# Experiments
* [x] row_col
## Row,Column Seperate
Model predict row and column respectively , and combine them to output. There are two ways : add and cross product .

Add :
    torch.add(row , col)
    Sigmoid
Mutiply :
    torch.bmm(row , col)
    Sigmoid

### Result
**Failed**
The method failed. When the optimizer wants to promote a block's probability , it higher both the column and row , which leads to a promotion of cross-like blocks. The backward works really worse in this method , *add* will leads to a score exceed 100% and *muti* will leads to a score around *0.5*.

## 3D-2D Model
### Head Record
**64**
* 4 * 3 * 3 -> 10 * 1 * 1 :　0.327
* 5 * 3 * 3 -> 9 * 1 * 1 :　0.328
* 1 * 1 * 1 -> 13 * 3 * 3 : 0.341

* 5 * 1 * 1 -> 10 * 3 * 3 :　0.344
* 14 * 1 * 1 -> 1 * 3 * 3 :　0.347

**32**
* 1 * 3 * 3 -> 13 * 1 * 1 :　0.328
* 2 * 3 * 3 -> 12 * 1 * 1 :　0.324
* 3 * 3 * 3 -> 11 * 1 * 1 :　0.315
* 4 * 3 * 3 -> 10 * 1 * 1 :　0.329
* 5 * 3 * 3 -> 9 * 1 * 1 :　0.333
* 6 * 3 * 3 -> 8 * 1 * 1 :　0.325
* 7 * 3 * 3 -> 7 * 1 * 1 :　0.322
* 8 * 3 * 3 -> 6 * 1 * 1 :　0.330
* 9 * 3 * 3 -> 5 * 1 * 1 :　0.320
* 10 * 3 * 3 -> 4 * 1 * 1 :　0.319
* 11 * 3 * 3 -> 3 * 1 * 1 :　0.319
* 12 * 3 * 3 -> 2 * 1 * 1 :　0.316
* 13 * 3 * 3 -> 1 * 1 * 1 :　0.317

**128**
* 1 * 1 * 1 -> 13 * 3 * 3 : 0.341

**256**
* 1 * 1 * 1 -> 13 * 3 * 3 : 0.338

## Map-in Map-out Model
Best :

from turtle import forward
from cv2 import merge
import torch
import torch.nn as nn
from torch.utils.data import Dataset, DataLoader
import torchvision.transforms as transforms
from torchviz import make_dot

from .blocks import SELayer3D, SELayer


class detect2D(nn.Module):
    def __init__(self, feature):
        super(detect2D, self).__init__()
        self.m1 = nn.Sequential(
            torch.nn.Conv2d(feature, feature, 3),
            torch.nn.BatchNorm2d(feature),
            nn.GELU()
        )
        self.m2 = nn.Sequential(
            torch.nn.Conv2d(feature, feature, 3),
            torch.nn.BatchNorm2d(feature),
            nn.GELU()
        )
        self.m3 = nn.Sequential(
            torch.nn.Conv2d(feature, feature, 3),
            torch.nn.BatchNorm2d(feature),
            nn.GELU()
        )
        self.m4 = nn.Sequential(
            torch.nn.Conv2d(feature, feature, 3),
            torch.nn.BatchNorm2d(feature),
            nn.GELU()
        )
        self.m5 = nn.Sequential(
            torch.nn.Conv2d(feature, feature, 3),
            torch.nn.BatchNorm2d(feature),
            nn.GELU()
        )
        self.se = SELayer(feature)
        self.flat = nn.Flatten()

    def addmap(self, big, small):
        B, C, H, W = big.size()
        return torch.nn.functional.interpolate(small, size=(H, W), mode='bilinear', align_corners=True) + big

    def forward(self, x):
        f1 = self.m1(x)
        f2 = self.m2(f1)
        f3 = self.m3(f2)
        f4 = self.m4(f3)
        f5 = self.m5(f4)
        #out = self.addmap(f1 , f2)
        #out = self.addmap(out , f3)
        out = self.addmap(f3 , f4)
        out = self.addmap(out , f5)
        #out = self.addmap(f4 , f5)
        #out = f3
        out = self.se(out)
        return out


class DCTT(nn.Module):
    def __init__(self, seq, blockn, stride=1):
        super(DCTT, self).__init__()
        self.seq = seq
        ochannel = 64
        self.ochannel = ochannel
        hidden = 64
        mapn = 64
        k1 = 1
        k2 = 3
        h1 = 14
        h2 = 15 - h1
        self.filter = nn.Sequential(
            nn.Conv3d(seq, hidden, (h1, k1, k1), stride=stride),
            nn.BatchNorm3d(hidden),
            nn.GELU(),
            nn.Conv3d(hidden, ochannel, (h2, k2, k2), stride=stride),
            nn.BatchNorm3d(ochannel),
            nn.GELU()
        )
        maphidden = 32
        self.mapping = nn.Sequential(
            nn.ConvTranspose2d(64,maphidden, 4),
            torch.nn.BatchNorm2d(maphidden),
            nn.GELU(),
            nn.ConvTranspose2d(maphidden, mapn, 4),
            torch.nn.BatchNorm2d(mapn),
            nn.GELU()
        )
        self.dts = nn.ModuleList(
            [detect2D(mapn) for i in range(blockn)])
        out_hidden = 32
        self.mapout = nn.Sequential(
            nn.ConvTranspose2d(mapn, out_hidden, 3),
            nn.BatchNorm2d(out_hidden) ,
            nn.GELU() ,
            nn.ConvTranspose2d(out_hidden, 1, 1),
            nn.Sigmoid()
        )

    def info(self, target):
        print("============ TT===========", file=target)
        print("Time_seq : ", self.seq, file=target)
        print("Detect Model 2-3D -> 3 , 2 , 2", file=target)
        print("DM number : ", len(self.dts), file=target)
        print("=====================================", file=target)

    def forward(self, x):
        x = self.filter(x)
        x = torch.squeeze(x, dim=2)
        x = self.mapping(x)
        #print(x.size())
        out = self.dts[0](x)
        for i in range(len(self.dts)):
            if i == 0:
                continue
            out += self.dts[i](x)
        #print(out.size())
        out = self.mapout(out)
        #print(out.size())
        return out