Accelerated transfer learning
Techniques for updating a model are described. An example includes determining a proper subset of weights of a first trained machine learning (ML) model to change during an optimization procedure; updating only the proper subset of weights of a second ML model on a first source dataset to generate an optimized ML model; and storing the optimized ML model.
1 . A method comprising:
in a provider network,
receiving a request to generate an optimized machine learning (ML) model;
determining, by a model training service of a provider network implemented as model training services code stored in memory and executed by one or more processors of the provider network, a proper subset of weights of a first trained ML model to change during an optimization procedure by generating a distribution of the weights of the first trained ML model and sampling the distribution of the weights based on the distribution, wherein a number of the weights in the proper subset of weights is set by the model training service;
updating only the proper subset of weights of a second trained ML model on a first source dataset to generate an optimized ML model; and
storing the optimized ML model according to the request.
2 . The computer-implemented method of claim 1 , wherein the first source dataset relates to a same task as a dataset used to train the second ML model to be optimized.
3 . The computer-implemented method of claim 1 , wherein the request includes one or more of an identification of a location of a ML model to optimize, a ML model to optimize, an indication of a type of optimization to perform, an indication of a stopping threshold, an indication of a number of training iterations to perform, an indication of a criteria for selecting weights, an indication of an algorithm to use to select parameters, and an identification of a location to store an optimized model.
4 . A computer-implemented method comprising:
determining, by a model training service of a provider network implemented as model training services code stored in memory and executed by one or more processors of the provider network, a proper subset of weights of a first trained machine learning (ML) model to change during an optimization procedure by generating a distribution of the weights of the first trained ML model and sampling the distribution of the weights based on the distribution, wherein a number of the weights in the proper subset of weights is set by the model training service;
updating only the proper subset of weights of a second ML model on a first source dataset to generate an optimized ML model; and
storing the optimized ML model.
5 . The computer-implemented method of claim 4 , further comprising training the first ML model on a second dataset.
6 . The computer-implemented method of claim 5 , wherein the second dataset is larger than the first source dataset.
7 . The computer-implemented method of claim 4 , wherein the determining a proper subset of weights of a first trained ML model to change during an optimization procedure comprises:
generating a distribution of weights of the ML model; and
sampling the weights based on the distribution to generate the proper subset of weights.
8 . The computer-implemented method of claim 4 , wherein the first source dataset relates to a same task as a dataset used to train the second ML model to be optimized.
9 . The computer-implemented method of claim 4 , wherein a number of weights in the proper subset is set by a service of a provider network.
10 . The computer-implemented method of claim 4 , further comprising receiving a request to optimize the second ML model, wherein the request includes one or more of an identification of a location of a ML model to optimize, a ML model to optimize, an indication of a type of optimization to perform, an indication of a stopping threshold, an indication of a number of training iterations to perform, an indication of a criteria for selecting weights, an indication of an algorithm to use to select parameters, and an identification of a location to store an optimized model.
11 . The computer-implemented method of claim 10 , wherein the request further comprises an indication of a number of weights to be sampled.
12 . The computer-implemented method of claim 4 , wherein the updating only the proper subset of weights of the second ML model on a first source dataset to generate an optimized ML model uses a variant of adaptive moment estimation.
13 . The computer-implemented method of claim 4 , wherein the updating only the proper subset of weights of the second ML model on a first source dataset to generate an optimized ML model uses a variant of stochastic gradient descent.
14 . The computer-implemented method of claim 4 , further comprising generating and providing information regarding the optimizing in comparison to other optimization techniques.
15 . A system comprising:
a first one or more electronic devices to implement a storage service in a multi-tenant provider network; and
a second one or more electronic devices to implement a model optimizing service in the multi-tenant provider network, the model optimizing service implemented as model optimizing service code, stored in memory, that upon execution by one or more processors of the provider network cause the model optimizing service to:
determine a proper subset of weights of a first trained machine learning (ML) model to change during an optimization procedure by generating a distribution of the weights of the first trained ML model and sampling the distribution of the weights based on the distribution, wherein a number of the weights in the proper subset of weights is set by the model optimizing service;
update only the proper subset of weights of a second ML model on a first source dataset to generate an optimized ML model; and
store the optimized ML model.
16 . The system of claim 15 , further comprising a training service to train the first trained ML model.
17 . The system of claim 15 , further comprising a model hosting service to host the optimized ML model.
18 . The system of claim 15 , wherein the model optimizing service is to receive a request to optimize the second ML model, wherein the request includes one or more of an identification of a location of a ML model to optimize, a ML model to optimize, an indication of a type of optimization to perform, an indication of a stopping threshold, an indication of a number of training iterations to perform, an indication of a criteria for selecting weights, an indication of an algorithm to use to select parameters, and an identification of a location to store an optimized model.
19 . The system of claim 18 , wherein the request further comprises an indication of a number of weights to be sampled.
20 . The system of claim 15 , wherein the update only the proper subset of weights of the second ML model on a first source dataset to generate an optimized ML model is to use a variant of adaptive moment estimation.