def get_data(self, video_paths, label_paths, gt_paths, clips, mode):
print("create generator....")
batch_gen = process_data.get_batch(video_paths,
label_paths,
gt_paths,
clips=clips,
batch_size=self.batch_size,
mode=mode,
width=self.width,
height=self.height)
return batch_gen
def evaluate(model, data_src):
model.eval()
total_loss = 0
src_mask = model.gen_square_subsequent_mask(bptt).to(device)
with torch.no_grad():
for ii in range(0, data_src.size(0) - 1, bptt):
data, targets = get_batch(data_src, ii)
if data.size(0) != bptt:
src_mask = model.gen_square_subsequent_mask(
data.size(0)).to(device)
output = model(data, src_mask)
output_flat = output.view(-1, ntokens)
total_loss += len(data) * criterion(output_flat,
targets).item()
return total_loss / (len(data_src) - 1)
def train(model):
# Turn on the train mode
model.train()
total_loss = 0
start_time = time.time()
src_mask = model.gen_square_subsequent_mask(bptt).to(device)
def _gen_train_info():
return ('| epoch {:3d} | {:5d}/{:5d} batches | ' +
'lr {:02.2f} | ms/batch {:5.2f} | ' +
'loss {:5.2f} | ppl {:8.2f}').format(
epoch, batch,
len(train_data) // bptt,
scheduler.get_last_lr()[0], elapsed * 1000 / log_interval,
cur_loss, math.exp(cur_loss))
for batch, ii in enumerate(range(0, train_data.size(0) - 1, bptt)):
data, targets = get_batch(train_data, ii)
optimizer.zero_grad()
if data.size(0) != bptt:
src_mask = model.gen_square_subsequent_mask(
data.size(0)).to(device)
output = model(data, src_mask)
loss = criterion(output.view(-1, ntokens), targets)
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), 0.5)
optimizer.step()
total_loss += loss.item()
log_interval = 200
if batch % log_interval == 0 and batch > 0:
cur_loss = total_loss / log_interval
elapsed = time.time() - start_time
print(_gen_train_info())
total_loss = 0
start_time = time.time()
print(y_conv)
# tf.nn.sofmax_cross_ entropy_ with_logits 函数 ， 它可以直接对
# Logit 定义交叉烟损失 ， 写法为
cross_entropy = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits_v2(labels=y_, logits=y_conv))
train_op = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
correct = tf.equal(tf.argmax(y_, axis=1), tf.argmax(y_conv, axis=1))
acc = tf.reduce_mean(tf.cast(correct, tf.float32))
saver = tf.train.Saver()
ck_path = "./model/hand"
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for i in range(100):
batch = get_batch(50)
if i % 10 == 0:
# 原来我写的是acc = sess.run(acc,feed_dict = {x:batch[0],y_:batch[1],keep_drop:1.0})
# 这样acc会变成常数导致出错
acc_v = sess.run(acc,
feed_dict={
x_image: batch[0],
y_: batch[1],
keep_drop: 1.0
})
print(acc_v)
if i % 10 == 0:
print("save")
saver.save(sess, ck_path, write_meta_graph=True)
sess.run(train_op,
feed_dict={
total = rep(data)
sc = MinMaxScaler((-0.95, 0.95))
total = sc.fit_transform(total)
#arr = total[60:60+batch_size, :]
#_, l = sess.run([training_op, loss], feed_dict = {X_in: arr, Y: arr, rate: 0.2})
#print("Batch loss: " + str(l))
for epoch in range(epochs):
epoch_loss = []
print("Epoch: " + str(epoch))
for i in range(len(total) // batch_size):
batch = process_data.get_batch(i, batch_size, total)
batch_loss = []
_, l = sess.run([training_op, loss],
feed_dict={
X_in: batch,
Y: batch,
rate: 0.2
})
batch_loss.append(l)
print("Curr Epoch: " + str(epoch))
print("Batch Loss: " + str(np.mean(batch_loss)))
epoch_loss.append(np.mean(batch_loss))
print("Epoch Avg Loss: " + str(np.mean(epoch_loss)))
save_path = saver.save(sess, "./Model/model1.ckpt")