dashboard_featurefinder.py

import jax
import flax
import matplotlib.pyplot as plt
import nmslib
import numpy as np
import os
import streamlit as st

from tempfile import NamedTemporaryFile
from torchvision.transforms import Compose, Resize, ToPILImage
from transformers import CLIPProcessor, FlaxCLIPModel
from PIL import Image


BASELINE_MODEL = "openai/clip-vit-base-patch32"
# MODEL_PATH = "/home/shared/models/clip-rsicd/bs128x8-lr5e-6-adam/ckpt-1"
MODEL_PATH = "flax-community/clip-rsicd-v2"

# IMAGE_VECTOR_FILE = "/home/shared/data/vectors/test-baseline.tsv"
# IMAGE_VECTOR_FILE = "/home/shared/data/vectors/test-bs128x8-lr5e-6-adam-ckpt-1.tsv"
IMAGE_VECTOR_FILE = "./vectors/test-bs128x8-lr5e-6-adam-ckpt-1.tsv"

# IMAGES_DIR = "/home/shared/data/rsicd_images"
IMAGES_DIR = "./images"

2
# @st.cache(allow_output_mutation=True)
# def load_index():
#     filenames, image_vecs = [], []
#     fvec = open(IMAGE_VECTOR_FILE, "r")
#     for line in fvec:
#         cols = line.strip().split('\t')
#         filename = cols[0]
#         image_vec = np.array([float(x) for x in cols[1].split(',')])
#         filenames.append(filename)
#         image_vecs.append(image_vec)
#     V = np.array(image_vecs)
#     index = nmslib.init(method='hnsw', space='cosinesimil')
#     index.addDataPointBatch(V)
#     index.createIndex({'post': 2}, print_progress=True)
#     return filenames, index


@st.cache(allow_output_mutation=True)
def load_model():
        # model = FlaxCLIPModel.from_pretrained(MODEL_PATH)
        # processor = CLIPProcessor.from_pretrained(BASELINE_MODEL)
    model = FlaxCLIPModel.from_pretrained("flax-community/clip-rsicd-v2")
    processor = CLIPProcessor.from_pretrained("flax-community/clip-rsicd-v2")
    return model, processor


def split_image(X):
  num_rows = X.shape[0] // 224
  num_cols = X.shape[1] // 224
  Xc = X[0 : num_rows * 224, 0 : num_cols * 224, :]
  patches = []
  for j in range(num_rows):
    for i in range(num_cols):
      patches.append(Xc[j * 224 : (j + 1) * 224, 
                        i * 224 : (i + 1) * 224,
                        :])
  return num_rows, num_cols, patches


def get_patch_probabilities(patches, searched_feature, 
                            image_preprocesor,
                            model, processor):
  images = [image_preprocesor(patch) for patch in patches]
  text = "An aerial image of {:s}".format(searched_feature)
  inputs = processor(images=images,
                    text=text,
                    return_tensors="jax",
                    padding=True)
  outputs = model(**inputs)
  probs = jax.nn.softmax(outputs.logits_per_text, axis=-1)
  probs_np = np.asarray(probs)[0]
  return probs_np


def get_image_ranks(probs):
  temp = np.argsort(-probs)
  ranks = np.empty_like(temp)
  ranks[temp] = np.arange(len(probs))
  return ranks


def app():
    model, processor = load_model()

    st.title("Find Features in Images")
    st.markdown("""
        The CLIP model from OpenAI is trained in a self-supervised manner using 
        contrastive learning to project images and caption text onto a common 
        embedding space. We have fine-tuned the model using the RSICD dataset
        (10k images and ~50k captions from the remote sensing domain).
        
        This demo shows the ability of the model to find specific features
        (specified as text queries) in the image. As an example, say you wish to
        find the parts of the following image that contain a `beach`, `houses`, 
        or `ships`. We partition the image into tiles of (224, 224) and report
        how likely each of them are to contain each text features.
    """)
    st.image("demo-images/st_tropez_1.png")
    st.image("demo-images/st_tropez_2.png")
    st.markdown("""
        For this image and the queries listed above, our model reports that the
        two left tiles are most likely to contain a `beach`, the two top right 
        tiles are most likely to contain `houses`, and the two bottom right tiles
        are likely to contain `boats`.

        You can try it yourself with your own photographs. 
        [Unsplash](https://unsplash.com/s/photos/aerial-view) has some good 
        aerial photographs. You will need to download from Unsplash to your 
        computer and upload it to the demo app.
    """)
    with st.form(key="form_3"):
        buf = st.file_uploader("Upload Image for Analysis")
        searched_feature = st.text_input(label="Feature to find")
        submit_button = st.form_submit_button("Find")

    if submit_button:
        ftmp = NamedTemporaryFile()
        ftmp.write(buf.getvalue())
        image = plt.imread(ftmp.name)
        if len(image.shape) != 3 and image.shape[2] != 3:
            st.error("Image should be an RGB image")
        if image.shape[0] < 224 or image.shape[1] < 224:
            st.error("Image should be at least (224 x 224")
        st.image(image, caption="Input Image")
        st.markdown("---")
        num_rows, num_cols, patches = split_image(image)
        image_preprocessor = Compose([
            ToPILImage(),
            Resize(224)
        ])
        num_rows, num_cols, patches = split_image(image)
        patch_probs = get_patch_probabilities(
            patches,
            searched_feature,
            image_preprocessor,
            model,
            processor)
        patch_ranks = get_image_ranks(patch_probs)
        for i in range(num_rows):
            row_patches = patches[i * num_cols : (i + 1) * num_cols]
            row_probs = patch_probs[i * num_cols : (i + 1) * num_cols]
            row_ranks = patch_ranks[i * num_cols : (i + 1) * num_cols]
            captions = ["p({:s})={:.3f}, rank={:d}".format(searched_feature, p, r + 1) 
                for p, r in zip(row_probs, row_ranks)]
            st.image(row_patches, caption=captions)