def train_model(model, train_iterator, val_iterator, optimizer, scheduler,
criterion, n_epochs, clip, model_path):
'''
开始训练我们的模型：
1.每一次epoch，都会检查模型是否达到的最佳的validation loss，如果达到了，就更新
最好的validation loss以及保存模型参数
2.打印每个epoch的loss以及困惑度。
'''
best_valid_loss = float('inf')
for epoch in range(n_epochs):
start_time = time.time()
train_loss = train(model, train_iterator, optimizer, criterion, clip)
valid_loss = evaluate(model, val_iterator, criterion)
end_time = time.time()
epoch_mins, epoch_secs = epoch_time(start_time, end_time)
if valid_loss < best_valid_loss:
best_valid_loss = valid_loss
torch.save(model.state_dict(), model_path)
scheduler.step()
print('epoch:{},time-mins:{},time-secs:{}'.format(
epoch + 1, epoch_mins, epoch_secs))
print('train loss:{},train perplexity:{}'.format(
train_loss, math.exp(train_loss)))
print('val loss:{}, val perplexity:{}'.format(valid_loss,
math.exp(valid_loss)))
def time_updated(self):
"""
Should be called once the system time is set correctly (ex. from NTP
or a browser client).
Updates the metadata now that we know the real time of the data we've been
collecting.
"""
self.start_time_offset = epoch_time() * 1000 - time.ticks_ms()
for log in self.our_logs_without_start_time:
log["start_time_offset"] = self.start_time_offset
self.our_logs_without_start_time = []
with open(self.meta_file, "r") as f:
with open(self.tmp_meta, "w") as f2:
# print("time_updated():")
for line in f.readlines():
# print("<%s>" % line.strip())
index, active, start_time, start_time_offset = line.strip().split(",")
if start_time_offset == "None":
start_time_offset = str(self.start_time_offset)
line2 = ",".join([index, active, start_time, start_time_offset])
# print("[%s]" % line2)
f2.write(line2)
f2.write("\n")
os.rename(self.tmp_meta, self.meta_file)
def collect(self):
time_start = datetime.utcnow()
collected_data = self.collect_fn()
time_end = datetime.utcnow()
data = {
'nick': self.nick,
'time_start': epoch_time(time_start),
'time_end': epoch_time(time_end),
'time_start_pretty': format_time(time_start),
'time_end_pretty': format_time(time_end),
'data': collected_data
}
log('%s: Collected the following data: %s' % (self.name, json.dumps(data, indent=2)), self.verbose)
self.upload_queue.append(data)
self.upload()
SAVE_DIR = 'models'
MODEL_SAVE_PATH = os.path.join(SAVE_DIR, 'transformer-seq2seq.pt')
best_valid_loss = float('inf')
if not os.path.isdir(f'{SAVE_DIR}'):
os.makedirs(f'{SAVE_DIR}')
train_total = []
valid_total = []
test_total = []
for epoch in range(N_EPOCHS):
start_time = time.time()
train_loss = trainer(model, train_iter, optimizer, criterion, CLIP)
valid_loss = evaluate(model, val_iter, criterion)
train_total.append(train_loss)
valid_total.append(valid_loss)
end_time = time.time()
epoch_mins, epoch_secs = epoch_time(start_time, end_time)
if valid_loss < best_valid_loss:
best_valid_loss = valid_loss
torch.save(model.state_dict(), MODEL_SAVE_PATH)
print(f'| Epoch: {epoch+1:03} | Time: {epoch_mins}m {epoch_secs}s| Train Loss: {train_loss:.3f} | Train PPL: {math.exp(train_loss):7.3f} | Val. Loss: {valid_loss:.3f} | Val. PPL: {math.exp(valid_loss):7.3f} |')
print(model.embedding.weight.data)
print('Just started:')
valid_loss, valid_acc = util.evaluate(model, valid_iterator, criterion,
labelName)
print(f'\t Val. Loss: {valid_loss:.3e} | Val. Acc: {valid_acc*100:.2e}%')
for epoch in range(N_EPOCHS):
start_time = time.time()
train_loss, train_acc = util.train(model, train_iterator, optimizer,
criterion, labelName)
valid_loss, valid_acc = util.evaluate(model, valid_iterator, criterion,
labelName)
end_time = time.time()
epoch_mins, epoch_secs = util.epoch_time(start_time, end_time)
if valid_loss < best_valid_loss:
best_valid_loss = valid_loss
torch.save(model.state_dict(), modelName)
print(f'Epoch: {epoch+1:02} | Epoch Time: {epoch_mins}m {epoch_secs}s')
print(
f'\tTrain Loss: {train_loss:.3e} | Train Acc: {train_acc*100:.2e}%'
)
print(
f'\t Val. Loss: {valid_loss:.3e} | Val. Acc: {valid_acc*100:.2e}%'
)
model.load_state_dict(torch.load(modelName))
test_loss, test_acc = util.evaluate(model, test_iterator, criterion,
labelName)
disp.fill_rect(0, 20, 128, 128, 0)
disp.text(
"In: %.2fA %.0fW" %
(sense.get_current("in") or 0, sense.get_power("in") or 0), 0,
22)
disp.text(
"Out: %.2fA %.0fW" %
(sense.get_current("out") or 0, sense.get_power("out") or 0),
0, 32)
disp.text(
"%02d:%02d:%02d UTC %3s" % (time.localtime()[3:6] +
(len(open_web_sockets), )), 0, 55)
disp.show()
ws_text = ",".join([
str(epoch_time() * 1000),
str(sense.get_power("in")),
str(sense.get_power("out")),
str(sense.get_voltage("in")),
str(sense.get_voltage("out")),
str(sense.get_current("in")),
str(sense.get_current("out")),
str(sense.get_available_power("in")),
])
# print(ws_text)
for ws in open_web_sockets:
start_time = time.ticks_ms()
ws.SendText(ws_text)
gc.collect()
