# -*- coding: utf-8 -*- # Copyright (c) Alibaba, Inc. and its affiliates. """Concise re-implementation of ``https://github.com/openai/CLIP'' and ``https://github.com/mlfoundations/open_clip''. """ import math from functools import partial from importlib import find_loader import torch import torch.nn as nn import torch.nn.functional as F from scepter.modules.model.base_model import BaseModel from scepter.modules.model.embedder.xlm_roberta import \ XLMRoberta # used in XLMRobertaCLIP (multilingual) from scepter.modules.model.registry import EMBEDDERS from scepter.modules.utils.config import dict_to_yaml from scepter.modules.utils.distribute import we from scepter.modules.utils.file_system import FS def map_dtype(m, dtype=torch.float16): if isinstance(m, (nn.Linear, nn.Conv2d)): _ = m.to(dtype) elif isinstance(m, LayerNorm): _ = m.float() elif hasattr(m, 'head') and isinstance(m.head, nn.Parameter): p = getattr(m, 'head') p.data = p.data.to(dtype) class QuickGELU(nn.Module): def forward(self, x): return x * torch.sigmoid(1.702 * x) class LayerNorm(nn.LayerNorm): def forward(self, x): return super().forward(x.float()).type_as(x) class SelfAttention(nn.Module): def __init__(self, dim, num_heads, causal=False, attn_dropout=0.0, proj_dropout=0.0, flash_dtype=torch.float16): assert dim % num_heads == 0 assert flash_dtype in (None, torch.float16, torch.bfloat16) super().__init__() self.dim = dim self.num_heads = num_heads self.head_dim = dim // num_heads self.causal = causal self.attn_dropout = attn_dropout self.proj_dropout = proj_dropout self.scale = math.pow(self.head_dim, -0.25) self.flash_dtype = flash_dtype # layers self.to_qkv = nn.Linear(dim, dim * 3) self.proj = nn.Linear(dim, dim) def forward(self, x): """x: [B, L, C]. """ b, s, c, n, d = *x.size(), self.num_heads, self.head_dim # compute query, key, value qkv = self.to_qkv(x).view(b, s, 3, n, d) # compute attention if x.device.type != 'cpu' and find_loader('flash_attn') and \ self.flash_dtype is not None: # flash implementation from flash_attn.flash_attn_interface import ( flash_attn_unpadded_qkvpacked_func, ) dtype = qkv.dtype if dtype != self.flash_dtype: qkv = qkv.type(self.flash_dtype) cu_seqlens = torch.arange(0, b * s + 1, s, dtype=torch.int32, device=x.device) x = flash_attn_unpadded_qkvpacked_func( qkv=qkv.reshape(-1, 3, n, d), cu_seqlens=cu_seqlens, max_seqlen=s, dropout_p=self.attn_dropout if self.training else 0.0, causal=self.causal, return_attn_probs=False).reshape(b, s, n, d).type(dtype) else: # torch implementation q, k, v = qkv.unbind(2) attn = torch.einsum('binc,bjnc->bnij', q * self.scale, k * self.scale) if self.causal: attn = attn.masked_fill( torch.tril(attn.new_ones(1, 1, s, s).float()).type_as(attn) == 0, float('-inf')) attn = F.softmax(attn.float(), dim=-1).type_as(attn) x = torch.einsum('bnij,bjnc->binc', attn, v) # output x = x.reshape(b, s, c) x = self.proj(x) x = F.dropout(x, self.proj_dropout, self.training) return x class AttentionBlock(nn.Module): def __init__(self, dim, mlp_ratio, num_heads, causal=False, activation='quick_gelu', attn_dropout=0.0, proj_dropout=0.0, flash_dtype=torch.float16): assert activation in ['quick_gelu', 'gelu'] super().__init__() self.dim = dim self.mlp_ratio = mlp_ratio self.num_heads = num_heads self.causal = causal self.flash_dtype = flash_dtype # layers self.norm1 = LayerNorm(dim) self.attn = SelfAttention(dim, num_heads, causal, attn_dropout, proj_dropout, flash_dtype) self.norm2 = LayerNorm(dim) self.mlp = nn.Sequential( nn.Linear(dim, int(dim * mlp_ratio)), QuickGELU() if activation == 'quick_gelu' else nn.GELU(), nn.Linear(int(dim * mlp_ratio), dim), nn.Dropout(proj_dropout)) def forward(self, x): x = x + self.attn(self.norm1(x)) x = x + self.mlp(self.norm2(x)) return x class VisionTransformer(nn.Module): def __init__(self, image_size=224, patch_size=16, dim=768, mlp_ratio=4, out_dim=512, num_heads=12, num_layers=12, activation='quick_gelu', attn_dropout=0.0, proj_dropout=0.0, embedding_dropout=0.0, flash_dtype=torch.float16): assert image_size % patch_size == 0 super().__init__() self.image_size = image_size self.patch_size = patch_size self.num_patches = (image_size // patch_size)**2 self.dim = dim self.mlp_ratio = mlp_ratio self.out_dim = out_dim self.num_heads = num_heads self.num_layers = num_layers self.flash_dtype = flash_dtype # embeddings gain = 1.0 / math.sqrt(dim) self.patch_embedding = nn.Conv2d(3, dim, kernel_size=patch_size, stride=patch_size, bias=False) self.cls_embedding = nn.Parameter(gain * torch.randn(1, 1, dim)) self.pos_embedding = nn.Parameter( gain * torch.randn(1, self.num_patches + 1, dim)) self.dropout = nn.Dropout(embedding_dropout) # transformer self.pre_norm = LayerNorm(dim) self.transformer = nn.Sequential(*[ AttentionBlock(dim, mlp_ratio, num_heads, False, activation, attn_dropout, proj_dropout, flash_dtype) for _ in range(num_layers) ]) self.post_norm = LayerNorm(dim) # head self.head = nn.Parameter(gain * torch.randn(dim, out_dim)) def forward(self, x): b, dtype = x.size(0), self.head.dtype x = x.type(dtype) # patch-embedding x = self.patch_embedding(x).flatten(2).permute(0, 2, 1) x = torch.cat([self.cls_embedding.repeat(b, 1, 1).type(dtype), x], dim=1) x = self.dropout(x + self.pos_embedding.type(dtype)) x = self.pre_norm(x) # transformer x = self.transformer(x) # head x = self.post_norm(x) x = torch.mm(x[:, 0, :], self.head) return x def fp16(self, dtype=torch.float16): return self.apply(partial(map_dtype, dtype=dtype)) class TextTransformer(nn.Module): def __init__(self, vocab_size, text_len, dim=512, mlp_ratio=4, out_dim=512, num_heads=8, num_layers=12, activation='quick_gelu', attn_dropout=0.0, proj_dropout=0.0, embedding_dropout=0.0, flash_dtype=torch.float16): super().__init__() self.vocab_size = vocab_size self.text_len = text_len self.dim = dim self.mlp_ratio = mlp_ratio self.out_dim = out_dim self.num_heads = num_heads self.num_layers = num_layers self.flash_dtype = flash_dtype # embeddings self.token_embedding = nn.Embedding(vocab_size, dim) self.pos_embedding = nn.Parameter(0.01 * torch.randn(1, text_len, dim)) self.dropout = nn.Dropout(embedding_dropout) # transformer self.transformer = nn.Sequential(*[ AttentionBlock(dim, mlp_ratio, num_heads, True, activation, attn_dropout, proj_dropout, flash_dtype) for _ in range(num_layers) ]) self.norm = LayerNorm(dim) # head gain = 1.0 / math.sqrt(dim) self.head = nn.Parameter(gain * torch.randn(dim, out_dim)) def forward(self, x): eot, dtype = x.argmax(dim=-1), self.head.dtype # embeddings x = self.dropout( self.token_embedding(x).type(dtype) + self.pos_embedding.type(dtype)) # transformer x = self.transformer(x) # head x = self.norm(x) x = torch.mm(x[torch.arange(x.size(0)), eot], self.head) return x def fp16(self, dtype=torch.float16): return self.apply(partial(map_dtype, dtype=dtype)) class CLIP(nn.Module): def __init__(self, embed_dim=512, image_size=224, patch_size=16, vision_dim=768, vision_mlp_ratio=4, vision_heads=12, vision_layers=12, vocab_size=49408, text_len=77, text_dim=512, text_mlp_ratio=4, text_heads=8, text_layers=12, activation='quick_gelu', attn_dropout=0.0, proj_dropout=0.0, embedding_dropout=0.0, flash_dtype=torch.float16, use_module=['visual', 'textual']): assert flash_dtype in (None, torch.float16, torch.bfloat16) super().__init__() self.embed_dim = embed_dim self.image_size = image_size self.patch_size = patch_size self.vision_dim = vision_dim self.vision_mlp_ratio = vision_mlp_ratio self.vision_heads = vision_heads self.vision_layers = vision_layers self.vocab_size = vocab_size self.text_len = text_len self.text_dim = text_dim self.text_mlp_ratio = text_mlp_ratio self.text_heads = text_heads self.text_layers = text_layers self.flash_dtype = flash_dtype self.use_module = use_module # models if 'visual' in use_module: self.visual = VisionTransformer( image_size=image_size, patch_size=patch_size, dim=vision_dim, mlp_ratio=vision_mlp_ratio, out_dim=embed_dim, num_heads=vision_heads, num_layers=vision_layers, activation=activation, attn_dropout=attn_dropout, proj_dropout=proj_dropout, embedding_dropout=embedding_dropout, flash_dtype=flash_dtype) self.scale = math.sqrt(self.visual.out_dim) else: self.visual = nn.Identity() if 'textual' in use_module: self.textual = TextTransformer(vocab_size=vocab_size, text_len=text_len, dim=text_dim, mlp_ratio=text_mlp_ratio, out_dim=embed_dim, num_heads=text_heads, num_layers=text_layers, activation=activation, attn_dropout=attn_dropout, proj_dropout=proj_dropout, embedding_dropout=embedding_dropout, flash_dtype=flash_dtype) else: self.textual = nn.Identity() self.log_scale = nn.Parameter(math.log(1 / 0.07) * torch.ones([])) # initialize weights self.init_weights() def forward(self, imgs, txt_tokens): """imgs: [B, 3, H, W] of torch.float32. mean: [0.48145466, 0.4578275, 0.40821073] std: [0.26862954, 0.26130258, 0.27577711] txt_tokens: [B, L] of torch.long. Encoded by data.CLIPTokenizer. """ xi = self.visual(imgs) xt = self.textual(txt_tokens) return xi, xt def encode_image(self, x, skip_layers=0): # clip inference b, dtype = x.size(0), self.visual.head.dtype x = x.type(dtype) # # patch-embedding x = self.visual.patch_embedding(x).flatten(2).permute(0, 2, 1) x = torch.cat( [self.visual.cls_embedding.repeat(b, 1, 1).type(dtype), x], dim=1) x = self.visual.dropout(x + self.visual.pos_embedding.type(dtype)) x = self.visual.pre_norm(x) # # transformer # assert skip_layers < 12 if skip_layers == 0: x = self.visual.transformer(x) else: for m in self.visual.transformer[:-skip_layers]: x = m(x) # # head x = self.visual.post_norm(x) x = torch.mm(x[:, 0, :], self.visual.head) x = self.scale * F.normalize(x, p=2, dim=1) return x def encode_text(self): pass def init_weights(self): # embeddings if 'textual' in self.use_module: nn.init.normal_(self.textual.token_embedding.weight, std=0.02) if 'visual' in self.use_module: nn.init.normal_(self.visual.patch_embedding.weight, std=0.1) # attentions for modality in self.use_module: dim = self.vision_dim if modality == 'visual' else self.text_dim transformer = getattr(self, modality).transformer proj_gain = (1.0 / math.sqrt(dim)) * ( 1.0 / math.sqrt(2 * len(transformer))) attn_gain = 1.0 / math.sqrt(dim) mlp_gain = 1.0 / math.sqrt(2.0 * dim) for block in transformer: nn.init.normal_(block.attn.to_qkv.weight, std=attn_gain) nn.init.normal_(block.attn.proj.weight, std=proj_gain) nn.init.normal_(block.mlp[0].weight, std=mlp_gain) nn.init.normal_(block.mlp[2].weight, std=proj_gain) def param_groups(self): groups = [{ 'params': [ p for n, p in self.named_parameters() if 'norm' in n or n.endswith('bias') ], 'weight_decay': 0.0 }, { 'params': [ p for n, p in self.named_parameters() if not ('norm' in n or n.endswith('bias')) ] }] return groups def fp16(self, dtype=torch.float16): return self.apply(partial(map_dtype, dtype=dtype)) def load_from_open_clip(self, checkpoint_or_path, **kwargs): """Load and remap state-dict from open-clip. """ # load state-dict device = next(self.parameters()).device state = checkpoint_or_path if isinstance(state, str): state = torch.load(state, map_location=device) # reorder prefix = [ 'logit_scale', 'visual.', 'position', 'text_proj', 'token', 'transformer.', 'ln_final.' ] state = type(state)([(k, v) for u in prefix for k, v in state.items() if k.startswith(u)]) # convert to target keys target = self.state_dict() target = { k: v.view(target[k].shape) for k, v in zip(target.keys(), state.values()) } return self.load_state_dict(target, **kwargs) class XLMRobertaWithHead(XLMRoberta): def __init__(self, **kwargs): self.out_dim = kwargs.pop('out_dim') super().__init__(**kwargs) # head mid_dim = (self.dim + self.out_dim) // 2 self.head = nn.Sequential(nn.Linear(self.dim, mid_dim, bias=False), nn.GELU(), nn.Linear(mid_dim, self.out_dim, bias=False)) def forward(self, tokens): # xlm-roberta x = super().forward(tokens) # average pooling mask = tokens.ne(self.pad_token).unsqueeze(-1).to(x) x = (x * mask).sum(dim=1) / mask.sum(dim=1) # head x = self.head(x) return x class XLMRobertaCLIP(nn.Module): def __init__(self, embed_dim=1024, image_size=224, patch_size=14, vision_dim=1280, vision_mlp_ratio=4, vision_heads=16, vision_layers=32, activation='gelu', vocab_size=250002, max_text_len=514, type_size=1, pad_token=1, text_dim=1024, text_heads=16, text_layers=24, text_eps=1e-5, text_dropout=0.1, attn_dropout=0.0, proj_dropout=0.0, embedding_dropout=0.0, flash_dtype=torch.float16, use_module=['visual', 'textual']): assert flash_dtype in (None, torch.float16, torch.bfloat16) super().__init__() self.embed_dim = embed_dim self.image_size = image_size self.patch_size = patch_size self.vision_dim = vision_dim self.vision_mlp_ratio = vision_mlp_ratio self.vision_heads = vision_heads self.vision_layers = vision_layers self.activation = activation self.vocab_size = vocab_size self.max_text_len = max_text_len self.type_size = type_size self.pad_token = pad_token self.text_dim = text_dim self.text_heads = text_heads self.text_layers = text_layers self.text_eps = text_eps self.flash_dtype = flash_dtype # models if 'visual' in use_module: self.visual = VisionTransformer( image_size=image_size, patch_size=patch_size, dim=vision_dim, mlp_ratio=vision_mlp_ratio, out_dim=embed_dim, num_heads=vision_heads, num_layers=vision_layers, activation=activation, attn_dropout=attn_dropout, proj_dropout=proj_dropout, embedding_dropout=embedding_dropout, flash_dtype=flash_dtype) else: self.visual = nn.Identity() if 'textual' in use_module: self.textual = XLMRobertaWithHead(vocab_size=vocab_size, max_seq_len=max_text_len, type_size=type_size, pad_token=pad_token, dim=text_dim, out_dim=embed_dim, num_heads=text_heads, num_layers=text_layers, dropout=text_dropout, eps=text_eps) else: self.textual = nn.Identity() self.log_scale = nn.Parameter(math.log(1 / 0.07) * torch.ones([])) def forward(self, imgs, txt_tokens): """imgs: [B, 3, H, W] of torch.float32. mean: [0.48145466, 0.4578275, 0.40821073] std: [0.26862954, 0.26130258, 0.27577711] txt_tokens: [B, L] of torch.long. Encoded by data.CLIPTokenizer. """ xi = self.visual(imgs) xt = self.textual(txt_tokens) return xi, xt def param_groups(self): groups = [{ 'params': [ p for n, p in self.named_parameters() if 'norm' in n or n.endswith('bias') ], 'weight_decay': 0.0 }, { 'params': [ p for n, p in self.named_parameters() if not ('norm' in n or n.endswith('bias')) ] }] return groups def fp16(self, dtype=torch.float16): return self.apply(partial(map_dtype, dtype=dtype)) def load_from_open_clip(self, checkpoint_or_path, **kwargs): """Load and remap state-dict from open-clip. """ # load state-dict device = next(self.parameters()).device state = checkpoint_or_path if isinstance(state, str): state = torch.load(state, map_location=device) if 'state_dict' in state: state = state['state_dict'] # reorder keys = [ 'logit_scale', 'visual.', 'word_embeddings', 'token_type_embeddings', 'position_embeddings', 'embeddings.LayerNorm', 'encoder.layer.', 'text.proj' ] state = type(state)([(k, v) for u in keys for k, v in state.items() if u in k]) # target state-dict target = self.state_dict() target = { k: v.view(target[k].shape) for k, v in zip(target.keys(), state.values()) } return self.load_state_dict(target, **kwargs) def _clip(pretrained=False, pretrained_path=None, model_cls=CLIP, **kwargs): model = model_cls(**kwargs) if pretrained and pretrained_path: pretrain_model = torch.load(pretrained_path, map_location='cpu') key_str = ' '.join(list(pretrain_model.keys())) have_load = False if 'use_module' in kwargs and len(kwargs['use_module']) < 2: for module_name in kwargs['use_module']: if hasattr(model, module_name) and module_name not in key_str: missing, unexpected = getattr( model, module_name).load_state_dict(pretrain_model, strict=False) if we.rank == 0: print(f'Restored from {pretrained_path} with' '{len(missing)} missing and {len(unexpected)}' 'unexpected keys') if len(missing) > 0: print(f'Missing Keys:\n {missing}') if len(unexpected) > 0: print(f'\nUnexpected Keys:\n {unexpected}') have_load = True if not have_load: missing, unexpected = model.load_state_dict(pretrain_model, strict=False) if we.rank == 0: print( f'Restored from {pretrained_path} with {len(missing)} missing and {len(unexpected)} unexpected keys' ) if len(missing) > 0: print(f'Missing Keys:\n {missing}') if len(unexpected) > 0: print(f'\nUnexpected Keys:\n {unexpected}') return model def clip_vit_b_32(**kwargs): cfg = dict(embed_dim=512, image_size=224, patch_size=32, vision_dim=768, vision_heads=12, vision_layers=12, vocab_size=49408, text_len=77, text_dim=512, text_heads=8, text_layers=12, activation='quick_gelu') cfg.update(**kwargs) return cfg, CLIP def clip_vit_b_16(**kwargs): cfg = dict(embed_dim=512, image_size=224, patch_size=16, vision_dim=768, vision_heads=12, vision_layers=12, vocab_size=49408, text_len=77, text_dim=512, text_heads=8, text_layers=12, activation='quick_gelu') cfg.update(**kwargs) return cfg, CLIP def clip_vit_l_14(**kwargs): cfg = dict(embed_dim=768, image_size=224, patch_size=14, vision_dim=1024, vision_heads=16, vision_layers=24, vocab_size=49408, text_len=77, text_dim=768, text_heads=12, text_layers=12, activation='quick_gelu') cfg.update(**kwargs) return cfg, CLIP def clip_vit_l_14_336px(**kwargs): cfg = dict(embed_dim=768, image_size=336, patch_size=14, vision_dim=1024, vision_heads=16, vision_layers=24, vocab_size=49408, text_len=77, text_dim=768, text_heads=12, text_layers=12, activation='quick_gelu') cfg.update(**kwargs) return cfg, CLIP def clip_vit_h_14(**kwargs): cfg = dict(embed_dim=1024, image_size=224, patch_size=14, vision_dim=1280, vision_heads=16, vision_layers=32, vocab_size=49408, text_len=77, text_dim=1024, text_heads=16, text_layers=24, activation='gelu') cfg.update(**kwargs) return cfg, CLIP def clip_vit_g_14(**kwargs): cfg = dict(embed_dim=1024, image_size=224, patch_size=14, vision_dim=1408, vision_mlp_ratio=4.3637, vision_heads=16, vision_layers=40, vocab_size=49408, text_len=77, text_dim=1024, text_heads=16, text_layers=24, activation='gelu') cfg.update(**kwargs) return cfg, CLIP def clip_vit_bigG_14(**kwargs): cfg = dict(embed_dim=1280, image_size=224, patch_size=14, vision_dim=1664, vision_mlp_ratio=4.9231, vision_heads=16, vision_layers=48, vocab_size=49408, text_len=77, text_dim=1280, text_heads=20, text_layers=32, activation='gelu') cfg.update(**kwargs) return cfg, CLIP def clip_xlm_roberta_vit_h_14(**kwargs): cfg = dict(embed_dim=1024, image_size=224, patch_size=14, vision_dim=1280, vision_mlp_ratio=4, vision_heads=16, vision_layers=32, activation='gelu', vocab_size=250002, max_text_len=514, type_size=1, pad_token=1, text_dim=1024, text_heads=16, text_layers=24, text_eps=1e-5, text_dropout=0.1, attn_dropout=0.0, proj_dropout=0.0, embedding_dropout=0.0, flash_dtype=torch.float16) cfg.update(**kwargs) return cfg, XLMRobertaCLIP clip_functions = { 'clip_vit_b_32': clip_vit_b_32, 'clip_vit_b_16': clip_vit_b_16, 'clip_vit_l_14': clip_vit_l_14, 'clip_vit_l_14_336px': clip_vit_l_14_336px, 'clip_vit_h_14': clip_vit_h_14, 'clip_vit_g_14': clip_vit_g_14, 'clip_vit_bigG_14': clip_vit_bigG_14, 'clip_xlm_roberta_vit_h_14': clip_xlm_roberta_vit_h_14 } @EMBEDDERS.register_class() class ClipEncoder(BaseModel): para_dict = { 'CLIP_FUNC': { 'value': 'clip_vit_b_32', 'description': f'Select clip model from {list(clip_functions.keys())}' }, 'PRETRAINED': { 'value': False, 'description': 'Wether load from pretrained model or not.' }, 'USE_GRAD': { 'value': False, 'description': '' }, 'PRETRAINED_PATH': { 'value': None, 'description': 'Pretrained model load from.' }, 'USE_MODULE': { 'value': ['visual', 'textual'], 'description': "Use module from visual or textual, default is ['visual', 'textual']." }, 'CLIP_SKIP': { 'value': 2, 'description': "Textuxl branch skip blocks' num. Default is 2." }, 'TOKEN_LENGTH': { 'value': 77, 'description': 'The input token length for text. Default is 77.' }, 'KWARGS': {} } def __init__(self, cfg, logger=None): super().__init__(cfg, logger=logger) clip_func = cfg.CLIP_FUNC pretrained = cfg.get('PRETRAINED', False) pretrained_path = cfg.get('PRETRAINED_PATH', None) use_module = cfg.get('USE_MODULE', ['visual', 'textual']) self.clip_skip = cfg.get('CLIP_SKIP', 2) self.use_grad = cfg.get('USE_GRAD', False) self.token_length = cfg.get('TOKEN_LENGTH', 77) kwargs = {k.lower(): v for k, v in cfg.get('KWARGS', {}).items()} if pretrained and pretrained_path: local_path = FS.get_from(pretrained_path, wait_finish=True) else: local_path = None assert clip_func in clip_functions if clip_func in clip_functions: conf, model_cls = clip_functions[clip_func](**kwargs) conf['use_module'] = use_module self.clip_model = _clip(pretrained, local_path, model_cls, **conf) for module in use_module: if hasattr(self.clip_model, module): setattr(self, module, getattr(self.clip_model, module)) def encode_image(self, image): if not self.use_grad: with torch.no_grad(): m = self.clip_model.visual return m(image) else: m = self.clip_model.visual return m(image) def encode_text(self, tokens, tokenizer=None, append_sentence_embedding=True): def fn(): m = self.clip_model.textual b, s = tokens.shape mask = tokens.ne(m.pad_token).long() # embeddings x = m.token_embedding(tokens) + \ m.type_embedding(torch.zeros_like(tokens)) + \ m.pos_embedding(m.pad_token + torch.cumsum(mask, dim=1) * mask) x = m.norm(x) x = m.dropout(x) # blocks for block in m.blocks[:-1]: x = block(x, mask.view(b, 1, 1, s)) words = x sentence = m.blocks[-1](x, mask.view(b, 1, 1, s)) mask = tokens.ne(m.pad_token).unsqueeze(-1).to(sentence) sentence = (sentence * mask).sum(dim=1) / mask.sum(dim=1) sentence = m.head(sentence) return {'crossattn': words, 'y': sentence} if not self.use_grad: with torch.no_grad(): return fn() else: return fn() def dynamic_encode_text(self, all_tokens, tokenizer=None, append_sentence_embedding=True): ''' m: clip model t: tokenzer tokens: tensor(1, N) ''' if tokenizer is None: tokenizer = self.tokenizer def fn(): m = self.clip_model.textual ret_data = {'crossattn': [], 'y': []} for tokens_id in range(all_tokens.shape[0]): tokens = all_tokens[tokens_id] text_len = self.token_length device = tokens.device dtype = m.type_embedding.weight.dtype # special tokens sos_emb, eos_emb, pad_emb = m.token_embedding( torch.LongTensor([ tokenizer.sos_token, tokenizer.eos_token, tokenizer.pad_token ]).to(device)).type(dtype).chunk(3) # get raw input tokens tokens = list(tokens.cpu().numpy()) while tokens[-1] == tokenizer.pad_token: tokens = tokens[:-1] tokens = tokens[1:-1] embeds = m.token_embedding( torch.LongTensor(tokens).to(device)).type(dtype) # split into chunks to support any-length text chunk_embeds, chunk_tokens = [], [] max_words = text_len - 2 if len(tokens) == 0: chunk = torch.cat([sos_emb, eos_emb]) chunk = torch.cat( [chunk, pad_emb.repeat(text_len - len(chunk), 1)]) chunk_embeds.append(chunk) chunk = torch.LongTensor([tokenizer.sos_token] + [tokenizer.eos_token]) chunk = torch.cat([ chunk, torch.LongTensor([tokenizer.pad_token] * (text_len - len(chunk))) ]) chunk_tokens.append(chunk) else: while len(tokens) > 0: # find splitting position if len(tokens) <= max_words: pos = len(tokens) else: pos = [ i for i, u in enumerate(tokens[:max_words]) if u == tokenizer.comma_token ] pos = max_words if len(pos) == 0 else pos[-1] + 1 # collect chunk chunk = torch.cat([sos_emb, embeds[:pos], eos_emb]) chunk = torch.cat( [chunk, pad_emb.repeat(text_len - len(chunk), 1)]) chunk_embeds.append(chunk) chunk = torch.LongTensor([tokenizer.sos_token] + tokens[:pos] + [tokenizer.eos_token]) chunk = torch.cat([ chunk, torch.LongTensor([tokenizer.pad_token] * (text_len - len(chunk))) ]) chunk_tokens.append(chunk) # update tokens = tokens[pos:] embeds = embeds[pos:] # loop over chunks words = [] sentences = [] for i, (chunk_token, chunk_embed) in enumerate( zip(chunk_tokens, chunk_embeds)): chunk_token = chunk_token.unsqueeze(0).to(device) chunk_embed = chunk_embed.unsqueeze(0).to(device) # embeddings mask = chunk_token.ne(tokenizer.pad_token).long() x = chunk_embed.type(dtype) + m.type_embedding( torch.zeros_like(chunk_token)) + m.pos_embedding( m.pad_token + torch.cumsum(mask, dim=1) * mask) x = m.norm(x) x = m.dropout(x) blocks = m.blocks[:-(self.clip_skip - 1)] for block in blocks: x = block(x, mask.view(1, 1, 1, -1)) print('word', torch.sum(x)) words.append(x.clone()) # if append_sentence_embedding: # last layers blocks = m.blocks[-(self.clip_skip - 1):] for block in blocks: x = block(x, mask.view(1, 1, 1, -1)) # get global embedding x = (x * mask.unsqueeze(2)).sum(dim=1) / mask.sum(dim=1) x = m.head(x) print('sentence', torch.sum(x)) # output sentences.append(x.unsqueeze(0)) sentence = torch.cat(sentences, dim=0).mean(dim=0) words = torch.cat(words, dim=1) ret_data['crossattn'] = words ret_data['y'] = sentence # ret_data['y'].append(sentence) # ret_data.append(torch.cat([sentence] + words, dim=1)) # ret_data['crossattn'].append(torch.cat(words, dim=1)) # ret_data['crossattn'] = torch.cat(ret_data['crossattn'], dim=0) # ret_data['y'] = torch.cat(ret_data['y'], dim=0) return ret_data if not self.use_grad: with torch.no_grad(): return fn() else: return fn() @staticmethod def get_config_template(): return dict_to_yaml('BACKBONE', __class__.__name__, ClipEncoder.para_dict, set_name=True)