| 1 | Inverted Residual Block: Before fusion
|
| 2 |
|
| 3 | Sequential(
|
| 4 | (0): ConvBNReLU(
|
| 5 | (0): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False)
|
| 6 | (1): BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
|
| 7 | (2): ReLU()
|
| 8 | )
|
| 9 | (1): Conv2d(32, 16, kernel_size=(1, 1), stride=(1, 1), bias=False)
|
| 10 | (2): BatchNorm2d(16, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
|
| 11 | )
|
| 12 |
|
| 13 | Inverted Residual Block: After fusion
|
| 14 |
|
| 15 | Sequential(
|
| 16 | (0): ConvBNReLU(
|
| 17 | (0): ConvReLU2d(
|
| 18 | (0): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32)
|
| 19 | (1): ReLU()
|
| 20 | )
|
| 21 | (1): Identity()
|
| 22 | (2): Identity()
|
| 23 | )
|
| 24 | (1): Conv2d(32, 16, kernel_size=(1, 1), stride=(1, 1))
|
| 25 | (2): Identity()
|
| 26 | )
|
| 27 | Size of baseline model
|
| 28 | Size (MB): 13.999657
|
| 29 | ........................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................Evaluation accuracy on 50000 images, 71.86
|
| 30 | QConfig(activation=functools.partial(<class 'torch.quantization.observer.MinMaxObserver'>, reduce_range=True), weight=functools.partial(<class 'torch.quantization.observer.MinMaxObserver'>, dtype=torch.qint8, qscheme=torch.per_tensor_symmetric))
|
| 31 | Post Training Quantization Prepare: Inserting Observers
|
| 32 |
|
| 33 | Inverted Residual Block:After observer insertion
|
| 34 |
|
| 35 | Sequential(
|
| 36 | (0): ConvBNReLU(
|
| 37 | (0): ConvReLU2d(
|
| 38 | (0): Conv2d(
|
| 39 | 32, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32
|
| 40 | (activation_post_process): MinMaxObserver(min_val=tensor([]), max_val=tensor([]))
|
| 41 | )
|
| 42 | (1): ReLU(
|
| 43 | (activation_post_process): MinMaxObserver(min_val=tensor([]), max_val=tensor([]))
|
| 44 | )
|
| 45 | )
|
| 46 | (1): Identity()
|
| 47 | (2): Identity()
|
| 48 | )
|
| 49 | (1): Conv2d(
|
| 50 | 32, 16, kernel_size=(1, 1), stride=(1, 1)
|
| 51 | (activation_post_process): MinMaxObserver(min_val=tensor([]), max_val=tensor([]))
|
| 52 | )
|
| 53 | (2): Identity()
|
| 54 | )
|
| 55 | ................................Post Training Quantization: Calibration done
|
| 56 | /projs/framework/zhaixiuchuan/train/venv/pytorch/lib/python3.6/site-packages/torch/quantization/observer.py:136: UserWarning: must run observer before calling calculate_qparams. Returning default scale and zero point
|
| 57 | Returning default scale and zero point "
|
| 58 | Post Training Quantization: Convert done
|
| 59 |
|
| 60 | Inverted Residual Block: After fusion and quantization, note fused modules:
|
| 61 |
|
| 62 | Sequential(
|
| 63 | (0): ConvBNReLU(
|
| 64 | (0): QuantizedConvReLU2d(32, 32, kernel_size=(3, 3), stride=(1, 1), scale=0.1571013331413269, zero_point=0, padding=(1, 1), groups=32)
|
| 65 | (1): Identity()
|
| 66 | (2): Identity()
|
| 67 | )
|
| 68 | (1): QuantizedConv2d(32, 16, kernel_size=(1, 1), stride=(1, 1), scale=0.190989688038826, zero_point=66)
|
| 69 | (2): Identity()
|
| 70 | )
|
| 71 | Size of model after quantization
|
| 72 | Size (MB): 3.631847
|
| 73 | ........................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................Evaluation accuracy on 50000 images, 56.39
|
| 74 | QConfig(activation=functools.partial(<class 'torch.quantization.observer.HistogramObserver'>, reduce_range=True), weight=functools.partial(<class 'torch.quantization.observer.PerChannelMinMaxObserver'>, dtype=torch.qint8, qscheme=torch.per_channel_symmetric))
|
| 75 | /projs/framework/zhaixiuchuan/train/venv/pytorch/lib/python3.6/site-packages/torch/quantization/observer.py:877: UserWarning: must run observer before calling calculate_qparams. Returning default scale and zero point
|
| 76 | Returning default scale and zero point "
|
| 77 |
|
| 78 | ........................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................Evaluation accuracy on 50000 images, 68.05
|
| 79 | Inverted Residual Block: After preparation for QAT, note fake-quantization modules
|
| 80 | Sequential(
|
| 81 | (0): ConvBNReLU(
|
| 82 | (0): ConvBnReLU2d(
|
| 83 | 32, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False
|
| 84 | (bn): BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
|
| 85 | (activation_post_process): FakeQuantize(
|
| 86 | fake_quant_enabled=tensor([1], dtype=torch.uint8), observer_enabled=tensor([1], dtype=torch.uint8), scale=tensor([1.]), zero_point=tensor([0])
|
| 87 | (activation_post_process): MovingAverageMinMaxObserver(min_val=tensor([]), max_val=tensor([]))
|
| 88 | )
|
| 89 | (weight_fake_quant): FakeQuantize(
|
| 90 | fake_quant_enabled=tensor([1], dtype=torch.uint8), observer_enabled=tensor([1], dtype=torch.uint8), scale=tensor([1.]), zero_point=tensor([0])
|
| 91 | (activation_post_process): MovingAveragePerChannelMinMaxObserver(min_val=tensor([]), max_val=tensor([]))
|
| 92 | )
|
| 93 | )
|
| 94 | (1): Identity()
|
| 95 | (2): Identity()
|
| 96 | )
|
| 97 | (1): ConvBn2d(
|
| 98 | 32, 16, kernel_size=(1, 1), stride=(1, 1), bias=False
|
| 99 | (bn): BatchNorm2d(16, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
|
| 100 | (activation_post_process): FakeQuantize(
|
| 101 | fake_quant_enabled=tensor([1], dtype=torch.uint8), observer_enabled=tensor([1], dtype=torch.uint8), scale=tensor([1.]), zero_point=tensor([0])
|
| 102 | (activation_post_process): MovingAverageMinMaxObserver(min_val=tensor([]), max_val=tensor([]))
|
| 103 | )
|
| 104 | (weight_fake_quant): FakeQuantize(
|
| 105 | fake_quant_enabled=tensor([1], dtype=torch.uint8), observer_enabled=tensor([1], dtype=torch.uint8), scale=tensor([1.]), zero_point=tensor([0])
|
| 106 | (activation_post_process): MovingAveragePerChannelMinMaxObserver(min_val=tensor([]), max_val=tensor([]))
|
| 107 | )
|
| 108 | )
|
| 109 | (2): Identity()
|
| 110 | )
|
| 111 | ....................Loss tensor(1.7456, grad_fn=<DivBackward0>)
|
| 112 | Training: * Acc@1 59.167 Acc@5 82.500
|
| 113 | ........................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................Epoch 0 :Evaluation accuracy on 50000 images, 67.67
|
| 114 | ....................Loss tensor(1.6714, grad_fn=<DivBackward0>)
|
| 115 | Training: * Acc@1 61.667 Acc@5 81.500
|
| 116 | ........................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................Epoch 1 :Evaluation accuracy on 50000 images, 67.33
|
| 117 | ....................Loss tensor(1.8121, grad_fn=<DivBackward0>)
|
| 118 | Training: * Acc@1 60.333 Acc@5 79.333
|
| 119 | ........................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................Epoch 2 :Evaluation accuracy on 50000 images, 66.42
|
| 120 | ....................Loss tensor(1.7800, grad_fn=<DivBackward0>)
|
| 121 | Training: * Acc@1 57.500 Acc@5 81.500
|
| 122 | ........................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................Epoch 3 :Evaluation accuracy on 50000 images, 67.03
|
| 123 | ....................Loss tensor(1.6966, grad_fn=<DivBackward0>)
|
| 124 | Training: * Acc@1 62.167 Acc@5 82.167
|
| 125 | ........................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................Epoch 4 :Evaluation accuracy on 50000 images, 67.30
|
| 126 | ....................Loss tensor(1.4494, grad_fn=<DivBackward0>)
|
| 127 | Training: * Acc@1 66.167 Acc@5 84.333
|
| 128 | ........................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................Epoch 5 :Evaluation accuracy on 50000 images, 67.59
|
| 129 | ....................Loss tensor(1.4924, grad_fn=<DivBackward0>)
|
| 130 | Training: * Acc@1 66.000 Acc@5 86.667
|
| 131 | ........................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................Epoch 6 :Evaluation accuracy on 50000 images, 67.77
|
| 132 | ....................Loss tensor(1.4817, grad_fn=<DivBackward0>)
|
| 133 | Training: * Acc@1 65.000 Acc@5 85.000
|
| 134 | ........................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................Epoch 7 :Evaluation accuracy on 50000 images, 67.88
|
| 135 | Elapsed time: 46 ms
|
| 136 | Elapsed time: 10 ms
|
