def processing():
dir_sent_train = f"../data/data_preprocessed/{domain}/train/"
dir_sent_test = f"../data/data_preprocessed/{domain}/test/"
# dir_doc = "../data/data_doc/"
dir_output = "../data/pkl/IMN/"
if not os.path.exists(dir_output):
os.mkdir(dir_output)
print("reading data file...")
dataset_sent_train = read_data1(dir_sent_train)
dataset_sent_test = read_data1(dir_sent_test)
# dataset_doc = read_data2(dir_doc)
# datasets = [dataset_sent_train, dataset_sent_test, dataset_doc]
datasets = [dataset_sent_train, dataset_sent_test]
print("making vocab...")
vocab = get_vocab(datasets)
word2idx, idx2word = get_word_to_idx(vocab)
# read the embedding files
print("making embedding")
emb_dir = "../data/embeddings/"
general_emb_file = emb_dir + "glove.840B.300d.txt"
domain_emb_file = emb_dir + f"{domain}.txt"
general_embeddings = read_emb_idx(general_emb_file, word2idx, 300)
domain_embeddings = read_emb_idx(domain_emb_file, word2idx, 100)
# dataset_total = {"train_sent": dataset_sent_train, "train_doc": dataset_doc, "test_sent": dataset_sent_test}
dataset_total = {"train_sent": dataset_sent_train, "test_sent": dataset_sent_test}
# transform sentence to word index
datasets = make_datasets(dataset_total, word2idx)
# dir_output the transformed files
print("making pickle file...")
utils.pkl_dump(datasets, dir_output + "/features.pkl")
utils.pkl_dump(word2idx, dir_output + "/word2idx.pkl")
utils.pkl_dump(general_embeddings, dir_output + "/general_embeddings.pkl")
utils.pkl_dump(domain_embeddings, dir_output + "/domain_embeddings.pkl")
utils.pkl_dump(idx2word, dir_output + "/idx2word.pkl")
def persist_experiment(t_experiment, i_episode, agent, scores):
os.chdir(t_experiment)
torch.save(agent.actor_local.state_dict(),
"checkpoint_actor_{}.pth".format(i_episode))
torch.save(agent.critic_local.state_dict(),
"checkpoint_critic_{}.pth".format(i_episode))
pkl_dump(scores, "scores_{}.pkl".format(i_episode))
os.chdir("..")
def get_results(test_ds, type):
dl = DataLoader(test_ds, batch_size=len(test_ds), num_workers=5)
retdict = train.eval_model(model, dl, args.OP_tgt, 3)
model_dir = os.path.join('models', args.modelname)
try:
os.makedirs(os.path.join(model_dir, f'{type}_eval_results'))
except:
shutil.rmtree(os.path.join(model_dir, f'{type}_eval_results'))
os.makedirs(os.path.join(model_dir, f'{type}_eval_results'))
retdict['modelname'] = args.modelname
# Print
pprint.pprint(retdict)
utils.pkl_dump(
retdict, os.path.join('models', args.modelname, f'{type}_report.dict'))
def df_to_training_pairs(self, seq_dir, auto_id, grp):
def decompose_slice(indices):
input_df = grp.loc[indices[0]:
indices[1]] # Slice out an input sequence
first_month = input_df.MONTH.iloc[0]
seq_name = f'{auto_id}_{first_month}'
# input_df = pd.DataFrame(self.scaler.transform(input_df), columns=input_df.columns)
x = input_df[self.col_dict['input']]
x_obs = x.shift(1).fillna(x.iloc[0]).values
x = x.values
m = input_df[self.col_dict['missing']].values
t = input_df[self.col_dict['delta']].values
return seq_name, np.array([x, m, t, x_obs])
input_start_ix = grp.index[0]
input_final_ix = grp.index[-1] - self.min_rows_reqd + 1
input_seq_df = [(pos, pos + self.history_len - 1)
for pos in range(input_start_ix, input_final_ix + 1)
] # pandas slicing is inclusive
target_start_ix = grp.index[0] + self.history_len
target_final_ix = grp.index[-1] - self.future_len + 1
target_seq_df = [(pos, pos + self.future_len - 1)
for pos in range(target_start_ix, target_final_ix + 1)
]
tgt_types = ['tgt1', 'tgt2']
label_dict = {}
target_loader = TargetFunc(grp)
# sliding window
for slice_ix in range(len(input_seq_df)):
name, x = decompose_slice(input_seq_df[slice_ix])
utils.pkl_dump(x, os.path.join(seq_dir, name + '.npy'))
label_dict[name] = {}
for tgt in tgt_types:
fn = eval(f"target_loader.{tgt}")
label_dict[name][tgt] = fn(target_seq_df[slice_ix])
return label_dict # return for collation with other patients
def train_model(train_iter, valid_iter, X_Mean, tgt_col, aux_cols, epochs, modelname, nb_classes, \
lr=0.001, aux_alpha=0, tr_alpha=0, class_weights=None, l2=None, model=None, print_every=100):
"""
Train a GRUD model
:param train_iter: Train DataLoader
:param valid_iter: Valid DataLoader
:param X_Mean: Empirical Mean values for each dimension in the input (only important for variables with missing data)
:param tgt_col: (str) Name of OP target
:param aux_cols: list(str) of names of Aux targets.
:param epochs: Int of epochs to run
:param modelname: Unique name for this model
:param nb_classes: Number of OP target classes
:param aux_alpha: Weight for Aux Loss
:param tr_alpha: Weight for TR Loss
:param class_weights (optional): Weights to scale OP Loss (for skewed datasets)
"""
device = utils.try_gpu()
# Set directory for model outputs
try:
os.makedirs(os.path.join('models',modelname))
os.makedirs(os.path.join('models',modelname, 'gradflow'))
except FileExistsError: pass
# Initialize plotter class for gradflow
plotter = TrainPlot(modelname)
# Initialize model and learners
class_weights = class_weights or [1]*nb_classes
l2 = l2 or 0
for X,y in train_iter: break
input_dim = X.size(-1)
aux_dim = [ (y[aux_c].size(-1) if len(y[aux_c].size())>1 else 1) for aux_c in aux_cols] # if-else for targets with single dimennsion. their size(-1) will be batchsize
model = StackedGRUDClassifier(input_dim, nb_classes, X_Mean, aux_dim).to(device=device, dtype=torch.float)
criterion = nn.CrossEntropyLoss(weight=torch.Tensor(class_weights).to(device=device))
optimizer = torch.optim.Adam(model.parameters(), lr=lr, weight_decay=l2)
scheduler = torch.optim.lr_scheduler.StepLR(optimizer, 30, 0.85)
# Store training metrics
train_meta = {}
train_meta['train_losses'] = []
train_meta['valid_losses'] = []
train_meta['min_valid_loss'] = sys.maxsize
train_meta['epoch_results'] = []
for epoch in range(epochs):
# TRAIN EPOCH
t0 = time.time()
metrics = train_epoch(model, train_iter, tgt_col, aux_cols, criterion, optimizer, aux_alpha, tr_alpha, scheduler, print_every=print_every, plotter=plotter)
if epoch<200:scheduler.step()
print(f"Epoch trained in {time.time()-t0}")
# EVALUATE AGAINST VALIDATION SET
t0 = time.time()
eval_scores = eval_model(model, valid_iter, tgt_col, nb_classes)
train_meta['epoch_results'].append(eval_scores)
print(f"Evaluation done in {time.time()-t0}")
t0 = time.time()
# SAVE CHECKPOINT
if eval_scores['loss'] < train_meta['min_valid_loss'] or epoch % 20 == 0:
train_meta['min_valid_loss'] = min(eval_scores['loss'], train_meta['min_valid_loss'])
checkpoint = {
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'optimizer': optimizer.state_dict()
}
save_ckp(checkpoint, True, os.path.join('models', modelname, 'checkpoint.pt'), os.path.join('models', modelname, 'best_model.pt'))
print(f"Checkpoint created")
# LOG PROGRESS
print("\n\n================================================================================================================\n")
print(f"Epoch: {epoch+1} TrainLoss: {metrics[1]/metrics[0]} ValidLoss: {eval_scores['loss']} ValidAcc:{eval_scores['accuracy']} WallTime: {datetime.datetime.now()}\n")
print(eval_scores['conf_matrix'])
print(pd.DataFrame.from_dict(eval_scores['clf_report']))
print(eval_scores['brier'])
print(eval_scores['roc'])
print("\n\n================================================================================================================\n")
# SAVE TRAINING PROGRESS DATA
train_meta['train_losses'].append(metrics[1]/metrics[0])
train_meta['valid_losses'].append(eval_scores['loss'])
utils.pkl_dump(train_meta, os.path.join('models', modelname, 'trainmeta.dict'))
print(f"post eval dumping took {time.time()-t0}")
# PLOT LOSSES
t0 = time.time()
plt.figure(1)
plt.plot(train_meta['train_losses'])
plt.plot(train_meta['valid_losses'])
plt.xlabel("Minibatch")
plt.ylabel("Loss")
plt.savefig(os.path.join('models', modelname, modelname+'_lossPlot.png'), bbox_inches='tight')
print(f"plotting took {time.time()-t0}")
return model
def train_test_split(self, dataset, x_mean, datadir, train_size):
"""
IBDSequenceGenerator.write_to_disk:
- Reads padded X, M, T matrices
- Generate overlapping sequences and {tgt_type} training labels for each patient
- Persist sequences as numpy arrays at data/{datadir}/{train, valid, test}/{ID}_{START_TS}.npy
- Persist labels for all sequences in data/{datadir}/{train, valid, test}/label_dict.pkl
- Persist empirical mean values at data/{datadir}/x_mean.pkl
"""
print(SequenceGenerator.write_to_disk.__doc__)
# Create out directories
if not datadir: datadir = f'm{self.history_len}p{self.future_len}'
out_fp = os.path.join('data', datadir)
train_seq_dir = os.path.join(out_fp, 'train')
valid_seq_dir = os.path.join(out_fp, 'valid')
test_seq_dir = os.path.join(out_fp, 'test')
if not os.path.exists(out_fp): os.makedirs(out_fp)
if not os.path.exists(train_seq_dir): os.makedirs(train_seq_dir)
if not os.path.exists(valid_seq_dir): os.makedirs(valid_seq_dir)
if not os.path.exists(test_seq_dir): os.makedirs(test_seq_dir)
groups = dataset.groupby('ID')
TRAIN_LABELS = {}
VALID_LABELS = {}
TEST_LABELS = {}
valid_size = (1 + train_size) / 2
patients_dropped = 0
for auto_id, grp in groups:
grplen = grp.shape[0]
if grplen < self.min_rows_reqd:
patients_dropped += 1
continue
train = grp.iloc[:int(train_size * grplen)]
valid = grp.iloc[int(train_size * grplen):int(valid_size * grplen)]
test = grp.iloc[int(valid_size * grplen):]
# if not enough observations for a validation sequence, use them either as test or train (don't waste any data)
if valid.shape[0] < self.min_rows_reqd:
if valid.shape[0] + test.shape[0] >= self.min_rows_reqd:
test = pd.concat([valid, test])
valid = None
else:
train = pd.concat([train, valid, test])
valid = None
test = None
TRAIN_LABELS.update(
self.df_to_training_pairs(train_seq_dir, auto_id, train))
if valid is not None:
VALID_LABELS.update(
self.df_to_training_pairs(valid_seq_dir, auto_id, valid))
if test is not None:
TEST_LABELS.update(
self.df_to_training_pairs(test_seq_dir, auto_id, test))
utils.pkl_dump(TRAIN_LABELS,
os.path.join(train_seq_dir, 'label_dict.pkl'))
utils.pkl_dump(VALID_LABELS,
os.path.join(valid_seq_dir, 'label_dict.pkl'))
utils.pkl_dump(TEST_LABELS, os.path.join(test_seq_dir,
'label_dict.pkl'))
utils.pkl_dump(x_mean[2:], os.path.join(
out_fp, 'x_mean.pkl')) # remove auto id and month ts
print(f"Patients dropped: {patients_dropped}")
all_data = pd.concat([df, m, t], axis=1)
col_dict = {'input': df.columns, 'missing': m.columns, 'delta': t.columns}
return all_data, col_dict, x_mean
if __name__ == "__main__":
# Generate dummy data
arr = pd.DataFrame(np.random.rand(5, 5, 5).reshape(25, 5),
columns=[f'col{i+1}' for i in range(5)])
arr['ID'] = sum([[i] * 5 for i in range(3, 8)], [])
arr['YEAR'] = 2020
arr['MONTH'] = [1, 4, 5, 7, 8] * 5
# Generate cleaned, resampled, imputed data with missing mask and delta + save to pickles folder
df, col_dict, x_mean = clean_longitudinal_inputs(arr)
# Init SequenceGenerator with look-back and look-forward durations
seqgen = SequenceGenerator(3, 1, col_dict)
# Generate time series inputs with their target labels
# X (input arrays) pickled to disk at /tmp/input_id.npy
# y returned as a dict {input_id : {tgt1: 0, tgt2: 1}}
c = 0
target_dict = {}
for auto_id, group in df.groupby('ID'):
target_dict.update(seqgen.df_to_training_pairs('tmp', auto_id, group))
utils.pkl_dump(target_dict, 'tmp/label_dict.pkl')
def ddpg_multiagent(env, agent1, agent2, cfg, db):
# Get configuration
n_episodes = cfg["Training"]["Number_episodes"]
max_t = cfg["Training"]["Max_timesteps"]
print_every = cfg["Training"]["Score_window"]
starting_random = cfg["Training"]["Starting_random"]
brain_index = cfg["Agent"]["Brain_index"]
persist_mongodb = cfg["Training"]["Persist_mongodb"]
success = cfg["Environment"]["Success"]
#Initialize score lists
scores_deque = deque(maxlen=print_every)
scores = []
# Create a directory to save the findings.
# experiment_dir = setup_experiment(cfg)
if persist_mongodb:
experiment_id = setup_experiment(db, cfg)
brain_name = env.brain_names[brain_index]
# Train for n_episodes
for i_episode in range(1, n_episodes + 1):
env_info = env.reset(train_mode=True)[brain_name]
n_agents = len(env_info.agents)
states = env_info.vector_observations
agent1.reset()
agent2.reset()
actions = np.zeros((n_agents, agent1.action_size))
score = np.zeros(n_agents)
for t in range(max_t):
if i_episode < starting_random:
actions = 2 * np.random.randn(n_agents,
agent1.action_size) - 1.0
else:
actions[0, :] = agent1.act(states[0, :])
actions[1, :] = agent2.act(states[1, :])
# for i_agent in range(n_agents):
# actions[i_agent, :] = agent.act(states[i_agent, :])
next_states, rewards, dones, _ = step_unity(
env, actions, brain_name)
for i_agent in range(n_agents):
# Add experience to both of the agent's replay buffers
agent1.step(states[i_agent, :], actions[i_agent, :],
rewards[i_agent], next_states[i_agent, :],
dones[i_agent])
agent2.step(states[i_agent, :], actions[i_agent, :],
rewards[i_agent], next_states[i_agent, :],
dones[i_agent])
states = next_states
score += rewards
if np.any(dones):
break
scores_deque.append(score)
scores.append(score)
mean_score = np.vstack(scores_deque).mean(axis=0).max()
print("\rEpisode {}\tAverage Score: {:.4f}\tNoise Modulation: {:.3f}".
format(i_episode, mean_score, agent1.noise_modulation),
end="")
# print("\rEpisode {}\tAverage Score: {}".format(i_episode, scores_deque), end="")
visualize = False
if i_episode % print_every == 0:
# persist_experiment(experiment_dir, i_episode, agent, scores)
if persist_mongodb:
persist_experiment(db, experiment_id, i_episode, agent1,
scores, print_every)
persist_experiment(db, experiment_id, i_episode, agent2,
scores, print_every)
else:
torch.save(agent1.actor_local.state_dict(),
f"checkpoint_actor_1_{i_episode}.pth")
torch.save(agent1.critic_local.state_dict(),
f"checkpoint_critic_1_{i_episode}.pth")
torch.save(agent2.actor_local.state_dict(),
f"checkpoint_actor_2_{i_episode}.pth")
torch.save(agent2.critic_local.state_dict(),
f"checkpoint_critic_2_{i_episode}.pth")
print("\rEpisode {}\tAverage Score: {:.4f}".format(
i_episode, mean_score))
if mean_score >= success:
# This is going to be the first thing I'd be digging in the database for,
# so here you go.
fpath_actor_1 = "checkpoint_actor_1_winner_{}.pth".format(
i_episode)
fpath_critic_1 = "checkpoint_critic_1_winner_{}.pth".format(
i_episode)
torch.save(agent1.actor_local.state_dict(), fpath_actor_1)
torch.save(agent1.critic_local.state_dict(), fpath_critic_1)
fpath_actor_2 = "checkpoint_actor_2_winner_{}.pth".format(
i_episode)
fpath_critic_2 = "checkpoint_critic_2_winner_{}.pth".format(
i_episode)
torch.save(agent2.actor_local.state_dict(), fpath_actor_2)
torch.save(agent2.critic_local.state_dict(), fpath_critic_2)
pkl_dump(scores, "scores_winner.pkl")
break
return scores
def ddpg_selfplay(env, agent, cfg, db):
# Get configuration
n_episodes = cfg["Training"]["Number_episodes"]
max_t = cfg["Training"]["Max_timesteps"]
print_every = cfg["Training"]["Score_window"]
starting_random = cfg["Training"]["Starting_random"]
brain_index = cfg["Agent"]["Brain_index"]
dump_agent = cfg["Training"]["Dump_agent"]
success = cfg["Environment"]["Success"]
persist_mongodb = cfg["Training"]["Persist_mongodb"]
pretrained = cfg["Training"]["Pretrained"]
agent.update_every = max_t * starting_random * 2
#Initialize score lists
scores_deque = deque(maxlen=print_every)
scores = []
# Create a directory to save the findings.
# experiment_dir = setup_experiment(cfg)
experiment_id = setup_experiment(db, cfg)
print("Experiment ID: {}".format(experiment_id))
brain_name = env.brain_names[brain_index]
# Train for n_episodes
for i_episode in range(1, n_episodes + 1):
env_info = env.reset(train_mode=True)[brain_name]
n_agents = len(env_info.agents)
states = env_info.vector_observations
agent.reset()
actions = np.zeros((n_agents, agent.action_size))
score = np.zeros(n_agents)
if i_episode == starting_random and pretrained:
print("Loading pre-trained weights")
agent = load_agent_weights(agent, cfg)
print("Pre-trained weights loaded!")
agent.update_every = cfg["Agent"]["Update_every"]
for t in range(max_t):
if i_episode < starting_random:
actions = 2 * np.random.rand(n_agents, agent.action_size) - 1.0
else:
for i_agent in range(n_agents):
actions[i_agent, :] = agent.act(states[i_agent, :])
next_states, rewards, dones, _ = step_unity(
env, actions, brain_name)
for i_agent in range(n_agents):
# Add experience to both of the agent's replay buffers
agent.step(states[i_agent, :], actions[i_agent, :],
rewards[i_agent], next_states[i_agent, :],
dones[i_agent])
states = next_states
score += rewards
if np.any(dones):
break
scores_deque.append(score)
scores.append(score)
mean_score = np.vstack(scores_deque).max(axis=1).mean()
print(
"\rEpisode {}\tAverage Score: {:.4f}\t Score: {:.3f} {:.3f} noise {:.2f}"
.format(i_episode, mean_score, score[0], score[1],
agent.noise_modulation),
end="")
# print("\rEpisode {}\tAverage Score: {}".format(i_episode, scores_deque), end="")
visualize = False
if i_episode % print_every == 0:
print("\rEpisode {}\tAverage Score: {:.4f}".format(
i_episode, mean_score))
if persist_mongodb:
persist_experiment(db, experiment_id, i_episode, agent, scores,
print_every)
else:
torch.save(agent.actor_local.state_dict(),
f"checkpoint_actor_{i_episode}.pth")
torch.save(agent.critic_local.state_dict(),
f"checkpoint_critic_{i_episode}.pth")
if i_episode % dump_agent == 0:
# To heck with it let's save some to disk even though I have utilities to load.
# It's important.
fpath_actor = "checkpoint_actor_{}.pth".format(i_episode)
fpath_critic = "checkpoint_critic_{}.pth".format(i_episode)
torch.save(agent.actor_local.state_dict(), fpath_actor)
torch.save(agent.critic_local.state_dict(), fpath_critic)
if mean_score >= success:
# This is going to be the first thing I'd be digging in the database for,
# so here you go.
fpath_actor = "checkpoint_actor_winner_{}.pth".format(i_episode)
fpath_critic = "checkpoint_critic_winner_{}.pth".format(i_episode)
torch.save(agent.actor_local.state_dict(), fpath_actor)
torch.save(agent.critic_local.state_dict(), fpath_critic)
pkl_dump(scores, "scores_winner.pkl")
break
return scores
