Module heyvi.model.face.recognition

import os.path
import torch
import torch.nn as nn
import math
import torch.nn.functional as F
import torchvision.transforms as transforms
from torch.autograd import Variable
import PIL
from PIL import Image, ImageFilter
import numpy as np
import pdb
import uuid


MEAN_RGB = np.array(([122.782, 117.001, 104.298]))

def convert_resnet101v4_image(img, mean_rgb=MEAN_RGB):
    """
    Convert an RGB byte image to a FloatTensor suitable for processing with the network.
    This function assumes the image has already been resized, cropped, jittered, etc.
    """
    # Subtract mean pixel value
    img_fp = img.convert('RGB') - MEAN_RGB

    # Permute dimensions so output is 3xHxW
    img_fp = np.moveaxis(img_fp, 2, 0)
    return torch.from_numpy(img_fp).float()


def unconvert_resnet101v4_image(img_in, mean_rgb=MEAN_RGB):
    """
    Convert a FloatTensor to an RGB byte Image
    """
    if img_in.dim() == 4:
        assert img_in.shape[0] == 1, "Expecting a single image as input"
        img_in = img_in.squeeze()
    img_fp = np.moveaxis(img_in.numpy(), 0, 2)
    img_fp += MEAN_RGB
    img_fp[img_fp < 0] = 0
    img_fp[img_fp > 255] = 255
    return Image.fromarray(img_fp.astype(np.uint8))



def conv3x3(in_planes, out_planes, stride=1):
    """3x3 convolution with padding"""
    return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride,
                     padding=1, bias=True)

class BasicBlock(nn.Module):
    expansion = 1

    def __init__(self, inplanes, planes, stride=1, downsample=None):
        super(BasicBlock, self).__init__()
        self.conv1 = conv3x3(inplanes, planes, stride)
        self.bn1 = nn.BatchNorm2d(planes, track_running_stats=True)
        self.relu = nn.ReLU(inplace=True)
        self.conv2 = conv3x3(planes, planes)
        self.bn2 = nn.BatchNorm2d(planes, track_running_stats=True) 
        self.downsample = downsample
        self.stride = stride

    def forward(self, x):
        residual = x

        out = self.conv1(x)
        out = self.bn1(out)
        out = self.relu(out)

        out = self.conv2(out)
        out = self.bn2(out)

        if self.downsample is not None:
            residual = self.downsample(x)

        out += residual
        out = self.relu(out)

        return out


class Bottleneck(nn.Module):
    expansion = 4

    def __init__(self, inplanes, planes, stride=1, downsample=None):
        super(Bottleneck, self).__init__()
        self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, stride=stride,
                               bias=True)
        self.bn1 = nn.BatchNorm2d(planes, track_running_stats=True, eps=1E-5)  
        self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, padding=1,
                               bias=True)
        self.bn2 = nn.BatchNorm2d(planes, track_running_stats=True, eps=1E-5)  
        self.conv3 = nn.Conv2d(planes, planes * 4, kernel_size=1, bias=True)
        self.bn3 = nn.BatchNorm2d(planes * 4, track_running_stats=True, eps=1E-5) 
        self.relu = nn.ReLU(inplace=True)
        self.downsample = downsample
        self.stride = stride

    def forward(self, x):
        residual = x

        out = self.conv1(x)
        out = self.bn1(out)
        out = self.relu(out)

        out = self.conv2(out)
        out = self.bn2(out)
        out = self.relu(out)

        out = self.conv3(out)
        out = self.bn3(out)

        if self.downsample is not None:
            residual = self.downsample(x)

        out += residual
        out = self.relu(out)
        return out


class ConcatChannels(nn.Module):
    def __init__(self, channels):
        super(ConcatChannels, self).__init__()
        self.channels = int(channels)
    def forward(self, x):
        return torch.cat((x, Variable(torch.zeros(x.size()).type_as(x.data).repeat(1,self.channels,1,1))), dim=1)

class Multiply(nn.Module): # ugh
    def __init__(self, n):
        super(Multiply, self).__init__()
        self.n = n
    def forward(self, x):
        return x*self.n


class ResNet(nn.Module):

    def __init__(self, block, layers, mode, num_classes):
        super(ResNet, self).__init__()
        valid_modes = {'encode','classify','both'}
        if mode not in valid_modes:
            raise Exception('mode should be one of ' + str(valid_modes))
        self.mode = mode
        self.inplanes = 64
        self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3,
                               bias=True)
        self.bn1 = nn.BatchNorm2d(64, track_running_stats=True, eps=1E-5)  
        self.relu = nn.ReLU(inplace=True)
        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
        self.layer1 = self._make_layer(block, 64, layers[0])
        self.layer2 = self._make_layer(block, 128, layers[1], stride=2)
        self.layer3 = self._make_layer(block, 256, layers[2], stride=2)
        self.layer4 = self._make_layer(block, 512, layers[3], stride=2)
        self.avgpool = nn.AvgPool2d(7, stride=7, count_include_pad=True)
        self.fc1 = nn.Linear(512 * block.expansion, 512)
        self.multiply = Multiply(50.0)
        self.fc2 = nn.Linear(512, num_classes)

        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
                m.weight.data.normal_(0, math.sqrt(2. / n))
                if m.bias is not None: m.bias.data.normal_(0, math.sqrt(2. / n)) # REVIEW
            elif isinstance(m, nn.BatchNorm2d):
                m.weight.data.fill_(1)
                m.bias.data.zero_()

    def _make_layer(self, block, planes, blocks, stride=1, use_conv=False):
        downsample = None
        if stride != 1 or self.inplanes != planes * block.expansion:
            if use_conv:
                downsample = nn.Sequential(
                    nn.Conv2d(self.inplanes, planes * block.expansion,
                              kernel_size=1, stride=stride, bias=False),
                    nn.BatchNorm2d(planes * block.expansion, track_running_stats=True, eps=1E-5), 
                )
            else:
                downsample = nn.Sequential(
                    nn.AvgPool2d(kernel_size=stride, stride=stride, count_include_pad=True), # VSI HACK: tying kernel_size and stride
                    ConcatChannels(planes*block.expansion//self.inplanes - 1)
                )

        layers = []
        layers.append(block(self.inplanes, planes, stride, downsample))
        self.inplanes = planes * block.expansion
        for i in range(1, blocks):
            layers.append(block(self.inplanes, planes))

        return nn.Sequential(*layers)


    def forward(self, x, mode='encode'):
        x = self.conv1(x)
        x = self.bn1(x)
        x = self.relu(x)
        x = self.maxpool(x)

        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)
        x = self.layer4(x)

        x = self.avgpool(x)  

        x = x.view(x.size(0), -1)
        x = self.fc1(x)

        xnorm = F.normalize(x, p=2, dim=1, eps=1E-5)  # torch vs. dlib normalize:  torch eps is clamped vs. dlib added to denominator
        xnorm = self.multiply(xnorm)

        mode = self.mode if None else mode
        if mode == 'encode':
            return xnorm
        else:
            scores = self.fc2(xnorm)
            if mode == 'classify':
                return scores
            elif mode == 'both':
                # else return the encoding and classification scores
                return xnorm, scores
            else:
                raise Exception('Invalid mode: ' + mode)


def resnet101v6(pthfile=None):
    """Construct resnet-101v6 model"""
    if pthfile is None:
        pthfile = os.path.join(os.path.dirname(os.path.abspath(__file__)), '..', 'models', 'recognition', 'resnet101_l2_d512_twocrop.pth')
        if not os.path.exists(pthfile):
            d = os.path.abspath(os.path.join(os.path.dirname(os.path.abspath(__file__)), '..'))
            raise ValueError('face resnet-101 face recognition models not downloaded; Run "cd %s; ./download_models.sh"' % d)
    model = ResNet(Bottleneck, [3, 4, 23, 3], mode='encode', num_classes=65359)
    model.load_state_dict(torch.load(pthfile))
    return model