Duplicate from akhaliq/CLIP_prefix_captioning

Co-authored-by: AK <akhaliq@users.noreply.huggingface.co>

.gitattributes

@@ -0,0 +1,27 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bin.* filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zstandard filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,38 @@
+---
+title: CLIP_prefix_captioning
+emoji: 💩
+colorFrom: red
+colorTo: indigo
+sdk: gradio
+app_file: app.py
+pinned: false
+duplicated_from: akhaliq/CLIP_prefix_captioning
+---
+
+# Configuration
+
+`title`: _string_  
+Display title for the Space
+
+`emoji`: _string_  
+Space emoji (emoji-only character allowed)
+
+`colorFrom`: _string_  
+Color for Thumbnail gradient (red, yellow, green, blue, indigo, purple, pink, gray)
+
+`colorTo`: _string_  
+Color for Thumbnail gradient (red, yellow, green, blue, indigo, purple, pink, gray)
+
+`sdk`: _string_  
+Can be either `gradio` or `streamlit`
+
+`sdk_version` : _string_  
+Only applicable for `streamlit` SDK.  
+See [doc](https://hf.co/docs/hub/spaces) for more info on supported versions.
+
+`app_file`: _string_  
+Path to your main application file (which contains either `gradio` or `streamlit` Python code).  
+Path is relative to the root of the repository.
+
+`pinned`: _boolean_  
+Whether the Space stays on top of your list.

app.py

@@ -0,0 +1,273 @@
+import os
+from huggingface_hub import hf_hub_download
+conceptual_weight = hf_hub_download(repo_id="akhaliq/CLIP-prefix-captioning-conceptual-weights", filename="conceptual_weights.pt")
+coco_weight = hf_hub_download(repo_id="akhaliq/CLIP-prefix-captioning-COCO-weights", filename="coco_weights.pt")
+import clip
+import os
+from torch import nn
+import numpy as np
+import torch
+import torch.nn.functional as nnf
+import sys
+from typing import Tuple, List, Union, Optional
+from transformers import GPT2Tokenizer, GPT2LMHeadModel, AdamW, get_linear_schedule_with_warmup
+from tqdm import tqdm, trange
+import skimage.io as io
+import PIL.Image
+import gradio as gr
+
+N = type(None)
+V = np.array
+ARRAY = np.ndarray
+ARRAYS = Union[Tuple[ARRAY, ...], List[ARRAY]]
+VS = Union[Tuple[V, ...], List[V]]
+VN = Union[V, N]
+VNS = Union[VS, N]
+T = torch.Tensor
+TS = Union[Tuple[T, ...], List[T]]
+TN = Optional[T]
+TNS = Union[Tuple[TN, ...], List[TN]]
+TSN = Optional[TS]
+TA = Union[T, ARRAY]
+
+
+D = torch.device
+CPU = torch.device('cpu')
+
+
+def get_device(device_id: int) -> D:
+    if not torch.cuda.is_available():
+        return CPU
+    device_id = min(torch.cuda.device_count() - 1, device_id)
+    return torch.device(f'cuda:{device_id}')
+
+
+CUDA = get_device
+
+class MLP(nn.Module):
+
+    def forward(self, x: T) -> T:
+        return self.model(x)
+
+    def __init__(self, sizes: Tuple[int, ...], bias=True, act=nn.Tanh):
+        super(MLP, self).__init__()
+        layers = []
+        for i in range(len(sizes) -1):
+            layers.append(nn.Linear(sizes[i], sizes[i + 1], bias=bias))
+            if i < len(sizes) - 2:
+                layers.append(act())
+        self.model = nn.Sequential(*layers)
+
+
+class ClipCaptionModel(nn.Module):
+
+    #@functools.lru_cache #FIXME
+    def get_dummy_token(self, batch_size: int, device: D) -> T:
+        return torch.zeros(batch_size, self.prefix_length, dtype=torch.int64, device=device)
+
+    def forward(self, tokens: T, prefix: T, mask: Optional[T] = None, labels: Optional[T] = None):
+        embedding_text = self.gpt.transformer.wte(tokens)
+        prefix_projections = self.clip_project(prefix).view(-1, self.prefix_length, self.gpt_embedding_size)
+        #print(embedding_text.size()) #torch.Size([5, 67, 768])
+        #print(prefix_projections.size()) #torch.Size([5, 1, 768])
+        embedding_cat = torch.cat((prefix_projections, embedding_text), dim=1)
+        if labels is not None:
+            dummy_token = self.get_dummy_token(tokens.shape[0], tokens.device)
+            labels = torch.cat((dummy_token, tokens), dim=1)
+        out = self.gpt(inputs_embeds=embedding_cat, labels=labels, attention_mask=mask)
+        return out
+
+    def __init__(self, prefix_length: int, prefix_size: int = 512):
+        super(ClipCaptionModel, self).__init__()
+        self.prefix_length = prefix_length
+        self.gpt = GPT2LMHeadModel.from_pretrained('gpt2')
+        self.gpt_embedding_size = self.gpt.transformer.wte.weight.shape[1]
+        if prefix_length > 10:  # not enough memory
+            self.clip_project = nn.Linear(prefix_size, self.gpt_embedding_size * prefix_length)
+        else:
+            self.clip_project = MLP((prefix_size, (self.gpt_embedding_size * prefix_length) // 2, self.gpt_embedding_size * prefix_length))
+
+
+class ClipCaptionPrefix(ClipCaptionModel):
+
+    def parameters(self, recurse: bool = True):
+        return self.clip_project.parameters()
+
+    def train(self, mode: bool = True):
+        super(ClipCaptionPrefix, self).train(mode)
+        self.gpt.eval()
+        return self
+        
+
+#@title Caption prediction
+
+def generate_beam(model, tokenizer, beam_size: int = 5, prompt=None, embed=None,
+                  entry_length=67, temperature=1., stop_token: str = '.'):
+
+    model.eval()
+    stop_token_index = tokenizer.encode(stop_token)[0]
+    tokens = None
+    scores = None
+    device = next(model.parameters()).device
+    seq_lengths = torch.ones(beam_size, device=device)
+    is_stopped = torch.zeros(beam_size, device=device, dtype=torch.bool)
+    with torch.no_grad():
+        if embed is not None:
+            generated = embed
+        else:
+            if tokens is None:
+                tokens = torch.tensor(tokenizer.encode(prompt))
+                tokens = tokens.unsqueeze(0).to(device)
+                generated = model.gpt.transformer.wte(tokens)
+        for i in range(entry_length):
+            outputs = model.gpt(inputs_embeds=generated)
+            logits = outputs.logits
+            logits = logits[:, -1, :] / (temperature if temperature > 0 else 1.0)
+            logits = logits.softmax(-1).log()
+            if scores is None:
+                scores, next_tokens = logits.topk(beam_size, -1)
+                generated = generated.expand(beam_size, *generated.shape[1:])
+                next_tokens, scores = next_tokens.permute(1, 0), scores.squeeze(0)
+                if tokens is None:
+                    tokens = next_tokens
+                else:
+                    tokens = tokens.expand(beam_size, *tokens.shape[1:])
+                    tokens = torch.cat((tokens, next_tokens), dim=1)
+            else:
+                logits[is_stopped] = -float(np.inf)
+                logits[is_stopped, 0] = 0
+                scores_sum = scores[:, None] + logits
+                seq_lengths[~is_stopped] += 1
+                scores_sum_average = scores_sum / seq_lengths[:, None]
+                scores_sum_average, next_tokens = scores_sum_average.view(-1).topk(beam_size, -1)
+                next_tokens_source = next_tokens // scores_sum.shape[1]
+                seq_lengths = seq_lengths[next_tokens_source]
+                next_tokens = next_tokens % scores_sum.shape[1]
+                next_tokens = next_tokens.unsqueeze(1)
+                tokens = tokens[next_tokens_source]
+                tokens = torch.cat((tokens, next_tokens), dim=1)
+                generated = generated[next_tokens_source]
+                scores = scores_sum_average * seq_lengths
+                is_stopped = is_stopped[next_tokens_source]
+            next_token_embed = model.gpt.transformer.wte(next_tokens.squeeze()).view(generated.shape[0], 1, -1)
+            generated = torch.cat((generated, next_token_embed), dim=1)
+            is_stopped = is_stopped + next_tokens.eq(stop_token_index).squeeze()
+            if is_stopped.all():
+                break
+    scores = scores / seq_lengths
+    output_list = tokens.cpu().numpy()
+    output_texts = [tokenizer.decode(output[:int(length)]) for output, length in zip(output_list, seq_lengths)]
+    order = scores.argsort(descending=True)
+    output_texts = [output_texts[i] for i in order]
+    return output_texts
+
+
+def generate2(
+        model,
+        tokenizer,
+        tokens=None,
+        prompt=None,
+        embed=None,
+        entry_count=1,
+        entry_length=67,  # maximum number of words
+        top_p=0.8,
+        temperature=1.,
+        stop_token: str = '.',
+):
+    model.eval()
+    generated_num = 0
+    generated_list = []
+    stop_token_index = tokenizer.encode(stop_token)[0]
+    filter_value = -float("Inf")
+    device = next(model.parameters()).device
+
+    with torch.no_grad():
+
+        for entry_idx in trange(entry_count):
+            if embed is not None:
+                generated = embed
+            else:
+                if tokens is None:
+                    tokens = torch.tensor(tokenizer.encode(prompt))
+                    tokens = tokens.unsqueeze(0).to(device)
+
+                generated = model.gpt.transformer.wte(tokens)
+
+            for i in range(entry_length):
+
+                outputs = model.gpt(inputs_embeds=generated)
+                logits = outputs.logits
+                logits = logits[:, -1, :] / (temperature if temperature > 0 else 1.0)
+                sorted_logits, sorted_indices = torch.sort(logits, descending=True)
+                cumulative_probs = torch.cumsum(nnf.softmax(sorted_logits, dim=-1), dim=-1)
+                sorted_indices_to_remove = cumulative_probs > top_p
+                sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[
+                                                    ..., :-1
+                                                    ].clone()
+                sorted_indices_to_remove[..., 0] = 0
+
+                indices_to_remove = sorted_indices[sorted_indices_to_remove]
+                logits[:, indices_to_remove] = filter_value
+                next_token = torch.argmax(logits, -1).unsqueeze(0)
+                next_token_embed = model.gpt.transformer.wte(next_token)
+                if tokens is None:
+                    tokens = next_token
+                else:
+                    tokens = torch.cat((tokens, next_token), dim=1)
+                generated = torch.cat((generated, next_token_embed), dim=1)
+                if stop_token_index == next_token.item():
+                    break
+
+            output_list = list(tokens.squeeze().cpu().numpy())
+            output_text = tokenizer.decode(output_list)
+            generated_list.append(output_text)
+
+    return generated_list[0]
+    
+is_gpu = False 
+device = CUDA(0) if is_gpu else "cpu"
+clip_model, preprocess = clip.load("ViT-B/32", device=device, jit=False)
+tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
+
+def inference(img,model_name):   
+  prefix_length = 10
+    
+  model = ClipCaptionModel(prefix_length)
+    
+  if model_name == "COCO":
+      model_path = coco_weight
+  else:
+      model_path = conceptual_weight
+  model.load_state_dict(torch.load(model_path, map_location=CPU)) 
+  model = model.eval() 
+  device = CUDA(0) if is_gpu else "cpu"
+  model = model.to(device)
+
+  use_beam_search = False 
+  image = io.imread(img.name)
+  pil_image = PIL.Image.fromarray(image)  
+  image = preprocess(pil_image).unsqueeze(0).to(device)
+  with torch.no_grad():
+      prefix = clip_model.encode_image(image).to(device, dtype=torch.float32)
+      prefix_embed = model.clip_project(prefix).reshape(1, prefix_length, -1)
+  if use_beam_search:
+      generated_text_prefix = generate_beam(model, tokenizer, embed=prefix_embed)[0]
+  else:
+      generated_text_prefix = generate2(model, tokenizer, embed=prefix_embed)
+  return generated_text_prefix
+  
+title = "CLIP prefix captioning"
+description = "Gradio demo for CLIP prefix captioning: a simple image captioning model. To use it, simply upload your image, or click one of the examples to load them. Read more at the links below."
+article = "<p style='text-align: center'><a href='https://github.com/rmokady/CLIP_prefix_caption' target='_blank'>Github Repo</a></p>"
+
+examples=[['water.jpeg',"COCO"]]
+gr.Interface(
+    inference, 
+    [gr.inputs.Image(type="file", label="Input"),gr.inputs.Radio(choices=["COCO","Conceptual captions"], type="value", default="COCO", label="Model")], 
+    gr.outputs.Textbox(label="Output"),
+    title=title,
+    description=description,
+    article=article,
+    enable_queue=True,
+    examples=examples
+    ).launch(debug=True)

requirements.txt

@@ -0,0 +1,8 @@
+transformers 
+gdown 
+torch 
+numpy 
+tqdm 
+Pillow 
+scikit-image
+git+https://github.com/openai/CLIP.git