Reconstruct the interface of model compression

[denghuilu]

Before this pull request, the model compression module within the deepmd-kit would require the users to provide the training script as well as the training data to compress a given frozen model. Herein, this pull request brings a more useful and robust interface of model compression, which removes the dependences on training script and data from the model compression.

The main changes are:

• Save the training script [Feature Request] Save the input script to DP model #727 as well as the min_nbor_dist value [Feature Request] Save the range of embedding matrix input in the DP model  #728 within the training process.
• Remove the dependences on training script and data from the model compression. Now the model compression command can be simply written as: dp compress -i original.pb -o compress.pb instead of dp compress input.json -i original.pb -o compress.pb.
• Update the UTs for model compression.
• Update the docs of model compression.

The next pull request will add an item within the convert-from interface to allow the users to upgrade their existing frozen models to adapt the new interface of model compression.

[codecov-commenter]

Codecov Report

Merging #921 (a1f087d) into devel (ee0ed99) will increase coverage by 0.11%.
The diff coverage is 94.11%.

@@            Coverage Diff             @@
##            devel     #921      +/-   ##
==========================================
+ Coverage   75.42%   75.53%   +0.11%     
==========================================
  Files          85       85              
  Lines        6730     6770      +40     
==========================================
+ Hits         5076     5114      +38     
- Misses       1654     1656       +2     

Impacted Files	Coverage Δ
deepmd/entrypoints/freeze.py	74.46% <ø> (ø)
deepmd/entrypoints/main.py	88.29% <ø> (-0.13%)	⬇️
deepmd/entrypoints/compress.py	91.11% <83.33%> (-1.39%)	⬇️
deepmd/common.py	83.44% <86.66%> (+0.35%)	⬆️
deepmd/train/trainer.py	71.92% <96.77%> (+0.85%)	⬆️
deepmd/entrypoints/train.py	87.97% <100.00%> (+1.21%)	⬆️
deepmd/utils/argcheck.py	87.58% <100.00%> (+0.04%)	⬆️
deepmd/utils/errors.py	100.00% <100.00%> (ø)
deepmd/utils/data.py	90.33% <0.00%> (-0.03%)	⬇️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update ee0ed99...a1f087d. Read the comment docs.

[njzjz]

I have another question. What is the behavior if one uses another JSON to restart the training? For example, in active learning cycles, we can continue training from the last iteration.

[denghuilu]

I have another question. What is the behavior if one uses another JSON to restart the training? For example, in active learning cycles, we can continue training from the last iteration.

The model will save the latest json file.

[amcadmus]

I have another question. What is the behavior if one uses another JSON to restart the training? For example, in active learning cycles, we can continue training from the last iteration.

In the init-model mode, overwriting the existing training script should be the expected behavior.

In the restart mode, the training script are expected to be the same.

[njzjz]

Also, please fix lint warnings.

[njzjz]

deepmd-kit/deepmd/train/trainer.py

Lines 288 to 307 in 425a896

	if self.is_compress == False:
	# Usually, the type number of the model should be equal to that of the data
	# However, nt_model > nt_data should be allowed, since users may only want to
	# train using a dataset that only have some of elements
	assert (self.ntypes >= data.get_ntypes()), "ntypes should match that found in data"
	self.stop_batch = stop_batch
	# self.batch_size = data.get_batch_size()
	if self.numb_fparam > 0 :
	log.info("training with %d frame parameter(s)" % self.numb_fparam)
	else:
	log.info("training without frame parameter")
	# self.type_map = data.get_type_map()
	if self.is_compress == False:
	# Usually, the type number of the model should be equal to that of the data
	# However, nt_model > nt_data should be allowed, since users may only want to
	# train using a dataset that only have some of elements
	assert (self.ntypes >= data.get_ntypes()), "ntypes should match that found in data"

Duplicated lines