"Class conv_transpose2d is not found" when exporting a pruning-optimized model about nncf

[vshampor]

@mkaglins @vanyalzr

[nsk-lab]

@vshampor @mkaglins @vanyalzr
I am testing NNCF tool for our development. Any suggestions would be greatly appreciated :)

[vshampor]

@nsk-lab indeed, nn.ConvTranspose2d pruning is not supported. However, the models that simply contain nn.ConvTranspose2d and don't prune it should still be prunable with respect to the supported prunable ops (which are convolutions, at the moment), so this seems to be a bug.

Please try #171 and see if it fixes your issue.

[lzrvch]

Hi @nsk-lab, does your model have a significant amount of ConvTranspose2D layers?
If that's the case and you wish to prune them along with standard Conv2d layers, we could plan to add ConvTranspose2D to the scope of prunable ops in filter pruning.

[vshampor]

@nsk-lab did applying #171 resolve your problem?

[nsk-lab]

@vshampor @vanyalzr

Yes. I could export onnx without error. thx.

But the final pruning rate was low.
my setting:

    "params": {
        "schedule": "baseline", // The type of scheduling to use for adjusting the target pruning level. Either `exponential`, `exponential_with_bias`,  or `baseline`, by default it is `baseline`"
        "pruning_init": 0.1, // Initial value of the pruning level applied to the model. 0.0 by default.
        "pruning_target": 0.4, // Target value of the pruning level for the model. 0.5 by default.
        "num_init_steps": 1, // Number of epochs for model pretraining before starting filter pruning. 0 by default.
        "pruning_steps": 4, // Number of epochs during which the pruning rate is increased from `pruning_init` to `pruning_target` value.
        "weight_importance": "L2", // The type of filter importance metric. Can be one of `L1`, `L2`, `geometric_median`. `L2` by default.
        "all_weights": false, // Whether to prune layers independently (choose filters with the smallest importance in each layer separately) or not. `False` by default.
        "prune_first_conv": false, // Whether to prune first Convolutional layers or not. First means that it is a convolutional layer such that there is a path from model input to this layer such that there are no other convolution operations on it. `False` by default.
        "prune_last_conv": false, // Whether to prune last Convolutional layers or not.  Last means that it is a Convolutional layer such that there is a path from this layer to the model output such that there are no other convolution operations on it. `False` by default.
        "prune_downsample_convs": false, // Whether to prune downsample Convolutional layers (with stride > 1) or not. `False` by default.
        "prune_batch_norms": false, // Whether to nullifies parameters of Batch Norm layer corresponds to zeroed filters of convolution corresponding to this Batch Norm. `False` by default.
        "zero_grad": true // Whether to setting gradients corresponding to zeroed filters to zero during training, `True` by default.
    },```

and the log at 5 epoch is

.
. (ommited)
.

+--------+--------+--------+--------+--------+--------+--------+-------+-------+
| XXXXXX | [64,   | [40,   | [64]   | [40]   | 147520 | 92200  | 0.375 | 0.375 |
| XXXX/Y | 256,   | 256,   |        |        |        |        |       |       |
| YYYYYY | 3, 3]  | 3, 3]  |        |        |        |        |       |       |
| YYY[ZZ |        |        |        |        |        |        |       |       |
| Z]/NNC |        |        |        |        |        |        |       |       |
| FConv2 |        |        |        |        |        |        |       |       |
| d[0]   |        |        |        |        |        |        |       |       |
+--------+--------+--------+--------+--------+--------+--------+-------+-------+
| XXXXXX | [64,   | [40,   | [64]   | [40]   | 147520 | 92200  | 0.375 | 0.375 |
| XXXX/Y | 256,   | 256,   |        |        |        |        |       |       |
| YYYYYY | 3, 3]  | 3, 3]  |        |        |        |        |       |       |
| YYY[ZZ |        |        |        |        |        |        |       |       |
| ]/NNCF |        |        |        |        |        |        |       |       |
| Conv2d |        |        |        |        |        |        |       |       |
| [0]    |        |        |        |        |        |        |       |       |
+--------+--------+--------+--------+--------+--------+--------+-------+-------+
INFO:nncf:Final Model Pruning Rate = 0.165
INFO:nncf:Total MAC pruning level = 0.124

0.165 is much lower than pruning target 0.4
Is it due to uncompressed ConvTranspose2D layer?

if so, i really want the support of pruning ConvTranspose2D