wehub-resource-sync
606 lines
39 KiB
Plaintext
606 lines
39 KiB
Plaintext
{
|
|
"cells": [
|
|
{
|
|
"cell_type": "markdown",
|
|
"metadata": {},
|
|
"source": [
|
|
"Following the calibrate_quant_resnet50 example, now we fine tune the model"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 1,
|
|
"metadata": {},
|
|
"outputs": [
|
|
{
|
|
"name": "stderr",
|
|
"output_type": "stream",
|
|
"text": [
|
|
"WARNING: Logging before flag parsing goes to stderr.\n",
|
|
"W0608 21:25:39.018203 140228493526848 tensor_quant.py:96] Meaning of axis has changed since v2.0. Make sure to update.\n",
|
|
"W0608 21:25:39.019082 140228493526848 tensor_quant.py:96] Meaning of axis has changed since v2.0. Make sure to update.\n",
|
|
"W0608 21:25:39.019555 140228493526848 tensor_quant.py:96] Meaning of axis has changed since v2.0. Make sure to update.\n",
|
|
"W0608 21:25:39.020030 140228493526848 tensor_quant.py:96] Meaning of axis has changed since v2.0. Make sure to update.\n",
|
|
"W0608 21:25:39.020492 140228493526848 tensor_quant.py:96] Meaning of axis has changed since v2.0. Make sure to update.\n",
|
|
"W0608 21:25:39.020947 140228493526848 tensor_quant.py:96] Meaning of axis has changed since v2.0. Make sure to update.\n",
|
|
"W0608 21:25:39.021392 140228493526848 tensor_quant.py:96] Meaning of axis has changed since v2.0. Make sure to update.\n"
|
|
]
|
|
}
|
|
],
|
|
"source": [
|
|
"import datetime\n",
|
|
"import os\n",
|
|
"import sys\n",
|
|
"import time\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"
|
|
]
|
|
},
|
|
{
|
|
"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",
|
|
"sys.path.append(\"/raid/skyw/models/torchvision/references/classification/\")\n",
|
|
"from train import evaluate, train_one_epoch, load_data"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 3,
|
|
"metadata": {},
|
|
"outputs": [
|
|
{
|
|
"data": {
|
|
"text/plain": [
|
|
"QuantResNet(\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=2.6387 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0000, 0.7817](64) calibrator=MaxCalibrator(track_amax=False) 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=2.9730 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0000, 0.7266](64) calibrator=MaxCalibrator(track_amax=False) 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=1.0971 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0000, 0.4679](64) calibrator=MaxCalibrator(track_amax=False) 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=1.3318 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0000, 0.3936](256) calibrator=MaxCalibrator(track_amax=False) 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=2.9730 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0000, 0.9879](256) calibrator=MaxCalibrator(track_amax=False) 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=1.4872 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0000, 0.2618](64) calibrator=MaxCalibrator(track_amax=False) 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=1.0466 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0564, 0.5201](64) calibrator=MaxCalibrator(track_amax=False) 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=1.5106 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0000, 0.2946](256) calibrator=MaxCalibrator(track_amax=False) 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=1.5250 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0533, 0.1921](64) calibrator=MaxCalibrator(track_amax=False) 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=0.9980 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0810, 0.2856](64) calibrator=MaxCalibrator(track_amax=False) 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=1.6532 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0197, 0.2752](256) calibrator=MaxCalibrator(track_amax=False) 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=1.5499 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0670, 0.3532](128) calibrator=MaxCalibrator(track_amax=False) 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=1.1606 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0694, 0.2993](128) calibrator=MaxCalibrator(track_amax=False) 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=1.1425 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0000, 0.3917](512) calibrator=MaxCalibrator(track_amax=False) 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=1.5499 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0000, 0.5662](512) calibrator=MaxCalibrator(track_amax=False) 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=1.4626 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0370, 0.2522](128) calibrator=MaxCalibrator(track_amax=False) 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=0.8304 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0142, 0.2998](128) calibrator=MaxCalibrator(track_amax=False) 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=1.1722 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0000, 0.3038](512) calibrator=MaxCalibrator(track_amax=False) 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=1.4864 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0653, 0.2383](128) calibrator=MaxCalibrator(track_amax=False) 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=0.9450 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0646, 0.2556](128) calibrator=MaxCalibrator(track_amax=False) 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=0.8535 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0162, 0.3522](512) calibrator=MaxCalibrator(track_amax=False) 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=1.5229 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0648, 0.2814](128) calibrator=MaxCalibrator(track_amax=False) 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=0.9247 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0595, 0.2210](128) calibrator=MaxCalibrator(track_amax=False) 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=0.9747 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0127, 0.2956](512) calibrator=MaxCalibrator(track_amax=False) 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=1.5941 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0842, 0.3425](256) calibrator=MaxCalibrator(track_amax=False) 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=1.3565 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0544, 0.2008](256) calibrator=MaxCalibrator(track_amax=False) 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=1.0293 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0000, 0.3212](1024) calibrator=MaxCalibrator(track_amax=False) 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=1.5941 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0000, 0.3460](1024) calibrator=MaxCalibrator(track_amax=False) 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=1.3305 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0407, 0.2942](256) calibrator=MaxCalibrator(track_amax=False) 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=0.9844 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0432, 0.2634](256) calibrator=MaxCalibrator(track_amax=False) 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=0.9333 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0000, 0.4969](1024) calibrator=MaxCalibrator(track_amax=False) 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=1.3388 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0469, 0.2715](256) calibrator=MaxCalibrator(track_amax=False) 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=0.8617 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0397, 0.2100](256) calibrator=MaxCalibrator(track_amax=False) 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=0.7507 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0000, 0.3538](1024) calibrator=MaxCalibrator(track_amax=False) 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=1.3554 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0553, 0.2390](256) calibrator=MaxCalibrator(track_amax=False) 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=0.9257 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0455, 0.2792](256) calibrator=MaxCalibrator(track_amax=False) 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=0.8117 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0000, 0.3126](1024) calibrator=MaxCalibrator(track_amax=False) 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=1.4199 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0598, 0.2722](256) calibrator=MaxCalibrator(track_amax=False) 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=0.9274 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0459, 0.1919](256) calibrator=MaxCalibrator(track_amax=False) 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=0.8702 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0160, 0.3161](1024) calibrator=MaxCalibrator(track_amax=False) 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=1.4258 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0618, 0.3995](256) calibrator=MaxCalibrator(track_amax=False) 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=1.2256 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0498, 0.2236](256) calibrator=MaxCalibrator(track_amax=False) 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=1.3560 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0180, 0.3288](1024) calibrator=MaxCalibrator(track_amax=False) 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=1.3915 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0744, 0.3415](512) calibrator=MaxCalibrator(track_amax=False) 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=1.1571 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0432, 0.3993](512) calibrator=MaxCalibrator(track_amax=False) 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=1.1295 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0198, 0.3546](2048) calibrator=MaxCalibrator(track_amax=False) 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=1.3915 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0197, 0.6413](2048) calibrator=MaxCalibrator(track_amax=False) 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=3.9348 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0482, 0.7003](512) calibrator=MaxCalibrator(track_amax=False) 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=1.1277 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0379, 0.2257](512) calibrator=MaxCalibrator(track_amax=False) 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=0.8992 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0211, 0.2427](2048) calibrator=MaxCalibrator(track_amax=False) 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=5.2181 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0601, 0.4541](512) calibrator=MaxCalibrator(track_amax=False) 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=1.2051 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0338, 0.1416](512) calibrator=MaxCalibrator(track_amax=False) 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=1.1045 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.0119, 0.2798](2048) calibrator=MaxCalibrator(track_amax=False) 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=5.5345 calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" (_weight_quantizer): TensorQuantizer(8bit fake axis=0 amax=[0.1716, 0.7371](1000) calibrator=MaxCalibrator(track_amax=False) quant)\n",
|
|
" )\n",
|
|
")"
|
|
]
|
|
},
|
|
"execution_count": 3,
|
|
"metadata": {},
|
|
"output_type": "execute_result"
|
|
}
|
|
],
|
|
"source": [
|
|
"from pytorch_quantization import quant_modules\n",
|
|
"quant_modules.initialize()\n",
|
|
"\n",
|
|
"# Create and load the calibrated model\n",
|
|
"model = torchvision.models.resnet50()\n",
|
|
"model.load_state_dict(torch.load(\"/tmp/quant_resnet50-calibrated.pth\"))\n",
|
|
"model.cuda()"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 4,
|
|
"metadata": {},
|
|
"outputs": [],
|
|
"source": [
|
|
"data_path = \"/raid/data/imagenet/imagenet_pytorch\"\n",
|
|
"\n",
|
|
"traindir = os.path.join(data_path, 'train')\n",
|
|
"valdir = os.path.join(data_path, 'val')\n",
|
|
"dataset, dataset_test, train_sampler, test_sampler = load_data(traindir, valdir, False, False)\n",
|
|
"\n",
|
|
"data_loader = torch.utils.data.DataLoader(\n",
|
|
" dataset, batch_size=256,\n",
|
|
" sampler=train_sampler, num_workers=4, pin_memory=True)\n",
|
|
"\n",
|
|
"data_loader_test = torch.utils.data.DataLoader(\n",
|
|
" dataset_test, batch_size=256,\n",
|
|
" sampler=test_sampler, num_workers=4, pin_memory=True)"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "markdown",
|
|
"metadata": {},
|
|
"source": [
|
|
"## Quantized fine tuning\n",
|
|
"Let's fine tune the model with fake quantization. We only fine tune for 1 epoch as an example."
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": null,
|
|
"metadata": {},
|
|
"outputs": [],
|
|
"source": [
|
|
"criterion = nn.CrossEntropyLoss()\n",
|
|
"optimizer = torch.optim.SGD(model.parameters(), lr=0.0001)\n",
|
|
"lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=1, gamma=0.1)\n",
|
|
"\n",
|
|
"data_loader = torch.utils.data.DataLoader(\n",
|
|
" dataset, batch_size=128,\n",
|
|
" sampler=train_sampler, num_workers=16, pin_memory=True)\n",
|
|
"\n",
|
|
"# Training takes about one and half hour per epoch on single V100\n",
|
|
"train_one_epoch(model, criterion, optimizer, data_loader, \"cuda\", 0, 100)"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "markdown",
|
|
"metadata": {},
|
|
"source": [
|
|
"## Evaluate the fine tuned model"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 6,
|
|
"metadata": {
|
|
"scrolled": true
|
|
},
|
|
"outputs": [
|
|
{
|
|
"name": "stdout",
|
|
"output_type": "stream",
|
|
"text": [
|
|
"Test: [ 0/196] eta: 0:09:36 loss: 0.4680 (0.4680) acc1: 85.9375 (85.9375) acc5: 98.0469 (98.0469) time: 2.9406 data: 2.1852 max mem: 16096\n",
|
|
"Test: [ 10/196] eta: 0:01:58 loss: 0.6694 (0.6522) acc1: 83.2031 (82.9545) acc5: 96.0938 (96.1293) time: 0.6346 data: 0.1988 max mem: 16096\n",
|
|
"Test: [ 20/196] eta: 0:01:30 loss: 0.6738 (0.6733) acc1: 82.0312 (82.4777) acc5: 95.7031 (95.7961) time: 0.3928 data: 0.0001 max mem: 16096\n",
|
|
"Test: [ 30/196] eta: 0:01:18 loss: 0.6219 (0.6322) acc1: 84.3750 (83.9718) acc5: 95.7031 (96.0181) time: 0.3859 data: 0.0001 max mem: 16096\n",
|
|
"Test: [ 40/196] eta: 0:01:10 loss: 0.6801 (0.6750) acc1: 81.6406 (82.5934) acc5: 95.7031 (95.9604) time: 0.3861 data: 0.0001 max mem: 16096\n",
|
|
"Test: [ 50/196] eta: 0:01:04 loss: 0.6937 (0.6724) acc1: 80.0781 (82.3529) acc5: 96.8750 (96.0938) time: 0.3834 data: 0.0001 max mem: 16096\n",
|
|
"Test: [ 60/196] eta: 0:00:58 loss: 0.7149 (0.6849) acc1: 80.0781 (81.9864) acc5: 96.4844 (96.1066) time: 0.3854 data: 0.0001 max mem: 16096\n",
|
|
"Test: [ 70/196] eta: 0:00:53 loss: 0.6616 (0.6716) acc1: 80.8594 (82.2843) acc5: 96.4844 (96.1983) time: 0.3859 data: 0.0001 max mem: 16096\n",
|
|
"Test: [ 80/196] eta: 0:00:48 loss: 0.6510 (0.6968) acc1: 81.2500 (81.7467) acc5: 95.7031 (95.9201) time: 0.3860 data: 0.0001 max mem: 16096\n",
|
|
"Test: [ 90/196] eta: 0:00:44 loss: 0.9469 (0.7444) acc1: 76.1719 (80.6834) acc5: 92.5781 (95.4370) time: 0.3868 data: 0.0001 max mem: 16096\n",
|
|
"Test: [100/196] eta: 0:00:39 loss: 1.1594 (0.7964) acc1: 70.7031 (79.5521) acc5: 90.6250 (94.8755) time: 0.3864 data: 0.0001 max mem: 16096\n",
|
|
"Test: [110/196] eta: 0:00:35 loss: 1.1594 (0.8214) acc1: 72.2656 (79.0365) acc5: 91.4062 (94.6298) time: 0.3836 data: 0.0001 max mem: 16096\n",
|
|
"Test: [120/196] eta: 0:00:31 loss: 0.9820 (0.8389) acc1: 76.1719 (78.7771) acc5: 92.1875 (94.3634) time: 0.3856 data: 0.0001 max mem: 16096\n",
|
|
"Test: [130/196] eta: 0:00:26 loss: 1.0825 (0.8705) acc1: 72.6562 (77.9610) acc5: 91.4062 (94.0303) time: 0.3866 data: 0.0001 max mem: 16096\n",
|
|
"Test: [140/196] eta: 0:00:22 loss: 1.1088 (0.8889) acc1: 72.2656 (77.6125) acc5: 91.4062 (93.8137) time: 0.3879 data: 0.0001 max mem: 16096\n",
|
|
"Test: [150/196] eta: 0:00:18 loss: 1.1069 (0.9059) acc1: 73.0469 (77.2998) acc5: 91.0156 (93.5586) time: 0.3914 data: 0.0002 max mem: 16096\n",
|
|
"Test: [160/196] eta: 0:00:14 loss: 1.1360 (0.9197) acc1: 73.0469 (77.0380) acc5: 90.2344 (93.3472) time: 0.3898 data: 0.0002 max mem: 16096\n",
|
|
"Test: [170/196] eta: 0:00:10 loss: 1.2171 (0.9371) acc1: 71.8750 (76.6265) acc5: 89.8438 (93.1721) time: 0.3845 data: 0.0002 max mem: 16096\n",
|
|
"Test: [180/196] eta: 0:00:06 loss: 1.2493 (0.9527) acc1: 68.7500 (76.2992) acc5: 90.2344 (93.0205) time: 0.3815 data: 0.0001 max mem: 16096\n",
|
|
"Test: [190/196] eta: 0:00:02 loss: 1.0816 (0.9511) acc1: 71.4844 (76.3089) acc5: 92.1875 (93.0465) time: 0.3736 data: 0.0001 max mem: 16096\n",
|
|
"Test: Total time: 0:01:17\n",
|
|
" * Acc@1 76.426 Acc@5 93.080\n"
|
|
]
|
|
}
|
|
],
|
|
"source": [
|
|
"with torch.no_grad():\n",
|
|
" evaluate(model, criterion, data_loader_test, device=\"cuda\")"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "markdown",
|
|
"metadata": {},
|
|
"source": [
|
|
"After only 1 epoch of quantized fine tuning, top-1 improved from ~76.1 to 76.426. Train longer with lr anealing can improve accuracy further"
|
|
]
|
|
}
|
|
],
|
|
"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
|
|
}
|