), scale=0.2838979959487915, zero_point=82, padding=(1, 1))
    )
  )
  (layer2): Module(
    (0): Module(
      (conv1): QuantizedConv2d(64, 128, kernel_size=(3, 3), stride=(2, 2), scale=0.12004681676626205, zero_point=71, padding=(1, 1))
      (drop_block): Identity()
      (act1): ReLU(inplace=True)
      (aa): Identity()
      (conv2): QuantizedConv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), scale=0.19042885303497314, zero_point=70, padding=(1, 1))
      (downsample): Module(
        (0): QuantizedConv2d(64, 128, kernel_size=(1, 1), stride=(2, 2), scale=0.16569757461547852, zero_point=65)
      )
    )
    (1): Module(
      (conv1): QuantizedConv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), scale=0.17097020149230957, zero_point=71, padding=(1, 1))
      (drop_block): Identity()
      (act1): ReLU(inplace=True)
      (aa): Identity()
      (conv2): QuantizedConv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), scale=0.14862027764320374, zero_point=78, padding=(1, 1))
    )
    (2): Module(
      (conv1): QuantizedConv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), scale=0.1713831126689911, zero_point=80, padding=(1, 1))
      (drop_block): Identity()
      (act1): ReLU(inplace=True)
      (aa): Identity()
      (conv2): QuantizedConv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), scale=0.23077945411205292, zero_point=84, padding=(1, 1))
    )
    (3): Module(
      (conv1): QuantizedConv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), scale=0.19932255148887634, zero_point=85, padding=(1, 1))
      (drop_block): Identity()
      (act1): ReLU(inplace=True)
      (aa): Identity()
      (conv2): QuantizedConv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), scale=0.3485361337661743, zero_point=63, padding=(1, 1))
    )
  )
  (layer3): Module(
    (0): Module(
      (conv1): QuantizedConv2d(128, 256, kernel_size=(3, 3), stride=(2, 2), scale=0.2729864716529846, zero_point=70, padding=(1, 1))
      (drop_block): Identity()
      (act1): ReLU(inplace=True)
      (aa): Identity()
      (conv2): QuantizedConv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), scale=0.3445647358894348, zero_point=61, padding=(1, 1))
      (downsample): Module(
        (0): QuantizedConv2d(128, 256, kernel_size=(1, 1), stride=(2, 2), scale=0.21365275979042053, zero_point=67)
      )
    )
    (1): Module(
      (conv1): QuantizedConv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), scale=0.28155988454818726, zero_point=75, padding=(1, 1))
      (drop_block): Identity()
      (act1): ReLU(inplace=True)
      (aa): Identity()
      (conv2): QuantizedConv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), scale=0.27959567308425903, zero_point=76, padding=(1, 1))
    )
    (2): Module(
      (conv1): QuantizedConv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), scale=0.32337626814842224, zero_point=92, padding=(1, 1))
      (drop_block): Identity()
      (act1): ReLU(inplace=True)
      (aa): Identity()
      (conv2): QuantizedConv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), scale=0.38359376788139343, zero_point=77, padding=(1, 1))
    )
    (3): Module(
      (conv1): QuantizedConv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), scale=0.22612670063972473, zero_point=79, padding=(1, 1))
      (drop_block): Identity()
      (act1): ReLU(inplace=True)
      (aa): Identity()
      (conv2): QuantizedConv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), scale=0.4580967426300049, zero_point=81, padding=(1, 1))
    )
    (4): Module(
      (conv1): QuantizedConv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), scale=0.21327464282512665, zero_point=77, padding=(1, 1))
      (drop_block): Identity()
      (act1): ReLU(inplace=True)
      (aa): Identity()
      (conv2): QuantizedConv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), scale=0.542456328868866, zero_point=78, padding=(1, 1))
    )
    (5): Module(
      (conv1): QuantizedConv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), scale=0.2356332689523697, zero_point=79, padding=(1, 1))
      (drop_block): Identity()
      (act1): ReLU(inplace=True)
      (aa): Identity()
      (conv2): QuantizedConv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), scale=0.9199671745300293, zero_point=90, padding=(1, 1))
    )
  )
  (layer4): Module(
    (0): Module(
      (conv1): QuantizedConv2d(256, 512, kernel_size=(3, 3), stride=(2, 2), scale=0.24926751852035522, zero_point=75, padding=(1, 1))
      (drop_block): Identity()
      (act1): ReLU(inplace=True)
      (aa): Identity()
      (conv2): QuantizedConv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), scale=0.2349451631307602, zero_point=87, padding=(1, 1))
      (downsample): Module(
        (0): QuantizedConv2d(256, 512, kernel_size=(1, 1), stride=(2, 2), scale=0.24794092774391174, zero_point=62)
      )
    )
    (1): Module(
      (conv1): QuantizedConv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), scale=0.19862131774425507, zero_point=85, padding=(1, 1))
      (drop_block): Identity()
      (act1): ReLU(inplace=True)
      (aa): Identity()
      (conv2): QuantizedConv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), scale=0.18632155656814575, zero_point=81, padding=(1, 1))
    )
    (2): Module(
      (conv1): QuantizedConv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), scale=0.20997950434684753, zero_point=90, padding=(1, 1))
      (drop_block): Identity()
      (act1): ReLU(inplace=True)
      (aa): Identity()
      (conv2): QuantizedConv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), scale=0.29171162843704224, zero_point=77, padding=(1, 1))
    )
  )
  (global_pool): Module(
    (pool): AdaptiveAvgPool2d(output_size=1)
    (flatten): Flatten(start_dim=1, end_dim=-1)
  )
  (fc): QuantizedLinear(in_features=512, out_features=1000, scale=1.0219743251800537, zero_point=57, qscheme=torch.per_channel_affine)
)

#model = resnet18()
#model.fc = nn.Linear(512, 2)

qt = Quantizer()

q_model = qt.quantize(model, dls); q_model

KeyboardInterrupt: