models.py

import torch
import torch.nn.functional as F
from torch.nn import Linear, Sequential, BatchNorm1d, ReLU, Dropout

'''
'''
import copy
import math
import torch
import torch.nn as nn
from timm.models.layers import DropPath, trunc_normal_
from pos_encoding import get_2d_sincos_pos_embed
import torch.nn.functional as F

from einops.layers.torch import Rearrange


def _no_grad_trunc_normal_(tensor, mean, std, a, b):
    # Cut & paste from PyTorch official master until it's in a few official releases - RW
    # Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf
    def norm_cdf(x):
        # Computes standard normal cumulative distribution function
        return (1. + math.erf(x / math.sqrt(2.))) / 2.

    with torch.no_grad():
        # Values are generated by using a truncated uniform distribution and
        # then using the inverse CDF for the normal distribution.
        # Get upper and lower cdf values
        l = norm_cdf((a - mean) / std)
        u = norm_cdf((b - mean) / std)

        # Uniformly fill tensor with values from [l, u], then translate to
        # [2l-1, 2u-1].
        tensor.uniform_(2 * l - 1, 2 * u - 1)

        # Use inverse cdf transform for normal distribution to get truncated
        # standard normal
        tensor.erfinv_()

        # Transform to proper mean, std
        tensor.mul_(std * math.sqrt(2.))
        tensor.add_(mean)

        # Clamp to ensure it's in the proper range
        tensor.clamp_(min=a, max=b)
        return tensor
    
def trunc_normal_(tensor, mean=0., std=1., a=-2., b=2.):
    # type: (Tensor, float, float, float, float) -> Tensor
    return _no_grad_trunc_normal_(tensor, mean, std, a, b)



## Transformers
class Mlp(nn.Module):
    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
        super().__init__()
        out_features = out_features or in_features
        hidden_features = hidden_features or in_features
        self.fc1 = nn.Linear(in_features, hidden_features)
        self.act = act_layer()
        self.fc2 = nn.Linear(hidden_features, out_features)
        self.drop = nn.Dropout(drop)

    def forward(self, x):
        x = self.fc1(x)
        x = self.act(x)
        x = self.drop(x)
        x = self.fc2(x)
        x = self.drop(x)
        return x


class Attention_with_padding(nn.Module):
    def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.):
        super().__init__()
        self.num_heads = num_heads
        head_dim = dim // num_heads
        self.scale = qk_scale or head_dim ** -0.5
        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
        self.attn_drop = nn.Dropout(attn_drop)
        self.proj = nn.Linear(dim, dim)
        self.proj_drop = nn.Dropout(proj_drop)

    def forward(self, x, n_valid_tokens):
        B, N, C = x.shape
        qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
        q, k, v = qkv[0], qkv[1], qkv[2]  # make torchscript happy (cannot use tensor as tuple)

        attn = (q @ k.transpose(-2, -1)) * self.scale
        
        # Create the attention mask
        attention_mask = torch.arange(N).unsqueeze(0).expand(B, N).to(x.device) < n_valid_tokens.unsqueeze(1)
        attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)  # (B, 1, 1, N)
        
        # Apply the attention mask
        attn = attn.masked_fill(~attention_mask, float('-inf'))
            
        attn = attn.softmax(dim=-1)
        attn = self.attn_drop(attn)
        
        x = (attn @ v).transpose(1, 2).reshape(B, N, C)
        x = self.proj(x)
        x = self.proj_drop(x)
        return x

class Block(nn.Module):
    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
                 drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm):
        super().__init__()
        self.norm1 = norm_layer(dim)
        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
        self.norm2 = norm_layer(dim)
        mlp_hidden_dim = int(dim * mlp_ratio)
        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)

        self.attn = Attention_with_padding(
            dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)

    def forward(self, x, n_valid_tokens):
        x = x + self.drop_path(self.attn(self.norm1(x), n_valid_tokens))
        x = x + self.drop_path(self.mlp(self.norm2(x)))
        return x
    
class Encoder_Block(nn.Module):
    def __init__(self, embed_dim=384, depth=12, num_heads=6, mlp_ratio=4., qkv_bias=False, qk_scale=None,
                 drop_rate=0., attn_drop_rate=0., drop_path_rate=0., norm_layer=nn.LayerNorm):
        super().__init__() 

        self.blocks = nn.ModuleList([
            Block(
                dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,
                drop=drop_rate, attn_drop=attn_drop_rate, 
                drop_path = drop_path_rate[i] if isinstance(drop_path_rate, list) else drop_path_rate,
                norm_layer=norm_layer)
            for i in range(depth)])

    def forward(self, x, n_valid_tokens, pos_embed):
        for i, block in enumerate(self.blocks):
            x = block(x+pos_embed, n_valid_tokens) + (x + pos_embed)
        
        return x

class Transformer(nn.Module):
    def __init__(
                self, 
                embed_dim=384,
                depth=12,
                num_heads=6,
                grid_size=50,
                patch_size=5,
                mlp_ratio=4.0,
                qkv_bias=False,
                qk_scale=None,
                cls_token=False,
                token='nonlinear',
                pos_type='sincos',
                drop_rate=0.0,
                attn_drop_rate=0.0,
                drop_path_rate=0.0,
                norm_layer=nn.LayerNorm,
                init_std=0.02,
                classification_head=False,
                num_classes=5
                ):
        super().__init__()

        self.init_std = init_std
        # self.mask_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
        self.grid_size = grid_size
        self.patch_size = patch_size
        self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim), requires_grad=True) if cls_token else None
        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)]
        self.encoder_blocks = Encoder_Block(embed_dim=embed_dim, depth=depth, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale, drop_rate=drop_rate, attn_drop_rate=attn_drop_rate, drop_path_rate=dpr, norm_layer=norm_layer)
        self.norm = nn.LayerNorm(embed_dim)
        self.embed_dim = embed_dim
        self.token = token
        if token == 'conv':
            self.tokenizer = nn.Conv2d(in_channels=4, out_channels=embed_dim, kernel_size=patch_size, stride=patch_size, padding=0)
        elif token == 'linear':
            patch_dim = 4*patch_size**2
            self.tokenizer = nn.Sequential(
                Rearrange('b c (h p1) (w p2) -> b (h w) (p1 p2 c)', p1=patch_size, p2=patch_size, c=4),
                nn.LayerNorm(patch_dim),
                nn.Linear(patch_dim, embed_dim),
                nn.LayerNorm(embed_dim),
            )
        elif token == 'nonlinear':
            patch_dim = 4*patch_size**2
            self.tokenizer = nn.Sequential(
                Rearrange('b c (h p1) (w p2) -> b (h w) (p1 p2 c)', p1=patch_size, p2=patch_size, c=4),
                nn.LayerNorm(patch_dim),
                nn.Linear(patch_dim, embed_dim//2),
                nn.LayerNorm(embed_dim//2),
                nn.GELU(),
                nn.Linear(embed_dim//2, embed_dim),
                nn.LayerNorm(embed_dim),
            )
        self.classification_head = classification_head
        self.decoder = nn.Sequential(
            nn.Linear(embed_dim, embed_dim),
            nn.GELU(),
            nn.Dropout(0.2),
            nn.Linear(embed_dim, num_classes)
            # nn.Linear(embed_dim, num_classes),

        )

        self.patch_num = grid_size // patch_size    
        self.pos_type = pos_type

        if pos_type == 'sincos':
            self.pos_embed = nn.Parameter(torch.zeros((self.patch_num)**2 + cls_token, embed_dim), requires_grad=False)
            pos_embed = get_2d_sincos_pos_embed(embed_dim, self.patch_num , self.patch_num, cls_token=cls_token)
            self.pos_embed.data.copy_(torch.from_numpy(pos_embed).float())

        elif pos_type == 'learnable':
            self.pos_embed = nn.Sequential(
                nn.Linear(2, embed_dim//2),
                nn.GELU(),
                nn.Linear(embed_dim//2, embed_dim)
            )

        # initialize the learnable token
        # trunc_normal_(self.mask_token, std=init_std)
        self.apply(self._init_weights)
        self.fix_init_weight()

    def apply_transform_to_batch(batch, transform):
        n, c, h, w = batch.shape
        transformed_batch = torch.zeros_like(batch)
        
        for i in range(n):
            transformed_batch[i] = transform(batch[i])
        
        return transformed_batch


    def fix_init_weight(self):
        def rescale(param, layer_id):
            param.div_(math.sqrt(2.0 * layer_id))

        for layer_id, layer in enumerate(self.encoder_blocks.blocks):
            rescale(layer.attn.proj.weight.data, layer_id + 1)
            rescale(layer.mlp.fc2.weight.data, layer_id + 1)

    def _init_weights(self, m):
        if isinstance(m, nn.Linear):
            trunc_normal_(m.weight, std=self.init_std)
            if isinstance(m, nn.Linear) and m.bias is not None:
                nn.init.constant_(m.bias, 0)
        elif isinstance(m, nn.LayerNorm):
            nn.init.constant_(m.bias, 0)
            nn.init.constant_(m.weight, 1.0)
        elif isinstance(m, nn.Conv1d):
            trunc_normal_(m.weight, std=self.init_std)
            if m.bias is not None:
                nn.init.constant_(m.bias, 0)

    def find_empty_tokens(self, quantized_rdgms, kernel_size, stride):
        kernel = torch.ones((1, 1, kernel_size, kernel_size), dtype=torch.float32, device=quantized_rdgms.device)
        convolved = F.conv2d(quantized_rdgms, kernel, stride=stride, padding=0) # (bs, 1, 10, 10)
        convolved = convolved.squeeze(1) # (bs, 10, 10)
        mask = (convolved == 0) # True if empty token

        return mask

    def slice_and_pad(self, x, mask):

        # Find the maximum number of True values in any mask[i]
        m = torch.max(mask.sum(dim=1)).item()
        
        # Initialize the output tensor with zeros
        bs = x.size(0)
        dim = x.size(-1)

        result = torch.zeros((bs, m, dim), dtype=x.dtype, device=x.device)
        n_valid_tokens = torch.zeros(bs, dtype=torch.int64, device=x.device, requires_grad=False)
        
        for i in range(bs):
            true_indices = torch.nonzero(mask[i], as_tuple=True)[0]
            sliced_x = x[i, true_indices]
            result[i, :len(true_indices)] = sliced_x
            n_valid_tokens[i] = len(true_indices)
        
        return result, n_valid_tokens

    def forward(self, data):
        '''
        data: torch_geometric.data.batch object, added with 'ppd' attribute
        ppd : (bs, 4, grid_size, grid_size)
        '''
        ppd = data.ppd
        ppd = ppd.reshape(-1, 4, self.grid_size, self.grid_size)
        bs = ppd.size(0)

        if self.token == 'conv':
            x = self.tokenizer(ppd).reshape(bs, -1, self.embed_dim) # (bs, patch_num^2, embed_dim)
        elif self.token == 'linear' or self.token == 'nonlinear':
            x = self.tokenizer(ppd)

        if self.pos_type == 'sincos':
            pos_embed = self.pos_embed.unsqueeze(0).expand(bs, -1, self.embed_dim) # (bs, 1 + patch_num^2, embed_dim)
            
        elif self.pos_type == 'learnable':
            pos_embed = self.pos_embed(ppd.sum(dim=1).unsqueeze(1).float())

        mask = self.find_empty_tokens(ppd.sum(dim=1).unsqueeze(1), self.patch_size, self.patch_size).reshape(bs, -1) # (bs, patch_num^2)
        cls_token = self.cls_token.expand(bs, -1, -1)
        
        x = torch.cat((cls_token, x), dim=1)
        cls_mask = torch.zeros(bs, 1, dtype=torch.bool, device=x.device)
        mask = torch.cat((cls_mask, mask), dim=1)


        x, n_valid_tokens = self.slice_and_pad(x, ~mask)
        n_valid_tokens += 1

        pos_embed, _ = self.slice_and_pad(pos_embed, ~mask)
        
        x = self.encoder_blocks(x, n_valid_tokens, pos_embed)

        # Apply normalization if specified
        if self.norm is not None:
            x = self.norm(x)

        x_cls = x[:, 0]

        if self.classification_head:
            # return self.decoder(x_cls + x[:,1:].max(dim=1)[0])
            return self.decoder(x_cls)
        else:
            return x_cls


def quantization(rdgms, grid_size=50):
    '''
    Read rotated persistence diagrams, and outputs a matrix of size (grid_size, grid_size).
    Rotated diagrams are quantized into grids, where the maximum values of x and y are given by 'x_ranges' and 'y_ranges'.
    Each entry of the matrix contains the number of points in the diagram that correspond to each grid.
    Outputs the matrix.
    
    rdgms : (bs, n_pts, 2)
    output : (bs, grid_size, grid_size)
    '''
    bs, n_pts, _ = rdgms.shape
    ppd = torch.zeros((bs, grid_size, grid_size), dtype=torch.float32)
    
    births = rdgms[:, :, 0]
    pers = rdgms[:, :, 1]
    
    # Filter out points with persistence 0
    valid_mask = (pers != 0)
    births = births * valid_mask
    pers = pers * valid_mask
    
    max_birth = births.max(dim=1, keepdim=True)[0]
    max_per = pers.max(dim=1, keepdim=True)[0]

    x_range = torch.where(max_birth == 0, torch.ones_like(max_birth), 1.1 * max_birth)
    y_range = 1.1 * max_per

    x_step = x_range / grid_size
    y_step = y_range / grid_size

    x_indices = ((births / x_step).floor())
    y_indices = ((pers / y_step).floor())

    x_indices = torch.clamp(x_indices, 0, grid_size - 1)
    y_indices = torch.clamp(y_indices, 0, grid_size - 1)

    for i in range(bs):
        idx_comb = (y_indices[i] * grid_size + x_indices[i]).to(torch.int64)
        idx_comb = idx_comb[valid_mask[i]]
        counts = torch.bincount(idx_comb, minlength=grid_size * grid_size).float()
        ppd[i] = counts.view(grid_size, grid_size)
    
    return ppd



class Transformer_orbit(nn.Module):
    def __init__(
                self, 
                embed_dim=192,
                depth=5,
                num_heads=8,
                grid_size=50,
                patch_size=5,
                mlp_ratio=4.0,
                qkv_bias=False,
                qk_scale=None,
                cls_token=False,
                pos_type='learnable',
                drop_rate=0.0,
                attn_drop_rate=0.0,
                drop_path_rate=0.0,
                norm_layer=nn.LayerNorm,
                init_std=0.02,
                n_classes=5
                ):
        super().__init__()

        self.init_std = init_std
        # self.mask_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
        self.grid_size = grid_size
        self.patch_size = patch_size
        self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim), requires_grad=True) if cls_token else None
        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)]
        self.encoder_blocks = Encoder_Block(embed_dim=embed_dim, depth=depth, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale, drop_rate=drop_rate, attn_drop_rate=attn_drop_rate, drop_path_rate=dpr, norm_layer=norm_layer)
        self.norm = nn.LayerNorm(embed_dim)
        self.embed_dim = embed_dim
        self.tokenizer0 = nn.Conv1d(in_channels=1, out_channels=embed_dim, kernel_size=patch_size, stride=patch_size, padding=0)
        self.tokenizer1 = nn.Conv2d(in_channels=1, out_channels=embed_dim, kernel_size=patch_size, stride=patch_size, padding=0)
        self.decoder = nn.Sequential(
            nn.Linear(embed_dim, n_classes),
        )

        self.patch_num = grid_size // patch_size    
        self.pos_type = pos_type

        if pos_type == 'learnable':
            self.pos_embed0 = nn.Sequential(
                nn.Linear(1, embed_dim//2),
                nn.GELU(),
                nn.Linear(embed_dim//2, embed_dim)
            )

            self.pos_embed1 = nn.Sequential(
                nn.Linear(2, embed_dim//2),
                nn.GELU(),
                nn.Linear(embed_dim//2, embed_dim)
            )
        elif pos_type == 'sincos':
            self.pos_embed = nn.Parameter(torch.zeros((self.patch_num)**2 + self.patch_num + cls_token, embed_dim), requires_grad=False)
            pos_embed = get_2d_sincos_pos_embed(embed_dim, self.patch_num + 1, self.patch_num, cls_token=cls_token)
            self.pos_embed.data.copy_(torch.from_numpy(pos_embed).float())

        # initialize the learnable token
        # trunc_normal_(self.mask_token, std=init_std)
        self.apply(self._init_weights)
        self.fix_init_weight()

    def fix_init_weight(self):
        def rescale(param, layer_id):
            param.div_(math.sqrt(2.0 * layer_id))

        for layer_id, layer in enumerate(self.encoder_blocks.blocks):
            rescale(layer.attn.proj.weight.data, layer_id + 1)
            rescale(layer.mlp.fc2.weight.data, layer_id + 1)

    def _init_weights(self, m):
        if isinstance(m, nn.Linear):
            trunc_normal_(m.weight, std=self.init_std)
            if isinstance(m, nn.Linear) and m.bias is not None:
                nn.init.constant_(m.bias, 0)
        elif isinstance(m, nn.LayerNorm):
            nn.init.constant_(m.bias, 0)
            nn.init.constant_(m.weight, 1.0)
        elif isinstance(m, nn.Conv1d):
            trunc_normal_(m.weight, std=self.init_std)
            if m.bias is not None:
                nn.init.constant_(m.bias, 0)

    def find_empty_tokens(self, quantized_rdgms, kernel_size, stride):
        if quantized_rdgms.dim() == 4:
            kernel = torch.ones((1, 1, kernel_size, kernel_size), dtype=torch.float32, device=quantized_rdgms.device)
            convolved = F.conv2d(quantized_rdgms, kernel, stride=stride, padding=0) # (bs, 1, 10, 10)
        
        elif quantized_rdgms.dim() == 3:
            kernel = torch.ones((1, 1, kernel_size), dtype=torch.float32, device=quantized_rdgms.device)
            convolved = F.conv1d(quantized_rdgms, kernel, stride=stride, padding=0)
        
        convolved = convolved.squeeze(1)
        mask = (convolved == 0)

        return mask

    def slice_and_pad(self, x, mask):

        # Find the maximum number of True values in any mask[i]
        m = torch.max(mask.sum(dim=1)).item()
        
        # Initialize the output tensor with zeros
        bs = x.size(0)
        dim = x.size(-1)

        result = torch.zeros((bs, m, dim), dtype=x.dtype, device=x.device)
        n_valid_tokens = torch.zeros(bs, dtype=torch.int64, device=x.device, requires_grad=False)
        
        for i in range(bs):
            true_indices = torch.nonzero(mask[i], as_tuple=True)[0]
            sliced_x = x[i, true_indices]
            result[i, :len(true_indices)] = sliced_x
            n_valid_tokens[i] = len(true_indices)
        
        return result, n_valid_tokens


    def forward(self, rdgms0, rdgms1):
        '''
        rdgms : (bs, n_pts, 2)
        '''
        
        bs,_,_ = rdgms1.shape

        quantized_rdgms0 = quantization(rdgms0, grid_size=self.grid_size) # (bs, 1, grid_size, grid_size): quantized diagrams
        quantized_rdgms0 = quantized_rdgms0.unsqueeze(1) # (bs, 1, grid_size, grid_size)
        quantized_rdgms0 = quantized_rdgms0[:,:,:,0].to('cuda') # (bs, 1, grid_size)
        x0 = self.tokenizer0(quantized_rdgms0).permute(0,2,1) # (bs, embed_dim, grid_size) -> (bs, patch_num, embed_dim)

        quantized_rdgms1 = quantization(rdgms1, grid_size=self.grid_size) # (bs, 1, grid_size, grid_size): quantized diagrams
        quantized_rdgms1 = quantized_rdgms1.unsqueeze(1).to('cuda') # (bs, 1, grid_size, grid_size)
        mask1 = self.find_empty_tokens(quantized_rdgms1, self.patch_size, self.patch_size) # (bs, 10, 10)

        x1 = self.tokenizer1(quantized_rdgms1).permute(0,2,3,1) # (bs, embed_dim, grid_size, grid_size) -> (bs, patch_num, patch_num, embed_dim)
        pos_embed = self.pos_embed.unsqueeze(0).expand(x1.size(0), -1, self.embed_dim) # (bs, 1 + patch_num^2 + patch_num, embed_dim)
        
        x1 = x1.reshape(bs, -1, self.embed_dim) # (bs, patch_num^2, embed_dim)
        mask1 = mask1.reshape(bs, -1)

        x1, n_valid_tokens1 = self.slice_and_pad(x1, ~mask1)


        cls_mask = torch.zeros(bs, 1, dtype=torch.bool, device=x1.device)
        rdgms0_mask = torch.zeros(bs, self.patch_num, dtype=torch.bool, device=x1.device)
        mask = torch.cat([cls_mask, rdgms0_mask, mask1], dim=1)
        
        # mask = torch.cat([cls_mask, mask0, mask1], dim=1)
        pos_embed, _ = self.slice_and_pad(pos_embed, ~mask)

        # Concat x0 and x1
        cls_token = self.cls_token.expand(bs, 1, -1)
        x = torch.concat([cls_token, x0, x1], dim=1)

        n_valid_tokens1_new = self.patch_num + n_valid_tokens1 + 1
        x = self.encoder_blocks(x, n_valid_tokens1_new, pos_embed)
        
        # Apply normalization if specified
        if self.norm is not None:
            x = self.norm(x)

        x_cls = x[:, 0]

        x_cls = self.decoder(x_cls)
        # x_cls = self.decoder(x_cls + x[:,1:].max(dim=1)[0])
        return x_cls

from torch_geometric.nn import GINConv, global_add_pool

class GIN(torch.nn.Module):
    def __init__(self, dim_h, num_node_features=1, num_classes=2):
        super(GIN, self).__init__()
        self.conv1 = GINConv(Sequential(Linear(num_node_features, dim_h), BatchNorm1d(dim_h), ReLU(), Linear(dim_h, dim_h), ReLU()))
        self.conv2 = GINConv(Sequential(Linear(dim_h, dim_h), BatchNorm1d(dim_h), ReLU(), Linear(dim_h, dim_h), ReLU()))
        self.conv3 = GINConv(Sequential(Linear(dim_h, dim_h), BatchNorm1d(dim_h), ReLU(), Linear(dim_h, dim_h), ReLU()))
        self.lin1 = Linear(3*dim_h, 3*dim_h)
        self.lin2 = Linear(3*dim_h, num_classes)

    def forward(self, data):
        x, edge_index, batch = data.node_feat, data.edge_index, data.batch
        # node embedding
        h1 = self.conv1(x, edge_index)
        h2 = self.conv2(h1, edge_index)
        h3 = self.conv3(h2, edge_index)

        # graph-level readout
        h1 = global_add_pool(h1, batch)
        h2 = global_add_pool(h2, batch)
        h3 = global_add_pool(h3, batch)

        # concatenate graph embeddings
        h = torch.cat([h1, h2, h3], dim=1) # (bs, 3*dim_h)

        # Classifier
        h = self.lin1(h) 
        h = h.relu()
        h = F.dropout(h, p=0.5, training=self.training)
        h = self.lin2(h)

        return F.log_softmax(h, dim=1)


class GIN_assisted(torch.nn.Module):
    def __init__(self, transformer, dim_h, num_node_features=1, num_classes=2, assist='concat'):
        super(GIN_assisted, self).__init__()
        self.conv1 = GINConv(Sequential(Linear(num_node_features, dim_h), BatchNorm1d(dim_h), ReLU(), Linear(dim_h, dim_h), ReLU()))
        self.conv2 = GINConv(Sequential(Linear(dim_h, dim_h), BatchNorm1d(dim_h), ReLU(), Linear(dim_h, dim_h), ReLU()))
        self.conv3 = GINConv(Sequential(Linear(dim_h, dim_h), BatchNorm1d(dim_h), ReLU(), Linear(dim_h, dim_h), ReLU()))
        self.transformer = transformer
        if assist == 'concat':
            self.lin1 = Linear(3*dim_h + 192, 3*dim_h)
        elif assist == 'sum':
            self.lin1 = Linear(3*dim_h, 3*dim_h)
        self.lin2 = Linear(3*dim_h, num_classes)
        self.assist = assist

    def forward(self, data):
        x, edge_index, batch, qepd = data.node_feat, data.edge_index, data.batch, data.ppd

        # node embedding
        h1 = self.conv1(x, edge_index)
        h2 = self.conv2(h1, edge_index)
        h3 = self.conv3(h2, edge_index)

        # graph-level readout
        h1 = global_add_pool(h1, batch)
        h2 = global_add_pool(h2, batch)
        h3 = global_add_pool(h3, batch)

        # concatenate graph embeddings
        h = torch.cat([h1, h2, h3], dim=1) # (bs, 3*dim_h)

        # transformer
        ppd = data.ppd
        cls_token = self.transformer(ppd)
        
        if self.assist == 'concat':
            h = torch.cat([h, cls_token], dim=1) # (bs, 3*dim_h + embed_dim)

        elif self.assist == 'sum':
            h = h + cls_token

        # Classifier
        h = self.lin1(h) 
        h = h.relu()
        h = F.dropout(h, p=0.5, training=self.training)
        h = self.lin2(h)

        return F.log_softmax(h, dim=1)