Exam DP-100 topic 6 question 2 discussion

TF supports static computational graph while pytorch supports dynamic Computational Graph. So the answer to the first question is pytorch since we are asked to use dynamic runtime graph computation the 2nd and 4th option are as described in the given solution The 3rd option is confusing, since SGD is offered by pytorch and gradient descent optimizer is offered by tensorflow. I will go with SGD, because it goes with the rest of the answers even though there is this "All penalty detection models show inference phases using a Stochastic Gradient Descent (SGD) are running too slow"

Box3: train.GradientDescentOptimizer belongs to TensorFlow, but the other boxes use Pytorch. https://www.tensorflow.org/api_docs/python/tf/compat/v1/train/GradientDescentOptimizer

-import pytorch as deeplearninglib f-train sampler = deeplearninglib.WeightedRandomSampler.(penalty video dataset) h-optimizer = deeplearninglib.optim. SGD(model. parameters).Ir=0,01) k-model = deeplearninglib.nn.parallel. DistributedDataParallelCPU(model) These options support the requirements of dynamic runtime graph computation, handling imbalance in the penalty detection classes, applying Stochastic Gradient Descent (SGD) optimizer, and employing parallel computations for the model respectively.

Box 1: A) import pytorch as deeplearninglib Explanation: Since the feature mentioned, dynamic runtime graph computation, is a feature of PyTorch, we should import PyTorch in this case. Box 2: C) train_sampler= deeplearninglib.WeightedRnadomSampler.(penalty_video_dataset) Explanation: A sampling strategy is required for the global penalty detection models. The WeightedRandomSampler allows for weighted sampling, which may be useful for ensuring that rarer samples are not overlooked in the training process. Box 3: A) optimizer= deeplearninglib.optim.SGD(model.parameters().lr=0.01) Explanation: The SGD optimizer is mentioned specifically for the penalty detection models, and the learning rate is set to 0.01. Box 4: A) model= deeplearninglib.parallel.DistributedDataParallel(model) Explanation: The DistributedDataParallel module allows for parallel processing of a single model across multiple devices or nodes, which can significantly speed up the training process. This is useful for the global penalty detection models, which are mentioned to have slow inference times.

No clue, the only thing I know of is that DistributedSampler, Optim.SGD, and nn.Parallel ... are all pytouch packages or classes ...

so hard to answer

its all messed up ......

Why Box 4 uses CPU?

Box 2 is not correct either, it says 'deeplearming' instead of 'deeplearning'...

I think, box3: optimizer = deeplearninglib.optim.SGD(model.parameters().lr=0,01) https://analyticsindiamag.com/how-ml-frameworks-like-tensorflow-and-pytorch-handle-gradient-descent/

Which is the correct one?

why pytorch not tensorflow? they both support Python