2 Modeling
In this guide, we recreate the model from the paper, Hierarchical Attention Networks for Document Classification in PyTorch, and walk through how to create and structure data appropriately to feed a HAN model.
2.1 Preparing the Data
# prepare and load data
def prepare_df(pkl_location):
# read pkl as pandas
df = pd.read_pickle(pkl_location)
# just keep us/kabul labels
df = df.loc[(df['target'] == 'US') | (df['target'] == 'Kabul')]
# mask DV to recode
us = df['target'] == 'US'
kabul = df['target'] == 'Kabul'
# apply mask
df.loc[us, 'target'] = 1
df.loc[kabul, 'target'] = 0
# reset index
df = df.reset_index(drop=True)
return df
df = prepare_df('C:\\Users\\Andrew\\Desktop\\df.pkl')
# prepare data
def clean_df(df):
# strip dash but keep a space
df['body'] = df['body'].str.replace('-', ' ')
# lower case the data
df['body'] = df['body'].apply(lambda x: x.lower())
# remove excess spaces near punctuation
df['body'] = df['body'].apply(lambda x: re.sub(r'\s([?.!"](?:\s|$))', r'\1', x))
# generate a word count
df['word_count'] = df['body'].apply(lambda x: len(x.split()))
# remove excess white spaces
df['body'] = df['body'].apply(lambda x: " ".join(x.split()))
return df
df = clean_df(df)
# lets remove rare words
def remove_rare_words(df):
# get counts of each word -- necessary for vocab
counts = Counter(" ".join(df['body'].values.tolist()).split(" "))
# remove low counts -- keep those above 2
counts = {key: value for key, value in counts.items() if value > 2}
# remove rare words from corpus
def remove_rare(x):
return ' '.join(list(filter(lambda x: x in counts.keys(), x.split())))
# apply funx
df['body'] = df['body'].apply(remove_rare)
return df
df = remove_rare_words(df)
# whats the length of the vocab?
counts = Counter(" ".join(df['body'].values.tolist()).split(" "))
vocab = sorted(counts, key=counts.get, reverse=True)
print(len(vocab))## 235222.2 GLoVe Embeddings
Now we are ready to load GloVe embeddings. The code below:
- Loads GloVe
- It creates one nearly empty vector for the padding tokens.
- It creates a randomly initialized vector between \([-0.14, 0.14]\) (mimicking the variance of 200d GLoVe) to account for unknown tokens.
- It extends the GloVe dimensions from 200 to 201 to account for this variation.
# load GloVe embeddings
def load_GloVe(file_path):
embeddings_dictionary = dict()
glove_file = open(file_path, encoding="utf8")
for line in glove_file:
records = line.split()
word = records[0]
vector_dimensions = np.asarray(records[1:], dtype='float32')
embeddings_dictionary[word] = vector_dimensions
glove_file.close()
return embeddings_dictionary
# load GloVe
file_path = 'C:\\Users\\Andrew\\Desktop\\glove.6B.200d.txt'
embeddings_dictionary = load_GloVe(file_path)
# useful computing to check vector values and indices
embeddings_dictionary.get('.') # get key value## array([ 1.2289e-01, 5.8037e-01, -6.9635e-02, -5.0288e-01, 1.0503e-01,
## 3.9945e-01, -3.8635e-01, -8.4279e-02, 1.2219e-01, 8.0312e-02,
## 3.2337e-01, 4.7579e-01, -3.8375e-02, -7.0900e-03, 4.1524e-01,
## 3.2121e-01, -2.1185e-01, 3.6144e-01, -5.5623e-02, -3.0512e-02,
## 4.2854e-01, 2.8547e+00, -1.4623e-01, -1.7557e-01, 3.1197e-01,
## -1.3118e-01, 3.3298e-02, 1.3093e-01, 8.9889e-02, -1.2417e-01,
## 2.3396e-03, -6.8954e-02, -1.0754e-01, -1.1551e-01, -3.1052e-01,
## -1.2097e-01, -4.6691e-01, -8.3600e-02, -3.7664e-02, -7.1779e-02,
## -1.1899e-01, -2.0381e-01, -1.2424e-01, 4.6339e-01, -1.9828e-01,
## -8.0365e-03, 5.3718e-01, 3.1739e-02, 3.4331e-01, 7.9704e-03,
## 4.8744e-03, 3.0592e-02, -1.7615e-01, 8.2342e-01, -1.3793e-01,
## -1.0075e-01, -1.2686e-01, 7.4735e-02, -8.8719e-02, -4.2719e-02,
## 7.6624e-02, 8.9263e-02, 6.4445e-02, -3.1958e-02, 1.5254e-01,
## -1.0384e-01, 7.6604e-02, 3.4099e-01, 2.4331e-01, -1.0452e-01,
## 4.0714e-01, -1.8260e-01, -4.0667e-02, 5.0878e-01, 8.0760e-02,
## 2.2759e-01, -4.2162e-02, -1.8171e-01, -9.5025e-02, 3.0334e-02,
## 8.8202e-02, -3.9843e-06, -3.9877e-03, 1.5724e-01, 3.3167e-01,
## 8.4710e-02, -2.5919e-01, -4.1384e-01, 2.9920e-01, -5.4255e-01,
## 3.2129e-02, 1.0030e-01, 4.4202e-01, 4.4682e-02, -9.0681e-02,
## -1.0481e-01, -1.1860e-01, -3.1972e-01, -2.0790e-01, -4.0203e-02,
## -2.2988e-02, 2.2824e-01, 5.5238e-03, 1.2568e-01, -1.4640e-01,
## -1.4904e-01, -1.1561e-01, 1.0517e+00, -1.9498e-01, 8.3958e-02,
## 4.4812e-02, -1.2965e-01, -9.3468e-02, 2.1237e-01, -8.8332e-02,
## -1.8680e-01, 2.6521e-01, 1.3097e-01, -4.8102e-02, -2.2467e-01,
## 2.8412e-01, 3.4907e-01, 3.4833e-01, 1.7877e-02, 3.0504e-01,
## -8.3453e-01, 4.8856e-02, -1.9330e-01, 2.0764e-01, -4.9701e-01,
## -1.8747e-01, -7.6801e-02, 1.5558e-01, -4.6844e-01, 4.0944e-01,
## 2.1386e-01, 8.2392e-02, -2.6491e-01, -2.1224e-01, -1.3293e-01,
## 1.4738e-01, -1.4192e-01, 1.8994e-01, -1.5587e-01, 1.0738e+00,
## 4.0789e-01, -2.7452e-01, -1.8431e-01, 6.8679e-04, -8.7115e-02,
## 1.9672e-01, 4.0918e-01, -3.5462e-01, -6.3260e-02, 4.4920e-01,
## -6.0568e-02, -4.1636e-02, 2.0531e-01, 1.7025e-02, -5.8448e-01,
## 7.5441e-02, 8.2116e-02, -4.6008e-01, 1.2393e-02, -2.5310e-02,
## 1.4177e-01, -9.2192e-02, 3.4505e-01, -5.2136e-01, 5.7304e-01,
## 1.1973e-02, 3.3196e-02, 2.9672e-01, -2.7899e-01, 1.9979e-01,
## 2.5666e-01, 8.2079e-02, -7.8436e-02, 9.3719e-02, 2.4202e-01,
## 1.3495e+00, -3.0434e-01, -3.0936e-01, 4.2047e-01, -7.9068e-02,
## -1.4819e-01, -8.9404e-02, 6.6800e-02, 2.2405e-01, 2.7226e-01,
## -3.5236e-02, 1.7688e-01, -5.3600e-02, 7.0031e-03, -3.3006e-02,
## -8.0021e-02, -2.4451e-01, -3.9174e-02, -1.6236e-01, -9.6652e-02],
## dtype=float32)list(embeddings_dictionary.keys()).index('.') # get key index
# create vectors for "unknown" and "padding" and add to GloVe## 2def modify_GloVe(embeddings_dictionary):
# create key values for unknown
unknown_vector = np.random.uniform(-0.14, 0.14, 201) # var of GloVe 200d
# create key values for padding
pad_vector = np.repeat(0, 200)
pad_vector = np.append(pad_vector, 1)
# turn dict into list to append easily
embeddings_tensor = list(embeddings_dictionary.values())
# extend GloVe dimension by 2
embeddings_tensor = [np.append(i, 0) for i in embeddings_tensor]
# add unknown and pad vectors via vstack
embeddings_tensor = np.vstack([embeddings_tensor, unknown_vector])
embeddings_tensor = np.vstack([embeddings_tensor, pad_vector])
# finalize transform into tensor
embeddings_tensor = torch.Tensor(embeddings_tensor)
return embeddings_tensor
# modify GloVe and turn into torch tensor
embeddings_tensor = modify_GloVe(embeddings_dictionary)
# check shape
print(embeddings_tensor.shape)## torch.Size([400002, 201])2.3 Prepare to Create 3D Data
# prepare data for 3 dimensional transformation
# create containers for storage
text_to_sentences = []
full_text = []
# for each document
for idx in range(df.shape[0]):
# retrieve the plain text
text = df['body'].iloc[idx]
# append it to the full_text container
full_text.append(text)
# use the nltk tokenizer to split document into sentences
sentences = nltk.tokenize.sent_tokenize(text)
# append them into the container
text_to_sentences.append(sentences)
# calculate the average number of sentences and words per sentence
n_sent = 0
n_words = 0
for i in range(df.shape[0]):
sentence = nltk.tokenize.sent_tokenize(df.loc[i, 'body'])
for word in sentence:
n_words += len(nltk.tokenize.word_tokenize(word))
n_sent += len(sentence)
print("Average number of words in each sentence: ", round(n_words / n_sent))## Average number of words in each sentence: 29## Average number of sentences in each document: 102.4 Create 3D Data Set
# since its the average, extend them slightly
MAX_SENT_LENGTH = 35
MAX_SENTS = 15
# instantiate a GloVe word map so we can tokenize based on GloVe IDs
word_map = dict(zip(embeddings_dictionary.keys(), range(len(embeddings_dictionary))))
# instantiate empty 3d data set
data_set_3d = np.zeros((len(full_text), MAX_SENTS, MAX_SENT_LENGTH), dtype='int32')
data_set_3d.shape
# for every document## (10045, 15, 35)for i, sentences in enumerate(text_to_sentences):
# for every split document into sentences
for j, sent in enumerate(sentences):
# if the step is less than max sentences
if j < MAX_SENTS:
# get the number of words in the sentence
wordTokens = text_to_word_sequence(sent)
# set k = 0
k = 0
# set container
words_in_sent = []
# for each word in the sentence
for _, word in enumerate(wordTokens):
# if k is less than max sentences
if k < MAX_SENT_LENGTH:
# if we can find the word in GloVe
if word in word_map.keys():
# append word token to position document_i, sentence_j, position_k in sentence
data_set_3d[i, j, k] = word_map.get(word)
# add the word to list of words found in the sentence
words_in_sent.append(word)
else:
# other wise, add unknown token
data_set_3d[i, j, k] = 400000
# add 1 to k
# keep looping until max sentence length is achieved
k = k + 1
# change np.zeros to the unknown token, 400001
# otherwise it will calculate a bunch of 'the' vectors
data_set_3d[data_set_3d == 0] = 400001
# check data
data_set_3d[0][0]## array([ 19, 9945, 249, 3, 400001, 547, 118, 400001,
## 959, 16119, 5, 400001, 104, 112, 33, 51,
## 1315, 53, 303, 4, 7255, 400001, 284, 226,
## 400001, 3485, 3, 400, 21, 400001, 104, 112,
## 5, 47, 205])2.5 Create Torch Tensor Data Sets
With the corpus tokenized, we now proceed to prepare our data for analysis in PyTorch. The code below creates a TensorDataset comprised of our features and our labels. It then proceeds to spit the data sets into train, validation, and test sets.
# prepare tensor data sets
def prepare_dataset(padded_tokens, target):
# prepare target into np array
target = np.array(target.values, dtype=np.int64).reshape(-1, 1)
# create tensor data sets
tensor_df = TensorDataset(torch.from_numpy(padded_tokens), torch.from_numpy(target))
# 80% of df
train_size = int(0.8 * len(df))
# 20% of df
val_size = len(df) - train_size
# 50% of validation
test_size = int(val_size - 0.5*val_size)
# divide the dataset by randomly selecting samples
train_dataset, val_dataset = random_split(tensor_df, [train_size, val_size])
# divide validation by randomly selecting samples
val_dataset, test_dataset = random_split(val_dataset, [test_size, test_size+1])
return train_dataset, val_dataset, test_dataset
# create tensor data sets
train_dataset, val_dataset, test_dataset = prepare_dataset(data_set_3d, df['target'])2.6 Data Loaders and Helper Functions
Since my corpus is imbalanced, we produce weighted samplers to help balance the distribution of data as it is fed via the data loaders.
# helper function to count target distribution inside tensor data sets
def target_count(tensor_dataset):
# set empty count containers
count0 = 0
count1 = 0
# set total container to turn into torch tensor
total = []
# for every item in the tensor data set
for i in tensor_dataset:
# if the target is equal to 0
if i[1].item() == 0:
count0 += 1
# if the target is equal to 1
elif i[1].item() == 1:
count1 += 1
total.append(count0)
total.append(count1)
return torch.tensor(total)
# prepare weighted sampling for imbalanced classification
def create_sampler(target_tensor, tensor_dataset):
# generate class distributions [x, y]
class_sample_count = target_count(tensor_dataset)
# weight
weight = 1. / class_sample_count.float()
# produce weights for each observation in the data set
samples_weight = torch.tensor([weight[t[1]] for t in tensor_dataset])
# prepare sampler
sampler = torch.utils.data.WeightedRandomSampler(weights=samples_weight,
num_samples=len(samples_weight),
replacement=True)
return sampler
# create samplers for each data set
train_sampler = create_sampler(target_count(train_dataset), train_dataset)
# prepare data loaders
def data_loading(batch_size, data_set, **kwargs):
# instantiate sampler and don't use last batch if uneven
temp_dataloader = DataLoader(data_set,
batch_size=batch_size,
drop_last=True)
return temp_dataloader
# create DataLoaders with samplers
train_dataloader = data_loading(batch_size=80,
data_set=train_dataset,
sampler=train_sampler,
shuffle=False)
valid_dataloader = data_loading(batch_size=80,
data_set=val_dataset,
shuffle=True)
test_dataloader = data_loading(batch_size=80,
data_set=test_dataset,
shuffle=True)
# time function
def format_time(elapsed):
'''
Takes a time in seconds and returns a string hh:mm:ss
'''
# round to the nearest second.
elapsed_rounded = int(round((elapsed)))
# format as hh:mm:ss
return str(datetime.timedelta(seconds=elapsed_rounded))2.7 Create Hierarchical Attention Network
# create HAN
class HAN(nn.Module):
def __init__(self, config):
super().__init__()
self.mode = config.mode
self.word_attention_rnn = WordLevelRNN(config)
self.sentence_attention_rnn = SentLevelRNN(config)
def forward(self, x, **kwargs):
x = x.permute(1, 2, 0) # Expected : # sentences, # words, batch size
num_sentences = x.size(0)
word_attentions = None
for i in range(num_sentences):
word_attn = self.word_attention_rnn(x[i, :, :])
if word_attentions is None:
word_attentions = word_attn
else:
word_attentions = torch.cat((word_attentions, word_attn), 0)
return self.sentence_attention_rnn(word_attentions)2.7.1 HAN: Create Sentence Level RNN
class SentLevelRNN(nn.Module):
def __init__(self, config):
super().__init__()
sentence_num_hidden = config.sentence_num_hidden
word_num_hidden = config.word_num_hidden
target_class = config.target_class
self.sentence_context_weights = nn.Parameter(torch.rand(2 * sentence_num_hidden, 1))
self.sentence_context_weights.data.uniform_(-0.1, 0.1)
self.sentence_gru = nn.GRU(2 * word_num_hidden, sentence_num_hidden, bidirectional=True)
self.sentence_linear = nn.Linear(2 * sentence_num_hidden, 2 * sentence_num_hidden, bias=True)
self.fc = nn.Linear(2 * sentence_num_hidden , target_class)
self.soft_sent = nn.Softmax(dim=1)
def forward(self,x):
sentence_h,_ = self.sentence_gru(x)
x = torch.relu(self.sentence_linear(sentence_h))
x = torch.matmul(x, self.sentence_context_weights)
x = x.squeeze(dim=2)
x = self.soft_sent(x.transpose(1,0))
x = torch.mul(sentence_h.permute(2, 0, 1), x.transpose(1, 0))
x = torch.sum(x, dim=1).transpose(1, 0).unsqueeze(0)
x = self.fc(x.squeeze(0))
return x2.7.2 HAN: Create Word Level RNN
class WordLevelRNN(nn.Module):
def __init__(self, config):
super().__init__()
pre_embed = config.pre_embed # embeddings
word_num_hidden = config.word_num_hidden
words_num = config.vocab_size
words_dim = config.words_dim
self.mode = config.mode
if self.mode == 'rand':
rand_embed_init = torch.Tensor(vocab_size, words_dim).uniform(-0.25, 0.25)
self.embed = nn.Embedding.from_pretrained(rand_embed_init, freeze=False)
elif self.mode == 'static':
self.static_embed = nn.Embedding.from_pretrained(pre_embed, freeze=True)
elif self.mode == 'non-static':
self.non_static_embed = nn.Embedding.from_pretrained(pre_embed, freeze=False)
else:
print("Unsupported order")
raise Exception
self.word_context_weights = nn.Parameter(torch.rand(2 * word_num_hidden, 1))
self.GRU = nn.GRU(words_dim, word_num_hidden, bidirectional=True)
self.linear = nn.Linear(2 * word_num_hidden, 2 * word_num_hidden, bias=True)
self.word_context_weights.data.uniform_(-0.25, 0.25)
self.soft_word = nn.Softmax(dim=1)
def forward(self, x):
# x expected to be of dimensions--> (num_words, batch_size)
if self.mode == 'rand':
x = self.embed(x)
elif self.mode == 'static':
x = self.static_embed(x)
elif self.mode == 'non-static':
x = self.non_static_embed(x)
else :
print("Unsupported mode")
raise Exception
h, _ = self.GRU(x)
x = torch.tanh(self.linear(h))
x = torch.matmul(x, self.word_context_weights)
x = x.squeeze(dim=2)
x = self.soft_word(x.transpose(1, 0))
x = torch.mul(h.permute(2, 0, 1), x.transpose(1, 0))
x = torch.sum(x, dim=1).transpose(1, 0).unsqueeze(0)
return x2.8 Create Training Functions
Now, we prepare functions to train, validate, and test our data.
def train(model, dataloader, optimizer, criterion):
# capture time
total_t0 = time.time()
# Perform one full pass over the training set.
print("")
print('======== Epoch {:} / {:} ========'.format(epoch + 1, epochs))
print('Training...')
# reset total loss for epoch
train_total_loss = 0
total_train_f1 = 0
# put model into traning mode
model.train()
# for each batch of training data...
for step, batch in enumerate(dataloader):
# progress update every 40 batches.
if step % 40 == 0 and not step == 0:
# Report progress.
print(' Batch {:>5,} of {:>5,}.'.format(step, len(dataloader)))
# Unpack this training batch from our dataloader:
#
# As we unpack the batch, we'll also copy each tensor to the GPU
#
# `batch` contains two pytorch tensors:
# [0]: input ids
# [1]: labels
b_input_ids = batch[0].cuda().long()
b_labels = batch[1].cuda().long()
# clear previously calculated gradients
optimizer.zero_grad()
with autocast():
# forward propagation (evaluate model on training batch)
logits = model(b_input_ids)
# calculate cross entropy loss
loss = criterion(logits.view(-1, 2), b_labels.view(-1))
# sum the training loss over all batches for average loss at end
# loss is a tensor containing a single value
train_total_loss += loss.item()
# Scales loss. Calls backward() on scaled loss to create scaled gradients.
# Backward passes under autocast are not recommended.
# Backward ops run in the same dtype autocast chose for corresponding forward ops.
scaler.scale(loss).backward()
# scaler.step() first unscales the gradients of the optimizer's assigned params.
# If these gradients do not contain infs or NaNs, optimizer.step() is then called,
# otherwise, optimizer.step() is skipped.
scaler.step(optimizer)
# Updates the scale for next iteration.
scaler.update()
# update the learning rate
scheduler.step()
# get preds
_, predicted = torch.max(logits, 1)
# move logits and labels to CPU
predicted = predicted.detach().cpu().numpy()
y_true = b_labels.detach().cpu().numpy()
# calculate f1
total_train_f1 += f1_score(predicted, y_true,
average='weighted',
labels=np.unique(predicted))
# calculate the average loss over all of the batches
avg_train_loss = train_total_loss / len(dataloader)
# calculate the average f1 over all of the batches
avg_train_f1 = total_train_f1 / len(dataloader)
# Record all statistics from this epoch.
training_stats.append(
{
'Train Loss': avg_train_loss,
'Train F1': avg_train_f1
}
)
# training time end
training_time = format_time(time.time() - total_t0)
# print result summaries
print("")
print("summary results")
print("epoch | trn loss | trn f1 | trn time ")
print(f"{epoch+1:5d} | {avg_train_loss:.5f} | {avg_train_f1:.5f} | {training_time:}")
torch.cuda.empty_cache()
return None
def validating(model, dataloader, criterion):
# capture validation time
total_t0 = time.time()
# After the completion of each training epoch, measure our performance on
# our validation set.
print("")
print("Running Validation...")
# put the model in evaluation mode
model.eval()
# track variables
total_valid_accuracy = 0
total_valid_loss = 0
total_valid_f1 = 0
total_valid_recall = 0
total_valid_precision = 0
# evaluate data for one epoch
for batch in dataloader:
# unpack batch from dataloader
b_input_ids = batch[0].cuda().long()
b_labels = batch[1].cuda().long()
# tell pytorch not to bother calculating gradients
# as its only necessary for training
with torch.no_grad():
# forward propagation (evaluate model on training batch)
logits = model(b_input_ids)
# calculate BCEWithLogitsLoss
loss = criterion(logits.view(-1, 2), b_labels.view(-1))
# calculate preds
_, predicted = torch.max(logits, 1)
# accumulate validation loss
total_valid_loss += loss.item()
# move logits and labels to CPU
predicted = predicted.detach().cpu().numpy()
y_true = b_labels.detach().cpu().numpy()
# calculate f1
total_valid_f1 += f1_score(predicted, y_true,
average='weighted',
labels=np.unique(predicted))
# calculate accuracy
total_valid_accuracy += accuracy_score(predicted, y_true)
# calculate precision
total_valid_precision += precision_score(predicted, y_true,
average='weighted',
labels=np.unique(predicted))
# calculate recall
total_valid_recall += recall_score(predicted, y_true,
average='weighted',
labels=np.unique(predicted))
# report final accuracy of validation run
avg_accuracy = total_valid_accuracy / len(dataloader)
# report final f1 of validation run
global avg_val_f1
avg_val_f1 = total_valid_f1 / len(dataloader)
# report final f1 of validation run
avg_precision = total_valid_precision / len(dataloader)
# report final f1 of validation run
avg_recall = total_valid_recall / len(dataloader)
# calculate the average loss over all of the batches.
avg_val_loss = total_valid_loss / len(dataloader)
# Record all statistics from this epoch.
valid_stats.append(
{
'Val Loss': avg_val_loss,
'Val Accur.': avg_accuracy,
'Val precision': avg_precision,
'Val recall': avg_recall,
'Val F1': avg_val_f1
}
)
# capture end validation time
training_time = format_time(time.time() - total_t0)
# print result summaries
print("")
print("summary results")
print("epoch | val loss | val f1 | val time")
print(f"{epoch+1:5d} | {avg_val_loss:.5f} | {avg_val_f1:.5f} | {training_time:}")
return None
def testing(model, dataloader, criterion):
print("")
print("Running Testing...")
# put the model in evaluation mode
model.eval()
# track variables
total_test_accuracy = 0
total_test_loss = 0
total_test_f1 = 0
total_test_recall = 0
total_test_precision = 0
# evaluate data for one epoch
for step, batch in enumerate(dataloader):
# progress update every 40 batches.
if step % 40 == 0 and not step == 0:
# Report progress.
print(' Batch {:>5,} of {:>5,}.'.format(step, len(dataloader)))
# unpack batch from dataloader
b_input_ids = batch[0].cuda().long()
b_labels = batch[1].cuda().long()
# tell pytorch not to bother calculating gradients
# only necessary for training
with torch.no_grad():
# forward propagation (evaluate model on training batch)
logits = model(b_input_ids)
# calculate cross entropy loss
loss = criterion(logits.view(-1, 2), b_labels.view(-1))
# calculate preds
_, predicted = torch.max(logits, 1)
# accumulate validation loss
total_test_loss += loss.item()
# move logits and labels to CPU
predicted = predicted.detach().cpu().numpy()
y_true = b_labels.detach().cpu().numpy()
# calculate f1
total_test_f1 += f1_score(predicted, y_true,
average='weighted',
labels=np.unique(predicted))
# calculate accuracy
total_test_accuracy += accuracy_score(predicted, y_true)
# calculate precision
total_test_precision += precision_score(predicted, y_true,
average='weighted',
labels=np.unique(predicted))
# calculate recall
total_test_recall += recall_score(predicted, y_true,
average='weighted',
labels=np.unique(predicted))
# report final accuracy of validation run
avg_accuracy = total_test_accuracy / len(dataloader)
# report final f1 of validation run
avg_test_f1 = total_test_f1 / len(dataloader)
# report final f1 of validation run
avg_precision = total_test_precision / len(dataloader)
# report final f1 of validation run
avg_recall = total_test_recall / len(dataloader)
# calculate the average loss over all of the batches.
avg_test_loss = total_test_loss / len(dataloader)
# Record all statistics from this epoch.
test_stats.append(
{
'Test Loss': avg_test_loss,
'Test Accur.': avg_accuracy,
'Test precision': avg_precision,
'Test recall': avg_recall,
'Test F1': avg_test_f1
}
)
return None2.9 Create Models and Training Necessities
In order to use our HAN, we need to specify a config class that sets a number of hyperparameters that the class is expecting. The rest of the objects are our usual optimizer, scheduler, epochs, and loss function.
# HAN config class
class HANconfig():
def __init__(self):
HANconfig.mode = 'static'
HANconfig.pre_embed = embeddings_tensor
HANconfig.word_num_hidden = 120
HANconfig.words_dim = 201
HANconfig.sentence_num_hidden = 115
HANconfig.vocab_size = len(vocab)
HANconfig.target_class = 2
return None
# instantiate HAN config
configHAN = HANconfig()
# instantiate model - attach to GPU
model = HAN(configHAN).cuda()
# set loss
criterion = nn.CrossEntropyLoss()
# set number of epochs
epochs = 5
# set optimizer
optimizer = AdamW(model.parameters(),
lr=0.0009515568738386746,
weight_decay=0.5
)
# set LR scheduler
total_steps = len(train_dataloader) * epochs
scheduler = get_linear_schedule_with_warmup(optimizer,
num_warmup_steps=0,
num_training_steps=total_steps)2.10 Train
Finally we are ready to train. Two containers are created to store the results of each training and validation epoch
# create gradient scaler for mixed precision
scaler = GradScaler()
# create training result storage
training_stats = []
valid_stats = []
best_valid_loss = float('inf')
# for each epoch
for epoch in range(epochs):
# train
train(model, train_dataloader, optimizer, criterion)
# validate
validating(model, valid_dataloader, criterion)
# check validation loss
if valid_stats[epoch]['Val Loss'] < best_valid_loss:
best_valid_loss = valid_stats[epoch]['Val Loss']
# save best model for use later
torch.save(model.state_dict(), 'han-model1.pt')##
## ======== Epoch 1 / 5 ========
## Training...
## Batch 40 of 100.
## Batch 80 of 100.
##
## summary results
## epoch | trn loss | trn f1 | trn time
## 1 | 0.35020 | 0.84468 | 0:00:07
##
## Running Validation...
##
## summary results
## epoch | val loss | val f1 | val time
## 1 | 0.25396 | 0.89940 | 0:00:00
##
## ======== Epoch 2 / 5 ========
## Training...
## Batch 40 of 100.
## Batch 80 of 100.
##
## summary results
## epoch | trn loss | trn f1 | trn time
## 2 | 0.28292 | 0.87081 | 0:00:07
##
## Running Validation...
##
## summary results
## epoch | val loss | val f1 | val time
## 2 | 0.24733 | 0.88715 | 0:00:00
##
## ======== Epoch 3 / 5 ========
## Training...
## Batch 40 of 100.
## Batch 80 of 100.
##
## summary results
## epoch | trn loss | trn f1 | trn time
## 3 | 0.26662 | 0.88033 | 0:00:07
##
## Running Validation...
##
## summary results
## epoch | val loss | val f1 | val time
## 3 | 0.24396 | 0.89047 | 0:00:00
##
## ======== Epoch 4 / 5 ========
## Training...
## Batch 40 of 100.
## Batch 80 of 100.
##
## summary results
## epoch | trn loss | trn f1 | trn time
## 4 | 0.25348 | 0.88681 | 0:00:07
##
## Running Validation...
##
## summary results
## epoch | val loss | val f1 | val time
## 4 | 0.24355 | 0.89599 | 0:00:00
##
## ======== Epoch 5 / 5 ========
## Training...
## Batch 40 of 100.
## Batch 80 of 100.
##
## summary results
## epoch | trn loss | trn f1 | trn time
## 5 | 0.24332 | 0.89265 | 0:00:07
##
## Running Validation...
##
## summary results
## epoch | val loss | val f1 | val time
## 5 | 0.24402 | 0.89838 | 0:00:00
##
## C:\Users\Andrew\Anaconda3\envs\my_ml\lib\site-packages\torch\optim\lr_scheduler.py:123: UserWarning: Detected call of `lr_scheduler.step()` before `optimizer.step()`. In PyTorch 1.1.0 and later, you should call them in the opposite order: `optimizer.step()` before `lr_scheduler.step()`. Failure to do this will result in PyTorch skipping the first value of the learning rate schedule. See more details at https://pytorch.org/docs/stable/optim.html#how-to-adjust-learning-rate
## "https://pytorch.org/docs/stable/optim.html#how-to-adjust-learning-rate", UserWarning)2.11 Show Results
After training, we organize the results nicely in pandas.
# organize results
pd.set_option('precision', 3)
df_train_stats = pd.DataFrame(data=training_stats)
df_valid_stats = pd.DataFrame(data=valid_stats)
df_stats = pd.concat([df_train_stats, df_valid_stats], axis=1)
df_stats.insert(0, 'Epoch', range(1, len(df_stats)+1))
df_stats = df_stats.set_index('Epoch')
df_stats## Train Loss Train F1 Val Loss ... Val precision Val recall Val F1
## Epoch ...
## 1 0.350 0.845 0.254 ... 0.905 0.898 0.899
## 2 0.283 0.871 0.247 ... 0.891 0.890 0.887
## 3 0.267 0.880 0.244 ... 0.893 0.893 0.890
## 4 0.253 0.887 0.244 ... 0.898 0.898 0.896
## 5 0.243 0.893 0.244 ... 0.901 0.900 0.898
##
## [5 rows x 7 columns]2.12 Plot Results
Then we plot our results like so:
# plot results above
def plot_results(df):
# styling from seaborn.
sns.set(style='darkgrid')
# uncrease the plot size and font size.
sns.set(font_scale=1.5)
plt.rcParams["figure.figsize"] = (12,6)
# plot the learning curve.
plt.plot(df_stats['Train Loss'], 'b-o', label="Training")
plt.plot(df_stats['Val Loss'], 'g-o', label="Validation")
# Label the plot.
plt.title("Training & Validation Loss")
plt.xlabel("Epoch")
plt.ylabel("Loss")
plt.legend()
plt.xticks(list(range(1, epochs+1)))
return plt.show()
plot_results(df_stats)
Training Results
2.13 Test the Model
And lastly run our final test:
## <All keys matched successfully>##
## Running Testing...## Test Loss Test Accur. Test precision Test recall Test F1
## 0 0.31 0.858 0.859 0.858 0.856