new features

Author: AstraBert

.gitignore

@@ -1,4 +1,4 @@
 flagged/
-scripts/__pycache__
 docker/__pycache__
-docker/flagged
+docker/flagged
+qdrant_storage/

README.md

@@ -57,6 +57,18 @@ Choose the task among:
 - *image-generation-pollinations*: stable diffusion, use Pollinations AI API; if you choose 'image-generation-pollinations', you do not need to specify anything else apart from the task - **MULTILINGUAL**
 - *image-classification*: classify an image, supports every image-classification model on HF Hub - **ENGLISH ONLY**
 - *image-to-text*:  describe an image, supports every image-to-text model on HF Hub - **ENGLISH ONLY**
+- *image-retrieval-search*: search an image database uploading a folder as database input. The folder should have the following structure:
+
+```
+./
+├── test/
+|   ├── label1/
+|   └── label2/
+└── train/
+    ├── label1/
+    └── label2/
+```
+You can query the database starting from your own pictures.
 
 ### 6. Go to `localhost:7860` and start using your assistant
 

docker/Dockerfile

@@ -7,7 +7,10 @@ WORKDIR /app
 # Add the current directory contents into the container at /app
 ADD . /app
 
+# Add new package
+RUN python3 -m pip install datasets==2.15.0
+
 # Expose the port that the application will run on
-EXPOSE 7860
+EXPOSE 8760
 
 ENTRYPOINT [ "python3", "select_and_run.py" ]

docker/image_classification.py

@@ -1,6 +1,7 @@
 from transformers import AutoModelForImageClassification, AutoImageProcessor, pipeline
 from PIL import Image
 from argparse import ArgumentParser
+import torch
 
 argparse = ArgumentParser()
 argparse.add_argument(
@@ -18,9 +19,11 @@
 
 model_checkpoint = mod
 
-model = AutoModelForImageClassification.from_pretrained(model_checkpoint)
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+model = AutoModelForImageClassification.from_pretrained(model_checkpoint).to(device)
 processor = AutoImageProcessor.from_pretrained(model_checkpoint)
 
+
 pipe = pipeline("image-classification", model=model, image_processor=processor)
 
 def get_results(image, ppln=pipe):

docker/image_to_text.py

@@ -19,7 +19,8 @@
 
 model_checkpoint = mod
 
-pipe = pipeline("image-to-text", model=model_checkpoint)
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+pipe = pipeline("image-to-text", model=model_checkpoint, device=device)
 
 def get_results(image, ppln=pipe):
     img = Image.fromarray(image)

docker/requirements.txt

@@ -11,4 +11,5 @@ diffusers==0.27.2
 pydantic==2.6.4
 qdrant_client==1.9.0
 pillow==10.2.0
-accelerate
+datasets==2.15.0
+accelerate

docker/retrieval_image_search.py

@@ -0,0 +1,74 @@
+from transformers import AutoImageProcessor, AutoModel
+from utils import ImageDB
+from PIL import Image
+from qdrant_client import QdrantClient
+import gradio as gr
+from argparse import ArgumentParser
+import torch
+
+argparse = ArgumentParser()
+argparse.add_argument(
+    "-m",
+    "--model",
+    help="HuggingFace Model identifier, such as 'google/flan-t5-base'",
+    required=True,
+)
+
+argparse.add_argument(
+    "-id",
+    "--image_dimension",
+    help="Dimension of the image (e.g. 512, 758, 384...)",
+    required=False,
+    default=512,
+    type=int
+)
+
+argparse.add_argument(
+    "-d",
+    "--directory",
+    help="Directory where all your pdfs of interest are stored",
+    required=False,
+    default="No directory"
+)
+
+
+args = argparse.parse_args()
+
+
+mod = args.model
+dirs = args.directory
+imd = args.image_dimension
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+processor = AutoImageProcessor.from_pretrained(mod)
+model = AutoModel.from_pretrained(mod).to(device)
+
+client = QdrantClient(host="host.docker.internal", port=6333)
+imdb = ImageDB(dirs, processor, model, client, imd)
+print(imdb.collection_name)
+imdb.create_dataset()
+imdb.to_collection()
+
+
+def see_images(dataset, results):
+    images = []
+    for i in range(len(results)):
+        img = dataset[results[0].id]['image']
+        images.append(img)
+    return images
+
+def process_img(image):
+    global imdb
+    results = imdb.searchDB(Image.fromarray(image))
+    images = see_images(imdb.dataset, results)
+    return images
+
+
+iface = gr.Interface(
+    title="everything-ai-retrievalimg",
+    fn=process_img,
+    inputs=gr.Image(label="Input Image"),
+    outputs=gr.Gallery(label="Matching Images"),  
+)
+
+iface.launch(server_name="0.0.0.0", share=False)

docker/retrieval_text_generation.py

@@ -4,7 +4,7 @@
 from sentence_transformers import SentenceTransformer
 from argparse import ArgumentParser
 from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
-import sys
+import torch
 import os
 
 argparse = ArgumentParser()
@@ -61,7 +61,8 @@
 pdfdb.collect_data()
 pdfdb.qdrant_collection_and_upload()
 
-model = AutoModelForCausalLM.from_pretrained(mod)
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+model = AutoModelForCausalLM.from_pretrained(mod).to(device)
 tokenizer = AutoTokenizer.from_pretrained(mod)
 
 pipe = pipeline("text-generation", model=model, tokenizer=tokenizer, max_new_tokens=2048, repetition_penalty=1.2, temperature=0.4)

docker/select_and_run.py

@@ -1,19 +1,22 @@
 import subprocess as sp
 import gradio as gr
 
-TASK_TO_SCRIPT = {"retrieval-text-generation": "retrieval_text_generation.py", "agnostic-text-generation": "agnostic_text_generation.py", "text-summarization": "text_summarization.py", "image-generation": "image_generation.py", "image-generation-pollinations": "image_generation_pollinations.py", "image-classification": "image_classification.py", "image-to-text": "image_to_text.py"}
+TASK_TO_SCRIPT = {"retrieval-text-generation": "retrieval_text_generation.py", "agnostic-text-generation": "agnostic_text_generation.py", "text-summarization": "text_summarization.py", "image-generation": "image_generation.py", "image-generation-pollinations": "image_generation_pollinations.py", "image-classification": "image_classification.py", "image-to-text": "image_to_text.py", "retrieval-image-search": "retrieval_image_search.py"}
 
 
-def build_command(tsk, mod="None", pdff="None", dirs="None", lan="None"):
-    if tsk != "retrieval-text-generation" and tsk != "image-generation-pollinations":
+def build_command(tsk, mod="None", pdff="None", dirs="None", lan="None", imdim="512"):
+    if tsk != "retrieval-text-generation" and tsk != "image-generation-pollinations" and tsk != "retrieval-image-search":
         sp.run(f"python3 {TASK_TO_SCRIPT[tsk]} -m {mod}", shell=True)
         return f"python3 {TASK_TO_SCRIPT[tsk]} -m {mod}"
     elif tsk == "retrieval-text-generation":
         sp.run(f"python3 {TASK_TO_SCRIPT[tsk]} -m {mod} -pf '{pdff}' -d '{dirs}' -l '{lan}'", shell=True)
         return f"python3 {TASK_TO_SCRIPT[tsk]} -m {mod} -pf '{pdff}' -d '{dirs}' -l '{lan}'"
-    else:
+    elif tsk == "image-generation-pollinations":
         sp.run(f"python3 {TASK_TO_SCRIPT[tsk]}", shell=True)
         return f"python3 {TASK_TO_SCRIPT[tsk]}"
+    else:
+        sp.run(f"python3 {TASK_TO_SCRIPT[tsk]} -d {dirs} -id {imdim} -m {mod}", shell=True)
+        return f"python3 {TASK_TO_SCRIPT[tsk]} -d {dirs} -id {imdim} -m {mod}"
 
 demo = gr.Interface(
     build_command,
@@ -38,7 +41,7 @@ def build_command(tsk, mod="None", pdff="None", dirs="None", lan="None"):
         ),
         gr.Textbox(
             label="Directory",
-            info="Directory where all your pdfs of interest are stored (only available with 'retrieval-text-generation')",
+            info="Directory where all your pdfs or images (.jpg, .jpeg, .png) of interest are stored (only available with 'retrieval-text-generation' for pdfs and 'retrieval-image-search' for images)",
             lines=3,
             value="None",
         ),
@@ -48,6 +51,12 @@ def build_command(tsk, mod="None", pdff="None", dirs="None", lan="None"):
             lines=3,
             value="None",
         ),
+        gr.Textbox(
+            label="Image dimension",
+            info="Dimension of the image (this is generally model and/or task-dependent!)",
+            lines=3,
+            value=f"e.g.: 512, 384, 758...",
+        ),
     ],
     outputs="textbox",
     theme=gr.themes.Base()

docker/text_summarization.py

@@ -5,6 +5,7 @@
 from utils import merge_pdfs
 import gradio as gr
 import time
+import torch
 
 histr = [[None, "Hi, I'm **everything-ai-summarization**🤖.\nI'm here to assist you and let you summarize _your_ texts and _your_ pdfs!\nCheck [my website](https://astrabert.github.io/everything-ai/) for troubleshooting and documentation reference\nHave fun!😊"]]
 
@@ -24,7 +25,8 @@
 
 model_checkpoint = mod
 
-summarizer = pipeline("summarization", model=model_checkpoint)
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+summarizer = pipeline("summarization", model=model_checkpoint, device=device)
 
 def convert_none_to_str(l: list):
     newlist = []

docker/utils.py

@@ -6,6 +6,10 @@
 from langchain.text_splitter import CharacterTextSplitter
 from langchain_community.document_loaders import PyPDFLoader
 import os
+from datasets import load_dataset, Dataset
+import torch
+import numpy as np
+
 
 def remove_items(test_list, item): 
     res = [i for i in test_list if i != item] 
@@ -92,3 +96,71 @@ def translatef(self):
         translation = translator.translate(self.text)
         return translation
 
+class ImageDB:
+    def __init__(self, imagesdir, processor, model, client, dimension):
+        self.imagesdir = imagesdir
+        self.processor = processor
+        self.model = model
+        self.client = client
+        self.dimension = dimension
+        if os.path.basename(self.imagesdir) != "":
+            self.collection_name = os.path.basename(self.imagesdir)+"_ImagesCollection"
+        else:
+            if "\\" in self.imagesdir:
+               self.collection_name = self.imagesdir.split("\\")[-2]+"_ImagesCollection" 
+            else:
+                self.collection_name = self.imagesdir.split("/")[-2]+"_ImagesCollection" 
+        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+        self.client.recreate_collection(
+            collection_name=self.collection_name,
+            vectors_config=models.VectorParams(size=self.dimension, distance=models.Distance.COSINE)
+        )
+    def get_embeddings(self, batch):
+        inputs = self.processor(images=batch['image'], return_tensors="pt").to(self.device)
+        with torch.no_grad():
+            outputs = self.model(**inputs).last_hidden_state.mean(dim=1).cpu().numpy()
+        batch['embeddings'] = outputs
+        return batch
+    def create_dataset(self):
+        self.dataset = load_dataset("imagefolder", data_dir=self.imagesdir, split="train")
+        self.dataset = self.dataset.map(self.get_embeddings, batched=True, batch_size=16)
+    def to_collection(self):
+        np.save(os.path.join(self.imagesdir, "vectors"), np.array(self.dataset['embeddings']), allow_pickle=False)
+
+        payload = self.dataset.select_columns([
+            "label"
+        ]).to_pandas().fillna(0).to_dict(orient="records")
+
+        ids = list(range(self.dataset.num_rows))
+        embeddings = np.load(os.path.join(self.imagesdir, "vectors.npy")).tolist()
+
+        batch_size = 1000
+
+        for i in range(0, self.dataset.num_rows, batch_size):
+
+            low_idx = min(i+batch_size, self.dataset.num_rows)
+
+            batch_of_ids = ids[i: low_idx]
+            batch_of_embs = embeddings[i: low_idx]
+            batch_of_payloads = payload[i: low_idx]
+
+            self.client.upsert(
+                collection_name = self.collection_name,
+                points=models.Batch(
+                    ids=batch_of_ids,
+                    vectors=batch_of_embs,
+                    payloads=batch_of_payloads
+                )
+            )
+    def searchDB(self, image):
+        dtst = {"image": [image], "label": ["None"]}
+        dtst = Dataset.from_dict(dtst)
+        dtst = dtst.map(self.get_embeddings, batched=True, batch_size=1)
+        img = dtst[0]
+        results = self.client.search(
+            collection_name=self.collection_name,
+            query_vector=img['embeddings'],
+            limit=4
+        )
+        return results
+