def main() -> None:
args = get_args()
config = get_bunch_config_from_json(args.config)
comet_experiment = Experiment(
api_key=config.comet_api_key,
project_name=config.comet_project_name,
workspace=config.comet_workspace,
disabled=not config.use_comet_experiments,
)
comet_experiment.set_name(config.experiment_name)
comet_experiment.log_parameters(config)
test_tweets = load_test_tweets(config.test_data_path)
client = LanguageServiceClient()
result = []
predictions = np.zeros(len(test_tweets), dtype=np.int32)
for i, tweet in enumerate(test_tweets):
start_iter_timestamp = time.time()
document = types.Document(
type=enums.Document.Type.PLAIN_TEXT, content=tweet, language="en"
)
response = client.analyze_sentiment(document=document)
response_dict = MessageToDict(response)
result.append(response_dict)
prediction_present = bool(response_dict["documentSentiment"])
if prediction_present:
# -1, 1 predictions
predictions[i] = 2 * (response.document_sentiment.score > 0) - 1
print("iteration", i, "took:", time.time() - start_iter_timestamp, "seconds")
comet_experiment.log_asset_data(result, name="google_nlp_api_response.json")
ids = np.arange(1, len(test_tweets) + 1).astype(np.int32)
predictions_table = np.column_stack((ids, predictions))
if comet_experiment.disabled:
save_path = build_save_path(config)
os.makedirs(save_path)
formatted_predictions_table = pd.DataFrame(
predictions_table, columns=["Id", "Prediction"], dtype=np.int32,
)
formatted_predictions_table.to_csv(
os.path.join(save_path, "google_nlp_api_predictions.csv"), index=False
)
else:
comet_experiment.log_table(
filename="google_nlp_api_predictions.csv",
tabular_data=predictions_table,
headers=["Id", "Prediction"],
)
percentage_predicted = np.sum(predictions != 0) / predictions.shape[0]
comet_experiment.log_metric(name="percentage predicted", value=percentage_predicted)
def init_experiment(experiment: Experiment, dataset: Dataset):
"""
Initializes an experiment by logging the template and the validation set ground truths if they have not already
been logged.
"""
api_experiment = APIExperiment(previous_experiment=experiment.id)
try:
api_experiment.get_asset("datatap/template.json")
except NotFound:
experiment.log_asset_data([
annotation.to_json()
for annotation in dataset.stream_split("validation")
],
name="datatap/validation/ground_truth.json")
experiment.log_asset_data(dataset.template.to_json(),
name="datatap/template.json")
class Trainer():
def __init__(self, log_dir, cfg):
self.path = log_dir
self.cfg = cfg
if cfg.TRAIN.FLAG:
self.model_dir = os.path.join(self.path, 'Model')
self.log_dir = os.path.join(self.path, 'Log')
mkdir_p(self.model_dir)
mkdir_p(self.log_dir)
self.writer = SummaryWriter(log_dir=self.log_dir)
self.logfile = os.path.join(self.path, "logfile.log")
sys.stdout = Logger(logfile=self.logfile)
self.data_dir = cfg.DATASET.DATA_DIR
self.max_epochs = cfg.TRAIN.MAX_EPOCHS
self.snapshot_interval = cfg.TRAIN.SNAPSHOT_INTERVAL
s_gpus = cfg.GPU_ID.split(',')
self.gpus = [int(ix) for ix in s_gpus]
self.num_gpus = len(self.gpus)
self.batch_size = cfg.TRAIN.BATCH_SIZE
self.lr = cfg.TRAIN.LEARNING_RATE
torch.cuda.set_device(self.gpus[0])
cudnn.benchmark = True
sample = cfg.SAMPLE
self.dataset = []
self.dataloader = []
self.use_feats = cfg.model.use_feats
eval_split = cfg.EVAL if cfg.EVAL else 'val'
train_split = cfg.DATASET.train_split
if cfg.DATASET.DATASET == 'clevr':
clevr_collate_fn = collate_fn
cogent = cfg.DATASET.COGENT
if cogent:
print(f'Using CoGenT {cogent.upper()}')
if cfg.TRAIN.FLAG:
self.dataset = ClevrDataset(data_dir=self.data_dir,
split=train_split + cogent,
sample=sample,
**cfg.DATASET.params)
self.dataloader = DataLoader(dataset=self.dataset,
batch_size=cfg.TRAIN.BATCH_SIZE,
shuffle=True,
num_workers=cfg.WORKERS,
drop_last=True,
collate_fn=clevr_collate_fn)
self.dataset_val = ClevrDataset(data_dir=self.data_dir,
split=eval_split + cogent,
sample=sample,
**cfg.DATASET.params)
self.dataloader_val = DataLoader(dataset=self.dataset_val,
batch_size=cfg.TEST_BATCH_SIZE,
drop_last=False,
shuffle=False,
num_workers=cfg.WORKERS,
collate_fn=clevr_collate_fn)
elif cfg.DATASET.DATASET == 'gqa':
if self.use_feats == 'spatial':
gqa_collate_fn = collate_fn_gqa
elif self.use_feats == 'objects':
gqa_collate_fn = collate_fn_gqa_objs
if cfg.TRAIN.FLAG:
self.dataset = GQADataset(data_dir=self.data_dir,
split=train_split,
sample=sample,
use_feats=self.use_feats,
**cfg.DATASET.params)
self.dataloader = DataLoader(dataset=self.dataset,
batch_size=cfg.TRAIN.BATCH_SIZE,
shuffle=True,
num_workers=cfg.WORKERS,
drop_last=True,
collate_fn=gqa_collate_fn)
self.dataset_val = GQADataset(data_dir=self.data_dir,
split=eval_split,
sample=sample,
use_feats=self.use_feats,
**cfg.DATASET.params)
self.dataloader_val = DataLoader(dataset=self.dataset_val,
batch_size=cfg.TEST_BATCH_SIZE,
shuffle=False,
num_workers=cfg.WORKERS,
drop_last=False,
collate_fn=gqa_collate_fn)
# load model
self.vocab = load_vocab(cfg)
self.model, self.model_ema = mac.load_MAC(cfg, self.vocab)
self.weight_moving_average(alpha=0)
if cfg.TRAIN.RADAM:
self.optimizer = RAdam(self.model.parameters(), lr=self.lr)
else:
self.optimizer = optim.Adam(self.model.parameters(), lr=self.lr)
self.start_epoch = 0
if cfg.resume_model:
location = 'cuda' if cfg.CUDA else 'cpu'
state = torch.load(cfg.resume_model, map_location=location)
self.model.load_state_dict(state['model'])
self.optimizer.load_state_dict(state['optim'])
self.start_epoch = state['iter'] + 1
state = torch.load(cfg.resume_model_ema, map_location=location)
self.model_ema.load_state_dict(state['model'])
if cfg.start_epoch is not None:
self.start_epoch = cfg.start_epoch
self.previous_best_acc = 0.0
self.previous_best_epoch = 0
self.previous_best_loss = 100
self.previous_best_loss_epoch = 0
self.total_epoch_loss = 0
self.prior_epoch_loss = 10
self.print_info()
self.loss_fn = torch.nn.CrossEntropyLoss().cuda()
self.comet_exp = Experiment(
project_name=cfg.COMET_PROJECT_NAME,
api_key=os.getenv('COMET_API_KEY'),
workspace=os.getenv('COMET_WORKSPACE'),
disabled=cfg.logcomet is False,
)
if cfg.logcomet:
exp_name = cfg_to_exp_name(cfg)
print(exp_name)
self.comet_exp.set_name(exp_name)
self.comet_exp.log_parameters(flatten_json_iterative_solution(cfg))
self.comet_exp.log_asset(self.logfile)
self.comet_exp.log_asset_data(json.dumps(cfg, indent=4),
file_name='cfg.json')
self.comet_exp.set_model_graph(str(self.model))
if cfg.cfg_file:
self.comet_exp.log_asset(cfg.cfg_file)
with open(os.path.join(self.path, 'cfg.json'), 'w') as f:
json.dump(cfg, f, indent=4)
def print_info(self):
print('Using config:')
pprint.pprint(self.cfg)
print("\n")
pprint.pprint("Size of train dataset: {}".format(len(self.dataset)))
# print("\n")
pprint.pprint("Size of val dataset: {}".format(len(self.dataset_val)))
print("\n")
print("Using MAC-Model:")
pprint.pprint(self.model)
print("\n")
def weight_moving_average(self, alpha=0.999):
for param1, param2 in zip(self.model_ema.parameters(),
self.model.parameters()):
param1.data *= alpha
param1.data += (1.0 - alpha) * param2.data
def set_mode(self, mode="train"):
if mode == "train":
self.model.train()
self.model_ema.train()
else:
self.model.eval()
self.model_ema.eval()
def reduce_lr(self):
epoch_loss = self.total_epoch_loss # / float(len(self.dataset) // self.batch_size)
lossDiff = self.prior_epoch_loss - epoch_loss
if ((lossDiff < 0.015 and self.prior_epoch_loss < 0.5 and self.lr > 0.00002) or \
(lossDiff < 0.008 and self.prior_epoch_loss < 0.15 and self.lr > 0.00001) or \
(lossDiff < 0.003 and self.prior_epoch_loss < 0.10 and self.lr > 0.000005)):
self.lr *= 0.5
print("Reduced learning rate to {}".format(self.lr))
for param_group in self.optimizer.param_groups:
param_group['lr'] = self.lr
self.prior_epoch_loss = epoch_loss
self.total_epoch_loss = 0
def save_models(self, iteration):
save_model(self.model,
self.optimizer,
iteration,
self.model_dir,
model_name="model")
save_model(self.model_ema,
None,
iteration,
self.model_dir,
model_name="model_ema")
def train_epoch(self, epoch):
cfg = self.cfg
total_loss = 0.
total_correct = 0
total_samples = 0
self.labeled_data = iter(self.dataloader)
self.set_mode("train")
dataset = tqdm(self.labeled_data, total=len(self.dataloader), ncols=20)
for data in dataset:
######################################################
# (1) Prepare training data
######################################################
image, question, question_len, answer = data['image'], data[
'question'], data['question_length'], data['answer']
answer = answer.long()
question = Variable(question)
answer = Variable(answer)
if cfg.CUDA:
if self.use_feats == 'spatial':
image = image.cuda()
elif self.use_feats == 'objects':
image = [e.cuda() for e in image]
question = question.cuda()
answer = answer.cuda().squeeze()
else:
question = question
image = image
answer = answer.squeeze()
############################
# (2) Train Model
############################
self.optimizer.zero_grad()
scores = self.model(image, question, question_len)
loss = self.loss_fn(scores, answer)
loss.backward()
if self.cfg.TRAIN.CLIP_GRADS:
torch.nn.utils.clip_grad_norm_(self.model.parameters(),
self.cfg.TRAIN.CLIP)
self.optimizer.step()
self.weight_moving_average()
############################
# (3) Log Progress
############################
correct = scores.detach().argmax(1) == answer
total_correct += correct.sum().cpu().item()
total_loss += loss.item() * answer.size(0)
total_samples += answer.size(0)
avg_loss = total_loss / total_samples
train_accuracy = total_correct / total_samples
# accuracy = correct.sum().cpu().numpy() / answer.shape[0]
# if avg_loss == 0:
# avg_loss = loss.item()
# train_accuracy = accuracy
# else:
# avg_loss = 0.99 * avg_loss + 0.01 * loss.item()
# train_accuracy = 0.99 * train_accuracy + 0.01 * accuracy
# self.total_epoch_loss += loss.item() * answer.size(0)
dataset.set_description(
'Epoch: {}; Avg Loss: {:.5f}; Avg Train Acc: {:.5f}'.format(
epoch + 1, avg_loss, train_accuracy))
self.total_epoch_loss = avg_loss
dict = {
"loss": avg_loss,
"accuracy": train_accuracy,
"avg_loss": avg_loss, # For commet
"avg_accuracy": train_accuracy, # For commet
}
return dict
def train(self):
cfg = self.cfg
print("Start Training")
for epoch in range(self.start_epoch, self.max_epochs):
with self.comet_exp.train():
dict = self.train_epoch(epoch)
self.reduce_lr()
dict['epoch'] = epoch + 1
dict['lr'] = self.lr
self.comet_exp.log_metrics(
dict,
epoch=epoch + 1,
)
with self.comet_exp.validate():
dict = self.log_results(epoch, dict)
dict['epoch'] = epoch + 1
dict['lr'] = self.lr
self.comet_exp.log_metrics(
dict,
epoch=epoch + 1,
)
if cfg.TRAIN.EALRY_STOPPING:
if epoch - cfg.TRAIN.PATIENCE == self.previous_best_epoch:
# if epoch - cfg.TRAIN.PATIENCE == self.previous_best_loss_epoch:
print('Early stop')
break
self.comet_exp.log_asset(self.logfile)
self.save_models(self.max_epochs)
self.writer.close()
print("Finished Training")
print(
f"Highest validation accuracy: {self.previous_best_acc} at epoch {self.previous_best_epoch}"
)
def log_results(self, epoch, dict, max_eval_samples=None):
epoch += 1
self.writer.add_scalar("avg_loss", dict["loss"], epoch)
self.writer.add_scalar("train_accuracy", dict["accuracy"], epoch)
metrics = self.calc_accuracy("validation",
max_samples=max_eval_samples)
self.writer.add_scalar("val_accuracy_ema", metrics['acc_ema'], epoch)
self.writer.add_scalar("val_accuracy", metrics['acc'], epoch)
self.writer.add_scalar("val_loss_ema", metrics['loss_ema'], epoch)
self.writer.add_scalar("val_loss", metrics['loss'], epoch)
print(
"Epoch: {epoch}\tVal Acc: {acc},\tVal Acc EMA: {acc_ema},\tAvg Loss: {loss},\tAvg Loss EMA: {loss_ema},\tLR: {lr}"
.format(epoch=epoch, lr=self.lr, **metrics))
if metrics['acc'] > self.previous_best_acc:
self.previous_best_acc = metrics['acc']
self.previous_best_epoch = epoch
if metrics['loss'] < self.previous_best_loss:
self.previous_best_loss = metrics['loss']
self.previous_best_loss_epoch = epoch
if epoch % self.snapshot_interval == 0:
self.save_models(epoch)
return metrics
def calc_accuracy(self, mode="train", max_samples=None):
self.set_mode("validation")
if mode == "train":
loader = self.dataloader
# elif (mode == "validation") or (mode == 'test'):
# loader = self.dataloader_val
else:
loader = self.dataloader_val
total_correct = 0
total_correct_ema = 0
total_samples = 0
total_loss = 0.
total_loss_ema = 0.
pbar = tqdm(loader, total=len(loader), desc=mode.upper(), ncols=20)
for data in pbar:
image, question, question_len, answer = data['image'], data[
'question'], data['question_length'], data['answer']
answer = answer.long()
question = Variable(question)
answer = Variable(answer)
if self.cfg.CUDA:
if self.use_feats == 'spatial':
image = image.cuda()
elif self.use_feats == 'objects':
image = [e.cuda() for e in image]
question = question.cuda()
answer = answer.cuda().squeeze()
with torch.no_grad():
scores = self.model(image, question, question_len)
scores_ema = self.model_ema(image, question, question_len)
loss = self.loss_fn(scores, answer)
loss_ema = self.loss_fn(scores_ema, answer)
correct = scores.detach().argmax(1) == answer
correct_ema = scores_ema.detach().argmax(1) == answer
total_correct += correct.sum().cpu().item()
total_correct_ema += correct_ema.sum().cpu().item()
total_loss += loss.item() * answer.size(0)
total_loss_ema += loss_ema.item() * answer.size(0)
total_samples += answer.size(0)
avg_acc = total_correct / total_samples
avg_acc_ema = total_correct_ema / total_samples
avg_loss = total_loss / total_samples
avg_loss_ema = total_loss_ema / total_samples
pbar.set_postfix({
'Acc': f'{avg_acc:.5f}',
'Acc Ema': f'{avg_acc_ema:.5f}',
'Loss': f'{avg_loss:.5f}',
'Loss Ema': f'{avg_loss_ema:.5f}',
})
return dict(acc=avg_acc,
acc_ema=avg_acc_ema,
loss=avg_loss,
loss_ema=avg_loss_ema)
print('Model loaded.')
n_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
log_dict = helpers.flatten_dict(config)
log_dict.update({'trainable_params': n_params})
exp.log_parameters(log_dict)
test_dataset = data.CSVDatasetsMerger(helpers.get_datasets_paths(config, 'test'))
test_dataloader = DataLoader(test_dataset,
batch_size=config['evaluation']['eval_batch_size'],
shuffle=False,
drop_last=False,
num_workers=config['evaluation']['n_eval_workers'],
collate_fn=text_proc)
evaluator = Evaluation(test_dataloader, config)
print('Testing ...')
results, assets, image_fns = evaluator.eval_model(model, finished_training=True)
print('Finished testing. Uploading ...')
exp.log_metrics(results, step=0, epoch=0)
[exp.log_asset_data(asset, step=0) for asset in assets]
[exp.log_image(fn, step=0) for fn in image_fns]
print('Finished uploading.')
def log_validation_proposals(experiment: Experiment,
proposals: Sequence[ImageAnnotation]):
experiment.log_asset_data(
[annotation.to_json() for annotation in proposals],
name="datatap/validation/proposals.json")
class Plotter:
"""
Handles plotting and logging to comet.
Args:
exp_args (args.parse_args): arguments for the experiment
agent_args (dict): arguments for the agent
agent (Agent): the agent
"""
def __init__(self, exp_args, agent_args, agent):
self.exp_args = exp_args
self.agent_args = agent_args
self.agent = agent
self.experiment = None
if self.exp_args.plotting:
self.experiment = Experiment(api_key=LOGGING_API_KEY,
project_name=PROJECT_NAME,
workspace=WORKSPACE)
self.experiment.disable_mp()
self.experiment.log_parameters(get_arg_dict(exp_args))
self.experiment.log_parameters(flatten_arg_dict(agent_args))
self.experiment.log_asset_data(json.dumps(get_arg_dict(exp_args)), name='exp_args')
self.experiment.log_asset_data(json.dumps(agent_args), name='agent_args')
if self.exp_args.checkpoint_exp_key is not None:
self.load_checkpoint()
self.result_dict = None
# keep a hard-coded list of returns in case Comet fails
self.returns = []
def _plot_ts(self, key, observations, statistics, label, color):
dim_obs = min(observations.shape[1], 9)
k = 1
for i in range(dim_obs):
plt.subplot(int(str(dim_obs) + '1' + str(k)))
observations_i = observations[:-1, i].cpu().numpy()
if key == 'action' and self.agent.postprocess_action:
observations_i = np.tanh(observations_i)
plt.plot(observations_i.squeeze(), 'o', label='observation', color='k', markersize=2)
if len(statistics) == 1: # Bernoulli distribution
probs = statistics['probs']
probs = probs.cpu().numpy()
plt.plot(probs, label=label, color=color)
elif len(statistics) == 2:
if 'loc' in statistics:
# Normal distribution
mean = statistics['loc']
std = statistics['scale']
mean = mean[:, i].cpu().numpy()
std = std[:, i].cpu().numpy()
mean = mean.squeeze()
std = std.squeeze()
x, plus, minus = mean, mean + std, mean - std
if key == 'action' and label == 'approx_post' and self.agent_args['approx_post_args']['dist_type'] in ['TanhNormal', 'TanhARNormal']:
# Tanh Normal distribution
x, plus, minus = np.tanh(x), np.tanh(plus), np.tanh(minus)
if key == 'action' and label == 'direct_approx_post' and self.agent_args['approx_post_args']['dist_type'] in ['TanhNormal', 'TanhARNormal']:
# Tanh Normal distribution
x, plus, minus = np.tanh(x), np.tanh(plus), np.tanh(minus)
if key == 'action' and label == 'prior' and self.agent_args['prior_args']['dist_type'] in ['TanhNormal', 'TanhARNormal']:
# Tanh Normal distribution
x, plus, minus = np.tanh(x), np.tanh(plus), np.tanh(minus)
if key == 'action' and self.agent.postprocess_action:
x, plus, minus = np.tanh(x), np.tanh(plus), np.tanh(minus)
if key == 'action' and label == 'prior' and self.agent_args['prior_args']['dist_type'] == 'NormalUniform':
# Normal + Uniform distribution
x, plus, minus = x, np.minimum(plus, 1.), np.maximum(minus, -1)
elif 'low' in statistics:
# Uniform distribution
low = statistics['low'][:, i].cpu().numpy()
high = statistics['high'][:, i].cpu().numpy()
x = low + (high - low) / 2
plus, minus = x + high, x + low
else:
raise NotImplementedError
plt.plot(x, label=label, color=color)
plt.fill_between(np.arange(len(x)), plus, minus, color=color, alpha=0.2, label=label)
else:
NotImplementedError
k += 1
def plot_states_and_rewards(self, states, rewards, step):
"""
Plots the states and rewards for a collected episode.
"""
# states
plt.figure()
dim_obs = states.shape[1]
for i in range(dim_obs):
plt.subplot(dim_obs, 1, i+1)
states_i = states[:-1, i].cpu().numpy()
plt.plot(states_i.squeeze(), 'o', label='state', color='k', markersize=2)
self.experiment.log_figure(figure=plt, figure_name='states_ts_'+str(step))
plt.close()
# rewards
plt.figure()
rewards = rewards[:-1, 0].cpu().numpy()
plt.plot(rewards.squeeze(), 'o', label='reward', color='k', markersize=2)
self.experiment.log_figure(figure=plt, figure_name='rewards_ts_'+str(step))
plt.close()
def plot_episode(self, episode, step):
"""
Plots a newly collected episode.
"""
if self.exp_args.plotting:
self.experiment.log_metric('cumulative_reward', episode['reward'].sum(), step)
def merge_legends():
handles, labels = plt.gca().get_legend_handles_labels()
newLabels, newHandles = [], []
for handle, label in zip(handles, labels):
if label not in newLabels:
newLabels.append(label)
newHandles.append(handle)
plt.legend(newHandles, newLabels)
for k in episode['distributions'].keys():
for i, l in enumerate(episode['distributions'][k].keys()):
color = COLORS[i]
self._plot_ts(k, episode[k], episode['distributions'][k][l], l, color)
plt.suptitle(k)
merge_legends()
self.experiment.log_figure(figure=plt, figure_name=k + '_ts_'+str(step))
plt.close()
self.plot_states_and_rewards(episode['state'], episode['reward'], step)
def log_eval(self, episode, eval_states, step):
"""
Plots an evaluation episode performance. Logs the episode.
Args:
episode (dict): dictionary containing agent's collected episode
eval_states (dict): dictionary of MuJoCo simulator states
step (int): the current step number in training
"""
# plot and log eval returns
eval_return = episode['reward'].sum()
print(' Eval. Return at Step ' + str(step) + ': ' + str(eval_return.item()))
self.returns.append(eval_return.item())
if self.exp_args.plotting:
self.experiment.log_metric('eval_cumulative_reward', eval_return, step)
json_str = json.dumps(self.returns)
self.experiment.log_asset_data(json_str, name='eval_returns', overwrite=True)
# log the episode itself
for ep_item_str in ['state', 'action', 'reward']:
ep_item = episode[ep_item_str].tolist()
json_str = json.dumps(ep_item)
item_name = 'episode_step_' + str(step) + '_' + ep_item_str
self.experiment.log_asset_data(json_str, name=item_name)
# log the MuJoCo simulator states
for sim_item_str in ['qpos', 'qvel']:
if len(eval_states[sim_item_str]) > 0:
sim_item = eval_states[sim_item_str].tolist()
json_str = json.dumps(sim_item)
item_name = 'episode_step_' + str(step) + '_' + sim_item_str
self.experiment.log_asset_data(json_str, name=item_name)
def plot_agent_kl(self, agent_kl, step):
if self.exp_args.plotting:
self.experiment.log_metric('agent_kl', agent_kl, step)
def log_results(self, results):
"""
Log the results dictionary.
"""
if self.result_dict is None:
self.result_dict = {}
for k, v in flatten_arg_dict(results).items():
if k not in self.result_dict:
self.result_dict[k] = [v]
else:
self.result_dict[k].append(v)
def plot_results(self, timestep):
"""
Plot/log the results to Comet.
"""
if self.exp_args.plotting:
for k, v in self.result_dict.items():
avg_value = np.mean(v)
self.experiment.log_metric(k, avg_value, timestep)
self.result_dict = None
def plot_model_eval(self, episode, predictions, log_likelihoods, step):
"""
Plot/log the results from model evaluation.
Args:
episode (dict): a collected episode
predictions (dict): predictions from each state, containing [n_steps, horizon, n_dims]
log_likelihoods (dict): log-likelihood evaluations of predictions, containing [n_steps, horizon, 1]
"""
if self.exp_args.plotting:
for variable, lls in log_likelihoods.items():
# average the log-likelihood estimates and plot the result at the horizon length
mean_ll = lls[:, -1].mean().item()
self.experiment.log_metric(variable + '_pred_log_likelihood', mean_ll, step)
# plot log-likelihoods as a function of rollout step
plt.figure()
mean = lls.mean(dim=0).view(-1)
std = lls.std(dim=0).view(-1)
plt.plot(mean.numpy())
lower = mean - std
upper = mean + std
plt.fill_between(np.arange(lls.shape[1]), lower.numpy(), upper.numpy(), alpha=0.2)
plt.xlabel('Rollout Step')
plt.ylabel('Prediction Log-Likelihood')
plt.xticks(np.arange(lls.shape[1]))
self.experiment.log_figure(figure=plt, figure_name=variable + '_pred_ll_' + str(step))
plt.close()
# plot predictions vs. actual values for an arbitrary time step
time_step = np.random.randint(predictions['state']['loc'].shape[0])
for variable, preds in predictions.items():
pred_loc = preds['loc'][time_step]
pred_scale = preds['scale'][time_step]
x = episode[variable][time_step+1:time_step+1+pred_loc.shape[0]]
plt.figure()
horizon, n_dims = pred_loc.shape
for plot_num in range(n_dims):
plt.subplot(n_dims, 1, plot_num + 1)
plt.plot(pred_loc[:, plot_num].numpy())
lower = pred_loc[:, plot_num] - pred_scale[:, plot_num]
upper = pred_loc[:, plot_num] + pred_scale[:, plot_num]
plt.fill_between(np.arange(horizon), lower.numpy(), upper.numpy(), alpha=0.2)
plt.plot(x[:, plot_num].numpy(), '.')
plt.xlabel('Rollout Step')
plt.xticks(np.arange(horizon))
self.experiment.log_figure(figure=plt, figure_name=variable + '_pred_' + str(step))
plt.close()
def save_checkpoint(self, step):
"""
Checkpoint the model by getting the state dictionary for each component.
"""
if self.exp_args.plotting:
print('Checkpointing the agent...')
state_dict = self.agent.state_dict()
cpu_state_dict = {k: v.cpu() for k, v in state_dict.items()}
ckpt_path = os.path.join('./ckpt_step_'+ str(step) + '.ckpt')
torch.save(cpu_state_dict, ckpt_path)
self.experiment.log_asset(ckpt_path)
os.remove(ckpt_path)
print('Done.')
def load_checkpoint(self, timestep=None):
"""
Loads a checkpoint from Comet.
Args:
timestep (int, optional): the checkpoint timestep, default is latest
"""
load_checkpoint(self.agent, self.exp_args.checkpoint_exp_key, timestep)
class Trainer:
def __init__(self, cfg, log_dir):
self.log_dir = log_dir
self.cfg = cfg
if log_dir:
self.log = True
self.model_dir = osp.join(log_dir, "checkpoints")
mkdir_p(self.model_dir)
self.logfile = osp.join(log_dir, "logfile.log")
sys.stdout = Logger(logfile=self.logfile)
self.summary_writer = tf.summary.create_file_writer(log_dir)
else:
self.log = False
self.generator = Generator(**cfg.model.generator)
self.discriminator = Discriminator(**cfg.model.discriminator)
self.g_optimizer = optimizers.Adam(**cfg.train.generator.optimizer)
self.d_optimizer = optimizers.Adam(**cfg.train.discriminator.optimizer)
self.bce = losses.BinaryCrossentropy(from_logits=True)
# TODO resume model
self.comet = Experiment(
api_key=os.getenv("COMET_API_KEY"),
workspace=os.getenv("COMET_WORKSPACE"),
project_name=cfg.comet_project_name,
disabled=cfg.logcomet is False or not self.log,
)
self.comet.set_name(f"{cfg.experiment_name}/{cfg.run_name}")
self.comet.log_parameters(flatten_json_iterative_solution(self.cfg))
self.comet.log_asset_data(json.dumps(self.cfg, indent=4), file_name="cfg.json")
if cfg.cfg_file:
self.comet.log_asset(cfg.cfg_file)
self.start_epoch = tf.Variable(0)
self.curr_step = tf.Variable(0)
self.ckpt = tf.train.Checkpoint(
generator=self.generator,
discriminator=self.discriminator,
g_optimizer=self.g_optimizer,
d_optimizer=self.d_optimizer,
start_epoch=self.start_epoch,
curr_step=self.curr_step,
)
if cfg.train.resume:
ckpt_resumer = tf.train.CheckpointManager(
self.ckpt, cfg.train.resume, max_to_keep=3,
)
# if a checkpoint exists, restore the latest checkpoint.
if ckpt_resumer.latest_checkpoint:
self.ckpt.restore(ckpt_resumer.latest_checkpoint)
print("Latest checkpoint restored!!", ckpt_resumer.latest_checkpoint)
print(
f"Last epoch trained:{self.start_epoch.numpy()}, Current step: {self.curr_step.numpy()}"
)
if self.log:
with open(osp.join(self.log_dir, "cfg.json"), "w") as f:
json.dump(cfg, f, indent=4)
self.ckpt_manager = tf.train.CheckpointManager(
self.ckpt, self.model_dir, max_to_keep=3
)
self.prepare_dataset(self.cfg.train.data_dir)
self.print_info()
if self.cfg.train.generator.fixed_z:
self.z_bg = sample_z(1, self.generator.z_dim_bg, num_objects=1)
self.z_fg = sample_z(1, self.generator.z_dim_fg, num_objects=1)
self.bg_view = sample_view(1, num_objects=1)
self.fg_view = sample_view(1, num_objects=1)
else:
self.z_bg = self.z_fg = self.bg_view = self.fg_view = None
def prepare_dataset(self, data_dir):
self.data_dir = data_dir
self.num_tr = len(glob.glob(osp.join(self.data_dir, "*.png")))
self.list_ds_train = tf.data.Dataset.list_files(
os.path.join(self.data_dir, "*.png")
)
self.labeled_ds = self.list_ds_train.map(
lambda x: process_path(
x, self.cfg.train.image_height, self.cfg.train.image_width
),
num_parallel_calls=AUTOTUNE,
)
self.steps_per_epoch = int(math.ceil(self.num_tr / self.cfg.train.batch_size))
def print_info(self):
print("Using config:")
pprint.pprint(self.cfg)
print("\n")
pprint.pprint("Size of train dataset: {}".format(self.num_tr))
print("\n")
# lossess
def discriminator_loss(self, real, generated):
real_loss = self.bce(tf.ones_like(real), real)
generated_loss = self.bce(tf.zeros_like(generated), generated)
total_disc_loss = real_loss + generated_loss
return total_disc_loss * 0.5
def generator_loss(self, generated):
return self.bce(tf.ones_like(generated), generated)
# def generate_random_noise(self, batch_size, num_objects=(3, 10)):
# z_bg = tf.random.uniform(
# (batch_size, self.generator.z_dim_bg), minval=-1, maxval=1
# )
# num_objs = tf.random.uniform(
# (batch_size,),
# minval=num_objects[0],
# maxval=num_objects[1] + 1,
# dtype=tf.int32,
# )
# tensors = []
# max_len = max(num_objs)
# for no in num_objs:
# _t = tf.random.uniform((no, self.generator.z_dim_fg), minval=-1, maxval=1)
# _z = tf.zeros((max_len - no, self.generator.z_dim_fg), dtype=tf.float32)
# _t = tf.concat((_t, _z), axis=0)
# tensors.append(_t)
# z_fg = tf.stack(tensors, axis=0)
# return z_bg, z_fg
def batch_logits(self, image_batch, z_bg, z_fg, bg_view, fg_view):
generated = self.generator(z_bg, z_fg, bg_view, fg_view)
d_fake_logits = self.discriminator(generated, training=True)
image_batch = (image_batch * 2) - 1
if self.cfg.train.discriminator.random_noise or self.curr_step <= 2000:
image_batch = image_batch + tf.random.normal(image_batch.shape, stddev=0.01)
d_real_logits = self.discriminator(image_batch, training=True,)
return d_fake_logits, d_real_logits, generated
# @tf.function
def train_epoch(self, epoch):
train_iter = prepare_for_training(
self.labeled_ds, self.cfg.train.batch_size, cache=False,
)
pbar = tqdm(
enumerate(train_iter),
total=self.steps_per_epoch,
ncols=20,
desc=f"Epoch {epoch}",
mininterval=10,
miniters=50,
)
total_d_loss = 0.0
total_g_loss = 0.0
counter = 1
real_are_real_samples_counter = 0
real_samples_counter = 0
fake_are_fake_samples_counter = 0
fake_samples_counter = 0
z_bg = z_fg = None
for it, image_batch in pbar:
bsz = image_batch.shape[0]
# generated random noise
if self.z_bg is not None:
# For overfitting one sample and debugging
z_bg = tf.repeat(self.z_bg, bsz, axis=0)
z_fg = tf.repeat(self.z_fg, bsz, axis=0)
bg_view = self.bg_view
fg_view = self.fg_view
else:
z_bg = sample_z(bsz, self.generator.z_dim_bg, num_objects=1)
z_fg = sample_z(
bsz,
self.generator.z_dim_fg,
num_objects=(3, min(10, 3 + 1 * (epoch // 2))),
)
bg_view = sample_view(
batch_size=bsz,
num_objects=1,
azimuth_range=(-20, 20),
elevation_range=(-10, 10),
scale_range=(0.9, 1.1),
)
fg_view = sample_view(batch_size=bsz, num_objects=z_fg.shape[1])
with tf.GradientTape(persistent=True) as tape:
# fake img
d_fake_logits, d_real_logits, generated = self.batch_logits(
image_batch, z_bg, z_fg, bg_view, fg_view
)
d_loss = self.discriminator_loss(d_real_logits, d_fake_logits)
g_loss = self.generator_loss(d_fake_logits)
total_d_loss += d_loss.numpy()
# total_g_loss += g_loss.numpy() / self.cfg.train.generator.update_freq
total_g_loss += g_loss.numpy()
d_variables = self.discriminator.trainable_variables
d_gradients = tape.gradient(d_loss, d_variables)
self.d_optimizer.apply_gradients(zip(d_gradients, d_variables))
g_variables = self.generator.trainable_variables
g_gradients = tape.gradient(g_loss, g_variables)
self.g_optimizer.apply_gradients(zip(g_gradients, g_variables))
del tape
real_samples_counter += d_real_logits.shape[0]
fake_samples_counter += d_fake_logits.shape[0]
real_are_real = (d_real_logits >= 0).numpy().sum()
real_are_real_samples_counter += real_are_real
fake_are_fake = (d_fake_logits < 0).numpy().sum()
fake_are_fake_samples_counter += fake_are_fake
# according to paper generator makes 2 steps per each step of the disc
# for _ in range(self.cfg.train.generator.update_freq - 1):
# with tf.GradientTape(persistent=True) as tape:
# # fake img
# d_fake_logits, _, generated = self.batch_logits(
# image_batch, z_bg, z_fg
# )
# g_loss = self.generator_loss(d_fake_logits)
# g_variables = self.generator.trainable_variables
# g_gradients = tape.gradient(g_loss, g_variables)
# self.g_optimizer.apply_gradients(zip(g_gradients, g_variables))
# total_g_loss += g_loss.numpy() / self.cfg.train.generator.update_freq
pbar.set_postfix(
g_loss=f"{g_loss.numpy():.2f} ({total_g_loss / (counter):.2f})",
d_loss=f"{d_loss.numpy():.4f} ({total_d_loss / (counter):.4f})",
rrr=f"{real_are_real / d_real_logits.shape[0]:.1f} ({real_are_real_samples_counter / real_samples_counter:.1f})",
frf=f"{fake_are_fake / d_fake_logits.shape[0]:.1f} ({fake_are_fake_samples_counter / fake_samples_counter:.1f})",
refresh=False,
)
if it % (self.cfg.train.it_log_interval) == 0:
self.log_training(
d_loss=total_d_loss / counter,
g_loss=total_g_loss / counter,
real_are_real=real_are_real_samples_counter / real_samples_counter,
fake_are_fake=fake_are_fake_samples_counter / fake_samples_counter,
fake_images=(generated + 1) / 2,
real_images=image_batch,
d_fake_logits=d_fake_logits,
d_real_logits=d_real_logits,
epoch=epoch,
it=it,
)
real_are_real_samples_counter = 0
fake_are_fake_samples_counter = 0
real_samples_counter = 0
fake_samples_counter = 0
total_d_loss = 0.0
total_g_loss = 0.0
counter = 0
counter += 1
gc.collect()
del train_iter
def log_training(
self,
d_loss,
g_loss,
fake_images,
real_images,
d_fake_logits,
d_real_logits,
epoch,
it,
real_are_real,
fake_are_fake,
):
if self.log:
curr_step = (self.curr_step + it).numpy()
real_are_real_images, real_are_fake_images = split_images_on_disc(
real_images, d_real_logits
)
fake_are_real_images, fake_are_fake_images = split_images_on_disc(
fake_images, d_fake_logits
)
with self.summary_writer.as_default():
tf.summary.scalar(
"losses/d_loss",
d_loss,
step=curr_step,
description="Average of predicting real images as real and fake as fake",
)
tf.summary.scalar(
"losses/g_loss",
g_loss,
step=curr_step,
description="Predicting fake images as real",
)
tf.summary.scalar(
"accuracy/real",
real_are_real,
step=curr_step,
description="Real images classified as real",
)
tf.summary.scalar(
"accuracy/fake",
fake_are_fake,
step=curr_step,
description="Fake images classified as fake",
)
tf.summary.image(
f"{epoch}-{curr_step}-fake/are_fake",
fake_are_fake_images,
max_outputs=25,
step=curr_step,
description="Fake images that the discriminator says are fake",
)
tf.summary.image(
f"{epoch}-{curr_step}-fake/are_real",
fake_are_real_images,
max_outputs=25,
step=curr_step,
description="Fake images that the discriminator says are real",
)
tf.summary.image(
f"{epoch}-{curr_step}-real/are_fake",
real_are_fake_images,
max_outputs=25,
step=curr_step,
description="Real images that the discriminator says are fake",
)
tf.summary.image(
f"{epoch}-{curr_step}-real/are_real",
real_are_real_images,
max_outputs=25,
step=curr_step,
description="Real images that the discriminator says are real",
)
self.comet.log_metrics(
{"d_loss": d_loss, "g_loss": g_loss}, step=curr_step, epoch=epoch
)
fig = show_batch(fake_images, labels=disc_preds_to_label(d_fake_logits))
self.comet.log_figure(
figure=fig, figure_name="" f"fake_{epoch}_{it}.jpg", step=curr_step,
)
plt.close(fig)
fig = show_batch(real_images, labels=disc_preds_to_label(d_real_logits))
self.comet.log_figure(
figure=fig, figure_name="" f"real_{epoch}_{it}.jpg", step=curr_step,
)
plt.close(fig)
def train(self):
print("Start training")
for epoch in range(self.start_epoch.numpy(), self.cfg.train.epochs):
with self.comet.train():
self.train_epoch(epoch)
self.curr_step.assign_add(self.steps_per_epoch)
self.start_epoch.assign_add(1)
if self.log and (((epoch + 1) % self.cfg.train.snapshot_interval) == 0):
self.ckpt_manager.save(epoch + 1)
def save_model(self, epoch):
pass