Spaces:
Build error
Build error
File size: 6,165 Bytes
7800904 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 | import tensorflow as tf
import numpy as np
import pickle
from preprocess import preprocess_image
from model import TransformerEncoder, TransformerDecoder, ImageCaptioningModel
from cnn import CNN_encoder
embed_dim=256
max_len=40
vocab_size=8002
cnn_model = CNN_encoder()
encoder = TransformerEncoder(embed_dim, 1)
decoder = TransformerDecoder(
num_layers=4,
embed_dim=256,
units=512,
num_heads=4,
vocab_size=vocab_size,
max_len=max_len
)
caption_model = ImageCaptioningModel(
cnn_model=cnn_model,
encoder=encoder,
decoder=decoder,
image_aug=None
)
dummy_img = tf.zeros((1, 224, 224, 3))
dummy_txt = tf.zeros((1, 40))
_ = caption_model([dummy_img, dummy_txt], training=False)
caption_model.load_weights("my_caption_model.weights.h5")
tokenizer = pickle.load(open('tokenizer.pkl', 'rb'))
vocab = tokenizer.get_vocabulary()
clean_vocab = [token for token in vocab if token not in ['', '[UNK]']]
word2idx = tf.keras.layers.StringLookup(
vocabulary = clean_vocab,
invert=False,
mask_token = '',
oov_token = '[UNK]'
)
idx2word = tf.keras.layers.StringLookup(
vocabulary = clean_vocab,
invert = True,
mask_token = '',
oov_token = '[UNK]'
)
def get_image_features(img_path):
img = preprocess_image(img_path)
img = tf.expand_dims(img, axis=0)
img_embed = caption_model.cnn_model(img)
return caption_model.encoder(img_embed, training=False)
def decode_ids_to_sentence(ids, idx2word):
words = []
for i in ids:
if i == 0:
continue
w = idx2word(tf.constant([i]))[0].numpy().decode("utf-8")
if w in ("[start]", "[end]"):
continue
words.append(w)
return " ".join(words)
def advanced_generate(encoder_output,
method="beam",
beam_width=3,
top_k=40,
temperature=1.0,
repetition_penalty=1.2,
alpha=0.6,
max_len=None):
if max_len is None:
max_len = 40
# --- Preprocessing & Encoding ---
# img = preprocess_image(img_path)
# img = tf.expand_dims(img, axis=0)
# img_embed = caption_model.cnn_model(img)
# encoder_output = caption_model.encoder(img_embed, training=False)
start_id = int(tokenizer(["[start]"])[0,0].numpy())
end_id = int(tokenizer(["[end]"])[0,0].numpy())
unk_id = 1
def process_logits(logits, current_seq):
logits[unk_id] = -1e10
if temperature != 1.0:
logits = logits / float(temperature)
if repetition_penalty != 1.0:
for tok in set(current_seq):
if logits[tok] > 0:
logits[tok] /= repetition_penalty
else:
logits[tok] *= repetition_penalty
return logits
if method == "greedy":
y = [start_id]
for _ in range(max_len-1):
inp = tf.constant([y], dtype=tf.int32)
mask = tf.cast(inp != 0, tf.int32)
preds = caption_model.decoder(inp, encoder_output, training=False, mask=mask)
logits = np.log(np.clip(preds[0, -1, :].numpy(), 1e-9, 1.0))
logits = process_logits(logits, y)
next_id = int(np.argmax(logits))
if next_id == end_id:
break
y.append(next_id)
return decode_ids_to_sentence(y, idx2word)
# ---------- BEAM SEARCH ----------
if method == "beam":
sequences = [([start_id], 0.0)]
for _step in range(max_len-1):
all_candidates = []
for seq, score in sequences:
if seq[-1] == end_id:
all_candidates.append((seq, score))
continue
inp = tf.constant([seq], dtype=tf.int32)
mask = tf.cast(inp != 0, tf.int32)
preds = caption_model.decoder(inp, encoder_output, training=False, mask=mask)
logits = np.log(np.clip(preds[0, -1, :].numpy(), 1e-9, 1.0))
logits = process_logits(logits, seq)
top_ids = np.argsort(logits)[-beam_width:]
for tok in top_ids:
candidate = seq + [int(tok)]
new_score = score + float(logits[tok])
all_candidates.append((candidate, new_score))
def get_lp_score(c_seq, c_score):
lp = ((5 + len(c_seq))**alpha) / ((5 + 1)**alpha)
return c_score / lp
sequences = sorted(all_candidates, key=lambda x: get_lp_score(x[0], x[1]), reverse=True)[:beam_width]
if all(seq[-1] == end_id for seq, sc in sequences):
break
return decode_ids_to_sentence(sequences[0][0], idx2word)
# ---------- SAMPLING ----------
if method == "sample":
y = [start_id]
for _ in range(max_len-1):
inp = tf.constant([y], dtype=tf.int32)
mask = tf.cast(inp != 0, tf.int32)
preds = caption_model.decoder(inp, encoder_output, training=False, mask=mask)
logits = np.log(np.clip(preds[0, -1, :].numpy(), 1e-9, 1.0))
logits = process_logits(logits, y)
topk_idx = np.argsort(logits)[-top_k:]
topk_logits = logits[topk_idx]
exp_logits = np.exp(topk_logits - np.max(topk_logits))
probs = exp_logits / np.sum(exp_logits)
next_id = int(np.random.choice(topk_idx, p=probs))
if next_id == end_id:
break
y.append(next_id)
return decode_ids_to_sentence(y, idx2word)
raise ValueError("Unknown method: choose 'greedy', 'beam', or 'sample'") |