我在Tensorflow中训练了一个模型，并且在PyTorch中复制它时遇到了麻烦。Tensorflow模型实现了接近100%的准确率(任务很简单)，但PyTorch模型是随机执行的。我花了一段时间想明白,和不明白的问题可以。
该模型是为二值分类任务训练的。给定描述象限和(x, y, z)坐标的输入话语，模型必须预测坐标的(x, z)部分是否在所描述的象限内。例如,如果输入文本"象限1";坐标为(0.5, -, 0.5)，则预测为真，但如果区域为象限2，则预测为真。对于相同的坐标，那么预测应该是假的。
我生成了一些数据，并使用以下代码在Tensorflow中训练模型:
x_data_placeholder = tf.placeholder(tf.float32, [FLAGS.batch_size, 1], name="x_data")
y_data_placeholder = tf.placeholder(tf.float32, [FLAGS.batch_size, 1], name="y_data")
z_data_placeholder = tf.placeholder(tf.float32, [FLAGS.batch_size, 1], name="z_data")
# text and labels placeholders
text_data = tf.placeholder(tf.int32, [FLAGS.batch_size, maxtextlength])
text_lengths = tf.placeholder(tf.int32, [FLAGS.batch_size])
y_labels_placeholder = tf.placeholder(tf.int64, [FLAGS.batch_size])
# encode text and coordinate
embeddings = tf.Variable(tf.random_uniform([100, embedding_size], -1, -1))
rnn_inputs = tf.nn.embedding_lookup(embeddings, text_data)
rnn_layers = [tf.compat.v1.nn.rnn_cell.LSTMCell(size, initializer=tf.compat.v1.keras.initializers.glorot_normal) for size in [256]]
multi_rnn_cell = tf.compat.v1.nn.rnn_cell.MultiRNNCell(rnn_layers, state_is_tuple=True)
text_outputs, text_fstate = tf.compat.v1.nn.dynamic_rnn(cell=multi_rnn_cell,
inputs=rnn_inputs,
dtype=tf.float32, sequence_length=text_lengths)
# have fully connected layers to map them the input coordinates into the same dimension as the LSTM output layer from above
x_output_layer = tf.compat.v1.layers.dense(x_data_placeholder, units=FLAGS.fc_column_size, activation=tf.nn.relu, name='x_coordinate')
y_output_layer = tf.compat.v1.layers.dense(y_data_placeholder, units=FLAGS.fc_column_size, activation=tf.nn.relu, name='y_coordinate')
z_output_layer = tf.compat.v1.layers.dense(z_data_placeholder, units=FLAGS.fc_column_size, activation=tf.nn.relu, name='z_coordinate')
# add the representations
total_output_layer = x_output_layer + y_output_layer + z_output_layer + lstm_output_layer
# make the predictions with two fully connected layers
fc_1 = tf.compat.v1.layers.dense(total_output_layer, units=FLAGS.hidden_layer_size, activation=tf.nn.relu, name='fc_1')
logits = tf.compat.v1.layers.dense(fc_1, units=FLAGS.output_dims, activation=None, name='logits')
# train the model
loss = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(labels=y_labels_placeholder, logits=logits))
optimizer = tf.train.AdamOptimizer(learning_rate=FLAGS.learning_rate, epsilon=1e-7)
gradients, variables = zip(*optimizer.compute_gradients(loss))
gradients, _ = tf.clip_by_global_norm(gradients, FLAGS.gradient_clip_threshold)
optimize = optimizer.apply_gradients(zip(gradients, variables))
# then it'll be trained with sess.run ...
现在对于PyTorch复制:
class BaselineModel(nn.Module):
def __init__(self):
super(BaselineModel, self).__init__()
self.encode_x = nn.Linear(1, embed_size)
self.encode_y = nn.Linear(1, embed_size)
self.encode_z = nn.Linear(1, embed_size)
self._embeddings = nn.Embedding(vocab_size, self.embedding_table_size)
nn.init.uniform_(self._embeddings.weight, -1.0, 1.0)
self.num_layers = 1
self.rnn = nn.LSTM(self.embedding_table_size, self.hidden_size, batch_first=True)
self.fc_after_text_lstm = nn.Linear(self.hidden_size, 100)
self.fc = nn.Linear(100, 256)
self.fc_final = nn.Linear(256, 2)
self.relu_activation = nn.ReLU()
self.softmax = nn.Softmax(dim=1)
def init_hidden(self, batch_size, device='cuda:0'):
# for LSTM, we need # of layers
h_0 = torch.zeros(1, batch_size, self.hidden_size).to(device)
c_0 = torch.zeros(1, batch_size, self.hidden_size).to(device)
return h_0, c_0
def forward(self, input_text, x_coordinate=None, y_coordinate=None, z_coordinate=None):
x_embed = self.relu_activation(self.encode_x(x_coordinate.cuda().to(torch.float32)).cuda())
y_embed = self.relu_activation(self.encode_y(y_coordinate.cuda().to(torch.float32))).cuda()
z_embed = self.relu_activation(self.encode_z(z_coordinate.cuda().to(torch.float32))).cuda()
embeds = self._embeddings(input_text)
embedding, hidden = self.rnn(embeds, self.hidden)
text_fc = self.relu_activation(self.fc_after_text_lstm(embedding[:, -1]))
representations_so_far_added = torch.sum(torch.stack([text_fc, x_embed, y_embed, z_embed]), dim=0)
pre_final_embedding = self.relu_activation(self.fc(representations_so_far_added))
return self.fc_final(pre_final_embedding )
### training code
optimizer = torch.optim.Adam(model.parameters(), lr=0.001, eps=1e-7)
criterion = nn.CrossEntropyLoss()
for input_text, x_coordinate, y_coordinate, z_coordinate, targets, train_data:
optimizer.zero_grad()
pred = model(input_text, x_coordinate=x_coordinate, y_coordinate=y_coordinate, z_coordinate=z_coordinate)
loss = criterion(pred.float(), targets)
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), 5.0)
optimizer.step()
scheduler.step()
# accuracy evaluation code, this is evaluated over the entire epoch
pred_idx = F.log_softmax(pred, dim=1)
target_labels = targets.cpu().int()
pred_labels = torch.argmax(pred_idx, dim=-1).cpu().data.int()
curr_acc = skm.accuracy_score(target_labels, pred_labels)
如果有人能发现PyTorch实现的任何问题，或者告诉我什么可能是错的，那将是非常感激的!我还尝试将Tensorflow模型的权重加载到所有适当的层中，PyTorch的性能仍然很挣扎!提前感谢!
编辑:我创造了一个最低限度可复制的例子，因为我仍然不知道问题是什么。任何帮助将仍然感激!
import torch
import torch.nn as nn
import numpy as np
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
lr = 0.0005
n_epochs = 10
input_dim = 4
hidden_dim = 128
layer_dim = 2
output_dim = 2
batch_size = 50
class FeatureDataSet(torch.utils.data.Dataset):
def __init__(self, x_train, y_train, x_coordinates):
self.x_train = torch.tensor(x_train, dtype=torch.long)
self.y_train = torch.tensor(y_train)
self.x_coordinates = torch.tensor(x_coordinates, dtype=torch.float32)
def __len__(self):
return len(self.y_train)
def __getitem__(self, idx):
return self.x_train[idx], self.y_train[idx], self.x_coordinates[idx]
class RNN(nn.Module):
def __init__(self, input_dim, hidden_dim, layer_dim, output_dim, batch_size):
super().__init__()
self.hidden_dim = hidden_dim
self.layer_dim = layer_dim
# linear layer to encode the coordinate
self.encode_x = nn.Linear(1, hidden_dim).cuda()
self._embeddings = nn.Embedding(40, 100).cuda()
# hidden_dim is 128
# layer_dim is 2
self.lstm = nn.LSTM(100, hidden_dim, layer_dim, batch_first=True).cuda()
self.fc = nn.Linear(2 * hidden_dim, output_dim).cuda()
self.batch_size = batch_size
self.hidden = None
def init_hidden(self, x):
h0 = torch.zeros(self.layer_dim, x.size(0), self.hidden_dim)
c0 = torch.zeros(self.layer_dim, x.size(0), self.hidden_dim)
return [t.cpu() for t in (h0, c0)]
def forward(self, x, x_coordinate):
#initializing the hidden states
h0, c0 = self.init_hidden(x)
embeds = self._embeddings(x)
out, (hn, cn) = self.lstm(embeds.cuda(), (h0.cuda(), c0.cuda()))
x_embed = F.relu(self.encode_x(x_coordinate.cuda().to(torch.float32)).cuda())
representations_so_far_added = torch.cat([out[:, -1, :], x_embed], dim=1)
out = self.fc(representations_so_far_added)
return out
model = RNN(input_dim, hidden_dim, layer_dim, output_dim, batch_size)
criterion = nn.CrossEntropyLoss()
opt = torch.optim.Adam(model.parameters(), lr=0.001)
print('Start model training')
import sklearn.metrics as skm
import torch.nn.functional as F
x_train = []
x_coordinates = []
y_train = []
for i in range(10000):
# create the data. if x_coordinate > 0 and the sentence says that (represented by [1, 5, 6, 8]), then we should predict positive else negative (if the x_coordinate > 0)
# same applies if the x_coordinate < 0, just that the sentence is now [1, 5, 6, 9]
if np.random.randint(0, 2) == 0:
if np.random.randint(0, 2) == 0:
# x coordinate > 0
x_train.append([1, 5, 6, 8])
x_coordinates.append([round(np.random.uniform(0.01, 1.00, 1)[0], 2)])
y_train.append(1.0)
else:
# x coordinate > 0 negative
x_train.append([1, 5, 6, 8])
x_coordinates.append([round(np.random.uniform(-1.00, 0.00, 1)[0], 2)])
y_train.append(0.0)
else:
if np.random.randint(0, 2) == 0:
# x coordinate < 0
x_train.append([1, 5, 6, 9])
x_coordinates.append([round(np.random.uniform(-1.00, 0.00, 1)[0], 2)])
y_train.append(1.0)
else:
# x coordinate < 0 negative
x_train.append([1, 5, 6, 9])
x_coordinates.append([round(np.random.uniform(0.01, 1.00, 1)[0], 2)])
y_train.append(0.0)
# print a sample of data
print(x_train[:10])
print(y_train[:10])
print(x_coordinates[:10])
# create a dataloader
trainingDataset = FeatureDataSet(x_train=x_train, y_train=y_train, x_coordinates=x_coordinates)
train_loader = torch.utils.data.DataLoader(dataset=trainingDataset, batch_size=batch_size, shuffle=True)
# for each epoch
for epoch in range(1, n_epochs + 1):
acc_all = []
# each batch
for i, (x_batch, y_batch, x_coord_batch) in enumerate(train_loader):
x_batch = x_batch.to(device)
y_batch = y_batch.to(device)
x_coord_batch = x_coord_batch.to(device)
opt.zero_grad()
# pass in the text (x_batch) and coordinate (x_coord_batch)
out = model(x_batch, x_coordinate=x_coord_batch)
loss = criterion(out.float(), y_batch.type(torch.LongTensor).cuda())
loss.backward()
opt.step()
pred_idx = F.log_softmax(out, dim=1)
target_labels = y_batch.cpu().int()
pred_labels = torch.argmax(pred_idx, dim=-1).cpu().data.int()
curr_acc = skm.accuracy_score(target_labels, pred_labels)
acc_all.append(curr_acc)
print(np.mean(acc_all))
我想也许有一些错误在你的数据集实现
我在你的"最低限度可复制的例子"中的两个数据集上尝试了你的pytorchBaselineModel;并根据您的描述构建了我自己的数据集，发现它工作得很好。
下面是我在自己的数据集上测试的代码。请注意，我在BaselineModel的代码中添加了几个超参数以使其运行。我的准确率超过99%。
import random
import torch
import torch.nn as nn
import numpy as np
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
lr = 0.0005
n_epochs = 100
input_dim = 4
hidden_dim = 128
layer_dim = 2
output_dim = 2
batch_size = 50
class AverageMeter(object):
"""Computes and stores the average and current value"""
def __init__(self):
self.reset()
def reset(self):
self.val = 0
self.avg = 0
self.sum = 0
self.count = 0
def update(self, val, n=1):
self.val = val
self.sum += val * n
self.count += n
self.avg = self.sum / self.count
class FeatureDataSet(torch.utils.data.Dataset):
def __init__(self, x_train, y_train, x_coordinates, y_coordinates, z_coordinates):
self.x_train = torch.tensor(x_train, dtype=torch.long)
self.y_train = torch.tensor(y_train)
self.x_coordinates = torch.tensor(x_coordinates, dtype=torch.float32)
self.y_coordinates = torch.tensor(y_coordinates, dtype=torch.float32)
self.z_coordinates = torch.tensor(z_coordinates, dtype=torch.float32)
def __len__(self):
return len(self.y_train)
def __getitem__(self, idx):
return self.x_train[idx], self.y_train[idx], self.x_coordinates[idx], self.y_coordinates[idx], self.z_coordinates[idx]
class BaselineModel(nn.Module):
def __init__(self):
super(BaselineModel, self).__init__()
vocab_size = 40
self.hidden_size = 100
self.embedding_table_size = self.hidden_size
self.encode_x = nn.Linear(1, self.hidden_size)
self.encode_y = nn.Linear(1, self.hidden_size)
self.encode_z = nn.Linear(1, self.hidden_size)
self._embeddings = nn.Embedding(vocab_size, self.embedding_table_size)
nn.init.uniform_(self._embeddings.weight, -1.0, 1.0)
self.num_layers = 1
self.rnn = nn.LSTM(self.embedding_table_size, self.hidden_size, batch_first=True)
self.fc_after_text_lstm = nn.Linear(self.hidden_size, 100)
self.fc = nn.Linear(100, 256)
self.fc_final = nn.Linear(256, 2)
self.relu_activation = nn.ReLU()
self.softmax = nn.Softmax(dim=1)
self.hidden = self.init_hidden(batch_size)
def init_hidden(self, batch_size, device='cuda:0'):
# for LSTM, we need # of layers
h_0 = torch.zeros(1, batch_size, self.hidden_size).to(device)
c_0 = torch.zeros(1, batch_size, self.hidden_size).to(device)
return h_0, c_0
def forward(self, input_text, x_coordinate=None, y_coordinate=None, z_coordinate=None):
x_embed = self.relu_activation(self.encode_x(x_coordinate.cuda().to(torch.float32)).cuda())
y_embed = self.relu_activation(self.encode_y(y_coordinate.cuda().to(torch.float32))).cuda()
z_embed = self.relu_activation(self.encode_z(z_coordinate.cuda().to(torch.float32))).cuda()
embeds = self._embeddings(input_text)
embedding, hidden = self.rnn(embeds, self.hidden)
text_fc = self.relu_activation(self.fc_after_text_lstm(embedding[:, -1]))
representations_so_far_added = torch.sum(torch.stack([text_fc, x_embed, y_embed, z_embed]), dim=0)
pre_final_embedding = self.relu_activation(self.fc(representations_so_far_added))
return self.fc_final(pre_final_embedding)
# model = RNN(input_dim, hidden_dim, layer_dim, output_dim, batch_size)
model = BaselineModel().cuda()
criterion = nn.CrossEntropyLoss()
opt = torch.optim.Adam(model.parameters(), lr=0.001)
print('Start model training')
import sklearn.metrics as skm
import torch.nn.functional as F
x_train = []
x_coordinates = []
y_coordinates = []
z_coordinates = []
y_train = []
for i in range(10000):
x_coordinate = round(np.random.uniform(-1, 1.00, 1)[0], 2)
y_coordinate = round(np.random.uniform(-1, 1.00, 1)[0], 2)
z_coordinate = round(np.random.uniform(-1, 1.00, 1)[0], 2)
x_coordinates.append([x_coordinate])
y_coordinates.append([y_coordinate])
z_coordinates.append([z_coordinate])
if np.random.randint(0, 2) == 0: # positive example
if x_coordinate <= 0 and z_coordinate <= 0:
x_train.append([1, 5, 6, 8])
elif x_coordinate <= 0 and z_coordinate > 0:
x_train.append([1, 5, 6, 9])
elif x_coordinate > 0 and z_coordinate <= 0:
x_train.append([1, 5, 6, 10])
elif x_coordinate > 0 and z_coordinate > 0:
x_train.append([1, 5, 6, 11])
y_train.append(1.0)
else:
if x_coordinate <= 0 and z_coordinate <= 0:
x_train.append(random.choice([[1, 5, 6, 9], [1, 5, 6, 10], [1, 5, 6, 11]]))
elif x_coordinate <= 0 and z_coordinate > 0:
x_train.append(random.choice([[1, 5, 6, 8], [1, 5, 6, 10], [1, 5, 6, 11]]))
elif x_coordinate > 0 and z_coordinate <= 0:
x_train.append(random.choice([[1, 5, 6, 8], [1, 5, 6, 9], [1, 5, 6, 11]]))
elif x_coordinate > 0 and z_coordinate > 0:
x_train.append(random.choice([[1, 5, 6, 8], [1, 5, 6, 9], [1, 5, 6, 10]]))
y_train.append(0.0)
# print a sample of data
print(x_train[:10])
print(y_train[:10])
print(x_coordinates[:10])
print(y_coordinates[:10])
print(z_coordinates[:10])
# create a dataloader
trainingDataset = FeatureDataSet(x_train=x_train, y_train=y_train, x_coordinates=x_coordinates, y_coordinates=y_coordinates, z_coordinates=z_coordinates)
train_loader = torch.utils.data.DataLoader(dataset=trainingDataset, batch_size=batch_size, shuffle=True)
# for each epoch
loss_meter = AverageMeter()
for epoch in range(1, n_epochs + 1):
acc_all = []
# each batch
loss_meter.reset()
for i, (x_batch, y_batch, x_coord_batch, y_coord_batch, z_coord_batch) in enumerate(train_loader):
x_batch = x_batch.to(device)
y_batch = y_batch.to(device)
x_coord_batch = x_coord_batch.to(device)
y_coord_batch = y_coord_batch.to(device)
z_coord_batch = z_coord_batch.to(device)
opt.zero_grad()
# pass in the text (x_batch) and coordinate (x_coord_batch)
out = model(x_batch, x_coordinate=x_coord_batch, y_coordinate=y_coord_batch, z_coordinate=z_coord_batch)
loss = criterion(out.float(), y_batch.type(torch.LongTensor).cuda())
loss.backward()
opt.step()
pred_idx = F.log_softmax(out, dim=1)
target_labels = y_batch.cpu().int()
pred_labels = torch.argmax(pred_idx, dim=-1).cpu().data.int()
curr_acc = skm.accuracy_score(target_labels, pred_labels)
acc_all.append(curr_acc)
loss_meter.update(loss.item())
print(np.mean(acc_all))
print("loss is %f" % loss_meter.val)
关于"最低限度可复制的例子"，我认为模型RNN不起作用是非常合理的，正如我在评论中所说的。我想tensorflow也不能很好地拟合，尽管我没有尝试过。你的"最低限度可复制的例子"可能与你的主要问题无关。