find learning rate from fast.ai

find_lr.py

import matplotlib.pyplot as plt
%matplotlib inline

def find_lr(net, criterion, optimizer, trn_loader, init_value = 1e-8, final_value=10., beta = 0.98):
    num = len(trn_loader)-1
    mult = (final_value / init_value) ** (1/num)
    lr = init_value
    optimizer.param_groups[0]['lr'] = lr
    avg_loss = 0.
    best_loss = 0.
    batch_num = 0
    losses = []
    log_lrs = []
    for data in trn_loader:
        batch_num += 1
        #As before, get the loss for this mini-batch of inputs/outputs
        inputs,labels = data
        inputs, labels = Variable(inputs.to(device)), Variable(labels.to(device))
        optimizer.zero_grad()
        pos, neg = net(labels, inputs)
        loss = criterion(pos.view(-1, 1), neg)

        #Compute the smoothed loss
        avg_loss = beta * avg_loss + (1-beta) *loss.data[0]
        smoothed_loss = avg_loss / (1 - beta**batch_num)
        #Stop if the loss is exploding
        if batch_num > 1 and smoothed_loss > 4 * best_loss:
            return log_lrs, losses
        #Record the best loss
        if smoothed_loss < best_loss or batch_num==1:
            best_loss = smoothed_loss
        #Store the values
        losses.append(smoothed_loss)
        log_lrs.append(torch.log10(torch.FloatTensor([lr]).to(device)))
        #Do the SGD step
        loss.backward()
        optimizer.step()
        #Update the lr for the next step
        lr *= mult
        optimizer.param_groups[0]['lr'] = lr
    return log_lrs, losses

'''
Example run. Select highest LR before graph looks erratic
model = BasicEmbModel(len(occupations), len(locations), len(industries), len(job_tags), len(user_ids), 25).to(device)
occu_warp = partial(warp_loss, num_labels=torch.FloatTensor([len(adids)]).to(device), device=device, limit_grad=False)
logs,losses = find_lr(model, occu_warp, torch.optim.RMSprop(model.parameters(), lr=0.05), train_loader, init_value=0.001, final_value=1000)
m = -5
plt.plot(logs[10:m],losses[10:m])
'''