Test your AI Agents with the all-new Agent to Agent Testing Platform.Learn More

How to use flop method in fMBT

Best Python code snippet using fMBT_python

search_main.py

Source:search_main.py

...48    top5.update       (prec5.item(), base_inputs.size(0))49    # update the architecture50    arch_optimizer.zero_grad()51    logits, expected_flop = network(arch_inputs)52    flop_cur  = network.module.get_flop('genotype', None, None)53    flop_loss, flop_loss_scale = get_flop_loss(expected_flop, flop_cur, flop_need, flop_tolerant)54    acls_loss = criterion(logits, arch_targets)55    arch_loss = acls_loss + flop_loss * flop_weight56    arch_loss.backward()57    arch_optimizer.step()58  59    # record60    arch_losses.update(arch_loss.item(), arch_inputs.size(0))61    arch_flop_losses.update(flop_loss_scale, arch_inputs.size(0))62    arch_cls_losses.update (acls_loss.item(), arch_inputs.size(0))63    64    # measure elapsed time65    batch_time.update(time.time() - end)66    end = time.time()...

search_main_v2.py

Source:search_main_v2.py

...46    top5.update       (prec5.item(), base_inputs.size(0))47    # update the architecture48    arch_optimizer.zero_grad()49    logits, expected_flop = network(arch_inputs)50    flop_cur  = network.module.get_flop('genotype', None, None)51    flop_loss, flop_loss_scale = get_flop_loss(expected_flop, flop_cur, flop_need, flop_tolerant)52    acls_loss = criterion(logits, arch_targets)53    arch_loss = acls_loss + flop_loss * flop_weight54    arch_loss.backward()55    arch_optimizer.step()56  57    # record58    arch_losses.update(arch_loss.item(), arch_inputs.size(0))59    arch_flop_losses.update(flop_loss_scale, arch_inputs.size(0))60    arch_cls_losses.update (acls_loss.item(), arch_inputs.size(0))61    62    # measure elapsed time63    batch_time.update(time.time() - end)64    end = time.time()...

cifar_resnet_flop.py

Source:cifar_resnet_flop.py

1def cifar_resnet_flop(layer=110, prune_rate=1):2    '''3    :param layer: the layer of Resnet for Cifar, including 110, 56, 32, 204    :param prune_rate: 1 means baseline5    :return: flop of the network6    '''7    flop = 08    channel = [16, 32, 64]9    width = [32, 16, 8]10    stage = int(layer / 3)11    for index in range(0, layer, 1):12        if index == 0:  # first conv layer before block13            flop += channel[0] * width[0] * width[0] * 9 * 3 * prune_rate14        elif index in [1, 2]:  # first block of first stage15            flop += channel[0] * width[0] * width[0] * 9 * channel[0] * (prune_rate ** 2)16        elif 2 < index <= stage:  # other blocks of first stage17            if index % 2 != 0:18                # first layer of block, only output channal reduced, input channel remain the same19                flop += channel[0] * width[0] * width[0] * 9 * channel[0] * (prune_rate)20            elif index % 2 == 0:21                # second layer of block, both input and output channal reduced22                flop += channel[0] * width[0] * width[0] * 9 * channel[0] * (prune_rate ** 2)23        elif stage < index <= stage * 2:  # second stage24            if index % 2 != 0:25                flop += channel[1] * width[1] * width[1] * 9 * channel[1] * (prune_rate)26            elif index % 2 == 0:27                flop += channel[1] * width[1] * width[1] * 9 * channel[1] * (prune_rate ** 2)28        elif stage * 2 < index <= stage * 3:  # third stage29            if index % 2 != 0:30                flop += channel[2] * width[2] * width[2] * 9 * channel[2] * (prune_rate)31            elif index % 2 == 0:32                flop += channel[2] * width[2] * width[2] * 9 * channel[2] * (prune_rate ** 2)33    # offset for dimension change between blocks34    offset1 = channel[1] * width[1] * width[1] * 9 * channel[1] * prune_rate - channel[1] * width[1] * width[1] * 9 * \35              channel[0] * prune_rate36    offset2 = channel[2] * width[2] * width[2] * 9 * channel[2] * prune_rate - channel[2] * width[2] * width[2] * 9 * \37              channel[1] * prune_rate38    flop = flop - offset1 - offset239    # print(flop)40    return flop41def cal_cifar_resnet_flop(layer, prune_rate):42    '''43    :param layer:  the layer of Resnet for Cifar, including 110, 56, 32, 2044    :param prune_rate: 1 means baseline45    :return:46    '''47    pruned_flop = cifar_resnet_flop(layer, prune_rate)48    baseline_flop = cifar_resnet_flop(layer, 1)49    print(50        "pruning rate of layer {:d} is {:.1f}, pruned FLOP is {:.0f}, "51        "baseline FLOP is {:.0f}, FLOP reduction rate is {:.4f}"52        .format(layer, prune_rate, pruned_flop, baseline_flop, 1 - pruned_flop / baseline_flop))53if __name__ == '__main__':54    layer_list = [110, 56, 32, 20]55    pruning_rate_list = [0.9, 0.8, 0.7]56    for layer in layer_list:57        for pruning_rate in pruning_rate_list:...

Automation Testing Tutorials

Learn to execute automation testing from scratch with LambdaTest Learning Hub. Right from setting up the prerequisites to run your first automation test, to following best practices and diving deeper into advanced test scenarios. LambdaTest Learning Hubs compile a list of step-by-step guides to help you be proficient with different test automation frameworks i.e. Selenium, Cypress, TestNG etc.