def predict_word(self, X):
crnn = CRNN()
crnn.build(dropout=False)
crnn.model.load_weights(self.weight_name + ".h5")
out = crnn.test_func([X])[0]
ret = []
for j in range(out.shape[0]):
out_best = list(np.argmax(out[j, 2:], 1))
out_best = [k for k, g in itertools.groupby(out_best)]
outstr = ''
for c in out_best:
if 0 <= c <= 9:
outstr += chr(c + ord('0'))
elif 10 <= c <= 35:
outstr += chr(c - 10 + ord('A'))
elif 36 <= c <= 61:
outstr += chr(c - 36 + ord('a'))
ret.append(outstr)
return ret
def train(args):
@tf.function
def train_step(x, y):
with tf.GradientTape() as tape:
y_pred = model(x["the_input"])
# loss = tf.reduce_mean(ctc_lambda_func((y_pred, x["the_labels"], x["input_length"].reshape((-1,1)), x["label_length"].reshape((-1,1)))))
loss = tf.reduce_mean(ctc_lambda_func((y_pred, x["the_labels"], tf.reshape(x["input_length"], [-1, 1]), tf.reshape(x["label_length"], [-1, 1]))))
# Compute gradients
trainable_vars = model.trainable_variables
gradients = tape.gradient(loss, trainable_vars)
# Update weights
model.optimizer.apply_gradients(zip(gradients, trainable_vars))
return loss
epochs = 1000
iter_per_epoch = 100
#model, test_func = get_CResRNN(weights=os.path.join("OUTPUT_DIR", "exp1", "weights06.h5"))
#model, test_func = get_CResRNN(weights=os.path.join("OUTPUT_DIR", "weights0995.h5"))
#model.load_weights(os.path.join("OUTPUT_DIR", "exp1", "weights15.h5"))
#model.load_weights(os.path.join("OUTPUT_DIR", "weights0995.h5"))
model2, test_func = CRNN_model()
train_generator = FakeImageGenerator(args).next_gen()
model = CRNN(ALPHABET)
model.build()
model.summary()
# model = tf.keras.load_model('checkpoints/checkpoint')
model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001, clipnorm=5))
loss_train = []
for epoch in range(1, epochs):
print(f"Start of epoch {epoch}")
pb = Progbar(iter_per_epoch, stateful_metrics="loss")
for iter in range(iter_per_epoch):
x, y = next(train_generator)
with tf.GradientTape() as tape:
y_pred = model(x["the_input"])
# loss = tf.reduce_mean(ctc_lambda_func((y_pred, x["the_labels"], x["input_length"].reshape((-1,1)), x["label_length"].reshape((-1,1)))))
loss = tf.reduce_mean(ctc_lambda_func((y_pred, x["the_labels"], tf.reshape(x["input_length"], [-1, 1]), tf.reshape(x["label_length"], [-1, 1]))))
# Compute gradients
trainable_vars = model.trainable_variables
gradients = tape.gradient(loss, trainable_vars)
# Update weights
model.optimizer.apply_gradients(zip(gradients, trainable_vars))
values = [('loss', loss)]
pb.add(1, values=values)
if epoch % 5 == 0:
model.save("checkpoints/base_crnn.h5")
# print("test2")
# x, y = next(train_generator)
# model.fit(x, y)
# print("test1")
x, y = next(train_generator)
print(model(x["the_input"]))
"""
