image-captioning / model.py
NicolasVana's picture
Upload 120 files
32162b0
import json
import os, shutil
import random
import streamlit as st
import os
from pathlib import Path
import numpy as np
from PIL import Image
import tensorflow as tf
from tensorflow.keras.applications.inception_v3 import preprocess_input
from tensorflow.keras.preprocessing import image
from tensorflow.keras.applications.inception_v3 import InceptionV3
from tensorflow.keras.models import Model
from tensorflow.keras.preprocessing.sequence import pad_sequences
root = Path(os.getcwd())
aux_pre = root / 'Inception' / 'PretrainedInceptionLSTM'
aux_re = root / 'Inception' / 'RetrainedInceptionLSTM'
model_re_path = root / 'Inception' / 'RetrainedInceptionLSTM' / 'Model'
model_inception_path = root / 'Inception' / 'RetrainedInceptionFeatureExtraction' / 'Model'
model_pre_path = root / 'Inception' / 'PretrainedInceptionLSTM' / 'Model'
# Must create
def get_pretrained_inceptionV3():
model = InceptionV3(weights='imagenet')
model2 = Model(model.input, model.layers[-2].output)
return model2
def fetch_auxiliary_files(type):
if type == 'Pretrained Inception':
word2Index = np.load(aux_pre / "word2Index.npy", allow_pickle=True).item()
index2Word = np.load(aux_pre / "index2Word.npy", allow_pickle=True).item()
variable_params = np.load(aux_pre / "variable_params.npy", allow_pickle=True).item()
return word2Index, index2Word, variable_params
if type == 'Retrained Inception':
word2Index = np.load(aux_re / "word2Index.npy", allow_pickle=True).item()
index2Word = np.load(aux_re / "index2Word.npy", allow_pickle=True).item()
variable_params = np.load(aux_re / "variable_params.npy", allow_pickle=True).item()
return word2Index, index2Word, variable_params
@st.cache(allow_output_mutation=True, show_spinner=False)
def fetch_model(type):
with st.spinner(text="Fetching Model"):
if type == 'Pretrained Inception':
model_pre = tf.keras.models.load_model(model_pre_path)
model_inc = get_pretrained_inceptionV3()
return model_inc, model_pre
if type == 'Retrained Inception':
model_re = tf.keras.models.load_model(model_re_path)
model_inc = tf.keras.models.load_model(model_inception_path)
return model_inc, model_re
def preprocess_image_inception(image):
if image.mode != "RGB":
image = image.convert(mode="RGB")
x = np.array(image)
x = np.expand_dims(x, axis = 0)
x = preprocess_input(x)
x = x.reshape(1, 299, 299, 3)
return x
def extract_features(model, image):
features = model.predict(image, verbose = 0)
return features
def generate_caption(model, features, max_len, word2Index, index2Word, beam_index = 3):
caption = beam_search(model, features, max_len, word2Index, index2Word, beam_index)
return caption
def beam_search(model, features, max_len, word2Index, index2Word, beam_index):
start = [word2Index["startseq"]]
start_word = [[start, 1]]
final_preds = []
live_seqs = beam_index
features = np.tile(features, (beam_index,1))
count = 0
while len(start_word) > 0:
#print(count)
count+=1
temp = []
padded_seqs = []
#Get padded seqs for each of the starting seqs so far, misnamed as start_word
for s in start_word:
par_caps = pad_sequences([s[0]], maxlen=max_len, padding='post')
padded_seqs.append(par_caps)
#Formatting input so that it can be used for a prediction
padded_seqs = np.array(padded_seqs).reshape(len(start_word), max_len)
preds = model.predict([features[:len(start_word)],padded_seqs], verbose=0)
#Getting the best branches for each of the start seqs that we had
for index, pred in enumerate(preds):
word_preds = np.argsort(pred)[-live_seqs:]
for w in word_preds:
next_cap, prob = start_word[index][0][:], start_word[index][1]
next_cap.append(w)
prob *= pred[w]
temp.append([next_cap, prob])
start_word = temp
# Sorting according to the probabilities
start_word = sorted(start_word, reverse=False, key=lambda l: l[1])
# Getting the top words from all branches
start_word = start_word[-live_seqs:]
for pair in start_word:
if index2Word[pair[0][-1]] == 'endseq':
final_preds.append([pair[0][:-1], pair[1]])
start_word = start_word[:-1]
live_seqs -= 1
if len(pair[0]) == max_len:
final_preds.append(pair)
start_word = start_word[:-1]
live_seqs -= 1
# Between all the finished sequences (either max len or predicted endseq), decide which is best
max_prob = 0
for index, pred in enumerate(final_preds):
if pred[1] > max_prob:
best_index = index
max_prob = pred[1]
# Convert to readable text
final_pred = final_preds[best_index]
final_caption = [index2Word[i] for i in final_pred[0]]
final_caption = ' '.join(final_caption[1:])
return final_caption
# # create target model directory
# model_dir = './models/'
# os.makedirs(model_dir, exist_ok=True)
#
# files_to_download = [
# "config.json",
# "flax_model.msgpack",
# "merges.txt",
# "special_tokens_map.json",
# "tokenizer.json",
# "tokenizer_config.json",
# "vocab.json",
# "preprocessor_config.json",
# ]
def _compile():
image_path = 'samples/ROCO_00929.jpg'
image = Image.open(image_path)
#predict(image)
image.close()
_compile()
sample_dir = './samples/'
sample_image_ids = tuple(["None"] + [int(f.replace('ROCO_', '').replace('.jpg', '')) for f in os.listdir(sample_dir) if f.startswith('ROCO_')])
with open(os.path.join(sample_dir, "Roco-img-ids.json"), "r", encoding="UTF-8") as fp:
roco_image_ids = json.load(fp)
def get_random_image_id():
image_id = random.sample(roco_image_ids, k=1)[0]
return image_id