model path about pocketflow HOT 9 CLOSED

1. We are considering unifying the save path arguments for model compression components. For instance, all model compression components should write the training graph (and variables) to mc_save_path and the evaluation graph (and variables) to mc_save_path_eval.

2. Yes, save_path means the save path of the full-precision uncompressed network, particularly the training graph (and variables). The evaluation graph (and variables) is written to save_path_eval.

The reason for saving both training and evaluation graph is that, to continue from previously-aborted training, you need the training graph, and to export a TF-Lite model from checkpoint files, you need the evaluation graph.

from pocketflow.

1. Can I understand the parameter save_path and save_path_eval is written in full-prec mode, and is read in other(compress model) mode? The parameter dcp_save_path and dcp_save_path_eval is written when compress the model?

2. If 1 is right. When I compress my custom model, can I don't provide the model in save_path_eval?
   save_path_eval is only read when exec the funciton forward_eval in ModelHelper class?

from pocketflow.

1. save_path & save_path_eval is written by the FullPrecLearner, and dcp_save_path & dcp_save_path_eval is written by the DisChnPrunedLearner. When running model compression components, only save_path is required to load a pre-trained uncompressed model.
2. As mentioned above, for model compression components, only save_path is required. You do not need to provide the save_path_eval. All save paths with a suffix "_eval" is only used for exporting TF-Lite models. They are not required in the training process.

from pocketflow.

I don't found the parameter that specify the save path of compressed model other than DisChnPrunedLearner mode, but I want to know you how to specify the path ? and I found some variable w.r.t path is specified to fixed value, such as channel_pruned_path = './models/pruned_model.ckpt' best_model_path = './models/best_model.ckpt', Can you tell me what does they mean?

from pocketflow.

For instance, UniformQuantTFLearner uses uqtf_save_path for saving the training graph and uqtf_save_path_eval for saving the evaluation graph.

For channel_pruned_path and best_model_path used in the ChannelPrunedLearner, @psyyz10 , could you please take a look and explain their actual usage?

from pocketflow.

@as754770178 in ChannelPrunedLearner, best_model_path is the parameter that specify the save path of compressed model and channel_pruned_path saved some intermediate results

from pocketflow.

I conclude the parameter about the path.

1. ChannelPrunedLearner
best_model_path is the parameter that specify the save path of compressed model and channel_pruned_path saved some intermediate results. Now, best_model_path changed to cp_best_paht; channel_pruned_path changed to cp_channel_pruned_path. But I doubt that what does cp_original_path mean, whether it conflict with save_path.

2. UniformQuanTFLearner
uqtf_save_path_eval is the parameter that specify the save path of compressed model and uqtf_save_path saved some intermediate results.

3. DisChnPrunedLearner
dcp_save_path_eval is the parameter that specify the save path of compressed model and dcp_save_path saved some intermediate results.

4. FullPrecLearner
save_path_eval is the parameter that specify the save path of compressed model and save_path saved some intermediate results.

5. NonUniformQuantLearner
nuql_save_quant_model_path is the parameter that specify the save path of compressed model.

6. UniformQuantLearner
uql_save_quant_model_path is the parameter that specify the save path of compressed model.

7. WeightSparseLearner
ws_save_path is the parameter that specify the save path of compressed model.
   Why NonUniformQuantLearner, UniformQuantLearner and WeightSparseLearner don't have the parameter to set the path saved some intermediate results.

I don't know whether my summary is correct?

from pocketflow.

@as754770178

For UniformQuanTFLearner, DisChnPrunedLearner, and FullPrecLearner:

• The model saving path without the "_eval" suffix is used to store checkpoint files from the training graph. You can recover a previous training process from it.
• The model saving path with the "_eval" suffix is used to store checkpoint files from the evaluation graph. You can export *.pb & *.tflite models from it.

For NonUniformQuantLearner, UniformQuantLearner, and WeightSparseLearner:

• The model saving path without the "_eval" suffix is used to store checkpoint files from the training graph. You can recover a previous training process from it.
• There is no model saving path with the "_eval" suffix is used to store checkpoint files from the evaluation graph, since these three learners does not support inference using *.tflite models.

For ChannelPrunedLearner, @psyyz10 , could you provide some more explanation?

from pocketflow.

@as754770178 cp_original_path is also a path to save temporal model, but the different temporal model from cp_channel_pruned_path. User may do some operations, for example changed model scope, and which can be saved to cp_original_path.

from pocketflow.