diff --git a/doc/fluid/design/concepts/cpp_data_feeding.md b/doc/fluid/design/concepts/cpp_data_feeding.md index 8607b40ccbbe01db77afed72c1efa780b520744c..aabc1ba75a67c5767d409bd6e7e6240dec86b16c 100644 --- a/doc/fluid/design/concepts/cpp_data_feeding.md +++ b/doc/fluid/design/concepts/cpp_data_feeding.md @@ -113,7 +113,7 @@ To solve this problem, we introduce `ReaderHolder` as a wrapper. It acts as an e To create and invoke readers, some new ops are introduced: -### CreateReaderOp +### Operators That Create Readers Each reader has its creation op. File readers' creation ops have no input and yield the created file reader as its output. Decorated readers' creation ops take the underlying readers as inputs and then yield new decorated readers. @@ -153,19 +153,52 @@ double_buffer_reader = create_double_buffer_op(batch_reader) The forwarding ops of the corresponding `main_program` would be like this: ``` -while_op { +not_completed = true +pass_count = 0 +while_op(not_completed) { has_next = has_next_op(double_buffer_reader) if_else_op(has_next) { batch_data = read_op(double_buffer_reader) ... (subsequent training ops) } else { reset_op(double_buffer_reader) + increase_op(pass_count) + not_completed = less_than_op(pass_count, reqiured_pass_num) } } ``` -Two important considerations for these programs are as follows: +A few important considerations for these programs are as follows: -1. The multiple\_reader is the batch\_reader's underlying reader, and the batch\_reader is the double\_buffer\_reader's underlying reader. `read_op`, `has_next_op` and other reader related ops will only invoke the top-most reader. In this case, it's the double\_buffer\_reader. +1. `not_completed`, `pass_count` and other variables shown above are all Fluid Variables. -2. All readers exist in both `startup_program` and `main_program`. And they are persistable. +2. The multiple\_reader is the batch\_reader's underlying reader, and the batch\_reader is the double\_buffer\_reader's underlying reader. `read_op`, `has_next_op` and other reader related ops will only invoke the top-most reader. In this case, it's the double\_buffer\_reader. + +3. All readers exist in both `startup_program` and `main_program`. And they are persistable. + +### Simplify Configuration by MultiPassReader + +The Program configuration mentioned above is complicated. Users need to be very familiar to concepts of Program and Block to prevent making mistakes in their code. To make the usage of C++ readers more friendly to new users, we introduce `MultiPassReader`. + +`MultiPassReader` is a decorated reader. A multi-pass reader is used to continuously yield data for several training passes. It takes the number of passes to run as one of its attributes('pass_num') and maintains a counter to record how many passes it has completed. Each time its underlying reader reaches the EOF, the multi-pass reader checks whether it has completed the training of given number of pass. If not, the underlying reader will be re-initialized and starts a new pass automatically. Before completing the whole training, the return of MultiPassReader's `HasNext()` will always be `true`. + +With `MultiPassReader`, the startup program would be like this: + +``` +multiple_reader = open_files_op(...) +batch_reader = create_batch_reader_op(multiple_reader) +multi_pass_reader = create_multi_pass_reader_op(batch_reader) +double_buffer_reader = create_double_buffer_op(multi_pass_reader) +... (other initializers) +``` + +The forwarding part of the corresponding `main_program` would be like this: + +``` +not_completed = true +while_op(not_completed) { + batch_data = read_op(double_buffer_reader) + ... (subsequent training ops) + not_completed = has_next_op(double_buffer_reader) +} +``` diff --git a/paddle/fluid/operators/reader/create_multi_pass_reader_op.cc b/paddle/fluid/operators/reader/create_multi_pass_reader_op.cc index 4d4e9fb909eafea5328491a4097276577f28a5ba..47d9989bc8748840ec2d39587fde24355d90b6b4 100644 --- a/paddle/fluid/operators/reader/create_multi_pass_reader_op.cc +++ b/paddle/fluid/operators/reader/create_multi_pass_reader_op.cc @@ -81,10 +81,10 @@ class CreateMultiPassReaderOpMaker : public DecoratedReaderMakerBase { This operator creates a multi-pass reader. A multi-pass reader is used to yield data for several pass training continuously. - It takes the the number of pass to run as one of its attributes + It takes the number of passes to run as one of its attributes ('pass_num'), and maintains a pass counter to record how many - passes it has completed. When the underlying reader reach the EOF, - the multi-pass reader checks whether it has completed training + passes it has completed. When the underlying reader reaches the + EOF, the multi-pass reader checks whether it has completed training of the given number of pass. If not, the underlying reader will be re-initialized and starts a new pass automatically. )DOC");