wehub-resource-sync
750 lines
40 KiB
Plaintext
750 lines
40 KiB
Plaintext
{
|
|
"cells": [
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 1,
|
|
"metadata": {},
|
|
"outputs": [],
|
|
"source": [
|
|
"import datetime\n",
|
|
"import os\n",
|
|
"import sys\n",
|
|
"import time\n",
|
|
"import collections\n",
|
|
"\n",
|
|
"import torch\n",
|
|
"import torch.utils.data\n",
|
|
"from torch import nn\n",
|
|
"\n",
|
|
"from tqdm import tqdm\n",
|
|
"\n",
|
|
"import torchvision\n",
|
|
"from torchvision import transforms\n",
|
|
"\n",
|
|
"from pytorch_quantization import nn as quant_nn\n",
|
|
"from pytorch_quantization import calib\n",
|
|
"from pytorch_quantization.tensor_quant import QuantDescriptor\n",
|
|
"\n",
|
|
"from absl import logging\n",
|
|
"logging.set_verbosity(logging.FATAL) # Disable logging as they are too noisy in notebook\n",
|
|
"\n"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 2,
|
|
"metadata": {},
|
|
"outputs": [],
|
|
"source": [
|
|
"# For simplicity, import train and eval functions from the train script from torchvision instead of copything them here\n",
|
|
"# Download torchvision from https://github.com/pytorch/vision\n",
|
|
"sys.path.append(\"/raid/skyw/models/torchvision/references/classification/\")\n",
|
|
"from train import evaluate, train_one_epoch, load_data"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "markdown",
|
|
"metadata": {},
|
|
"source": [
|
|
"## Set default QuantDescriptor to use histogram based calibration for activation"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 3,
|
|
"metadata": {},
|
|
"outputs": [],
|
|
"source": [
|
|
"quant_desc_input = QuantDescriptor(calib_method='histogram')\n",
|
|
"quant_nn.QuantConv2d.set_default_quant_desc_input(quant_desc_input)\n",
|
|
"quant_nn.QuantLinear.set_default_quant_desc_input(quant_desc_input)"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "markdown",
|
|
"metadata": {},
|
|
"source": [
|
|
"## Initialize quantized modules"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 4,
|
|
"metadata": {},
|
|
"outputs": [],
|
|
"source": [
|
|
"from pytorch_quantization import quant_modules\n",
|
|
"quant_modules.initialize()"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "markdown",
|
|
"metadata": {},
|
|
"source": [
|
|
"## Create model with pretrained weight"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 5,
|
|
"metadata": {
|
|
"scrolled": true
|
|
},
|
|
"outputs": [
|
|
{
|
|
"data": {
|
|
"text/plain": [
|
|
"ResNet(\n",
|
|
" (conv1): QuantConv2d(\n",
|
|
" 3, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (relu): ReLU(inplace=True)\n",
|
|
" (maxpool): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False)\n",
|
|
" (layer1): Sequential(\n",
|
|
" (0): Bottleneck(\n",
|
|
" (conv1): QuantConv2d(\n",
|
|
" 64, 64, kernel_size=(1, 1), stride=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (conv2): QuantConv2d(\n",
|
|
" 64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (conv3): QuantConv2d(\n",
|
|
" 64, 256, kernel_size=(1, 1), stride=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn3): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (relu): ReLU(inplace=True)\n",
|
|
" (downsample): Sequential(\n",
|
|
" (0): QuantConv2d(\n",
|
|
" 64, 256, kernel_size=(1, 1), stride=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" )\n",
|
|
" )\n",
|
|
" (1): Bottleneck(\n",
|
|
" (conv1): QuantConv2d(\n",
|
|
" 256, 64, kernel_size=(1, 1), stride=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (conv2): QuantConv2d(\n",
|
|
" 64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (conv3): QuantConv2d(\n",
|
|
" 64, 256, kernel_size=(1, 1), stride=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn3): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (relu): ReLU(inplace=True)\n",
|
|
" )\n",
|
|
" (2): Bottleneck(\n",
|
|
" (conv1): QuantConv2d(\n",
|
|
" 256, 64, kernel_size=(1, 1), stride=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (conv2): QuantConv2d(\n",
|
|
" 64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (conv3): QuantConv2d(\n",
|
|
" 64, 256, kernel_size=(1, 1), stride=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn3): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (relu): ReLU(inplace=True)\n",
|
|
" )\n",
|
|
" )\n",
|
|
" (layer2): Sequential(\n",
|
|
" (0): Bottleneck(\n",
|
|
" (conv1): QuantConv2d(\n",
|
|
" 256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (conv2): QuantConv2d(\n",
|
|
" 128, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (conv3): QuantConv2d(\n",
|
|
" 128, 512, kernel_size=(1, 1), stride=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn3): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (relu): ReLU(inplace=True)\n",
|
|
" (downsample): Sequential(\n",
|
|
" (0): QuantConv2d(\n",
|
|
" 256, 512, kernel_size=(1, 1), stride=(2, 2), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" )\n",
|
|
" )\n",
|
|
" (1): Bottleneck(\n",
|
|
" (conv1): QuantConv2d(\n",
|
|
" 512, 128, kernel_size=(1, 1), stride=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (conv2): QuantConv2d(\n",
|
|
" 128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (conv3): QuantConv2d(\n",
|
|
" 128, 512, kernel_size=(1, 1), stride=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn3): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (relu): ReLU(inplace=True)\n",
|
|
" )\n",
|
|
" (2): Bottleneck(\n",
|
|
" (conv1): QuantConv2d(\n",
|
|
" 512, 128, kernel_size=(1, 1), stride=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (conv2): QuantConv2d(\n",
|
|
" 128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (conv3): QuantConv2d(\n",
|
|
" 128, 512, kernel_size=(1, 1), stride=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn3): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (relu): ReLU(inplace=True)\n",
|
|
" )\n",
|
|
" (3): Bottleneck(\n",
|
|
" (conv1): QuantConv2d(\n",
|
|
" 512, 128, kernel_size=(1, 1), stride=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (conv2): QuantConv2d(\n",
|
|
" 128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (conv3): QuantConv2d(\n",
|
|
" 128, 512, kernel_size=(1, 1), stride=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn3): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (relu): ReLU(inplace=True)\n",
|
|
" )\n",
|
|
" )\n",
|
|
" (layer3): Sequential(\n",
|
|
" (0): Bottleneck(\n",
|
|
" (conv1): QuantConv2d(\n",
|
|
" 512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (conv2): QuantConv2d(\n",
|
|
" 256, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (conv3): QuantConv2d(\n",
|
|
" 256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn3): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (relu): ReLU(inplace=True)\n",
|
|
" (downsample): Sequential(\n",
|
|
" (0): QuantConv2d(\n",
|
|
" 512, 1024, kernel_size=(1, 1), stride=(2, 2), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (1): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" )\n",
|
|
" )\n",
|
|
" (1): Bottleneck(\n",
|
|
" (conv1): QuantConv2d(\n",
|
|
" 1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (conv2): QuantConv2d(\n",
|
|
" 256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (conv3): QuantConv2d(\n",
|
|
" 256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn3): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (relu): ReLU(inplace=True)\n",
|
|
" )\n",
|
|
" (2): Bottleneck(\n",
|
|
" (conv1): QuantConv2d(\n",
|
|
" 1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (conv2): QuantConv2d(\n",
|
|
" 256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (conv3): QuantConv2d(\n",
|
|
" 256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn3): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (relu): ReLU(inplace=True)\n",
|
|
" )\n",
|
|
" (3): Bottleneck(\n",
|
|
" (conv1): QuantConv2d(\n",
|
|
" 1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (conv2): QuantConv2d(\n",
|
|
" 256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (conv3): QuantConv2d(\n",
|
|
" 256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn3): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (relu): ReLU(inplace=True)\n",
|
|
" )\n",
|
|
" (4): Bottleneck(\n",
|
|
" (conv1): QuantConv2d(\n",
|
|
" 1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (conv2): QuantConv2d(\n",
|
|
" 256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (conv3): QuantConv2d(\n",
|
|
" 256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn3): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (relu): ReLU(inplace=True)\n",
|
|
" )\n",
|
|
" (5): Bottleneck(\n",
|
|
" (conv1): QuantConv2d(\n",
|
|
" 1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (conv2): QuantConv2d(\n",
|
|
" 256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (conv3): QuantConv2d(\n",
|
|
" 256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn3): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (relu): ReLU(inplace=True)\n",
|
|
" )\n",
|
|
" )\n",
|
|
" (layer4): Sequential(\n",
|
|
" (0): Bottleneck(\n",
|
|
" (conv1): QuantConv2d(\n",
|
|
" 1024, 512, kernel_size=(1, 1), stride=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (conv2): QuantConv2d(\n",
|
|
" 512, 512, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn2): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (conv3): QuantConv2d(\n",
|
|
" 512, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn3): BatchNorm2d(2048, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (relu): ReLU(inplace=True)\n",
|
|
" (downsample): Sequential(\n",
|
|
" (0): QuantConv2d(\n",
|
|
" 1024, 2048, kernel_size=(1, 1), stride=(2, 2), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (1): BatchNorm2d(2048, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" )\n",
|
|
" )\n",
|
|
" (1): Bottleneck(\n",
|
|
" (conv1): QuantConv2d(\n",
|
|
" 2048, 512, kernel_size=(1, 1), stride=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (conv2): QuantConv2d(\n",
|
|
" 512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn2): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (conv3): QuantConv2d(\n",
|
|
" 512, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn3): BatchNorm2d(2048, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (relu): ReLU(inplace=True)\n",
|
|
" )\n",
|
|
" (2): Bottleneck(\n",
|
|
" (conv1): QuantConv2d(\n",
|
|
" 2048, 512, kernel_size=(1, 1), stride=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (conv2): QuantConv2d(\n",
|
|
" 512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn2): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (conv3): QuantConv2d(\n",
|
|
" 512, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
" (bn3): BatchNorm2d(2048, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n",
|
|
" (relu): ReLU(inplace=True)\n",
|
|
" )\n",
|
|
" )\n",
|
|
" (avgpool): AdaptiveAvgPool2d(output_size=(1, 1))\n",
|
|
" (fc): QuantLinear(\n",
|
|
" in_features=2048, out_features=1000, bias=True\n",
|
|
" (_input_quantizer): TensorQuantizer(8bit fake per-tensor amax=dynamic calibrator=HistogramCalibrator quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=dynamic calibrator=MaxCalibrator quant)\n",
|
|
" )\n",
|
|
")"
|
|
]
|
|
},
|
|
"execution_count": 5,
|
|
"metadata": {},
|
|
"output_type": "execute_result"
|
|
}
|
|
],
|
|
"source": [
|
|
"model = torchvision.models.resnet50(pretrained=True, progress=False)\n",
|
|
"model.cuda()"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "markdown",
|
|
"metadata": {},
|
|
"source": [
|
|
"## Create data loader"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 6,
|
|
"metadata": {},
|
|
"outputs": [
|
|
{
|
|
"name": "stdout",
|
|
"output_type": "stream",
|
|
"text": [
|
|
"Loading data\n",
|
|
"Loading training data\n",
|
|
"Took 3.580507755279541\n",
|
|
"Loading validation data\n",
|
|
"Creating data loaders\n"
|
|
]
|
|
}
|
|
],
|
|
"source": [
|
|
"data_path = \"/raid/data/imagenet/imagenet_pytorch\"\n",
|
|
"batch_size = 512\n",
|
|
"\n",
|
|
"traindir = os.path.join(data_path, 'train')\n",
|
|
"valdir = os.path.join(data_path, 'val')\n",
|
|
"_args = collections.namedtuple('mock_args', ['model', 'distributed', 'cache_dataset'])\n",
|
|
"dataset, dataset_test, train_sampler, test_sampler = load_data(traindir, valdir, _args(model='resnet50', distributed=False, cache_dataset=False))\n",
|
|
"\n",
|
|
"data_loader = torch.utils.data.DataLoader(\n",
|
|
" dataset, batch_size=batch_size,\n",
|
|
" sampler=train_sampler, num_workers=4, pin_memory=True)\n",
|
|
"\n",
|
|
"data_loader_test = torch.utils.data.DataLoader(\n",
|
|
" dataset_test, batch_size=batch_size,\n",
|
|
" sampler=test_sampler, num_workers=4, pin_memory=True)\n"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "markdown",
|
|
"metadata": {},
|
|
"source": [
|
|
"## Calibrate the model"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 7,
|
|
"metadata": {},
|
|
"outputs": [],
|
|
"source": [
|
|
"def collect_stats(model, data_loader, num_batches):\n",
|
|
" \"\"\"Feed data to the network and collect statistic\"\"\"\n",
|
|
"\n",
|
|
" # Enable calibrators\n",
|
|
" for name, module in model.named_modules():\n",
|
|
" if isinstance(module, quant_nn.TensorQuantizer):\n",
|
|
" if module._calibrator is not None:\n",
|
|
" module.disable_quant()\n",
|
|
" module.enable_calib()\n",
|
|
" else:\n",
|
|
" module.disable()\n",
|
|
"\n",
|
|
" for i, (image, _) in tqdm(enumerate(data_loader), total=num_batches):\n",
|
|
" model(image.cuda())\n",
|
|
" if i >= num_batches:\n",
|
|
" break\n",
|
|
"\n",
|
|
" # Disable calibrators\n",
|
|
" for name, module in model.named_modules():\n",
|
|
" if isinstance(module, quant_nn.TensorQuantizer):\n",
|
|
" if module._calibrator is not None:\n",
|
|
" module.enable_quant()\n",
|
|
" module.disable_calib()\n",
|
|
" else:\n",
|
|
" module.enable()\n",
|
|
" \n",
|
|
"def compute_amax(model, **kwargs):\n",
|
|
" # Load calib result\n",
|
|
" for name, module in model.named_modules():\n",
|
|
" if isinstance(module, quant_nn.TensorQuantizer):\n",
|
|
" if module._calibrator is not None:\n",
|
|
" if isinstance(module._calibrator, calib.MaxCalibrator):\n",
|
|
" module.load_calib_amax()\n",
|
|
" else:\n",
|
|
" module.load_calib_amax(**kwargs)\n",
|
|
"# print(F\"{name:40}: {module}\")\n",
|
|
" model.cuda()"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 8,
|
|
"metadata": {
|
|
"scrolled": true
|
|
},
|
|
"outputs": [
|
|
{
|
|
"name": "stderr",
|
|
"output_type": "stream",
|
|
"text": [
|
|
"100%|██████████| 2/2 [04:50<00:00, 111.13s/it]"
|
|
]
|
|
}
|
|
],
|
|
"source": [
|
|
"# It is a bit slow since we collect histograms on CPU\n",
|
|
"with torch.no_grad():\n",
|
|
" collect_stats(model, data_loader, num_batches=2)\n",
|
|
" compute_amax(model, method=\"percentile\", percentile=99.99)"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "markdown",
|
|
"metadata": {},
|
|
"source": [
|
|
"## Now evaluate the calibrated model"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 9,
|
|
"metadata": {},
|
|
"outputs": [
|
|
{
|
|
"name": "stdout",
|
|
"output_type": "stream",
|
|
"text": [
|
|
"Test: [ 0/98] eta: 0:05:53 loss: 0.5656 (0.5656) acc1: 85.7422 (85.7422) acc5: 96.0938 (96.0938) time: 3.6079 data: 2.8152 max mem: 5880\n",
|
|
"Test: [20/98] eta: 0:01:07 loss: 0.6741 (0.6825) acc1: 82.8125 (82.4219) acc5: 95.8984 (95.7682) time: 0.7343 data: 0.0002 max mem: 5882\n",
|
|
"Test: [40/98] eta: 0:00:46 loss: 0.6995 (0.7157) acc1: 80.0781 (81.4024) acc5: 96.0938 (95.7412) time: 0.7226 data: 0.0002 max mem: 5882\n",
|
|
"Test: [60/98] eta: 0:00:29 loss: 1.1064 (0.8590) acc1: 71.4844 (78.2627) acc5: 91.0156 (94.1150) time: 0.7259 data: 0.0002 max mem: 5882\n",
|
|
"Test: [80/98] eta: 0:00:13 loss: 1.1220 (0.9372) acc1: 72.4609 (76.7072) acc5: 89.6484 (93.1375) time: 0.7220 data: 0.0002 max mem: 5882\n",
|
|
"Test: Total time: 0:01:13\n",
|
|
" * Acc@1 76.138 Acc@5 92.916\n"
|
|
]
|
|
}
|
|
],
|
|
"source": [
|
|
"criterion = nn.CrossEntropyLoss()\n",
|
|
"with torch.no_grad():\n",
|
|
" evaluate(model, criterion, data_loader_test, device=\"cuda\", print_freq=20)\n",
|
|
" \n",
|
|
"# Save the model\n",
|
|
"torch.save(model.state_dict(), \"/tmp/quant_resnet50-calibrated.pth\")"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "markdown",
|
|
"metadata": {},
|
|
"source": [
|
|
"## We can also try different calibrations and see which one works the best"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 10,
|
|
"metadata": {},
|
|
"outputs": [
|
|
{
|
|
"name": "stdout",
|
|
"output_type": "stream",
|
|
"text": [
|
|
"Test: [ 0/98] eta: 0:05:27 loss: 0.6037 (0.6037) acc1: 84.9609 (84.9609) acc5: 95.3125 (95.3125) time: 3.3411 data: 2.6190 max mem: 5882\n",
|
|
"Test: [20/98] eta: 0:01:06 loss: 0.6760 (0.7041) acc1: 81.2500 (81.7522) acc5: 95.7031 (95.4892) time: 0.7243 data: 0.0002 max mem: 5882\n",
|
|
"Test: [40/98] eta: 0:00:45 loss: 0.7241 (0.7351) acc1: 79.1016 (80.7784) acc5: 95.8984 (95.4459) time: 0.7243 data: 0.0002 max mem: 5882\n",
|
|
"Test: [60/98] eta: 0:00:29 loss: 1.1162 (0.8793) acc1: 71.4844 (77.6383) acc5: 90.8203 (93.7948) time: 0.7204 data: 0.0002 max mem: 5882\n",
|
|
"Test: [80/98] eta: 0:00:13 loss: 1.1498 (0.9603) acc1: 71.4844 (76.0368) acc5: 89.4531 (92.7156) time: 0.7164 data: 0.0002 max mem: 5882\n",
|
|
"Test: Total time: 0:01:12\n",
|
|
" * Acc@1 75.438 Acc@5 92.486\n"
|
|
]
|
|
}
|
|
],
|
|
"source": [
|
|
"with torch.no_grad():\n",
|
|
" compute_amax(model, method=\"percentile\", percentile=99.9)\n",
|
|
" evaluate(model, criterion, data_loader_test, device=\"cuda\", print_freq=20)"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 11,
|
|
"metadata": {},
|
|
"outputs": [
|
|
{
|
|
"name": "stdout",
|
|
"output_type": "stream",
|
|
"text": [
|
|
"mse calibration\n",
|
|
"Test: [ 0/98] eta: 0:06:34 loss: 0.5700 (0.5700) acc1: 85.1562 (85.1562) acc5: 96.2891 (96.2891) time: 4.0243 data: 3.3231 max mem: 5882\n",
|
|
"Test: [20/98] eta: 0:01:08 loss: 0.6758 (0.6838) acc1: 82.8125 (82.5707) acc5: 96.0938 (95.7868) time: 0.7204 data: 0.0002 max mem: 5882\n",
|
|
"Test: [40/98] eta: 0:00:46 loss: 0.7047 (0.7163) acc1: 80.2734 (81.4834) acc5: 96.2891 (95.7746) time: 0.7178 data: 0.0002 max mem: 5882\n",
|
|
"Test: [60/98] eta: 0:00:29 loss: 1.1127 (0.8585) acc1: 71.0938 (78.3395) acc5: 90.8203 (94.1278) time: 0.7192 data: 0.0002 max mem: 5882\n",
|
|
"Test: [80/98] eta: 0:00:13 loss: 1.1261 (0.9367) acc1: 72.6562 (76.7530) acc5: 89.8438 (93.1785) time: 0.7176 data: 0.0002 max mem: 5882\n",
|
|
"Test: Total time: 0:01:13\n",
|
|
" * Acc@1 76.186 Acc@5 92.926\n",
|
|
"entropy calibration\n",
|
|
"Test: [ 0/98] eta: 0:05:28 loss: 0.5648 (0.5648) acc1: 85.3516 (85.3516) acc5: 96.0938 (96.0938) time: 3.3558 data: 2.6268 max mem: 5882\n",
|
|
"Test: [20/98] eta: 0:01:05 loss: 0.6724 (0.6815) acc1: 82.8125 (82.5428) acc5: 95.8984 (95.7589) time: 0.7196 data: 0.0002 max mem: 5882\n",
|
|
"Test: [40/98] eta: 0:00:45 loss: 0.7090 (0.7149) acc1: 80.6641 (81.4929) acc5: 96.0938 (95.7269) time: 0.7214 data: 0.0002 max mem: 5882\n",
|
|
"Test: [60/98] eta: 0:00:29 loss: 1.1077 (0.8571) acc1: 72.0703 (78.3779) acc5: 90.6250 (94.0798) time: 0.7198 data: 0.0002 max mem: 5882\n",
|
|
"Test: [80/98] eta: 0:00:13 loss: 1.1253 (0.9356) acc1: 72.2656 (76.7626) acc5: 90.0391 (93.1231) time: 0.7192 data: 0.0002 max mem: 5882\n",
|
|
"Test: Total time: 0:01:12\n",
|
|
" * Acc@1 76.206 Acc@5 92.900\n"
|
|
]
|
|
}
|
|
],
|
|
"source": [
|
|
"with torch.no_grad():\n",
|
|
" for method in [\"mse\", \"entropy\"]:\n",
|
|
" print(F\"{method} calibration\")\n",
|
|
" compute_amax(model, method=method)\n",
|
|
" evaluate(model, criterion, data_loader_test, device=\"cuda\", print_freq=20)"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": null,
|
|
"metadata": {},
|
|
"outputs": [],
|
|
"source": []
|
|
}
|
|
],
|
|
"metadata": {
|
|
"kernelspec": {
|
|
"display_name": "Python 3",
|
|
"language": "python",
|
|
"name": "python3"
|
|
},
|
|
"language_info": {
|
|
"codemirror_mode": {
|
|
"name": "ipython",
|
|
"version": 3
|
|
},
|
|
"file_extension": ".py",
|
|
"mimetype": "text/x-python",
|
|
"name": "python",
|
|
"nbconvert_exporter": "python",
|
|
"pygments_lexer": "ipython3",
|
|
"version": "3.6.10"
|
|
}
|
|
},
|
|
"nbformat": 4,
|
|
"nbformat_minor": 2
|
|
}
|