Upload teacher_code/llava_arch.py with huggingface_hub

teacher_code/llava_arch.py

@@ -0,0 +1,434 @@
+#    Copyright 2023 Haotian Liu
+#
+#    Licensed under the Apache License, Version 2.0 (the "License");
+#    you may not use this file except in compliance with the License.
+#    You may obtain a copy of the License at
+#
+#        http://www.apache.org/licenses/LICENSE-2.0
+#
+#    Unless required by applicable law or agreed to in writing, software
+#    distributed under the License is distributed on an "AS IS" BASIS,
+#    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#    See the License for the specific language governing permissions and
+#    limitations under the License.
+
+
+from abc import ABC, abstractmethod
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from diffusers.models.embeddings import PixArtAlphaTextProjection
+
+from .multimodal_llava_encoder.builder import build_vision_tower
+from .multimodal_llava_projector.builder import build_vision_projector
+from .multimodal_projector.builder import build_down_projector
+from .multimodal_decoder.builder import build_vae, build_sana
+from diffusers import FlowMatchEulerDiscreteScheduler
+from diffusers.models.normalization import RMSNorm
+import math
+
+from blip3o.constants import DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN, IMAGE_TOKEN_IDX, UND_IMAGE_TOKEN_IDX, DEFAULT_IMAGE_PATCH_TOKEN
+
+
+class DiffusionConnector(nn.Module):
+    def __init__(self, input_dim=896, hidden_dim=1024, output_dim=2304, eps=1e-5):
+        super().__init__()
+        self.linear1 = nn.Linear(input_dim, hidden_dim)
+        self.act = nn.GELU(approximate="tanh")
+        self.linear2 = nn.Linear(hidden_dim, output_dim)
+        self.norm = RMSNorm(output_dim, eps=eps, elementwise_affine=True)
+
+        nn.init.xavier_uniform_(self.linear1.weight)
+        nn.init.zeros_(self.linear1.bias)
+        nn.init.xavier_uniform_(self.linear2.weight)
+        nn.init.zeros_(self.linear2.bias)
+        with torch.no_grad():
+            self.norm.weight.fill_(math.sqrt(5.5))
+
+    def forward(self, x):
+        x = self.linear1(x)
+        x = self.act(x)
+        x = self.linear2(x)
+        x = self.norm(x)
+        return x
+
+
+class LlavaMetaModel:
+
+    def __init__(self, config):
+        super(LlavaMetaModel, self).__init__(config)
+
+        if hasattr(config, "mm_vision_tower"):
+            self.vision_tower = build_vision_tower(config, delay_load=True)
+            self.mm_projector = build_vision_projector(config)
+        if hasattr(config, "diffusion_name_or_path"):
+            self.dit, _, self.noise_scheduler, _ = build_sana(config, load_teacher=False)
+            self.vae = build_vae(config)
+            self.diffusion_connector = DiffusionConnector(input_dim=self.config.hidden_size,hidden_dim=1024,output_dim=2304)
+            '''
+            norm = RMSNorm(896, eps=1e-5, elementwise_affine=True)
+            with torch.no_grad():
+                norm.weight.fill_(math.sqrt(5.5))
+            self.diffusion_connector = nn.Sequential(
+                nn.Linear(config.hidden_size, 896),
+                nn.GELU(approximate="tanh"),
+                nn.Linear(896, 896),
+                norm,
+            )
+            '''
+            self.latent_queries = nn.Parameter(torch.randn(1, self.config.n_query, self.config.hidden_size))
+            
+        
+
+    def get_vision_tower(self):
+        vision_tower = getattr(self, 'vision_tower', None)
+        if type(vision_tower) is list:
+            vision_tower = vision_tower[0]
+        return vision_tower
+    
+    def get_sana(self):
+        dit = getattr(self, 'dit', None)
+        if type(dit) is list:
+            dit = dit[0]
+        if dit is not None:
+            dit.to(self.device)
+        return dit
+
+    def get_sana_vae(self):
+        vae = getattr(self, 'vae', None)
+        if type(vae) is list:
+            vae = vae[0]
+        if vae is not None:
+            vae.to(self.device)
+        return vae
+
+    def initialize_vision_modules(self, model_args, fsdp=None):
+        vision_tower = model_args.vision_tower
+        mm_vision_select_layer = model_args.mm_vision_select_layer
+        mm_vision_select_feature = model_args.mm_vision_select_feature
+        mm_patch_merge_type = model_args.mm_patch_merge_type
+        
+        self.config.mm_vision_tower = vision_tower
+        self.config.vision_tower_pretrained = getattr(model_args, "vision_tower_pretrained", "")
+
+        if self.get_sana() is None:
+            dit, self.dit_teacher, self.noise_scheduler, self.text_encoder = build_sana(model_args, device=self.device)
+
+            if fsdp is not None and len(fsdp) > 0:
+                self.dit = [dit]
+            else:
+                self.dit = dit
+        else:
+            if fsdp is not None and len(fsdp) > 0:
+                dit = self.dit[0]
+            else:
+                dit = self.dit
+                # dit_teacher = self.dit_teacher
+        for p in self.text_encoder.parameters():
+            p.requires_grad = False
+            
+        for p in self.dit_teacher.parameters():
+            p.requires_grad = False
+        
+        if self.get_sana_vae() is None:
+            vae = build_vae(model_args)
+
+            if fsdp is not None and len(fsdp) > 0:
+                self.vae = [vae]
+            else:
+                self.vae = vae
+        else:
+            if fsdp is not None and len(fsdp) > 0:
+                vae = self.vae[0]
+            else:
+                vae = self.vae
+        for p in vae.parameters():
+            p.requires_grad = False
+    
+
+        if self.get_vision_tower() is None:
+            print("=" * 20, "Building vision tower", "=" * 20)
+            vision_tower = build_vision_tower(model_args)
+            
+
+            if fsdp is not None and len(fsdp) > 0:
+                self.vision_tower = [vision_tower]
+            else:
+                self.vision_tower = vision_tower
+        else:
+            if fsdp is not None and len(fsdp) > 0:
+                vision_tower = self.vision_tower[0]
+            else:
+                vision_tower = self.vision_tower
+            vision_tower.load_model()
+        
+        
+        if getattr(self, 'diffusion_connector', None) is None:
+            self.diffusion_connector = DiffusionConnector(input_dim=self.config.hidden_size,hidden_dim=1024,output_dim=2304)
+
+            '''
+            norm = RMSNorm(2304, eps=1e-5, elementwise_affine=True)
+            with torch.no_grad():
+                norm.weight.fill_(math.sqrt(5.5))
+            self.diffusion_connector = nn.Sequential(
+                nn.Linear(self.config.hidden_size, 1024),
+                nn.GELU(approximate="tanh"),
+                nn.Linear(1024, 2304),
+                norm,
+            )
+            '''
+        else:
+            for p in self.diffusion_connector.parameters():
+                p.requires_grad = True
+
+            for p in self.diffusion_connector_teacher.parameters():
+                p.requires_grad = True
+        
+        # freeze all parameters in dit except for caption_projection
+        for name, param in self.dit.named_parameters():
+            if "caption" in name:
+                param.requires_grad = True
+            else:
+                param.requires_grad = False
+
+
+        #for p in dit.parameters():
+        #    p.requires_grad = True
+            #if param.ndim > 1:
+            #    nn.init.xavier_uniform_(param)
+            #else:
+            #    nn.init.zeros_(param)
+        
+        #for p in dit.parameters():
+        #    p.requires_grad = True
+
+        '''
+        for p in dit_teacher.parameters():
+            p.requires_grad = False
+
+        for name, param in self.dit_teacher.named_parameters():
+            if "caption" in name:
+                param.requires_grad = True
+            else:
+                param.requires_grad = False
+        '''
+
+        self.config.use_mm_proj = True
+        self.config.mm_projector_type = getattr(model_args, 'mm_projector_type', 'linear')
+        self.config.mm_vision_select_layer = mm_vision_select_layer
+        self.config.mm_vision_select_feature = mm_vision_select_feature
+        self.config.mm_patch_merge_type = mm_patch_merge_type
+        self.config.n_query = model_args.n_query
+        self.config.gen_pooling = model_args.gen_pooling
+        self.config.diffusion_name_or_path = model_args.diffusion_name_or_path
+        
+
+
+        if getattr(self, 'down_projector', None) is None:
+            self.down_projector = build_down_projector(self.config)
+        else:
+            # In case it is frozen by LoRA
+            for p in self.down_projector.parameters():
+                p.requires_grad = True
+
+        if getattr(self, 'latent_queries', None) is None:
+            print("random initiation the latent_queries !!!")
+            self.latent_queries = nn.Parameter(torch.randn(1, self.config.n_query, self.config.hidden_size))
+        else:
+            print("latent_queries load from checkpoint!!!")
+            self.latent_queries.requires_grad = True
+        if not hasattr(self, 'dit_teacher') or self.dit_teacher is None:
+            print("Teacher model not properly initialized!")
+
+
+
+def unpad_image(tensor, original_size):
+    """
+    Unpads a PyTorch tensor of a padded and resized image.
+
+    Args:
+    tensor (torch.Tensor): The image tensor, assumed to be in CxHxW format.
+    original_size (tuple): The original size of PIL image (width, height).
+
+    Returns:
+    torch.Tensor: The unpadded image tensor.
+    """
+    original_width, original_height = original_size
+    current_height, current_width = tensor.shape[1:]
+
+    original_aspect_ratio = original_width / original_height
+    current_aspect_ratio = current_width / current_height
+
+    if original_aspect_ratio > current_aspect_ratio:
+        scale_factor = current_width / original_width
+        new_height = int(original_height * scale_factor)
+        padding = (current_height - new_height) // 2
+        unpadded_tensor = tensor[:, padding:current_height - padding, :]
+    else:
+        scale_factor = current_height / original_height
+        new_width = int(original_width * scale_factor)
+        padding = (current_width - new_width) // 2
+        unpadded_tensor = tensor[:, :, padding:current_width - padding]
+
+    return unpadded_tensor
+
+
+class LlavaMetaForCausalLM(ABC):
+
+    @abstractmethod
+    def get_model(self):
+        pass
+
+    def get_vision_tower(self):
+        return self.get_model().get_vision_tower()
+
+
+    def encode_image(self, images):
+        vision_tower = self.get_vision_tower()
+        device = vision_tower.device
+        images = images.to(device)
+        prompt_image_embeds = vision_tower(images)
+        if 'early' in self.get_gen_pooling():
+            prompt_image_embeds = self.pool_img(prompt_image_embeds)
+
+        # ------------- compute similarity -------
+        all_dist = 0
+        count = 0
+        for i in range(2, prompt_image_embeds.shape[1]-1):
+            diff = (prompt_image_embeds[:,i,:].unsqueeze(1) -  prompt_image_embeds[:,:i,:])
+            dist = torch.sqrt(diff.square().sum(-1)).min().item()
+            all_dist+=dist
+            count+=1
+        all_dist /= count
+        return prompt_image_embeds
+    
+    def get_mm_projector(self):
+        return self.get_model().mm_projector
+
+    def get_gen_projector(self):
+        return None
+    
+
+    def get_n_query(self):
+        return self.get_model().config.n_query
+
+    def get_gen_pooling(self):
+        return self.get_model().config.gen_pooling
+
+    def pool_img(self, image_features):
+        num_img, n, c = image_features.shape
+        gen_pooling = self.get_gen_pooling()
+        stride = int(gen_pooling.split('_')[-1])
+        sqrt_n = int(n**0.5)
+        image_features = image_features.permute(0, 2, 1).view(num_img, c, sqrt_n, sqrt_n)
+        image_features = F.avg_pool2d(image_features, kernel_size=(stride, stride), stride=stride)
+        return image_features
+
+    def get_sigmas(self, timesteps, device, n_dim=4, dtype=torch.float32):
+        sigmas = self.get_model().noise_scheduler.sigmas.to(device=device, dtype=dtype)
+        schedule_timesteps = self.get_model().noise_scheduler.timesteps.to(device=device)
+        timesteps = timesteps.to(device)
+        step_indices = [(schedule_timesteps == t).nonzero().item() for t in timesteps]
+
+        sigma = sigmas[step_indices].flatten()
+        while len(sigma.shape) < n_dim:
+            sigma = sigma.unsqueeze(-1)
+        return sigma
+
+    def mask_drop(self, latents, drop_prob=0.1):
+        if drop_prob <= 0:
+            return latents
+        mask = torch.bernoulli(torch.zeros(latents.shape[0], device=latents.device, dtype=latents.dtype) + drop_prob)
+        while len(mask.shape) < len(latents.shape):
+            mask = mask.unsqueeze(-1)
+        mask = 1 - mask  # need to flip 0 <-> 1
+        return latents * mask
+
+    def prepare_inputs_labels_for_multimodal(
+        self, input_ids, position_ids, attention_mask, past_key_values, labels,
+        gen_images, und_images, grid_thw, i_s_pos, image_sizes=None
+    ):
+        vision_tower = self.visual
+        if (gen_images is None and und_images is None) or input_ids.shape[1] == 1:
+            return input_ids, position_ids, attention_mask, past_key_values, None, labels, None, None, None
+        
+        vae = self.get_model().get_sana_vae()
+        vae_device = vae.device
+        prompt_image_embeds = vae.encode(gen_images.to(vae_device)).latent if gen_images is not None else None
+        prompt_image_embeds = prompt_image_embeds * vae.config.scaling_factor if prompt_image_embeds is not None else None
+        target_image_embeds = torch.clone(prompt_image_embeds).detach()
+        latent_queries = self.get_model().latent_queries.repeat(input_ids.shape[0], 1, 1)
+        H = latent_queries.shape[-1]
+        latent_queries = latent_queries.contiguous().view(-1, H)
+    
+        # vocab_size = self.get_model().embed_tokens.num_embeddings
+        # max_token_id = input_ids.max().item()
+        # min_token_id = input_ids.min().item()
+        # print(f"Vocab size: {vocab_size}")
+        # print(f"Max token ID: {max_token_id}")
+        # print(f"Min token ID: {min_token_id}")
+
+        if not und_images is None:
+            und_image_embeds = vision_tower(und_images, grid_thw=grid_thw)
+
+        image_idx = (input_ids == IMAGE_TOKEN_IDX)
+        und_image_idx = (input_ids == UND_IMAGE_TOKEN_IDX)
+        output_indicator = labels != -100
+        input_indicator = labels == -100
+        text_embeds = self.get_model().embed_tokens(input_ids)
+        gen_img_idx = torch.logical_and(output_indicator, image_idx)
+        text_embeds = text_embeds.clone() 
+        text_embeds[gen_img_idx] = latent_queries.to(text_embeds.dtype)
+        und_img_idx = torch.logical_and(input_indicator, und_image_idx)
+
+        if not und_images is None:
+            text_embeds[und_img_idx] = und_image_embeds.to(text_embeds.device)[:und_img_idx.sum(), :]
+
+        labels[image_idx] = -100
+        return None, position_ids, attention_mask, past_key_values, text_embeds, labels, target_image_embeds
+
+
+    def initialize_vision_tokenizer(self, model_args, tokenizer):
+        if model_args.mm_use_im_patch_token:
+            tokenizer.add_tokens([DEFAULT_IMAGE_PATCH_TOKEN], special_tokens=True)
+            self.resize_token_embeddings(len(tokenizer))
+
+        if model_args.mm_use_im_start_end:
+            num_new_tokens = tokenizer.add_tokens([DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN], special_tokens=True)
+            self.resize_token_embeddings(len(tokenizer))
+
+            if num_new_tokens > 0:
+                input_embeddings = self.get_input_embeddings().weight.data
+                output_embeddings = self.get_output_embeddings().weight.data
+
+                input_embeddings_avg = input_embeddings[:-num_new_tokens].mean(
+                    dim=0, keepdim=True)
+                output_embeddings_avg = output_embeddings[:-num_new_tokens].mean(
+                    dim=0, keepdim=True)
+
+                input_embeddings[-num_new_tokens:] = input_embeddings_avg
+                output_embeddings[-num_new_tokens:] = output_embeddings_avg
+
+            if model_args.tune_mm_mlp_adapter:
+                for p in self.get_input_embeddings().parameters():
+                    p.requires_grad = True
+                for p in self.get_output_embeddings().parameters():
+                    p.requires_grad = False
+
+            if model_args.pretrain_mm_mlp_adapter:
+                mm_projector_weights = torch.load(model_args.pretrain_mm_mlp_adapter, map_location='cpu')
+                embed_tokens_weight = mm_projector_weights['model.embed_tokens.weight']
+                assert num_new_tokens == 2
+                if input_embeddings.shape == embed_tokens_weight.shape:
+                    input_embeddings[-num_new_tokens:] = embed_tokens_weight[-num_new_tokens:]
+                elif embed_tokens_weight.shape[0] == num_new_tokens:
+                    input_embeddings[-num_new_tokens:] = embed_tokens_weight
+                else:
+                    raise ValueError(f"Unexpected embed_tokens_weight shape. Pretrained: {embed_tokens_weight.shape}. Current: {input_embeddings.shape}. Numer of new tokens: {num_new_tokens}.")
+        elif model_args.mm_use_im_patch_token:
+            if model_args.tune_mm_mlp_adapter:
+                for p in self.get_input_embeddings().parameters():
+                    p.requires_grad = False
+                for p in self.get_output_embeddings().parameters():
+                    p.requires_grad = False