def test_early_stop_histories_are_unequal(self):
""" Should return False, because both EarlyStopHistories contain unequal indices. """
net = Linear(1, 1)
history1, history2 = EarlyStopHistory(), EarlyStopHistory()
history1.setup(1)
history2.setup(1)
history1.state_dicts[0], history2.state_dicts[0] = net.state_dict(
), net.state_dict()
history1.indices[0], history2.indices[0] = 3, 42
self.assertIs(EarlyStopHistory.__eq__(history1, history2), False)
def train(data_loader):
net = Linear(3, 1)
lr = 1e-3
optimizer = AdamW(net.parameters(), lr)
scheduler = CosineAnnealingWarmRestarts(optimizer, T_0=1, T_mult=2)
max_epochs = 40
for ep_id in range(max_epochs):
net.train()
for b_id, batch in enumerate(data_loader):
optimizer.zero_grad()
output = net(batch['data'].to(data_loader))
CEloss = CrossEntropyLoss()
loss = CEloss(output, batch['ground_truth'])
writer = SummaryWriter(
log_dir=
'C:\Users\andre\OneDrive\Рабочий стол\train\checkpoints',
comment="Batch loss")
writer.add_graph(loss)
loss.backward()
optimiser.step()
scheduler.step()
torch.save(
{
'epoch': ep_id,
'model_state_dict': net.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'loss': loss,
}, '/checkpoints/checkpoints.txt')
def test_pretrainer(tmpdir, device):
import torch
from torch.nn import Linear
# save a model in tmpdir/original/model.ckpt
first_model = Linear(32, 32).to(device)
pretrained_dir = tmpdir / "original"
pretrained_dir.mkdir()
with open(pretrained_dir / "model.ckpt", "wb") as fo:
torch.save(first_model.state_dict(), fo)
# Make a new model and Pretrainer
pretrained_model = Linear(32, 32).to(device)
assert not torch.all(torch.eq(pretrained_model.weight, first_model.weight))
from speechbrain.utils.parameter_transfer import Pretrainer
pt = Pretrainer(collect_in=tmpdir / "reused",
loadables={"model": pretrained_model})
pt.collect_files(default_source=pretrained_dir)
pt.load_collected()
assert torch.all(torch.eq(pretrained_model.weight, first_model.weight))
def from_module(cls, src: nn.Linear) -> 'Linear':
with_bias = src.bias is not None
new_lin = Linear(src.in_features, src.out_features, with_bias)
new_lin.load_state_dict(src.state_dict())
return new_lin
maint_logits = torch.stack([torch.sum(maint_logits_daily[slice]) \
for slice in maint_hist_slices])
logits = mission_logits + maint_logits
bce = loss_fun(logits, y)
bce = loss_fun(mission_logits, y)
train_losses.append(bce.cpu().detach().numpy())
n_seen += 1
time_elapsed += dt.now() - start_time
update_str = 'Elapsed Time: {0}, Batches Seen: {1}/{2}, Avg. Loss: {3:.4f}'
elapsed_time_str = ''.join(str(time_elapsed).split('.')[0])
format_args = (elapsed_time_str, n_seen, len(train_loader),
np.mean(train_losses[-5:]))
print(update_str.format(*format_args))
if not n_seen % val_every:
val_acc = evaluate(mission_model, maint_model, val_loader, device)
val_accs.append(val_acc)
if not n_seen % save_every:
ckpt = dict(mission_state_dict=mission_model.state_dict(),
maint_state_dict=maint_model.state_dict(),
train_losses=train_losses,
val_accs=val_accs,
time_elapsed=time_elapsed)
torch.save(ckpt, save_path)
print('\nTraining complete!\n')
"""
A Simple Linear model using Linear module in pytorch
"""
import torch
from torch.nn import Linear
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
# set random seed
torch.manual_seed(1)
lr = Linear(in_features=2, out_features=1, bias=True)
print("Parameter w and b:", list(lr.parameters()))
# A method state_dict() Returns a Python dictionary
# object corresponding to the layers of each parameter tensor
print("Layers of each parameter:", lr.state_dict())
# make prediction now for the linear model
x = torch.tensor([[1.0, 2.0], [2.0, 2.0]])
yhat = lr(x)
print("The prediction:", yhat)
print(x.data)
class LR(nn.Module): #LR is defined as a child class of nn.Module
def __init__(self, in_size, out_size): #Constructor
#Inheriting methods from parent class nn.Module
super(LR, self).__init__()
self.linear = nn.Linear(in_size, out_size)
def forward(self, x): #Prediction funciton
out = self.linear(x)
return out
model = LR(1, 1)
#We use state_dict to initialize the weight and bias of the model.
#State_dict is a python dictionary. It can be used to map the relationship of linear layers to their parameters.
model.state_dict()['linear.weight'].data[0] = torch.tensor(
[0.5153]) #Initializing weight
model.state_dict()['linear.bias'].data[0] = torch.tensor(
[-0.4414]) #initializing bias
print("The Parameters: ",
list(model.parameters())) #Lets see the initialized weight and bias
#The above function works because the class LR, is a subclass of Modules.
#Lets use the custom model to make predictions
x = torch.tensor([[1.0], [2.0]]) #2D tensor
print("The shape of x: ", x.shape)
yhat = model(x)
print(yhat)
x = torch.tensor([[1.0], [2.0], [3.0]])
print("The shape of x: ", x.shape)
def forward(x):
yhat = w * x + b
return yhat
x = torch.tensor([[1.0], [2.0], [3.0]])
yhat = forward(x)
print("The Prediction: ", yhat)
print("Y size", yhat.shape)
torch.manual_seed(1)
lr = Linear(in_features=1, out_features=1, bias=True)
print("Parameters w and b: ", list(lr.parameters()))
print("Python Dictionary", lr.state_dict())
print("keys:", lr.state_dict().keys())
print("values:", lr.state_dict().values())
print("weight:", lr.weight)
print("bias:", lr.bias)
x = torch.tensor([[1.0]])
yhat = lr(x)
print("The prediction: ", yhat)
x = torch.tensor([[1.0], [2.0]])
yhat = lr(x)
print("The prediction: ", yhat)
x = torch.tensor([[1.0], [2.0], [3.0]])
def main():
cmd_args = add_argument()
path_to_file_tr = cmd_args.path_to_file_tr
path_to_file_ts = cmd_args.path_to_file_ts
min_len_mol = cmd_args.min_len_mol
max_len_mol = cmd_args.max_len_mol
num_examples_tr = cmd_args.num_examples_tr
num_examples_ts = cmd_args.num_examples_ts
train_batch_size = json.load(open(cmd_args.ds_conf))['train_batch_size']
gradient_accumulation_steps = json.load(open(
cmd_args.ds_conf))['gradient_accumulation_steps']
deepspeed_optimizer = True if json.load(open(cmd_args.ds_conf)).get(
'optimizer', None) is not None else False
epochs = cmd_args.epochs
emb_dim = cmd_args.emb_dim
dim = cmd_args.dim
bucket_size = cmd_args.bucket_size
depth = cmd_args.depth
heads = cmd_args.heads
n_hashes = cmd_args.n_hashes
ff_chunks = cmd_args.ff_chunks
attn_chunks = cmd_args.attn_chunks
validate_every = cmd_args.validate_every
save_every = cmd_args.save_every
output_folder = cmd_args.output_folder
use_full_attn = cmd_args.use_full_attn
mrpc_test = cmd_args.mrpc_test
use_deepspeed = cmd_args.use_deepspeed
os.makedirs(output_folder, exist_ok=True)
pickle.dump(cmd_args,
open(os.sep.join([output_folder, 'training_conf.pkl']), 'wb'))
MIN_LENGTH_MOL = min_len_mol
MAX_LENGTH_MOL = max_len_mol # 2048
NUM_EXAMPLES_TR = num_examples_tr # 1024
NUM_EXAMPLES_TS = num_examples_ts # 1024
N_EPOCHS = epochs # 10
VALIDATE_EVERY = validate_every
SAVE_EVERY = save_every
MOL_SEQ_LEN = MAX_LENGTH_MOL # output_lang.max_len if (output_lang.max_len % 2) == 0 else output_lang.max_len + 1 # ??
saved_mol_lang = os.sep.join([output_folder, 'mol_lang.pkl'])
MAX_LENGTH_MOL = cmd_args.max_len_mol
saved_target_lang = os.sep.join([output_folder, 'mol_lang.pkl'])
if mrpc_test:
mol_lang, tr_samples, ts_samples = readMRPC(
molecule_file_tr=path_to_file_tr,
molecule_file_ts=path_to_file_ts,
saved_molecule_lang=saved_target_lang,
num_examples_tr=NUM_EXAMPLES_TR,
num_examples_ts=NUM_EXAMPLES_TS,
min_len_molecule=MIN_LENGTH_MOL,
max_len_molecule=MAX_LENGTH_MOL,
shuffle=True)
else:
mol_lang, tr_samples, ts_samples = readMolecules(
molecule_file_tr=path_to_file_tr,
molecule_file_ts=path_to_file_ts,
saved_molecule_lang=saved_target_lang,
num_examples_tr=NUM_EXAMPLES_TR,
num_examples_ts=NUM_EXAMPLES_TS,
min_len_molecule=MIN_LENGTH_MOL,
max_len_molecule=MAX_LENGTH_MOL,
shuffle=True)
pickle.dump(mol_lang, open(saved_mol_lang, 'wb'))
train_dataset = MolecularSimilarityDataset(
tr_samples, mol_lang, train_batch_size if device == 'cuda' else 1)
test_dataset = MolecularSimilarityDataset(
ts_samples, mol_lang, train_batch_size if device == 'cuda' else 1)
MAX_SEQ_LEN = MOL_SEQ_LEN * 2
print('Axial Embedding shape:', compute_axial_position_shape(MAX_SEQ_LEN))
model = ReformerLM(
num_tokens=mol_lang.n_words,
dim=dim,
bucket_size=bucket_size,
depth=depth,
heads=heads,
n_hashes=n_hashes,
max_seq_len=MAX_SEQ_LEN,
ff_chunks=ff_chunks,
attn_chunks=attn_chunks,
weight_tie=True,
weight_tie_embedding=True,
axial_position_emb=True,
axial_position_shape=compute_axial_position_shape(MAX_SEQ_LEN),
axial_position_dims=(dim // 2, dim // 2),
return_embeddings=True,
use_full_attn=use_full_attn).to(device)
linear_regressor = Linear(512, 2).to(device)
model = TrainingWrapper(model, ignore_index=PAD_IDX,
pad_value=PAD_IDX).to(device)
model_params = filter(lambda p: p.requires_grad, model.parameters())
linear_params = filter(lambda p: p.requires_grad,
linear_regressor.parameters())
SAVE_DIR = os.sep.join([output_folder, 'saved_model'])
os.makedirs(SAVE_DIR, exist_ok=True)
try:
model_ckp_max = np.max(
[int(ckp) for ckp in os.listdir(os.sep.join([SAVE_DIR, 'model']))])
except:
model_ckp_max = 0
gpus_mini_batch = (train_batch_size // gradient_accumulation_steps
) // torch.cuda.device_count()
print('gpus_mini_batch:', gpus_mini_batch,
'with gradient_accumulation_steps:', gradient_accumulation_steps)
log_file = open(os.sep.join([output_folder, 'training_log.log']), 'a')
log_file.write(
"\n\n\n{}\tStarting new training from chekpoint: EncoderDecoder-{}\n".
format(datetime.datetime.now(), model_ckp_max))
log_file.flush()
if use_deepspeed:
if deepspeed_optimizer == False:
print('No DeepSpeed optimizer found. Using RangerLars.')
model_optimizer = RangerLars(model.parameters())
linear_optimizer = RangerLars(linear_regressor.parameters())
model_engine, model_optimizer, trainloader, _ = deepspeed.initialize(
args=cmd_args,
model=model,
optimizer=model_optimizer,
model_parameters=model_params,
training_data=train_dataset)
linear_engine, linear_optimizer, _, _ = deepspeed.initialize(
args=cmd_args,
model=linear_regressor,
optimizer=linear_optimizer,
model_parameters=linear_params)
else:
print('Found optimizer in the DeepSpeed configurations. Using it.')
model_engine, model_optimizer, trainloader, _ = deepspeed.initialize(
args=cmd_args,
model=model,
model_parameters=model_params,
training_data=train_dataset)
linear_engine, linear_optimizer, _, _ = deepspeed.initialize(
args=cmd_args,
model=linear_regressor,
model_parameters=linear_params)
_, model_client_sd = model_engine.load_checkpoint(
os.sep.join([SAVE_DIR, 'model']), model_ckp_max)
testloader = model_engine.deepspeed_io(test_dataset)
######TO DO
for eph in range(epochs):
print('Starting Epoch: {}'.format(eph))
for i, pair in enumerate(tqdm(trainloader)):
tr_step = ((eph * len(trainloader)) + i) + 1
src = pair[0]
trg = pair[1]
pickle.dump(src, open('src.pkl', 'wb'))
pickle.dump(trg, open('trg.pkl', 'wb'))
model_engine.train()
linear_engine.train()
#enc_dec.train()
src = src.to(model_engine.local_rank)
trg = trg.to(linear_engine.local_rank)
print("Sample:", src)
print("Target:", trg)
print("Target Shape:", trg.shape)
print("len Samples:", len(src))
## Need to learn how to use masks correctly
enc_input_mask = torch.tensor(
[[1 if idx != PAD_IDX else 0 for idx in smpl]
for smpl in src]).bool().to(model_engine.local_rank)
# context_mask = torch.tensor([[1 for idx in smpl if idx != PAD_IDX] for smpl in trg]).bool().to(device)
#################
enc_keys = model_engine(
src, return_loss=False, input_mask=enc_input_mask
) #enc_input_mask)#, context_mask=context_mask)
#loss = enc_dec(src, trg, return_loss = True, enc_input_mask = None)#enc_input_mask)#, context_mask=context_mask)
print('enc_keys shape', enc_keys.shape)
#enc_keys_cls = enc_keys[:,0:1,:].to(linear_engine.local_rank)#torch.tensor([s[0] for s in enc_keys]).to(linear_engine.local_rank)
#print('enc_keys_cls shape', enc_keys_cls.shape)
preds = torch.softmax(linear_engine(enc_keys),
dim=1).to(linear_engine.local_rank)
print('preds shape', preds.shape)
#preds = np.array([r[0] for r in results])
#print('Pred:', preds.shape)
loss = F.cross_entropy(preds, trg).to(linear_engine.local_rank)
loss.backward()
model_engine.step()
linear_engine.step()
print('Training Loss:', loss.item())
if tr_step % validate_every == 0:
val_loss = []
for pair in tqdm(
testloader
): #Can't use the testloader or I will mess up with the model assignment and it won't learn during training, need to use normal validation instead of parallel one
model_engine.eval()
linear_engine.eval()
with torch.no_grad():
ts_src = pair[0]
ts_trg = pair[1]
pickle.dump(ts_src, open('ts_src.pkl', 'wb'))
pickle.dump(ts_trg, open('ts_trg.pkl', 'wb'))
ts_src = ts_src.to(model_engine.local_rank)
ts_trg = ts_trg.to(linear_engine.local_rank)
#ts_src = torch.tensor(np.array([pair[0].numpy()])).to(device)
#ts_trg = torch.tensor(np.array([pair[1].numpy()])).to(device)
## Need to learn how to use masks correctly
ts_enc_input_mask = torch.tensor([
[1 if idx != PAD_IDX else 0 for idx in smpl]
for smpl in ts_src
]).bool().to(model_engine.local_rank)
#ts_context_mask = torch.tensor([[1 for idx in smpl if idx != PAD_IDX] for smpl in ts_trg]).bool().to(device)
# loss = model_engine(
# ts_src,
# ts_trg,
# return_loss=True,
# enc_input_mask=ts_enc_input_mask
# ) #ts_enc_input_mask)#, context_mask=ts_context_mask)
# #loss = enc_dec(ts_src, ts_trg, return_loss = True, enc_input_mask = None)
ts_enc_keys = model_engine(
ts_src,
return_loss=False,
input_mask=ts_enc_input_mask)
ts_pred = torch.softmax(
linear_engine(ts_enc_keys),
dim=1).to(linear_engine.local_rank)
loss = F.cross_entropy(ts_pred, ts_trg).to(
linear_engine.local_rank)
val_loss.append(loss.item())
print(
f'\tValidation Loss: AVG: {np.mean(val_loss)}, MEDIAN: {np.median(val_loss)}, STD: {np.std(val_loss)} '
)
log_file.write(
'Step: {}\tTraining Loss:{}\t Validation LOSS: AVG: {}| MEDIAN: {}| STD: {}\n'
.format(i, loss.item(), np.mean(val_loss),
np.median(val_loss), np.std(val_loss)))
else:
log_file.write('Step: {}\tTraining Loss:{}\n'.format(
i, loss.item()))
log_file.flush()
if tr_step % save_every == 0:
print('\tSaving Checkpoint')
model_ckpt_id = str(model_ckp_max + tr_step + 1)
model_engine.save_checkpoint(
os.sep.join([SAVE_DIR, 'model']), model_ckpt_id)
log_file.close()
print('\tSaving Final Checkpoint')
model_ckpt_id = str(model_ckp_max + tr_step + 1)
model_engine.save_checkpoint(os.sep.join([SAVE_DIR, 'model']),
model_ckpt_id)
else:
#model_optimizer = torch.optim.Adam(model.parameters()) # RangerLars(model.parameters())
#linear_optimizer = torch.optim.Adam(linear_regressor.parameters()) # RangerLars(linear_regressor.parameters())
model_optimizer = torch.optim.Adam(
list(model.parameters()) + list(linear_regressor.parameters())
) #RangerLars(list(model.parameters())+list(linear_regressor.parameters())) #
PATH = os.sep.join(
[SAVE_DIR, 'model',
str(model_ckp_max), 'sts_model.pt'])
if os.path.exists(PATH):
print('********** Found Checkpoint. Loading:', PATH)
checkpoint = torch.load(PATH)
model.load_state_dict(checkpoint['model_state_dict'])
linear_regressor.load_state_dict(checkpoint['linear_state_dict'])
model_optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
trainloader = DataLoader(train_dataset,
batch_size=train_batch_size,
shuffle=False)
testloader = DataLoader(test_dataset,
batch_size=train_batch_size,
shuffle=False)
######TO DO
train_loss_list = []
for eph in range(epochs):
print('Starting Epoch: {}'.format(eph))
for i, pair in enumerate(tqdm(trainloader)):
tr_step = ((eph * len(trainloader)) + i) + 1
src = pair[0]
trg = pair[1]
pickle.dump(src, open('src.pkl', 'wb'))
pickle.dump(trg, open('trg.pkl', 'wb'))
model.train()
linear_regressor.train()
#enc_dec.train()
src = src.to(device)
trg = trg.to(device)
#print("Sample:", src)
#print("Target:", trg)
#print("Target Shape:", trg.shape)
#print("len Samples:", len(src))
## Need to learn how to use masks correctly
enc_input_mask = torch.tensor(
[[1 if idx != PAD_IDX else 0 for idx in smpl]
for smpl in src]).bool().to(device)
# context_mask = torch.tensor([[1 for idx in smpl if idx != PAD_IDX] for smpl in trg]).bool().to(device)
#################
enc_keys = model(
src, return_loss=False, input_mask=enc_input_mask
) #enc_input_mask)#, context_mask=context_mask)
#loss = enc_dec(src, trg, return_loss = True, enc_input_mask = None)#enc_input_mask)#, context_mask=context_mask)
#print('enc_keys shape', enc_keys.shape)
enc_keys_cls = enc_keys[:, 0, :].to(
device
) #torch.tensor([s[0] for s in enc_keys]).to(linear_engine.local_rank)
#print('enc_keys_cls shape', enc_keys_cls.shape)
preds = torch.softmax(linear_regressor(enc_keys_cls),
dim=1).to(device)
#print('preds shape', preds.shape)
#preds = np.array([r[0] for r in results])
#print('Pred:', preds.shape)
loss = F.cross_entropy(preds, trg).to(device)
loss.backward()
model_optimizer.step()
#linear_optimizer.step()
train_loss_list.append(loss.item())
#print('Training Loss:', loss.item())
if tr_step % validate_every == 0:
val_loss = []
ACC_list = []
MCC_list = []
for pair in tqdm(
testloader
): #Can't use the testloader or I will mess up with the model assignment and it won't learn during training, need to use normal validation instead of parallel one
model.eval()
linear_regressor.eval()
with torch.no_grad():
ts_src = pair[0]
ts_trg = pair[1]
pickle.dump(ts_src, open('ts_src.pkl', 'wb'))
pickle.dump(ts_trg, open('ts_trg.pkl', 'wb'))
ts_src = ts_src.to(device)
ts_trg = ts_trg.to(device)
#ts_src = torch.tensor(np.array([pair[0].numpy()])).to(device)
#ts_trg = torch.tensor(np.array([pair[1].numpy()])).to(device)
## Need to learn how to use masks correctly
ts_enc_input_mask = torch.tensor(
[[1 if idx != PAD_IDX else 0 for idx in smpl]
for smpl in ts_src]).bool().to(device)
#ts_context_mask = torch.tensor([[1 for idx in smpl if idx != PAD_IDX] for smpl in ts_trg]).bool().to(device)
# loss = model_engine(
# ts_src,
# ts_trg,
# return_loss=True,
# enc_input_mask=ts_enc_input_mask
# ) #ts_enc_input_mask)#, context_mask=ts_context_mask)
# #loss = enc_dec(ts_src, ts_trg, return_loss = True, enc_input_mask = None)
ts_enc_keys = model(ts_src,
return_loss=False,
input_mask=ts_enc_input_mask)
ts_enc_keys_cls = ts_enc_keys[:, 0, :].to(device)
ts_pred = torch.softmax(
linear_regressor(ts_enc_keys_cls),
dim=1).to(device)
loss = F.cross_entropy(ts_pred, ts_trg).to(device)
ACC, MCC = compute_simple_metrics(ts_pred, ts_trg)
ACC_list.append(ACC)
MCC_list.append(MCC)
val_loss.append(loss.item())
print(
f'\Train Loss: LAST: {train_loss_list[-1]}, AVG: {np.mean(train_loss_list)}, MEDIAN: {np.median(train_loss_list)}, STD: {np.std(train_loss_list)} '
)
print(
f'\tValidation Loss: AVG: {np.mean(val_loss)}, MEDIAN: {np.median(val_loss)}, STD: {np.std(val_loss)} '
)
print(
f'\tValidation ACC: AVG: {np.mean(ACC_list)}, MEDIAN: {np.median(ACC_list)}, STD: {np.std(ACC_list)} '
)
print(
f'\tValidation MCC: AVG: {np.mean(MCC_list)}, MEDIAN: {np.median(MCC_list)}, STD: {np.std(MCC_list)} '
)
log_file.write(
'Step: {}\tTraining Loss:{}\t Validation LOSS: AVG: {}| MEDIAN: {}| STD: {}\n'
.format(i, loss.item(), np.mean(val_loss),
np.median(val_loss), np.std(val_loss)))
else:
log_file.write('Step: {}\tTraining Loss:{}\n'.format(
i, loss.item()))
log_file.flush()
if tr_step % save_every == 0:
print('\tSaving Checkpoint')
model_ckpt_id = str(model_ckp_max + tr_step + 1)
#model_engine.save_checkpoint(os.sep.join([SAVE_DIR, 'model']),
# model_ckpt_id)
PATH = os.sep.join([
SAVE_DIR, 'model',
str(model_ckpt_id), 'sts_model.pt'
])
os.makedirs(os.sep.join(PATH.split(os.sep)[:-1]),
exist_ok=True)
torch.save(
{
'step': tr_step,
'model_state_dict': model.state_dict(),
'linear_state_dict': linear_regressor.state_dict(),
'optimizer_state_dict':
model_optimizer.state_dict(),
}, PATH)
log_file.close()
print('\tSaving Final Checkpoint')
model_ckpt_id = str(model_ckp_max + tr_step + 1)
#model_engine.save_checkpoint(os.sep.join([SAVE_DIR, 'model']),
# model_ckpt_id)
PATH = os.sep.join(
[SAVE_DIR, 'model',
str(model_ckpt_id), 'sts_model.pt'])
os.makedirs(os.sep.join(PATH.split(os.sep)[:-1]), exist_ok=True)
torch.save(
{
'step': tr_step,
'model_state_dict': model.state_dict(),
'linear_state_dict': linear_regressor.state_dict(),
'optimizer_state_dict': model_optimizer.state_dict(),
}, PATH)
