I’m trying to solve a VQA classification problem. my training loss is not changing at all while training the model.
I put in comment the CNN model and try to run it with the text only, but still, no change in the loss function.
I pass through those models:
class question_lstm(nn.Module):
def __init__(self, input_dim, emb_dim, hid_dim, n_layers, dropout, output_dim, que_size):
super(question_lstm, self).__init__()
self.hid_dim = hid_dim
self.n_layers = n_layers
self.embedding = nn.Embedding(input_dim, emb_dim)
self.tanh = nn.Tanh()
self.lstm = nn.LSTM(emb_dim, hid_dim, n_layers, dropout = dropout)
self.dropout = nn.Dropout(dropout)
#self.fc1=nn.Linear(n_layers*hid_dim,que_size)
self.fc1=nn.Linear(n_layers*output_dim,que_size)
def forward(self, question):
emb_question=self.embedding(question) #(batchsize, input_dim, emb_dim=256)
emb_question=self.dropout(emb_question)
emb_question=self.tanh(emb_question)
emb_question = emb_question.transpose(0, 1) #(input_dim, batchsize, emb_dim)
output, (hidden, cell) = self.lstm(emb_question)
qu_feature = torch.cat((hidden, cell), dim=2)
qu_feature = qu_feature.transpose(0, 1) #(batchsize=100, num_layer=2, hid_dim=2048)
question_output =self.fc1(qu_feature)
return question_output
class vqamodel(nn.Module):
def __init__(self, output_dim,input_dim, emb_dim, hid_dim, n_layers, dropout, answer_len, que_size,):
super(vqamodel,self).__init__()
#self.image=img_CNN(img_size,image_feature)
self.question=question_lstm(input_dim, emb_dim, hid_dim, n_layers, dropout,output_dim,que_size)
self.tanh=nn.Tanh()
self.relu=nn.ReLU()
self.dropout=nn.Dropout(dropout)
self.fc1=nn.Linear(que_size,output_dim)
self.fc2=nn.Linear(output_dim,answer_len)
def forward(self, image, question):
question_emb=self.question(question)
combine =question_emb #*img_emb
out_feature=self.fc1(combine) #(batchsize=100, output_dim=2048)
out_feature=self.relu(out_feature)
out_feature=self.dropout(out_feature)
out_feature=self.fc2(out_feature) #(batchsize=100, answer_len=1000)
return (out_feature)
I’m using cross entropy loss and Adam:
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(vqa_model.parameters(),lr=0.001)
scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=7, gamma=0.1)
any idea what can cause this constant loss value?