Spaces:
Sleeping
Sleeping
Update model.py
Browse files
model.py
CHANGED
@@ -280,25 +280,25 @@ def get_caption_model():
|
|
280 |
|
281 |
cnn_model = CNN_Encoder()
|
282 |
|
283 |
-
|
284 |
cnn_model=cnn_model, encoder=encoder, decoder=decoder, image_aug=None,
|
285 |
)
|
286 |
|
287 |
def call_fn(batch, training):
|
288 |
return batch
|
289 |
|
290 |
-
|
291 |
sample_x, sample_y = tf.random.normal((1, 299, 299, 3)), tf.zeros((1, 40))
|
292 |
|
293 |
-
|
294 |
|
295 |
-
sample_img_embed =
|
296 |
-
sample_enc_out =
|
297 |
-
|
298 |
|
299 |
-
|
300 |
|
301 |
-
return
|
302 |
|
303 |
def load_image_from_path(img_path):
|
304 |
img = tf.io.read_file(img_path)
|
@@ -317,14 +317,14 @@ def generate_caption(img_path, add_noise=False):
|
|
317 |
img = (img - tf.reduce_min(img))/(tf.reduce_max(img) - tf.reduce_min(img))
|
318 |
|
319 |
img = tf.expand_dims(img, axis=0)
|
320 |
-
img_embed =
|
321 |
-
img_encoded =
|
322 |
|
323 |
y_inp = '[start]'
|
324 |
for i in range(MAX_LENGTH-1):
|
325 |
tokenized = tokenizer([y_inp])[:, :-1]
|
326 |
mask = tf.cast(tokenized != 0, tf.int32)
|
327 |
-
pred =
|
328 |
tokenized, img_encoded, training=False, mask=mask)
|
329 |
|
330 |
pred_idx = np.argmax(pred[0, i, :])
|
|
|
280 |
|
281 |
cnn_model = CNN_Encoder()
|
282 |
|
283 |
+
caption_model = ImageCaptioningModel(
|
284 |
cnn_model=cnn_model, encoder=encoder, decoder=decoder, image_aug=None,
|
285 |
)
|
286 |
|
287 |
def call_fn(batch, training):
|
288 |
return batch
|
289 |
|
290 |
+
caption_model.call = call_fn
|
291 |
sample_x, sample_y = tf.random.normal((1, 299, 299, 3)), tf.zeros((1, 40))
|
292 |
|
293 |
+
caption_model((sample_x, sample_y))
|
294 |
|
295 |
+
sample_img_embed = caption_model.cnn_model(sample_x)
|
296 |
+
sample_enc_out = caption_model.encoder(sample_img_embed, training=False)
|
297 |
+
caption_model.decoder(sample_y, sample_enc_out, training=False)
|
298 |
|
299 |
+
caption_model.load_weights('model.h5')
|
300 |
|
301 |
+
return caption_model
|
302 |
|
303 |
def load_image_from_path(img_path):
|
304 |
img = tf.io.read_file(img_path)
|
|
|
317 |
img = (img - tf.reduce_min(img))/(tf.reduce_max(img) - tf.reduce_min(img))
|
318 |
|
319 |
img = tf.expand_dims(img, axis=0)
|
320 |
+
img_embed = caption_model.cnn_model(img)
|
321 |
+
img_encoded = caption_model.encoder(img_embed, training=False)
|
322 |
|
323 |
y_inp = '[start]'
|
324 |
for i in range(MAX_LENGTH-1):
|
325 |
tokenized = tokenizer([y_inp])[:, :-1]
|
326 |
mask = tf.cast(tokenized != 0, tf.int32)
|
327 |
+
pred = caption_model.decoder(
|
328 |
tokenized, img_encoded, training=False, mask=mask)
|
329 |
|
330 |
pred_idx = np.argmax(pred[0, i, :])
|