models.py

import os
import numpy as np
from torchvision import transforms
import torch
import torch.nn as nn
import PIL
import clip
import open_clip
from functools import partial
import random
import json
from tqdm import tqdm
import utils

# class BrainMLP(nn.Module):
#     def __init__(self, out_dim=257*768, in_dim=15724, clip_size=768, h=4096):
#         super().__init__()
#         self.lin0 = nn.Sequential(
#             nn.Linear(in_dim, h, bias=False),
#             nn.LayerNorm(h),
#             nn.GELU(inplace=True),
#             nn.Dropout(0.5))
#         self.mlp = nn.ModuleList([
#             nn.Sequential(
#                 nn.Linear(h, h),
#                 nn.LayerNorm(h),
#                 nn.GELU(inplace=True),
#                 nn.Dropout(0.15)
#             ) for _ in range(4)])
#         self.lin1 = nn.Linear(h, out_dim, bias=True)
#         self.proj = nn.Sequential(
#             nn.LayerNorm(clip_size),
#             nn.GELU(),
#             nn.Linear(clip_size, 2048),
#             nn.LayerNorm(2048),
#             nn.GELU(),
#             nn.Linear(2048, 2048),
#             nn.LayerNorm(2048),
#             nn.GELU(),
#             nn.Linear(2048, clip_size))
#     def forward(self, x):
#         x = self.lin0(x)
#         residual = x
#         for res_block in range(self.n_blocks):
#             x = self.mlp[res_block](x)
#             x += residual
#             residual = x
#         diffusion_prior_input = self.lin1(x.reshape(len(x), -1))
#         disjointed_clip_fmri = self.proj(diffusion_prior_input.reshape(
#                                         len(x),-1, self.clip_size))
#         return diffusion_prior_input, disjointed_clip_fmri



class Clipper(torch.nn.Module):
    def __init__(self, clip_variant, clamp_embs=False, norm_embs=False,
                 hidden_state=False, device=torch.device('cpu')):
        super().__init__()
        assert clip_variant in ("RN50", "ViT-L/14", "ViT-B/32", "RN50x64"), \
            "clip_variant must be one of RN50, ViT-L/14, ViT-B/32, RN50x64"
        print(clip_variant, device)
        
        if clip_variant=="ViT-L/14" and hidden_state:
            # from transformers import CLIPVisionModelWithProjection
            # image_encoder = CLIPVisionModelWithProjection.from_pretrained("openai/clip-vit-large-patch14",cache_dir="/fsx/proj-medarc/fmri/cache")
            from transformers import CLIPVisionModelWithProjection
            sd_cache_dir = '/fsx/proj-fmri/shared/cache/models--shi-labs--versatile-diffusion/snapshots/2926f8e11ea526b562cd592b099fcf9c2985d0b7'
            image_encoder = CLIPVisionModelWithProjection.from_pretrained(sd_cache_dir, subfolder='image_encoder').eval()
            image_encoder = image_encoder.to(device)
            for param in image_encoder.parameters():
                param.requires_grad = False # dont need to calculate gradients
            self.image_encoder = image_encoder
        elif hidden_state:
            raise Exception("hidden_state embeddings only works with ViT-L/14 right now")
        
        clip_model, preprocess = clip.load(clip_variant, device=device)
        clip_model.eval() # dont want to train model
        for param in clip_model.parameters():
            param.requires_grad = False # dont need to calculate gradients
            
        self.clip = clip_model
        self.clip_variant = clip_variant
        if clip_variant == "RN50x64":
            self.clip_size = (448,448)
        else:
            self.clip_size = (224,224)
            
        preproc = transforms.Compose([
            transforms.Resize(size=self.clip_size[0], interpolation=transforms.InterpolationMode.BICUBIC),
            transforms.CenterCrop(size=self.clip_size),
            transforms.Normalize(mean=(0.48145466, 0.4578275, 0.40821073), std=(0.26862954, 0.26130258, 0.27577711))
        ])
        self.preprocess = preproc
        self.hidden_state = hidden_state
        self.mean = np.array([0.48145466, 0.4578275, 0.40821073])
        self.std = np.array([0.26862954, 0.26130258, 0.27577711])
        self.normalize = transforms.Normalize(self.mean, self.std)
        self.denormalize = transforms.Normalize((-self.mean / self.std).tolist(), (1.0 / self.std).tolist())
        self.clamp_embs = clamp_embs
        self.norm_embs = norm_embs
        self.device= device
        
        def versatile_normalize_embeddings(encoder_output):
            embeds = encoder_output.last_hidden_state
            embeds = image_encoder.vision_model.post_layernorm(embeds)
            embeds = image_encoder.visual_projection(embeds)
            return embeds
        self.versatile_normalize_embeddings = versatile_normalize_embeddings

    def resize_image(self, image):
        # note: antialias should be False if planning to use Pinkney's Image Variation SD model
        return transforms.Resize(self.clip_size)(image.to(self.device))

    def embed_image(self, image):
        """Expects images in -1 to 1 range"""
        if self.hidden_state:
            # clip_emb = self.preprocess((image/1.5+.25).to(self.device)) # for some reason the /1.5+.25 prevents oversaturation
            clip_emb = self.preprocess((image).to(self.device))
            clip_emb = self.image_encoder(clip_emb)
            clip_emb = self.versatile_normalize_embeddings(clip_emb)
        else:
            clip_emb = self.preprocess(image.to(self.device))
            clip_emb = self.clip.encode_image(clip_emb)
        # input is now in CLIP space, but mind-reader preprint further processes embeddings:
        if self.clamp_embs:
            clip_emb = torch.clamp(clip_emb, -1.5, 1.5)
        if self.norm_embs:
            if self.hidden_state:        
                # normalize all tokens by cls token's norm
                clip_emb = clip_emb / torch.norm(clip_emb[:, 0], dim=-1).reshape(-1, 1, 1)
            else:
                clip_emb = nn.functional.normalize(clip_emb, dim=-1)
        return clip_emb

    def embed_text(self, text_samples):
        clip_text = clip.tokenize(text_samples).to(self.device)
        clip_text = self.clip.encode_text(clip_text)
        if self.clamp_embs:
            clip_text = torch.clamp(clip_text, -1.5, 1.5)
        if self.norm_embs:
            clip_text = nn.functional.normalize(clip_text, dim=-1)
        return clip_text

    def embed_curated_annotations(self, annots):
        for i,b in enumerate(annots):
            t = ''
            while t == '':
                rand = torch.randint(5,(1,1))[0][0]
                t = b[0,rand]
            if i==0:
                txt = np.array(t)
            else:
                txt = np.vstack((txt,t))
        txt = txt.flatten()
        return self.embed_text(txt)

# for prior
from dalle2_pytorch import DiffusionPrior
from dalle2_pytorch.dalle2_pytorch import l2norm, default, exists
from dalle2_pytorch.train_configs import DiffusionPriorNetworkConfig
# vd prior
from dalle2_pytorch.dalle2_pytorch import RotaryEmbedding, CausalTransformer, SinusoidalPosEmb, MLP, Rearrange, repeat, rearrange, prob_mask_like, LayerNorm, RelPosBias, Attention, FeedForward


class BrainDiffusionPrior(DiffusionPrior):
    """ 
    Differences from original:
    - Allow for passing of generators to torch random functions
    - Option to include the voxel2clip model and pass voxels into forward method
    - Return predictions when computing loss
    - Load pretrained model from @nousr trained on LAION aesthetics
    """
    def __init__(self, *args, **kwargs):
        voxel2clip = kwargs.pop('voxel2clip', None)
        super().__init__(*args, **kwargs)
        self.voxel2clip = voxel2clip

    @torch.no_grad()
    def p_sample(self, x, t, text_cond = None, self_cond = None, clip_denoised = True, cond_scale = 1.,
                generator=None):
        b, *_, device = *x.shape, x.device
        model_mean, _, model_log_variance, x_start = self.p_mean_variance(x = x, t = t, text_cond = text_cond, self_cond = self_cond, clip_denoised = clip_denoised, cond_scale = cond_scale)
        if generator is None:
            noise = torch.randn_like(x)
        else:
            noise = torch.randn_like(x)
            # noise = torch.randn(x.size(), device=x.device, dtype=x.dtype, generator=generator)
        # no noise when t == 0
        nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1)))
        pred = model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
        return pred, x_start

    @torch.no_grad()
    def p_sample_loop(self, *args, timesteps = None, **kwargs):
        timesteps = default(timesteps, self.noise_scheduler.num_timesteps)
        assert timesteps <= self.noise_scheduler.num_timesteps
        is_ddim = timesteps < self.noise_scheduler.num_timesteps

        if not is_ddim:
            normalized_image_embed = self.p_sample_loop_ddpm(*args, **kwargs)
        else:
            normalized_image_embed = self.p_sample_loop_ddim(*args, **kwargs, timesteps = timesteps)

        # print("PS removed all image_embed_scale instances!")
        image_embed = normalized_image_embed #/ self.image_embed_scale
        return image_embed

    @torch.no_grad()
    def p_sample_loop_ddpm(self, shape, text_cond, cond_scale = 1., generator=None):
        batch, device = shape[0], self.device

        if generator is None:
            image_embed = torch.randn(shape, device = device)
        else:
            image_embed = torch.randn(shape, device = device, generator=generator)
        x_start = None # for self-conditioning

        if self.init_image_embed_l2norm:
            image_embed = l2norm(image_embed) * self.image_embed_scale

        for i in tqdm(reversed(range(0, self.noise_scheduler.num_timesteps)), desc='sampling loop time step', total=self.noise_scheduler.num_timesteps, disable=True):
            times = torch.full((batch,), i, device = device, dtype = torch.long)

            self_cond = x_start if self.net.self_cond else None
            image_embed, x_start = self.p_sample(image_embed, times, text_cond = text_cond, self_cond = self_cond, cond_scale = cond_scale, 
                                                 generator=generator)

        if self.sampling_final_clamp_l2norm and self.predict_x_start:
            image_embed = self.l2norm_clamp_embed(image_embed)

        return image_embed

    def p_losses(self, image_embed, times, text_cond, noise = None):
        noise = default(noise, lambda: torch.randn_like(image_embed))

        image_embed_noisy = self.noise_scheduler.q_sample(x_start = image_embed, t = times, noise = noise)

        self_cond = None
        if self.net.self_cond and random.random() < 0.5:
            with torch.no_grad():
                self_cond = self.net(image_embed_noisy, times, **text_cond).detach()

        pred = self.net(
            image_embed_noisy,
            times,
            self_cond = self_cond,
            text_cond_drop_prob = self.text_cond_drop_prob,
            image_cond_drop_prob = self.image_cond_drop_prob,
            **text_cond
        )

        if self.predict_x_start and self.training_clamp_l2norm:
            pred = self.l2norm_clamp_embed(pred)

        if self.predict_v:
            target = self.noise_scheduler.calculate_v(image_embed, times, noise)
        elif self.predict_x_start:
            target = image_embed
        else:
            target = noise

        loss = nn.functional.mse_loss(pred, target) # mse
        # print("1", loss)
        # loss += (1 - nn.functional.cosine_similarity(pred, target).mean())
        # print("2", (1 - nn.functional.cosine_similarity(pred, target).mean()))
        return loss, pred

    def forward(
        self,
        text = None,
        image = None,
        voxel = None,
        text_embed = None,      # allow for training on preprocessed CLIP text and image embeddings
        image_embed = None,
        text_encodings = None,  # as well as CLIP text encodings
        *args,
        **kwargs
    ):
        assert exists(text) ^ exists(text_embed) ^ exists(voxel), 'either text, text embedding, or voxel must be supplied'
        assert exists(image) ^ exists(image_embed), 'either image or image embedding must be supplied'
        assert not (self.condition_on_text_encodings and (not exists(text_encodings) and not exists(text))), 'text encodings must be present if you specified you wish to condition on it on initialization'

        if exists(voxel):
            assert exists(self.voxel2clip), 'voxel2clip must be trained if you wish to pass in voxels'
            assert not exists(text_embed), 'cannot pass in both text and voxels'
            if self.voxel2clip.use_projector:
                clip_voxels_mse, clip_voxels = self.voxel2clip(voxel)
                text_embed = clip_voxels_mse
            else:
                clip_voxels = self.voxel2clip(voxel)
                text_embed = clip_voxels_mse = clip_voxels
            # text_embed = self.voxel2clip(voxel)

        if exists(image):
            image_embed, _ = self.clip.embed_image(image)

        # calculate text conditionings, based on what is passed in

        if exists(text):
            text_embed, text_encodings = self.clip.embed_text(text)

        text_cond = dict(text_embed = text_embed)

        if self.condition_on_text_encodings:
            assert exists(text_encodings), 'text encodings must be present for diffusion prior if specified'
            text_cond = {**text_cond, 'text_encodings': text_encodings}

        # timestep conditioning from ddpm

        batch, device = image_embed.shape[0], image_embed.device
        times = self.noise_scheduler.sample_random_times(batch)

        # PS: I dont think we need this? also if uncommented this does in-place global variable change
        # scale image embed (Katherine)
        # image_embed *= self.image_embed_scale

        # calculate forward loss

        loss, pred = self.p_losses(image_embed, times, text_cond = text_cond, *args, **kwargs)
        
        # undo the scaling so we can directly use it for real mse loss and reconstruction
        return loss, pred

    
class VersatileDiffusionPriorNetwork(nn.Module):
    def __init__(
        self,
        dim,
        num_timesteps = None,
        num_time_embeds = 1,
        # num_image_embeds = 1,
        # num_brain_embeds = 1,
        num_tokens = 257,
        causal = True,
        learned_query_mode = 'none',
        **kwargs
    ):
        super().__init__()
        self.dim = dim
        self.num_time_embeds = num_time_embeds
        self.continuous_embedded_time = not exists(num_timesteps)
        self.learned_query_mode = learned_query_mode

        self.to_time_embeds = nn.Sequential(
            nn.Embedding(num_timesteps, dim * num_time_embeds) if exists(num_timesteps) else nn.Sequential(SinusoidalPosEmb(dim), MLP(dim, dim * num_time_embeds)), # also offer a continuous version of timestep embeddings, with a 2 layer MLP
            Rearrange('b (n d) -> b n d', n = num_time_embeds)
        )

        if self.learned_query_mode == 'token':
            self.learned_query = nn.Parameter(torch.randn(num_tokens, dim))
        if self.learned_query_mode == 'pos_emb':
            scale = dim ** -0.5
            self.learned_query = nn.Parameter(torch.randn(num_tokens, dim) * scale)
        if self.learned_query_mode == 'all_pos_emb':
            scale = dim ** -0.5
            self.learned_query = nn.Parameter(torch.randn(num_tokens*2+1, dim) * scale)
        self.causal_transformer = FlaggedCausalTransformer(dim = dim, causal=causal, **kwargs)

        self.null_brain_embeds = nn.Parameter(torch.randn(num_tokens, dim))
        self.null_image_embed = nn.Parameter(torch.randn(num_tokens, dim))

        self.num_tokens = num_tokens
        self.self_cond = False

    def forward_with_cond_scale(
        self,
        *args,
        cond_scale = 1.,
        **kwargs
    ):
        logits = self.forward(*args, **kwargs)

        if cond_scale == 1:
            return logits

        null_logits = self.forward(*args, brain_cond_drop_prob = 1., image_cond_drop_prob = 1, **kwargs)
        return null_logits + (logits - null_logits) * cond_scale

    def forward(
        self,
        image_embed,
        diffusion_timesteps,
        *,
        self_cond=None,
        brain_embed=None,
        text_embed=None,
        brain_cond_drop_prob = 0.,
        text_cond_drop_prob = None,
        image_cond_drop_prob = 0.
    ):
        if text_embed is not None:
            brain_embed = text_embed
        if text_cond_drop_prob is not None:
            brain_cond_drop_prob = text_cond_drop_prob
        
        # image_embed = image_embed.view(len(image_embed),-1,16*16)
        # text_embed = text_embed.view(len(text_embed),-1,768)
        # brain_embed = brain_embed.view(len(brain_embed),-1,16*16)
        # print(*image_embed.shape)
        # print(*image_embed.shape, image_embed.device, image_embed.dtype)
        
        batch, _, dim, device, dtype = *image_embed.shape, image_embed.device, image_embed.dtype
        # num_time_embeds, num_image_embeds, num_brain_embeds = self.num_time_embeds, self.num_image_embeds, self.num_brain_embeds
        
        # classifier free guidance masks
        brain_keep_mask = prob_mask_like((batch,), 1 - brain_cond_drop_prob, device = device)
        brain_keep_mask = rearrange(brain_keep_mask, 'b -> b 1 1')

        image_keep_mask = prob_mask_like((batch,), 1 - image_cond_drop_prob, device = device)
        image_keep_mask = rearrange(image_keep_mask, 'b -> b 1 1')

        # mask out brain embeddings with null brain embeddings

        # import pdb; pdb.set_trace()
        null_brain_embeds = self.null_brain_embeds.to(brain_embed.dtype)
        brain_embed = torch.where(
            brain_keep_mask,
            brain_embed,
            null_brain_embeds[None]
        )

        # mask out image embeddings with null image embeddings
        null_image_embed = self.null_image_embed.to(image_embed.dtype)
        image_embed = torch.where(
            image_keep_mask,
            image_embed,
            null_image_embed[None]
        )

        # whether brain embedding is used for conditioning depends on whether brain encodings are available for attention (for classifier free guidance, even though it seems from the paper it was not used in the prior ddpm, as the objective is different)
        # but let's just do it right
        if self.continuous_embedded_time:
            # if continuous cast to flat, else keep int for indexing embeddings
            diffusion_timesteps = diffusion_timesteps.type(dtype)
        time_embed = self.to_time_embeds(diffusion_timesteps)

        if self.learned_query_mode == 'token':
            learned_queries = repeat(self.learned_query, 'n d -> b n d', b = batch)
        elif self.learned_query_mode == 'pos_emb':
            pos_embs = repeat(self.learned_query, 'n d -> b n d', b = batch)
            image_embed = image_embed + pos_embs
            learned_queries = torch.empty((batch, 0, dim), device=brain_embed.device)
        elif self.learned_query_mode == 'all_pos_emb':
            pos_embs = repeat(self.learned_query, 'n d -> b n d', b = batch)
            learned_queries = torch.empty((batch, 0, dim), device=brain_embed.device)
        else:
            learned_queries = torch.empty((batch, 0, dim), device=brain_embed.device)
        
        tokens = torch.cat((
            brain_embed,  # 257
            time_embed,  # 1
            image_embed,  # 257
            learned_queries  # 257
        ), dim = -2)
        if self.learned_query_mode == 'all_pos_emb':
            tokens = tokens + pos_embs

        # attend
        tokens = self.causal_transformer(tokens)

        # get learned query, which should predict the image embedding (per DDPM timestep)
        pred_image_embed = tokens[..., -self.num_tokens:, :]

        return pred_image_embed

class FlaggedCausalTransformer(nn.Module):
    def __init__(
        self,
        *,
        dim,
        depth,
        dim_head = 64,
        heads = 8,
        ff_mult = 4,
        norm_in = False,
        norm_out = True,
        attn_dropout = 0.,
        ff_dropout = 0.,
        final_proj = True,
        normformer = False,
        rotary_emb = True,
        causal=True
    ):
        super().__init__()
        self.init_norm = LayerNorm(dim) if norm_in else nn.Identity() # from latest BLOOM model and Yandex's YaLM

        self.rel_pos_bias = RelPosBias(heads = heads)

        rotary_emb = RotaryEmbedding(dim = min(32, dim_head)) if rotary_emb else None

        self.layers = nn.ModuleList([])
        for _ in range(depth):
            self.layers.append(nn.ModuleList([
                Attention(dim = dim, causal = causal, dim_head = dim_head, heads = heads, dropout = attn_dropout, rotary_emb = rotary_emb),
                FeedForward(dim = dim, mult = ff_mult, dropout = ff_dropout, post_activation_norm = normformer)
            ]))

        self.norm = LayerNorm(dim, stable = True) if norm_out else nn.Identity()  # unclear in paper whether they projected after the classic layer norm for the final denoised image embedding, or just had the transformer output it directly: plan on offering both options
        self.project_out = nn.Linear(dim, dim, bias = False) if final_proj else nn.Identity()

    def forward(self, x):
        n, device = x.shape[1], x.device

        x = self.init_norm(x)

        attn_bias = self.rel_pos_bias(n, n + 1, device = device)

        for attn, ff in self.layers:
            x = attn(x, attn_bias = attn_bias) + x
            x = ff(x) + x

        out = self.norm(x)
        return self.project_out(out)
    
#Subclass for GNET
class TrunkBlock(nn.Module):
    def __init__(self, feat_in, feat_out):
        super(TrunkBlock, self).__init__()
        self.conv1 = nn.Conv2d(feat_in, int(feat_out*1.), kernel_size=3, stride=1, padding=1, dilation=1)
        self.drop1 = nn.Dropout2d(p=0.5, inplace=False)
        self.bn1 = nn.BatchNorm2d(feat_in, eps=1e-05, momentum=0.25, affine=True, track_running_stats=True)

        torch.nn.init.xavier_normal_(self.conv1.weight, gain=torch.nn.init.calculate_gain('relu'))
        torch.nn.init.constant_(self.conv1.bias, 0.0) # current
        
    def forward(self, x):
        return torch.nn.functional.relu(self.conv1(self.drop1(self.bn1(x))))

#Subclass for GNET
class PreFilter(nn.Module):
    def __init__(self):
        super(PreFilter, self).__init__()
        self.conv1 = nn.Sequential(
            nn.Conv2d(3, 64, kernel_size=11, stride=4, padding=2),
            nn.ReLU(inplace=True),
            nn.MaxPool2d(kernel_size=3, stride=2)
        )
        self.conv2 = nn.Sequential(
            nn.Conv2d(64, 192, kernel_size=5, padding=2),
            nn.ReLU(inplace=True)
        )        
        
    def forward(self, x):
        c1 = self.conv1(x)
        y = self.conv2(c1)
        return y 

#Subclass for GNET
class EncStage(nn.Module):
    def __init__(self, trunk_width=64, pass_through=64):
        super(EncStage, self).__init__()
        self.conv3  = nn.Conv2d(192, 128, kernel_size=3, stride=1, padding=0)
        self.drop1  = nn.Dropout2d(p=0.5, inplace=False) ##
        self.bn1    = nn.BatchNorm2d(192, eps=1e-05, momentum=0.25, affine=True, track_running_stats=True) ##
        self.pool1  = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
        ##
        self.tw = int(trunk_width)
        self.pt = int(pass_through)
        ss = (self.tw + self.pt)
        self.conv4a  = TrunkBlock(128, ss)
        self.conv5a  = TrunkBlock(ss, ss)
        self.conv6a  = TrunkBlock(ss, ss)
        self.conv4b  = TrunkBlock(ss, ss)
        self.conv5b  = TrunkBlock(ss, ss)
        self.conv6b  = TrunkBlock(ss, self.tw)
        ##
        torch.nn.init.xavier_normal_(self.conv3.weight, gain=torch.nn.init.calculate_gain('relu'))        
        torch.nn.init.constant_(self.conv3.bias, 0.0)
        
    def forward(self, x):
        c3 = (torch.nn.functional.relu(self.conv3(self.drop1(self.bn1(x))), inplace=False))
        c4a = self.conv4a(c3)
        c4b = self.conv4b(c4a)
        c5a = self.conv5a(self.pool1(c4b))
        c5b = self.conv5b(c5a)
        c6a = self.conv6a(c5b)
        c6b = self.conv6b(c6a)

        return [torch.cat([c3, c4a[:,:self.tw], c4b[:,:self.tw]], dim=1), 
                torch.cat([c5a[:,:self.tw], c5b[:,:self.tw], c6a[:,:self.tw], c6b], dim=1)], c6b
    
#Subclass for GNET
class GEncoder(nn.Module):
    def __init__(self, mu, trunk_width, pass_through=64 ):
        super(GEncoder, self).__init__()
        self.mu = nn.Parameter(torch.from_numpy(mu), requires_grad=False) #.to(device)
        self.pre = PreFilter()
        self.enc = EncStage(trunk_width, pass_through) 

    def forward(self, x):
        fmaps, h = self.enc(self.pre(x - self.mu))
        return x, fmaps, h

#Main GNET model class
class Torch_LayerwiseFWRF(nn.Module):
    def __init__(self, fmaps, nv=1, pre_nl=None, post_nl=None, dtype=np.float32):
        super(Torch_LayerwiseFWRF, self).__init__()
        self.fmaps_shapes = [list(f.size()) for f in fmaps]
        self.nf = np.sum([s[1] for s in self.fmaps_shapes])
        self.pre_nl  = pre_nl
        self.post_nl = post_nl
        self.nv = nv
        ##
        self.rfs = []
        self.sm = nn.Softmax(dim=1)
        for k,fm_rez in enumerate(self.fmaps_shapes):
            rf = nn.Parameter(torch.tensor(np.ones(shape=(self.nv, fm_rez[2], fm_rez[2]), dtype=dtype), requires_grad=True))
            self.register_parameter('rf%d'%k, rf)
            self.rfs += [rf,]
        self.w  = nn.Parameter(torch.tensor(np.random.normal(0, 0.01, size=(self.nv, self.nf)).astype(dtype=dtype), requires_grad=True))
        self.b  = nn.Parameter(torch.tensor(np.random.normal(0, 0.01, size=(self.nv,)).astype(dtype=dtype), requires_grad=True))
        
    def forward(self, fmaps):
        phi = []
        for fm,rf in zip(fmaps, self.rfs): #, self.scales):
            g = self.sm(torch.flatten(rf, start_dim=1))
            f = torch.flatten(fm, start_dim=2)  # *s
            if self.pre_nl is not None:          
                f = self.pre_nl(f)
            # fmaps : [batch, features, space]
            # v     : [nv, space]
            phi += [torch.tensordot(g, f, dims=[[1],[2]]),] # apply pooling field and add to list.
            # phi : [nv, batch, features] 
        Phi = torch.cat(phi, dim=2)
        if self.post_nl is not None:
            Phi = self.post_nl(Phi)
        vr = torch.squeeze(torch.bmm(Phi, torch.unsqueeze(self.w,2))).t() + torch.unsqueeze(self.b,0)
        return vr
    
class GNet8_Encoder():
    
    def __init__(self, subject = 1, device = "cuda", model_path = "gnet_multisubject.pt"):
        
        # Setting up Cuda
        self.device = torch.device(device)
        torch.backends.cudnn.enabled=True
        # Subject number
        self.subject = subject
        
        # Vector type
        self.vector = "images"
        
        # x size
        subject_sizes = [0, 15724, 14278, 15226, 13153, 13039, 17907, 12682, 14386]
        self.x_size = subject_sizes[self.subject]
        
        # Reload joined GNet model files
        self.joined_checkpoint = torch.load(model_path, map_location=self.device)
        
        self.subjects = list(self.joined_checkpoint['voxel_mask'].keys())
        self.gnet8j_voxel_mask = self.joined_checkpoint['voxel_mask']
        self.gnet8j_voxel_roi  = self.joined_checkpoint['voxel_roi']
        self.gnet8j_voxel_index= self.joined_checkpoint['voxel_index']
        self.gnet8j_brain_nii_shape= self.joined_checkpoint['brain_nii_shape']
        self.gnet8j_val_cc = self.joined_checkpoint['val_cc']
        
            
    
    def load_image(self, image_path):
        
        image = PIL.Image.open(image_path).convert('RGB')
        
        w, h = 227, 227  # resize to integer multiple of 64
        imagePil = image.resize((w, h), resample=PIL.Image.Resampling.LANCZOS)
        image = np.array(imagePil).astype(np.float32) / 255.0
        
        return image  
    
    # Rebuild Model
    def _model_fn(self, _ext, _con, _x):
        '''model consists of an extractor (_ext) and a connection model (_con)'''
        _y, _fm, _h = _ext(_x)
        return _con(_fm)

    def _pred_fn(self, _ext, _con, xb):
        return self._model_fn(_ext, _con, torch.from_numpy(xb).to(self.device))  
                    
    def subject_pred_pass(self, _pred_fn, _ext, _con, x, batch_size):
        pred = _pred_fn(_ext, _con, x[:batch_size]) # this is just to get the shape
        pred = np.zeros(shape=(len(x), pred.shape[1]), dtype=np.float32) # allocate
        for rb,_ in utils.iterate_range(0, len(x), batch_size):
            pred[rb] = utils.get_value(_pred_fn(_ext, _con, x[rb]))
        return pred

    def gnet8j_predictions(self, image_data, _pred_fn, trunk_width, pass_through, checkpoint, mask, batch_size, device=torch.device("cuda:0")):
        
        subjects = list(image_data.keys())

        if(mask is None):
            subject_nv = {s: len(v) for s,v in checkpoint['val_cc'].items()} 
        else:
            subject_nv = {s: len(v) for s,v in checkpoint['val_cc'].items()}    
            subject_nv[subjects[0]] = int(torch.sum(mask == True)) 

        # allocate
        subject_image_pred = {s: np.zeros(shape=(len(image_data[s]), subject_nv[s]), dtype=np.float32) for s in subjects}
        # print(subject_image_pred)
        _log_act_fn = lambda _x: torch.log(1 + torch.abs(_x))*torch.tanh(_x)
        
        best_params = checkpoint['best_params']
        # print(best_params)
        shared_model = GEncoder(np.array(checkpoint['input_mean']).astype(np.float32), trunk_width=trunk_width, pass_through=pass_through).to(device)
        shared_model.load_state_dict(best_params['enc'])
        shared_model.eval() 

        # example fmaps
        rec, fmaps, h = shared_model(torch.from_numpy(image_data[list(image_data.keys())[0]][:20]).to(device))                                     
        for s in subjects:
            sd = Torch_LayerwiseFWRF(fmaps, nv=subject_nv[s], pre_nl=_log_act_fn, post_nl=_log_act_fn, dtype=np.float32).to(device) 
            params = best_params['fwrfs'][s]
            
            if(mask is None):
                sd.load_state_dict(params)
            
            else:
                masked_params = {}
                for key, value in params.items():
                    masked_params[key] = value[mask]
                    
                sd.load_state_dict(masked_params)
                
            # print(params['w'].shape)
            # print(params['b'].shape)
            # sd.load_state_dict(best_params['fwrfs'][s])
            sd.eval() 
            # print(sd)
            
            subject_image_pred[s] = self.subject_pred_pass(_pred_fn, shared_model, sd, image_data[s], batch_size)

        return subject_image_pred

    def predict(self, images, mask = None):
        self.stim_data = {}
        data = []
        w, h = 227, 227  # resize to integer multiple of 64
        
        if(isinstance(images, list)):
            for i in range(len(images)):
                
                imagePil = images[i].convert("RGB").resize((w, h), resample=PIL.Image.Resampling.LANCZOS)
                image = np.array(imagePil).astype(np.float32) / 255.0
                data.append(image)
            
        elif(isinstance(images, torch.Tensor)):
            for i in range(images.shape[0]):
                
                imagePil = utils.process_image(images[i], w, h)
                image = np.array(imagePil).astype(np.float32) / 255.0
                data.append(image)
            
        
        self.stim_data[self.subject] = np.moveaxis(np.array(data), 3, 1)

        gnet8j_image_pred = self.gnet8j_predictions(self.stim_data, self._pred_fn, 64, 192, self.joined_checkpoint, mask, batch_size=100, device=self.device)

        return torch.from_numpy(gnet8j_image_pred[self.subject])